fix(ai): keep OpenAI image output settings and Z.ai URL expiry (#51380 )

fix(ai): harden OpenAI transcription stream parsing (#51378 )
fix(ai): tighten media error consistency (#51374 )
2026-09-25 18:17:35 +00:00 · 2026-09-25 22:51:46 +05:30 · 2026-09-25 22:51:22 +05:30 · 2026-09-25 22:46:04 +05:30 · 2026-09-25 22:37:20 +05:30 · 2026-09-25 22:36:02 +05:30
820 changed files with 65213 additions and 7093 deletions
@@ -53,8 +53,6 @@ runs:
      with:
        path: ${{ steps.cache.outputs.dir }}
        key: ${{ runner.os }}-bun-${{ hashFiles('**/bun.lock') }}
-        restore-keys: |
-          ${{ runner.os }}-bun-

    - name: Install setuptools for distutils compatibility
      run: python3 -m pip install setuptools || pip install setuptools || true
@@ -66,9 +64,9 @@ runs:
        # e.g. ./patches/ for standard-openapi
        # https://github.com/oven-sh/bun/issues/28147
        if [ "$RUNNER_OS" = "Windows" ]; then
-          bun install --linker hoisted ${{ inputs.install-flags }}
+          bun install --frozen-lockfile --linker hoisted ${{ inputs.install-flags }}
        else
-          bun install ${{ inputs.install-flags }}
+          bun install --frozen-lockfile ${{ inputs.install-flags }}
        fi
      shell: bash

@@ -32,7 +32,7 @@
    },
    "packages/ai": {
      "name": "@opencode/ai",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@aws-sdk/credential-providers": "3.1057.0",
        "@opencode/schema": "workspace:*",
@@ -54,7 +54,7 @@
    },
    "packages/app": {
      "name": "@opencode/app",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@corvu/drawer": "catalog:",
        "@dnd-kit/abstract": "0.5.0",
@@ -112,8 +112,9 @@
    },
    "packages/cli": {
      "name": "@opencode/cli",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "bin": {
+        "opencode": "./bin/opencode.cjs",
        "opencode2": "./bin/opencode2.cjs",
      },
      "dependencies": {
@@ -176,7 +177,7 @@
    },
    "packages/client": {
      "name": "@opencode/client",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@opencode/protocol": "workspace:*",
        "@opencode/schema": "workspace:*",
@@ -202,7 +203,7 @@
    },
    "packages/codemode": {
      "name": "@opencode/codemode",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "acorn": "8.15.0",
        "effect": "catalog:",
@@ -215,7 +216,7 @@
    },
    "packages/console/app": {
      "name": "@opencode/console-app",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@cloudflare/vite-plugin": "1.15.2",
        "@ibm/plex": "6.4.1",
@@ -251,7 +252,7 @@
    },
    "packages/console/core": {
      "name": "@opencode/console-core",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@aws-sdk/client-sts": "3.782.0",
        "@jsx-email/render": "1.1.1",
@@ -278,7 +279,7 @@
    },
    "packages/console/function": {
      "name": "@opencode/console-function",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@openauthjs/openauth": "0.0.0-20250322224806",
        "@opencode/console-core": "workspace:*",
@@ -295,7 +296,7 @@
    },
    "packages/console/mail": {
      "name": "@opencode/console-mail",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@jsx-email/all": "2.2.3",
        "@jsx-email/cli": "1.4.3",
@@ -319,7 +320,7 @@
    },
    "packages/console/support": {
      "name": "@opencode/console-support",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@cloudflare/vite-plugin": "1.15.2",
        "@opencode/console-core": "workspace:*",
@@ -339,7 +340,7 @@
    },
    "packages/core": {
      "name": "@opencode/core",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@ai-sdk/cohere": "3.0.27",
        "@ai-sdk/gateway": "3.0.104",
@@ -367,7 +368,7 @@
        "drizzle-orm": "catalog:",
        "effect": "catalog:",
        "fuzzysort": "3.1.0",
-        "gitlab-ai-provider": "6.12.1",
+        "gitlab-ai-provider": "6.18.0",
        "google-auth-library": "10.5.0",
        "gray-matter": "4.0.3",
        "htmlparser2": "8.0.2",
@@ -407,7 +408,7 @@
    },
    "packages/desktop": {
      "name": "@opencode/desktop",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@zip.js/zip.js": "2.7.62",
        "electron-context-menu": "5.0.0",
@@ -456,7 +457,7 @@
    },
    "packages/enterprise": {
      "name": "@opencode/enterprise",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@hono/standard-validator": "catalog:",
        "@opencode-ai/sdk": "1.18.21",
@@ -493,7 +494,7 @@
    },
    "packages/function": {
      "name": "@opencode/function",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@octokit/auth-app": "8.0.1",
        "@octokit/rest": "catalog:",
@@ -509,7 +510,7 @@
    },
    "packages/http-recorder": {
      "name": "@opencode/http-recorder",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@effect/platform-node-shared": "4.0.0-rc.112",
      },
@@ -528,7 +529,7 @@
    },
    "packages/httpapi-codegen": {
      "name": "@opencode/httpapi-codegen",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "effect": "catalog:",
        "prettier": "3.6.2",
@@ -541,7 +542,7 @@
    },
    "packages/latex": {
      "name": "@opencode/latex",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@opencode/plugin": "workspace:*",
        "@opentui/core": "catalog:",
@@ -555,7 +556,7 @@
    },
    "packages/merman": {
      "name": "@opencode/merman",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@opencode/plugin": "workspace:*",
        "@opentui/core": "catalog:",
@@ -570,7 +571,7 @@
    },
    "packages/plugin": {
      "name": "@opencode/plugin",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@ai-sdk/provider": "3.0.8",
        "@opencode/ai": "workspace:*",
@@ -596,8 +597,8 @@
      },
      "peerDependencies": {
        "@opencode/theme": "workspace:*",
-        "@opentui/core": ">=0.5.10",
-        "@opentui/solid": ">=0.5.10",
+        "@opentui/core": ">=0.5.12",
+        "@opentui/solid": ">=0.5.12",
        "solid-js": ">=1.9.0",
      },
      "optionalPeers": [
@@ -609,7 +610,7 @@
    },
    "packages/plugin-browser": {
      "name": "@opencode/plugin-browser",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@opencode/plugin": "workspace:*",
        "@opencode/schema": "workspace:*",
@@ -639,7 +640,7 @@
    },
    "packages/protocol": {
      "name": "@opencode/protocol",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@opencode/schema": "workspace:*",
        "effect": "catalog:",
@@ -654,7 +655,7 @@
    },
    "packages/schema": {
      "name": "@opencode/schema",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@standard-schema/spec": "catalog:",
        "effect": "catalog:",
@@ -678,7 +679,7 @@
    },
    "packages/sdk": {
      "name": "@opencode/sdk",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@opencode/client": "workspace:*",
        "@opencode/core": "workspace:*",
@@ -699,7 +700,7 @@
    },
    "packages/server": {
      "name": "@opencode/server",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@effect/platform-node": "catalog:",
        "@effect/platform-node-shared": "catalog:",
@@ -721,7 +722,7 @@
    },
    "packages/session-ui": {
      "name": "@opencode/session-ui",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@kobalte/core": "catalog:",
        "@opencode/client": "workspace:*",
@@ -756,7 +757,7 @@
    },
    "packages/simulation": {
      "name": "@opencode/simulation",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@opencode/ai": "workspace:*",
        "@opencode/core": "workspace:*",
@@ -776,7 +777,7 @@
    },
    "packages/stats/app": {
      "name": "@opencode/stats-app",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@ibm/plex": "6.4.1",
        "@kobalte/core": "catalog:",
@@ -810,7 +811,7 @@
    },
    "packages/stats/core": {
      "name": "@opencode/stats-core",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@aws-sdk/client-athena": "3.933.0",
        "@planetscale/database": "1.19.0",
@@ -829,7 +830,7 @@
    },
    "packages/stats/server": {
      "name": "@opencode/stats-server",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@aws-sdk/client-firehose": "3.933.0",
        "@effect/platform-node": "catalog:",
@@ -875,7 +876,7 @@
    },
    "packages/theme": {
      "name": "@opencode/theme",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@opentui/core": "catalog:",
        "effect": "catalog:",
@@ -889,7 +890,7 @@
    },
    "packages/tui": {
      "name": "@opencode/tui",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@opencode/client": "workspace:*",
        "@opencode/core": "workspace:*",
@@ -924,7 +925,7 @@
    },
    "packages/ui": {
      "name": "@opencode/ui",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@kobalte/core": "catalog:",
        "@pierre/diffs": "catalog:",
@@ -959,7 +960,7 @@
    },
    "packages/util": {
      "name": "@opencode/util",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@effect/opentelemetry": "catalog:",
        "@effect/platform-node": "catalog:",
@@ -967,10 +968,14 @@
        "@npmcli/arborist": "catalog:",
        "@npmcli/config": "10.8.1",
        "@opentelemetry/api": "1.9.0",
-        "@opentelemetry/context-async-hooks": "2.6.1",
-        "@opentelemetry/exporter-trace-otlp-http": "0.214.0",
-        "@opentelemetry/sdk-trace-base": "2.6.1",
-        "@opentelemetry/sdk-trace-node": "2.6.1",
+        "@opentelemetry/api-logs": "0.219.0",
+        "@opentelemetry/context-async-hooks": "2.8.0",
+        "@opentelemetry/exporter-trace-otlp-http": "0.219.0",
+        "@opentelemetry/resources": "2.8.0",
+        "@opentelemetry/sdk-logs": "0.219.0",
+        "@opentelemetry/sdk-metrics": "2.8.0",
+        "@opentelemetry/sdk-trace-base": "2.8.0",
+        "@opentelemetry/sdk-trace-node": "2.8.0",
        "cross-spawn": "catalog:",
        "effect": "catalog:",
        "glob": "13.0.5",
@@ -992,7 +997,7 @@
    },
    "packages/web": {
      "name": "@opencode/web",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "@astrojs/cloudflare": "12.6.3",
        "@astrojs/markdown-remark": "6.3.1",
@@ -1033,7 +1038,7 @@
    },
    "services/update": {
      "name": "@opencode/update",
-      "version": "2.0.14",
+      "version": "2.0.16",
      "dependencies": {
        "jose": "6.0.11",
        "semver": "catalog:",
@@ -1068,13 +1073,13 @@
  "trustedDependencies": [
    "electron",
    "esbuild",
-    "protobufjs",
  ],
  "patchedDependencies": {
    "@pierre/trees@1.0.0-beta.4": "patches/@pierre%2Ftrees@1.0.0-beta.4.patch",
    "@tanstack/virtual-core@3.17.8": "patches/@tanstack%2Fvirtual-core@3.17.8.patch",
    "ghostty-web@github:anomalyco/ghostty-web#83c0a07": "patches/ghostty-web@0.3.0.patch",
    "@modelcontextprotocol/client@2.0.0": "patches/@modelcontextprotocol%2Fclient@2.0.0.patch",
+    "pacote@21.5.1": "patches/pacote@21.5.1.patch",
    "@standard-community/standard-openapi@0.2.9": "patches/@standard-community%2Fstandard-openapi@0.2.9.patch",
    "@npmcli/agent@4.0.2": "patches/@npmcli%2Fagent@4.0.2.patch",
    "@silvia-odwyer/photon-node@0.3.4": "patches/@silvia-odwyer%2Fphoton-node@0.3.4.patch",
@@ -1109,9 +1114,9 @@
    "@npmcli/arborist": "9.4.0",
    "@octokit/rest": "22.0.0",
    "@openauthjs/openauth": "0.0.0-20250322224806",
-    "@opentui/core": "0.5.10",
-    "@opentui/keymap": "0.5.10",
-    "@opentui/solid": "0.5.10",
+    "@opentui/core": "0.5.12",
+    "@opentui/keymap": "0.5.12",
+    "@opentui/solid": "0.5.12",
    "@pierre/diffs": "1.2.10",
    "@playwright/test": "1.59.1",
    "@sentry/solid": "10.71.0",
@@ -2219,55 +2224,55 @@

    "@opentelemetry/api": ["@opentelemetry/api@1.9.0", "", {}, "sha512-3giAOQvZiH5F9bMlMiv8+GSPMeqg0dbaeo58/0SlA9sxSqZhnUtxzX9/2FzyhS9sWQf5S0GJE0AKBrFqjpeYcg=="],

-    "@opentelemetry/api-logs": ["@opentelemetry/api-logs@0.214.0", "", { "dependencies": { "@opentelemetry/api": "^1.3.0" } }, "sha512-40lSJeqYO8Uz2Yj7u94/SJWE/wONa7rmMKjI1ZcIjgf3MHNHv1OZUCrCETGuaRF62d5pQD1wKIW+L4lmSMTzZA=="],
+    "@opentelemetry/api-logs": ["@opentelemetry/api-logs@0.219.0", "", { "dependencies": { "@opentelemetry/api": "^1.3.0" } }, "sha512-FFx7YnaYJlIjqWW/AG/yAZ0L/NEY724PipXXXQLdtZPbLwBGbUMTGL1i/esI56TWfTUXxhLfpgrnWJCG8aUJyg=="],

-    "@opentelemetry/context-async-hooks": ["@opentelemetry/context-async-hooks@2.6.1", "", { "peerDependencies": { "@opentelemetry/api": ">=1.0.0 <1.10.0" } }, "sha512-XHzhwRNkBpeP8Fs/qjGrAf9r9PRv67wkJQ/7ZPaBQQ68DYlTBBx5MF9LvPx7mhuXcDessKK2b+DcxqwpgkcivQ=="],
+    "@opentelemetry/context-async-hooks": ["@opentelemetry/context-async-hooks@2.8.0", "", { "peerDependencies": { "@opentelemetry/api": ">=1.0.0 <1.10.0" } }, "sha512-/3FIraneMcng67SUJCxvyInk/oxzwsxyadufk0wwfOBLf5wqtAGX4MoQASwSbndBPeARzBryUM9Azr5kHIdWLw=="],

-    "@opentelemetry/core": ["@opentelemetry/core@2.6.1", "", { "dependencies": { "@opentelemetry/semantic-conventions": "^1.29.0" }, "peerDependencies": { "@opentelemetry/api": ">=1.0.0 <1.10.0" } }, "sha512-8xHSGWpJP9wBxgBpnqGL0R3PbdWQndL1Qp50qrg71+B28zK5OQmUgcDKLJgzyAAV38t4tOyLMGDD60LneR5W8g=="],
+    "@opentelemetry/core": ["@opentelemetry/core@2.8.0", "", { "dependencies": { "@opentelemetry/semantic-conventions": "^1.29.0" }, "peerDependencies": { "@opentelemetry/api": ">=1.0.0 <1.10.0" } }, "sha512-hd1Lfh8p545nNz+jq1Ejfz+Mn1hyLuxYn1YzTfFNrxr8urEWMNQLPf1Th8kjOH+HxwawCrtgBp8JpBUR4ZSgww=="],

-    "@opentelemetry/exporter-trace-otlp-http": ["@opentelemetry/exporter-trace-otlp-http@0.214.0", "", { "dependencies": { "@opentelemetry/core": "2.6.1", "@opentelemetry/otlp-exporter-base": "0.214.0", "@opentelemetry/otlp-transformer": "0.214.0", "@opentelemetry/resources": "2.6.1", "@opentelemetry/sdk-trace-base": "2.6.1" }, "peerDependencies": { "@opentelemetry/api": "^1.3.0" } }, "sha512-kIN8nTBMgV2hXzV/a20BCFilPZdAIMYYJGSgfMMRm/Xa+07y5hRDS2Vm12A/z8Cdu3Sq++ZvJfElokX2rkgGgw=="],
+    "@opentelemetry/exporter-trace-otlp-http": ["@opentelemetry/exporter-trace-otlp-http@0.219.0", "", { "dependencies": { "@opentelemetry/core": "2.8.0", "@opentelemetry/otlp-exporter-base": "0.219.0", "@opentelemetry/otlp-transformer": "0.219.0", "@opentelemetry/resources": "2.8.0", "@opentelemetry/sdk-trace-base": "2.8.0" }, "peerDependencies": { "@opentelemetry/api": "^1.3.0" } }, "sha512-9t6SvBXXBEjOBcIzgozvBbd3jWrv3Gt3ngGhl1fhdZ/zRc7oZDVOFEqbi2zlBpW9BXhgDMKv422J0DL/3iQWfw=="],

    "@opentelemetry/instrumentation": ["@opentelemetry/instrumentation@0.220.0", "", { "dependencies": { "@opentelemetry/api-logs": "0.220.0", "import-in-the-middle": "^3.0.0", "require-in-the-middle": "^8.0.0" }, "peerDependencies": { "@opentelemetry/api": "^1.3.0" } }, "sha512-xQx3E2WxP1mDvKzxLxX+CTCtNLa560YJZ3087qYHerl2YmiKpv7AH+dAy7vmx+eVrZ5BwhfWUAVoKOoxCNHcpw=="],

-    "@opentelemetry/otlp-exporter-base": ["@opentelemetry/otlp-exporter-base@0.214.0", "", { "dependencies": { "@opentelemetry/core": "2.6.1", "@opentelemetry/otlp-transformer": "0.214.0" }, "peerDependencies": { "@opentelemetry/api": "^1.3.0" } }, "sha512-u1Gdv0/E9wP+apqWf7Wv2npXmgJtxsW2XL0TEv9FZloTZRuMBKmu8cYVXwS4Hm3q/f/3FuCnPTgiwYvIqRSpRg=="],
+    "@opentelemetry/otlp-exporter-base": ["@opentelemetry/otlp-exporter-base@0.219.0", "", { "dependencies": { "@opentelemetry/core": "2.8.0", "@opentelemetry/otlp-transformer": "0.219.0" }, "peerDependencies": { "@opentelemetry/api": "^1.3.0" } }, "sha512-zvIxQX/AZUVKDU+hCuYx+7UkiP7GRdnk1ZbFQRYzHvYp47cAWR4j3IhoPhV9KaeXEv2xdGq3IA6PnpzDmLcmSA=="],

-    "@opentelemetry/otlp-transformer": ["@opentelemetry/otlp-transformer@0.214.0", "", { "dependencies": { "@opentelemetry/api-logs": "0.214.0", "@opentelemetry/core": "2.6.1", "@opentelemetry/resources": "2.6.1", "@opentelemetry/sdk-logs": "0.214.0", "@opentelemetry/sdk-metrics": "2.6.1", "@opentelemetry/sdk-trace-base": "2.6.1", "protobufjs": "^7.0.0" }, "peerDependencies": { "@opentelemetry/api": "^1.3.0" } }, "sha512-DSaYcuBRh6uozfsWN3R8HsN0yDhCuWP7tOFdkUOVaWD1KVJg8m4qiLUsg/tNhTLS9HUYUcwNpwL2eroLtsZZ/w=="],
+    "@opentelemetry/otlp-transformer": ["@opentelemetry/otlp-transformer@0.219.0", "", { "dependencies": { "@opentelemetry/api-logs": "0.219.0", "@opentelemetry/core": "2.8.0", "@opentelemetry/resources": "2.8.0", "@opentelemetry/sdk-logs": "0.219.0", "@opentelemetry/sdk-metrics": "2.8.0", "@opentelemetry/sdk-trace-base": "2.8.0" }, "peerDependencies": { "@opentelemetry/api": "^1.3.0" } }, "sha512-aaYKAyXhw9VchKZVGOopD3Gw/kPsyrX2c6IQ0AW32mTjqmZOh5Y6Gf5OYqTNqVktAeBjmFinhyFaCwW6GYK9YQ=="],

-    "@opentelemetry/resources": ["@opentelemetry/resources@2.6.1", "", { "dependencies": { "@opentelemetry/core": "2.6.1", "@opentelemetry/semantic-conventions": "^1.29.0" }, "peerDependencies": { "@opentelemetry/api": ">=1.3.0 <1.10.0" } }, "sha512-lID/vxSuKWXM55XhAKNoYXu9Cutoq5hFdkbTdI/zDKQktXzcWBVhNsOkiZFTMU9UtEWuGRNe0HUgmsFldIdxVA=="],
+    "@opentelemetry/resources": ["@opentelemetry/resources@2.8.0", "", { "dependencies": { "@opentelemetry/core": "2.8.0", "@opentelemetry/semantic-conventions": "^1.29.0" }, "peerDependencies": { "@opentelemetry/api": ">=1.3.0 <1.10.0" } }, "sha512-qmXQ27ilDbUK/vGMqwL8D4/rhn76C+sherM4wTbjlfknR8Nvfc/hCxjRJPhkzZzUsPiNg16SA31NxMabwttRjg=="],

-    "@opentelemetry/sdk-logs": ["@opentelemetry/sdk-logs@0.214.0", "", { "dependencies": { "@opentelemetry/api-logs": "0.214.0", "@opentelemetry/core": "2.6.1", "@opentelemetry/resources": "2.6.1", "@opentelemetry/semantic-conventions": "^1.29.0" }, "peerDependencies": { "@opentelemetry/api": ">=1.4.0 <1.10.0" } }, "sha512-zf6acnScjhsaBUU22zXZ/sLWim1dfhUAbGXdMmHmNG3LfBnQ3DKsOCITb2IZwoUsNNMTogqFKBnlIPPftUgGwA=="],
+    "@opentelemetry/sdk-logs": ["@opentelemetry/sdk-logs@0.219.0", "", { "dependencies": { "@opentelemetry/api-logs": "0.219.0", "@opentelemetry/core": "2.8.0", "@opentelemetry/resources": "2.8.0", "@opentelemetry/semantic-conventions": "^1.29.0" }, "peerDependencies": { "@opentelemetry/api": ">=1.4.0 <1.10.0" } }, "sha512-s6lTKRakaPClvKoWHRChxnXjDMkM/TQ30ff78jN6EBGf7MI7VzANE5PU3f4z9qDUudWjvZjOLHG0rBnBKYvoXA=="],

-    "@opentelemetry/sdk-metrics": ["@opentelemetry/sdk-metrics@2.6.1", "", { "dependencies": { "@opentelemetry/core": "2.6.1", "@opentelemetry/resources": "2.6.1" }, "peerDependencies": { "@opentelemetry/api": ">=1.9.0 <1.10.0" } }, "sha512-9t9hJHX15meBy2NmTJxL+NJfXmnausR2xUDvE19XQce0Qi/GBtDGamU8nS1RMbdgDmhgpm3VaOu2+fiS/SfTpQ=="],
+    "@opentelemetry/sdk-metrics": ["@opentelemetry/sdk-metrics@2.8.0", "", { "dependencies": { "@opentelemetry/core": "2.8.0", "@opentelemetry/resources": "2.8.0" }, "peerDependencies": { "@opentelemetry/api": ">=1.9.0 <1.10.0" } }, "sha512-UDBGaj6W0Rgy5rTTaoxs8gVGF/aGkAKyjurJv7se6wjRxJu7FoquTLT/vt54DZfo4crbprYfhX/SOK9+BPw1qg=="],

    "@opentelemetry/sdk-trace": ["@opentelemetry/sdk-trace@2.11.0", "", { "dependencies": { "@opentelemetry/core": "2.11.0", "@opentelemetry/resources": "2.11.0", "@opentelemetry/semantic-conventions": "^1.29.0" }, "peerDependencies": { "@opentelemetry/api": ">=1.3.0 <1.10.0" } }, "sha512-fFnTqGm8/G73GQVnxYi7LXa1ZVYEUvgL6XI1LpvV0bPC7WQ/ZGgKxCSl8FnlZBKto9JHHEFTO6s6CUpvvtwFrA=="],

-    "@opentelemetry/sdk-trace-base": ["@opentelemetry/sdk-trace-base@2.6.1", "", { "dependencies": { "@opentelemetry/core": "2.6.1", "@opentelemetry/resources": "2.6.1", "@opentelemetry/semantic-conventions": "^1.29.0" }, "peerDependencies": { "@opentelemetry/api": ">=1.3.0 <1.10.0" } }, "sha512-r86ut4T1e8vNwB35CqCcKd45yzqH6/6Wzvpk2/cZB8PsPLlZFTvrh8yfOS3CYZYcUmAx4hHTZJ8AO8Dj8nrdhw=="],
+    "@opentelemetry/sdk-trace-base": ["@opentelemetry/sdk-trace-base@2.8.0", "", { "dependencies": { "@opentelemetry/core": "2.8.0", "@opentelemetry/resources": "2.8.0", "@opentelemetry/semantic-conventions": "^1.29.0" }, "peerDependencies": { "@opentelemetry/api": ">=1.3.0 <1.10.0" } }, "sha512-mhU4jp+vW0mGbFRd+GeXHvmfA4aDqWjBjLC3pE5XMpLs0IE2ryYb019Ts2AQrOq67gaTF25D91+fgvEHDZEnuQ=="],

-    "@opentelemetry/sdk-trace-node": ["@opentelemetry/sdk-trace-node@2.6.1", "", { "dependencies": { "@opentelemetry/context-async-hooks": "2.6.1", "@opentelemetry/core": "2.6.1", "@opentelemetry/sdk-trace-base": "2.6.1" }, "peerDependencies": { "@opentelemetry/api": ">=1.0.0 <1.10.0" } }, "sha512-Hh2i4FwHWRFhnO2Q/p6svMxy8MPsNCG0uuzUY3glqm0rwM0nQvbTO1dXSp9OqQoTKXcQzaz9q1f65fsurmOhNw=="],
+    "@opentelemetry/sdk-trace-node": ["@opentelemetry/sdk-trace-node@2.8.0", "", { "dependencies": { "@opentelemetry/context-async-hooks": "2.8.0", "@opentelemetry/core": "2.8.0", "@opentelemetry/sdk-trace-base": "2.8.0" }, "peerDependencies": { "@opentelemetry/api": ">=1.0.0 <1.10.0" } }, "sha512-nZt9OGufioAc3AfoLTqA9bsAeaMJAictYDdI2VcNQ+PmT+3rfKjAZDZvgPfd8VPX0O5Bw1hdQF6kDK8VSpZiWg=="],

    "@opentelemetry/semantic-conventions": ["@opentelemetry/semantic-conventions@1.43.0", "", {}, "sha512-eSYWTm620tTk45EKSedaUL8MFYI8hW164hIXsgIHyxu3VobUB3fFCu5t0hQby6OoWRPsG1KkKUG2M5UadiLiVg=="],

-    "@opentui/core": ["@opentui/core@0.5.10", "", { "dependencies": { "bun-ffi-structs": "0.3.1", "diff": "9.0.0", "marked": "17.0.1", "string-width": "7.2.0", "strip-ansi": "7.1.2" }, "optionalDependencies": { "@opentui/core-darwin-arm64": "0.5.10", "@opentui/core-darwin-x64": "0.5.10", "@opentui/core-linux-arm64": "0.5.10", "@opentui/core-linux-arm64-musl": "0.5.10", "@opentui/core-linux-x64": "0.5.10", "@opentui/core-linux-x64-musl": "0.5.10", "@opentui/core-win32-arm64": "0.5.10", "@opentui/core-win32-x64": "0.5.10" }, "peerDependencies": { "web-tree-sitter": "0.25.10" } }, "sha512-C3a2UbmefeAjIxAgm4BqjuSxKT4oqutfvYFwVvUgMxmGRHkNbBc/s7sukV0JgwcxFcV3uMFrXxo+E+BQtvuOiw=="],
+    "@opentui/core": ["@opentui/core@0.5.12", "", { "dependencies": { "bun-ffi-structs": "0.3.1", "diff": "9.0.0", "marked": "17.0.1", "string-width": "7.2.0", "strip-ansi": "7.1.2" }, "optionalDependencies": { "@opentui/core-darwin-arm64": "0.5.12", "@opentui/core-darwin-x64": "0.5.12", "@opentui/core-linux-arm64": "0.5.12", "@opentui/core-linux-arm64-musl": "0.5.12", "@opentui/core-linux-x64": "0.5.12", "@opentui/core-linux-x64-musl": "0.5.12", "@opentui/core-win32-arm64": "0.5.12", "@opentui/core-win32-x64": "0.5.12" }, "peerDependencies": { "web-tree-sitter": "0.25.10" } }, "sha512-ZXBE5gmvdovmV8zJQrOQf6E44v1tJRDEgrM2MYhEglzgXZ+smIUp95O8zeRYGsuIzQIiMPMgQqKtTJuzvAb7BQ=="],

-    "@opentui/core-darwin-arm64": ["@opentui/core-darwin-arm64@0.5.10", "", { "os": "darwin", "cpu": "arm64" }, "sha512-Vyb+nTbhab8ZcRy5gg1loEEGwRcIbjAeVRIBfHBcbFDqmITBOg7x2gqJ+x/TnoOy4uwMhCmICUN2wiyREw3r1Q=="],
+    "@opentui/core-darwin-arm64": ["@opentui/core-darwin-arm64@0.5.12", "", { "os": "darwin", "cpu": "arm64" }, "sha512-YdVnP0tAyerBNl0mIcmQEOotPeZzW1VnSXKBl5cyZ5e6nDd2Y+ui/8eRPpn1oqcamf1NCnzS4ohMgejOvna8Zg=="],

-    "@opentui/core-darwin-x64": ["@opentui/core-darwin-x64@0.5.10", "", { "os": "darwin", "cpu": "x64" }, "sha512-tTFLcM7Oj1gTyhm/bUdAt3C6grZdCxPk6+/g2azcZBUlI3/62LwbeRS6HbQKFFmm+1fUmX8cq6kWrtul885mVg=="],
+    "@opentui/core-darwin-x64": ["@opentui/core-darwin-x64@0.5.12", "", { "os": "darwin", "cpu": "x64" }, "sha512-uRrQJdHmLUSj3PV23QPi3WSimYTTxcXnVouxF6U4xMXlOv4N3SxnHfVwMRQkPqbGOfvVWHeLE6FdK4C+ubU0sQ=="],

-    "@opentui/core-linux-arm64": ["@opentui/core-linux-arm64@0.5.10", "", { "os": "linux", "cpu": "arm64" }, "sha512-ncJXcgudhBf2GdJyF3xVQN/Ec+1F7GOL+pRrURmgBYSj2v1w6EyoDQFAACtPTK2c3R38W6fvZwL4JSLlm4EFXQ=="],
+    "@opentui/core-linux-arm64": ["@opentui/core-linux-arm64@0.5.12", "", { "os": "linux", "cpu": "arm64" }, "sha512-XeKhuIaEtgipvuPHbl4qPOBj+Ut+2zObmsxMVM1jDcjz/FatG9PGeGQPx1G1SnvH2AgpT4K+eCu7DUF0+yIqoQ=="],

-    "@opentui/core-linux-arm64-musl": ["@opentui/core-linux-arm64-musl@0.5.10", "", { "os": "linux", "cpu": "arm64" }, "sha512-dGMphDKexSdeYqwl0wgoFBP88Ta/cdi1Zc1mk29/ENkSCGz+74zlCHgqTHRNGLmI8W5TfuUtCyktQH11/Z+TBQ=="],
+    "@opentui/core-linux-arm64-musl": ["@opentui/core-linux-arm64-musl@0.5.12", "", { "os": "linux", "cpu": "arm64" }, "sha512-VZ2sNMw1d/r1SLPjUbOP9LKscKz1CQjID8adTL6gG8Lrrq+mYcIUxutyB+P/eG0J/7oRZLPR6OMt7dUOap6RTg=="],

-    "@opentui/core-linux-x64": ["@opentui/core-linux-x64@0.5.10", "", { "os": "linux", "cpu": "x64" }, "sha512-5qtYaOgwVycZD1GaGshTRsi0rXPAmVExO03N1JQaHu+NYxK/vXSOc7Bu4QW0sPXx3Sp0SpzpP+FHjXABfoK66g=="],
+    "@opentui/core-linux-x64": ["@opentui/core-linux-x64@0.5.12", "", { "os": "linux", "cpu": "x64" }, "sha512-eZiCjEzwbb6qClPPfk32Nha9xmr9obt69Xj0+9SKsXxWLBKkjQEGOMRoh/R9ObaQF4aq8If1xV3VEY0sD9W9vg=="],

-    "@opentui/core-linux-x64-musl": ["@opentui/core-linux-x64-musl@0.5.10", "", { "os": "linux", "cpu": "x64" }, "sha512-Oj4H9hApuvuTKPWxh4SoZAgGJorR7vbvnrZA/cAkSMAk2VGSoHRRcqeXQbcH8IcdjVZ0KFpv8Zkl/D5Ye+2mew=="],
+    "@opentui/core-linux-x64-musl": ["@opentui/core-linux-x64-musl@0.5.12", "", { "os": "linux", "cpu": "x64" }, "sha512-WWW0hVBoSYZ3D6AgZ4u2Y5/u/IyIq2pDb+4yI3WgJ70Wyt6ofHy+6kRGRgbXFn1p+rPInAHjCXD2v6C7iEKSrA=="],

-    "@opentui/core-win32-arm64": ["@opentui/core-win32-arm64@0.5.10", "", { "os": "win32", "cpu": "arm64" }, "sha512-A9VhgvTxQoUdZ+8LmUumEng1sQNbj9QQQT3NYG9mSxI54qTANi7vOWNSphMiY6RMVsr22pgm6nUvSSvJXv7Jog=="],
+    "@opentui/core-win32-arm64": ["@opentui/core-win32-arm64@0.5.12", "", { "os": "win32", "cpu": "arm64" }, "sha512-aLbm6870Ybls6CYL4zMOCImTBPLZHZMUXJFGqMI44lIWxitkAtT6zg5lYA4oRqFRzzryDclxr29+hDgT3p3Blw=="],

-    "@opentui/core-win32-x64": ["@opentui/core-win32-x64@0.5.10", "", { "os": "win32", "cpu": "x64" }, "sha512-u3KHa7kEeWrmKVDRJYpxSGO+g5E9cMGlrmTsPN3GVPHUmQMiREUawLXUvsU8+IHaQnqG3Q5nuE1yf4fPBzS+Qw=="],
+    "@opentui/core-win32-x64": ["@opentui/core-win32-x64@0.5.12", "", { "os": "win32", "cpu": "x64" }, "sha512-KTwtwpfd2zF9opVh3SyRJYDd1o3Xv4XL8OZb8Zi+CqWUel6Y2IDCiVivCv8fGJt3J7wOIXXtuZI9ZUkLyKJCiQ=="],

-    "@opentui/keymap": ["@opentui/keymap@0.5.10", "", { "dependencies": { "@opentui/core": "0.5.10" }, "peerDependencies": { "@opentui/react": "0.5.10", "@opentui/solid": "0.5.10", "react": ">=19.2.0", "solid-js": "1.9.12" }, "optionalPeers": ["@opentui/react", "@opentui/solid", "react", "solid-js"] }, "sha512-8vDJF+ltXscSnLEv3rgCa4m7PcoYZeUT9BngugpFCmVoNevbaRtYijjdfiUuLmXfT61lO5QbR6nEhn2RZMK8ow=="],
+    "@opentui/keymap": ["@opentui/keymap@0.5.12", "", { "dependencies": { "@opentui/core": "0.5.12" }, "peerDependencies": { "@opentui/react": "0.5.12", "@opentui/solid": "0.5.12", "react": ">=19.2.0", "solid-js": "1.9.12" }, "optionalPeers": ["@opentui/react", "@opentui/solid", "react", "solid-js"] }, "sha512-yWPvJjRhJTRoRSUucQq9Ua8ZW7n/2YQ/j6JxWq5Qekm4WuFiTplEkebR/Aj2/xA8tX68NOE5qv1LrY0Jk3NLNQ=="],

-    "@opentui/solid": ["@opentui/solid@0.5.10", "", { "dependencies": { "@babel/core": "7.28.0", "@babel/preset-typescript": "7.27.1", "@opentui/core": "0.5.10", "babel-plugin-module-resolver": "5.0.2", "babel-preset-solid": "1.9.12", "entities": "7.0.1", "s-js": "^0.4.9" }, "peerDependencies": { "solid-js": "1.9.12" } }, "sha512-KrmMIsHiKBHOABTC0brOwqWm+sGq1ZX2sGCAx6WgtBbE3STMup9n8TAy/6gUYhwcjC9zugT53ytfSVwCwVWZUg=="],
+    "@opentui/solid": ["@opentui/solid@0.5.12", "", { "dependencies": { "@babel/core": "7.28.0", "@babel/preset-typescript": "7.27.1", "@opentui/core": "0.5.12", "babel-plugin-module-resolver": "5.0.2", "babel-preset-solid": "1.9.12", "entities": "7.0.1", "s-js": "^0.4.9" }, "peerDependencies": { "solid-js": "1.9.12" } }, "sha512-hAiVlVMtT7AkHGblKwcW1YAuXtxkSy1XSf/RRc4j3IlG3mTNX0bhJdnGOo3Xw14EqeZMp41Mcp5WzHAzMm/DzA=="],

    "@oslojs/asn1": ["@oslojs/asn1@1.0.0", "", { "dependencies": { "@oslojs/binary": "1.0.0" } }, "sha512-zw/wn0sj0j0QKbIXfIlnEcTviaCzYOY3V5rAyjR6YtOByFtJiT574+8p9Wlach0lZH9fddD4yb9laEAIl4vXQA=="],

@@ -2553,24 +2558,6 @@

    "@protobuf-ts/runtime-rpc": ["@protobuf-ts/runtime-rpc@2.11.1", "", { "dependencies": { "@protobuf-ts/runtime": "^2.11.1" } }, "sha512-4CqqUmNA+/uMz00+d3CYKgElXO9VrEbucjnBFEjqI4GuDrEQ32MaI3q+9qPBvIGOlL4PmHXrzM32vBPWRhQKWQ=="],

-    "@protobufjs/aspromise": ["@protobufjs/aspromise@1.1.2", "", {}, "sha512-j+gKExEuLmKwvz3OgROXtrJ2UG2x8Ch2YZUxahh+s1F2HZ+wAceUNLkvy6zKCPVRkU++ZWQrdxsUeQXmcg4uoQ=="],
-
-    "@protobufjs/base64": ["@protobufjs/base64@1.1.2", "", {}, "sha512-AZkcAA5vnN/v4PDqKyMR5lx7hZttPDgClv83E//FMNhR2TMcLUhfRUBHCmSl0oi9zMgDDqRUJkSxO3wm85+XLg=="],
-
-    "@protobufjs/codegen": ["@protobufjs/codegen@2.0.5", "", {}, "sha512-zgXFLzW3Ap33e6d0Wlj4MGIm6Ce8O89n/apUaGNB/jx+hw+ruWEp7EwGUshdLKVRCxZW12fp9r40E1mQrf/34g=="],
-
-    "@protobufjs/eventemitter": ["@protobufjs/eventemitter@1.1.1", "", {}, "sha512-vW1GmwMZNnL+gMRaovlh9yZX74kc+TTU3FObkkurpMaRtBfLP3ldjS9KQWlwZgraRE0+dheEEoAxdzcJQ8eXZg=="],
-
-    "@protobufjs/fetch": ["@protobufjs/fetch@1.1.1", "", { "dependencies": { "@protobufjs/aspromise": "^1.1.1" } }, "sha512-GpptLrs57adMSuHi3VNj0mAF8dwh36LMaYF6XyJ6JMWlVsc+t42tm1HSEDmOs3A8fC9yyeisgLhsTVQokOZ0zw=="],
-
-    "@protobufjs/float": ["@protobufjs/float@1.0.2", "", {}, "sha512-Ddb+kVXlXst9d+R9PfTIxh1EdNkgoRe5tOX6t01f1lYWOvJnSPDBlG241QLzcyPdoNTsblLUdujGSE4RzrTZGQ=="],
-
-    "@protobufjs/path": ["@protobufjs/path@1.1.2", "", {}, "sha512-6JOcJ5Tm08dOHAbdR3GrvP+yUUfkjG5ePsHYczMFLq3ZmMkAD98cDgcT2iA1lJ9NVwFd4tH/iSSoe44YWkltEA=="],
-
-    "@protobufjs/pool": ["@protobufjs/pool@1.1.0", "", {}, "sha512-0kELaGSIDBKvcgS4zkjz1PeddatrjYcmMWOlAuAPwAeccUrPHdUqo/J6LiymHHEiJT5NrF1UVwxY14f+fy4WQw=="],
-
-    "@protobufjs/utf8": ["@protobufjs/utf8@1.1.2", "", {}, "sha512-b1UQwcEZ4yCnMCD8DAL1VlbvBJE9/IX4FTIp7BG1xYpf29SLazLSrqUkj4w7Y5y7cCVP6E5tcqqcI0xemPkHug=="],
-
    "@puppeteer/browsers": ["@puppeteer/browsers@3.2.1", "", { "dependencies": { "modern-tar": "^0.8.0", "yargs": "^18.0.0" }, "peerDependencies": { "proxy-agent": ">=8.0.1", "yauzl": "^2.10.0 || ^3.4.0" }, "optionalPeers": ["proxy-agent", "yauzl"], "bin": { "browsers": "lib/main-cli.js" } }, "sha512-KDz+3qDRdBAlRlMjmKyj6dEs33YHTk/xRHEENSXq6TNnhgoU15ruSHtEBeVF6OZ9tBDY55Se4P0nFMNsipzU9A=="],

    "@radix-ui/colors": ["@radix-ui/colors@1.0.1", "", {}, "sha512-xySw8f0ZVsAEP+e7iLl3EvcBXX7gsIlC1Zso/sPBW9gIWerBTgz6axrjU+MZ39wD+WFi5h5zdWpsg3+hwt2Qsg=="],
@@ -4127,7 +4114,7 @@

    "github-slugger": ["github-slugger@2.0.0", "", {}, "sha512-IaOQ9puYtjrkq7Y0Ygl9KDZnrf/aiUJYUpVf89y8kyaxbRG7Y1SrX/jaumrv81vc61+kiMempujsM3Yw7w5qcw=="],

-    "gitlab-ai-provider": ["gitlab-ai-provider@6.12.1", "", { "dependencies": { "@anthropic-ai/sdk": "^0.71.0", "@anycable/core": "^0.9.2", "graphql-request": "^6.1.0", "isomorphic-ws": "^5.0.0", "openai": "^6.16.0", "socket.io-client": "^4.8.1", "vscode-jsonrpc": "^8.2.1", "zod": "^3.25.76" }, "peerDependencies": { "@ai-sdk/provider": ">=3.0.0", "@ai-sdk/provider-utils": ">=4.0.0" } }, "sha512-Qn5iHqvjG8yktI5MWaUgdRR94l7O4WtYW0CAbhsCh1Tj0Fei/DeprOYPVyf4Nht1Ix6U2PXSYM32QOHI6Z2TDw=="],
+    "gitlab-ai-provider": ["gitlab-ai-provider@6.18.0", "", { "dependencies": { "@anthropic-ai/sdk": "^0.71.0", "@anycable/core": "^0.9.2", "graphql-request": "^6.1.0", "isomorphic-ws": "^5.0.0", "openai": "^6.16.0", "socket.io-client": "^4.8.1", "vscode-jsonrpc": "^8.2.1", "zod": "^3.25.76" }, "peerDependencies": { "@ai-sdk/provider": ">=3.0.0", "@ai-sdk/provider-utils": ">=4.0.0" } }, "sha512-dXTXkNt1SFCL7jGlqazHL6iUJEug21qq0tx2s+Tui7jWNUqIAHEzY9WY1+PnZlhxvB56jPPBcFAT0attpjnopg=="],

    "glob": ["glob@13.0.5", "", { "dependencies": { "minimatch": "^10.2.1", "minipass": "^7.1.2", "path-scurry": "^2.0.0" } }, "sha512-BzXxZg24Ibra1pbQ/zE7Kys4Ua1ks7Bn6pKLkVPZ9FZe4JQS6/Q7ef3LG1H+k7lUf5l4T3PLSyYyYJVYUvfgTw=="],

@@ -5047,8 +5034,6 @@

    "proto-list": ["proto-list@1.2.4", "", {}, "sha512-vtK/94akxsTMhe0/cbfpR+syPuszcuwhqVjJq26CuNDgFGj682oRBXOP5MJpv2r7JtE8MsiepGIqvvOTBwn2vA=="],

-    "protobufjs": ["protobufjs@7.6.5", "", { "dependencies": { "@protobufjs/aspromise": "^1.1.2", "@protobufjs/base64": "^1.1.2", "@protobufjs/codegen": "^2.0.5", "@protobufjs/eventemitter": "^1.1.1", "@protobufjs/fetch": "^1.1.1", "@protobufjs/float": "^1.0.2", "@protobufjs/path": "^1.1.2", "@protobufjs/pool": "^1.1.0", "@protobufjs/utf8": "^1.1.1", "@types/node": ">=13.7.0", "long": "^5.3.2" } }, "sha512-/FPD0nUc9jH6rfFjji9IBqOz4pcSE3CsT1m7Ep6Mdb0LxSUMj8hgl6GomOvZzpNpAqqGaXA0P3VSrZLFzIhQrw=="],
-
    "proxy-from-env": ["proxy-from-env@1.1.0", "", {}, "sha512-D+zkORCbA9f1tdWRK0RaCR3GPv50cMxcrz4X8k5LTSUD1Dkw47mKJEZQNunItRTkWwgtaUSo1RVFRIG9ZXiFYg=="],

    "pump": ["pump@3.0.4", "", { "dependencies": { "end-of-stream": "^1.1.0", "once": "^1.3.1" } }, "sha512-VS7sjc6KR7e1ukRFhQSY5LM2uBWAUPiOPa/A3mkKmiMwSmRFUITt0xuj+/lesgnCv+dPIEYlkzrcyXgquIHMcA=="],
@@ -6245,6 +6230,8 @@

    "@opencode/www/wrangler": ["wrangler@4.110.0", "", { "dependencies": { "@cloudflare/kv-asset-handler": "0.5.0", "@cloudflare/unenv-preset": "2.16.1", "blake3-wasm": "2.1.5", "esbuild": "0.28.1", "miniflare": "4.20260708.1", "path-to-regexp": "6.3.0", "unenv": "2.0.0-rc.24", "workerd": "1.20260708.1" }, "optionalDependencies": { "fsevents": "2.3.3" }, "peerDependencies": { "@cloudflare/workers-types": "^5.20260708.1" }, "optionalPeers": ["@cloudflare/workers-types"], "bin": { "wrangler": "bin/wrangler.js", "wrangler2": "bin/wrangler.js", "cf-wrangler": "bin/cf-wrangler.js" } }, "sha512-xZeXKYi7hxQRF5anL+v77RkufJNpF9f3Eqeyqq2QBsETpLZgh0Agj0jJ6JPtkbgn6ukZdh8OK5egsGPWIditgg=="],

+    "@opentelemetry/api-logs/@opentelemetry/api": ["@opentelemetry/api@1.9.1", "", {}, "sha512-gLyJlPHPZYdAk1JENA9LeHejZe1Ti77/pTeFm/nMXmQH/HFZlcS/O2XJB+L8fkbrNSqhdtlvjBVjxwUYanNH5Q=="],
+
    "@opentelemetry/instrumentation/@opentelemetry/api-logs": ["@opentelemetry/api-logs@0.220.0", "", { "dependencies": { "@opentelemetry/api": "^1.3.0" } }, "sha512-CmVa4ImJ+ynfrPMNaAXHET6Bhb44SwzmfyVJFq9ni2jgXJR/l7C6gfVFddNmHP+ZOkP9cf4f9DBe68qVLTHc9w=="],

    "@opentelemetry/sdk-trace/@opentelemetry/core": ["@opentelemetry/core@2.11.0", "", { "dependencies": { "@opentelemetry/semantic-conventions": "^1.29.0" }, "peerDependencies": { "@opentelemetry/api": ">=1.0.0 <1.10.0" } }, "sha512-7YP44XH0tV6+Mb54x2YGf84i7yi+31MBZlE8JwvozkxyTvXbSp10X7cI7YE49ChJ3shMJoBmCJF3+1QFBJctGA=="],
@@ -1,8 +1,8 @@
 {
  "nodeModules": {
-    "x86_64-linux": "sha256-WBGj4wU5PQs+0zbDbScFhJs0GL8qWnswIsJb/Kee5Qc=",
-    "aarch64-linux": "sha256-Z8hkyNMdqCpuv24q50QhO2RECINvENjBhLDY5FKcnFc=",
-    "aarch64-darwin": "sha256-HZHb9HcXMOc7bKR7eADQSyfWxX37GAcoRi/EMSpRiP4=",
-    "x86_64-darwin": "sha256-CcWCOAcW4t9ADs48fMhumSbDcQWdk91LpkG6fe/gWlk="
+    "x86_64-linux": "sha256-aQQQhaUlAhpfqzH0vNi0IJ1cg7FQHIKYzxeq5d8PZoU=",
+    "aarch64-linux": "sha256-r9aDFu3UYmudmmYPhzCrpFvQlaejXc8V1IzLtG3jZPc=",
+    "aarch64-darwin": "sha256-B0m41LelD7d61vPHGIZZSO/cU7gbjHDJHt6oxNRRM8Q=",
+    "x86_64-darwin": "sha256-9TWJsyI3Y6BMomtGSgqA1th9LpxpqP4F5Tl/GyexVYw="
  }
 }
@@ -2,7 +2,7 @@
  "$schema": "https://json.schemastore.org/package.json",
  "name": "opencode",
  "description": "AI-powered development tool",
-  "version": "2.0.14",
+  "version": "2.0.16",
  "private": true,
  "type": "module",
  "packageManager": "bun@1.4.2",
@@ -52,9 +52,9 @@
      "@octokit/rest": "22.0.0",
      "@hono/standard-validator": "0.2.0",
      "@hono/zod-validator": "0.4.2",
-      "@opentui/core": "0.5.10",
-      "@opentui/keymap": "0.5.10",
-      "@opentui/solid": "0.5.10",
+      "@opentui/core": "0.5.12",
+      "@opentui/keymap": "0.5.12",
+      "@opentui/solid": "0.5.12",
      "@tanstack/solid-virtual": "3.13.37",
      "@shikijs/stream": "4.4.3",
      "@standard-schema/spec": "1.1.0",
@@ -173,7 +173,7 @@
    "solid-js@1.9.15": "patches/solid-js@1.9.15.patch",
    "@ai-sdk/mistral@3.0.51": "patches/@ai-sdk%2Fmistral@3.0.51.patch",
    "gcp-metadata@8.1.2": "patches/gcp-metadata@8.1.2.patch",
-    "pacote@21.5.0": "patches/pacote@21.5.0.patch",
+    "pacote@21.5.1": "patches/pacote@21.5.1.patch",
    "@ai-sdk/google@3.0.73": "patches/@ai-sdk%2Fgoogle@3.0.73.patch",
    "@pierre/trees@1.0.0-beta.4": "patches/@pierre%2Ftrees@1.0.0-beta.4.patch",
    "@tanstack/virtual-core@3.17.8": "patches/@tanstack%2Fvirtual-core@3.17.8.patch",
@@ -10,11 +10,11 @@

 ## Conventions

-Per-type constructors live on the type, not as top-level re-exports. Use `Message.system(...)`, `Message.user(...)`, `Message.assistant(...)`, `Message.tool(...)`, `Message.media(...)`, `LanguageModel.make(...)`, `ToolDefinition.make(...)`, `ToolCallPart.make(...)`, `ToolResultPart.make(...)`, `ToolChoice.make(...)`, `ToolChoice.named(...)`, `SystemPart.make(...)`, and `GenerationOptions.make(...)` directly. The top-level `LLM` namespace is reserved for request-shaped call APIs: `LLM.request`, `LLM.generate`, `LLM.stream`, and `LLM.generateObject`. Use `LLMRequest.update(...)` when deriving canonical request data; do not add a duplicate `LLM.updateRequest(...)` path. Two ways to construct the same thing is one too many.
+Per-type constructors live on the type, not as top-level re-exports. Use `Message.system(...)`, `Message.user(...)`, `Message.assistant(...)`, `Message.tool(...)`, `Message.media(...)`, `LanguageModel.make(...)`, `ToolDefinition.make(...)`, `ToolCallPart.make(...)`, `ToolResultPart.make(...)`, `ToolChoice.make(...)`, `ToolChoice.named(...)`, `SystemPart.make(...)`, and `GenerationOptions.make(...)` directly. The top-level `LLM` namespace is reserved for request-shaped call APIs: `LLM.request`, `LLM.generate`, `LLM.stream`, and `LLM.generateObject`. `LLM.generate`/`LLM.stream` and Promise `ai.llm.generate`/`ai.llm.stream` accept ergonomic input or an `LLMRequest`; both paths use the same canonical request. Core still builds, logs, replays, and updates that durable `LLMRequest` boundary. Use `LLMRequest.update(...)` when deriving canonical request data; do not add a duplicate `LLM.updateRequest(...)` path.

-Modality namespaces mirror `LLM` exactly: `Image.request`, `Image.generate`, `Image.stream` (later `Video`, `Speech`, `Transcription`). Common request fields (`images`, `mask`, `n`, `size`, `aspectRatio`, `seed`, `format`) lower natively or fail with a typed `AIError`; provider-native controls always live under `providerOptions`, never under a modality-specific `options` key.
+Modality namespaces mirror `LLM` exactly: `Image.request`, `Image.generate`, `Image.stream`, and the same for `Video`, `Speech`, and `Transcription`. Common request fields (`images`, `mask`, `n`, `size`, `aspectRatio`, `seed`, `format`) lower natively or fail with a typed `AIError`; provider-native controls always live under `providerOptions`, never under a modality-specific `options` key.

-Media payloads are always `Media.Asset` (`src/media.ts`). Construct them with `Media.bytes`, `Media.base64`, `Media.url`, `Media.ref`, `Media.fromDataUrl`, or `Media.file`; never introduce a parallel `data: string | Uint8Array` shape. `MediaPart.media`, `ImageRequest.images`/`mask`, `ImageResponse.images`, and the `media` `LLMEvent` all share it. Protocols branch on `asset.source.type` and `asset.kind` and use `ProviderShared.inlineMedia` / `requireInlineMedia` / `mediaUrl` / `MediaInput.refID` rather than re-deriving base64 or URL handling.
+Media payloads are always `Media.Asset` (`src/media.ts`). Construct them with `Media.bytes`, `Media.base64`, `Media.url`, `Media.ref`, `Media.fromDataUrl`, or `Media.file`; never introduce a parallel `data: string | Uint8Array` shape. `MediaPart.media`, `ImageRequest.images`/`mask`, `ImageResponse.images`, and the `media` `LLMEvent` all share it. Protocols branch on `asset.source.type` and `asset.kind` and use `ProviderShared.requireInlineMedia` / `inlineRequired` / `mediaUrl` / `mediaReference` and `MediaInput.inlineBytes` / `refID` rather than re-deriving base64 or URL handling.

 `schema/messages.ts → media.ts → route/executor-service.ts` is an accepted runtime dependency from the schema layer on the executor service tag: `Media.Asset.bytes()` must be able to download `url` sources, and the tag lives in that leaf module precisely so the schema barrel never imports the executor implementation (which imports the schema barrel back). Do not move the tag into `route/executor.ts` or import `route/executor.ts` from `src/schema/*` or `src/media.ts`.

@@ -55,7 +55,7 @@ const request = LLM.request({
  prompt: "Say hello.",
 })

-const response = yield * LLMClient.generate(request)
+const response = yield * LLMClient.generate(request) // inside Effect.gen
 ```

 `LLM.request(...)` builds an `LLMRequest`. `LLMClient.generate(...)` reads the executable route carried by `request.model.route`, builds the provider-native body, asks the route's transport for a real `HttpClientRequest.HttpClientRequest`, sends it through `RequestExecutor.Service`, parses the provider stream into common `LLMEvent`s, and finally returns an `LLMResponse`.
@@ -96,7 +96,13 @@ When a provider supports multiple physical transports, selection remains executi

 ### Media Routes

-Media does not fit the SSE-frames-to-event-state-machine LLM route. `MediaRoute.make(...)` (`src/route/media.ts`) composes a `MediaProtocol` kind with `Endpoint` and `Auth` and owns the transport plumbing: `http` option merging, URL/query rendering, auth headers, JSON vs multipart encoding, and handing the response back to the protocol. `MediaProtocol.inline` (`src/route/media-protocol.ts`) is `body.from(request)` plus `response.decode(response, context)`; use `MediaProtocol.decodeJson` / `text` / `bytes` so decode failures retain the raw body and HTTP context. `Generation` (`src/generation.ts`) is the provider-neutral handle for a queued generation over a `GenerationRoute` (`status`, `result`, `cancel`, `pollHint`); the first video route implements it. Image protocol files follow the same section order as LLM protocols and declare unsupported common fields once through `MediaInput.rejectUnsupported`.
+Media does not fit the SSE-frames-to-event-state-machine LLM route. `MediaRoute.inline(...)` / `queued(...)` / `stream(...)` (`src/route/media.ts`) compose a `MediaProtocol` kind with `Endpoint` and `Auth` and own the transport plumbing: `http` option merging, URL/query rendering, auth headers, JSON vs multipart encoding, and handing the response back to the protocol. `MediaProtocol.inline` (`src/route/media-protocol.ts`) is `body.from(request)` plus `response.decode(response, context)`; each protocol declares `const route = MediaProtocol.identity({ id, name, provider })` once and decodes through `route.decodeJson` / `route.text` / `route.decodeStarted` so decode failures retain the raw body and HTTP context, raising `route.unsupported(operation, message)` for requests it cannot lower, and passes `route` as the first argument to `MediaProtocol.inline` / `queued` / `stream`. `Generation` (`src/generation.ts`) is the provider-neutral handle for a queued generation over a `GenerationRoute` (`status`, `result`, `cancel`). Image protocol files follow the same section order as LLM protocols and declare unsupported common fields once through the protocol's `unsupported` list.
+
+`MediaProtocol.queued` is the submit-then-poll kind every video route uses: `start` (body + decode into `{ token, snapshot }`), `status`, `result`, and optional `cancel` (with `activeOnly` when the provider's cancel endpoint deletes finished work, as Runway's does: the route refreshes status first and skips terminal generations), each addressed by a route-owned `token` whose `Schema.Codec` makes it serializable. `MediaRoute.inline` and `MediaRoute.queued` compose the two kinds with `Endpoint` and `Auth`; the queued route decodes the token once at the boundary (`start` output or `resume` input) and closes over it in a token-free `GenerationRoute` (`status`/`result`/`cancel` are plain Effects), so `Generation` never sees the token's shape and only carries the encoded JSON for persistence. Polls reuse the route's auth and deployment headers plus the request's `http` overlay after `start`, and resolve relative paths against the route base URL (provider-issued absolute URLs such as fal's `status_url` pass through). `result` is always its own GET even when the provider returns output inside the status document, so `Generation.await` behaves the same after `start` and after `resume`. `PollContext.auth` carries only what `Auth` added or changed so protocols can hand download credentials to output assets as transient `Media.Asset.headers` (Veo) — never part of `source` or JSON. Status strings map through a per-protocol `STATUS` table via `MediaProtocol.status`; terminal generations without output fail through `output.ended` / `output.contentPolicy` with the provider document on `reason.body`. `GenerationAwaitOptions` (`AwaitOptions` in `src/generation.ts`, `{ poll?: Poll }`) is the one options type for `await`, `events`, `Video.generate`, and `Video.stream`.
+
+`MediaProtocol.stream` is the incremental kind every speech route uses, with the same discipline as LLM protocols. `MediaRoute.stream` submits the caller's request as `MediaProtocol.Addressed<Request>` (`{ ...request, mode }`, `mode: "generate" | "stream"`), so one provider stays one protocol: `body.from`, the endpoint path, and `frames` read `request.mode` to pick the body, path, and framing. `frames(bytes, context)` returns frames — `Framing.sse`, `Framing.lines`, `Framing.document` (a single-document response shaped like a streamed record), or the raw `bytes` for chunked audio. `initial()` is fresh per-response parser state; `step` folds each frame into it and emits modality events; `finish(state, context)` runs once after the last frame with the request, body, and observed `http` (header-only usage lives there) and emits exactly one terminal event or fails with `route.incomplete()`. Keep parser state to real accumulators and derive anything the request or body determines in `finish`. `generate` runs the same stream and folds it with the modality's `collect`. Request-derived URL parameters go on the body's `query` (array values repeat the parameter), applied before route and caller `http.query`. Decode frames with `route.decodeFrame` and raise stream-time failures with `route.frameError` (the frame stays on `reason.body`); protocols never thread HTTP context, because the route fills `reason.http` on stream errors that lack it. Speech protocols share `protocols/utils/speech-stream.ts` for deltas, timestamps, voice ids, PCM and container descriptions, and the terminal asset.
+
+Every modality route is the inline | stream | queued union (transcription uses all three: OpenAI and Gemini stream, Deepgram is inline, AssemblyAI is queued), every client is `MediaClient.make(Service, { modality, responseEvents })` (`src/media-client.ts`), which dispatches on the route's `kind`, and every model composes through `composeRoute`. fal queue protocols come from `protocols/utils/fal-queue.ts`, bodies are `json`, `multipart`, or `binary` (a raw upload), and a queued protocol that must upload media before submitting implements `start.prepare` (`MediaProtocol.Prepare`; AssemblyAI `/v2/upload`).

 ### URL Construction

@@ -106,7 +112,7 @@ For providers where the URL is derived from typed inputs (Azure resource name, B

 ### Provider Facades

-Provider-facing APIs are configured facades over route values. Endpoint/auth/resource/API-version setup happens before model selection, and model selectors accept only a model or deployment id. Media models use per-modality selectors on the same facade (`openai.image(id)`, later `.video` / `.speech` / `.transcription`) that mirror `openai.responses(id)`; the one-word overlap with the request namespace is accepted over a second construction path:
+Provider-facing APIs are configured facades over route values. Endpoint/auth/resource/API-version setup happens before model selection, and model selectors accept only a model or deployment id. Media models use per-modality selectors on the same facade (`openai.image(id)`, `.speech(id)`, `.transcription(id)`, `google.video(id)`) that mirror `openai.responses(id)`; the one-word overlap with the request namespace is accepted over a second construction path:

 ```ts
 const openai = OpenAI.configure({ apiKey, baseURL })
@@ -269,6 +275,7 @@ Use this order for every protocol module:
 ### Rules

 - Keep protocol files focused on the protocol. Move provider-specific projection, signing, media normalization, or other bulky transformations into `src/protocols/utils/*`.
+- Send `tool.inputSchema` as given. `prepareRequest` applies the tool schema rules (`ToolSchemaProjection.tools`) once per request, including tools in namespaces. A protocol whose API needs a model family's rules for every model declares `sanitizer` instead of transforming schemas itself.
 - Use `Effect.fn("Provider.fromRequest")` for request body construction entrypoints. Use `Effect.fn(...)` for event handlers that yield effects; keep purely synchronous handlers as plain functions returning a `StepResult` that the dispatcher lifts via `Effect.succeed(...)`.
 - Parser state owns terminal information. The state machine records finish reason, usage, and pending tool calls; emit one terminal `finish` event (or `provider-error`) for each completed response. If a provider splits reason and usage across events, merge them in parser state before flushing.
 - Emit exactly one terminal `finish` event for a completed response, normally after a matching `step-finish`. Use `stream.terminal` to stop reading when the provider has a completion sentinel; use `stream.onHalt` when the final event must be flushed after the framed stream ends.
@@ -1,40 +1,38 @@
 # @opencode/ai

-Schema-first language model and image-generation APIs built with Effect.
+Schema-first APIs for text, images, video, speech, and transcription, built with Effect.

 ```ts
-import { Effect, Layer } from "effect"
-import { LLM, LLMClient } from "@opencode/ai"
-import { RequestExecutor } from "@opencode/ai/route"
+import { Effect } from "effect"
+import { AIClient, LLM } from "@opencode/ai"
 import { OpenAI } from "@opencode/ai/providers"

 const openai = OpenAI.configure({ apiKey: process.env.OPENAI_API_KEY })

-const request = LLM.request({
-  model: openai.responses("gpt-4o-mini"), // `.chat(...)` selects the Chat Completions API instead
-  system: "You are concise.",
-  prompt: "Say hello in one short sentence.",
-  generation: { maxTokens: 40 },
-})
-
 const program = Effect.gen(function* () {
-  const response = yield* LLMClient.generate(request)
+  const response = yield* LLM.generate({
+    model: openai.responses("gpt-4o-mini"), // `.chat(...)` selects the Chat Completions API instead
+    system: "You are concise.",
+    prompt: "Say hello in one short sentence.",
+    generation: { maxTokens: 40 },
+  })
  console.log(response.text)
 })

-const llmLayer = LLMClient.layer.pipe(Layer.provide(RequestExecutor.fetchLayer))
-
-await Effect.runPromise(program.pipe(Effect.provide(llmLayer)))
+// Every modality client plus the HTTP request executor; `AIClient.layerWith(executor)` swaps the executor.
+await Effect.runPromise(program.pipe(Effect.provide(AIClient.layer)))
 ```

-Run `LLMClient.stream(request)` instead of `generate` when you want incremental `LLMEvent`s. The event stream is provider-neutral — same shape across OpenAI Chat, OpenAI Responses, Anthropic Messages, Gemini, Bedrock Converse, and any OpenAI-compatible deployment.
+Run `LLM.stream(...)` instead of `generate` when you want incremental `LLMEvent`s. Both accept input or a prebuilt
+`LLM.request(...)`. The event stream is provider-neutral — same shape across OpenAI Chat, OpenAI Responses,
+Anthropic Messages, Gemini, Bedrock Converse, and any OpenAI-compatible deployment.

-The same configured facade names image models. `Image.request` resolves the provider's image route from the ref and
-returns `Media.Asset`s with lazily decoded bytes:
+The same configured facade names image, video, speech, and transcription models. `Image.generate` resolves the
+provider's image route from the model and returns `Media.Asset`s with lazily decoded bytes:

 ```ts
 import { NodeFileSystem } from "@effect/platform-node"
-import { Image, ImageClient, Media } from "@opencode/ai"
+import { Image, Media } from "@opencode/ai"

 const image = Effect.gen(function* () {
  const response = yield* Image.generate({
@@ -46,21 +44,38 @@ const image = Effect.gen(function* () {
  yield* Media.write(response.image, "./garden.png")
 })

-// `asset.bytes()` / `Media.write` also need the executor, so merge it into the environment instead of hiding it.
-const imageLayer = ImageClient.layer.pipe(Layer.provideMerge(RequestExecutor.fetchLayer))
-
-await Effect.runPromise(image.pipe(Effect.provide(imageLayer), Effect.provide(NodeFileSystem.layer)))
+// `Media.file` / `Media.write` use the Effect `FileSystem` service; provide your platform's layer.
+await Effect.runPromise(image.pipe(Effect.provide(AIClient.layer), Effect.provide(NodeFileSystem.layer)))
 ```

-Prefer promises? `@opencode/ai/promise` exposes the same LLM and image APIs over one managed runtime:
+Advanced: each client also has its own `layer`, which requires `RequestExecutor.Service`. Compose client layers with
+`Layer.provideMerge`, not `Layer.provide`: `asset.bytes()`, `Media.write`, and Gemini's `media` output parts need the
+executor too, and hiding it fails type-checking with `RequestExecutorService` left in the requirements.
+
+To share a policy such as logging across every client, wrap the executor once with `RequestExecutor.middleware`:
+
+```ts
+import { RequestExecutor } from "@opencode/ai/route"
+
+const logged = RequestExecutor.middleware((request, next) =>
+  Effect.log(`${request.method} ${request.url}`).pipe(Effect.andThen(next(request))),
+)
+
+const everything = AIClient.layerWith(logged) // or AI.make({ layer: logged })
+```
+
+Prefer promises? `@opencode/ai/promise` exposes the same LLM and media APIs over one managed runtime, plus asset
+helpers; `ai.file` and `ai.write` load `node:fs/promises` on first use, so no Effect `FileSystem` is needed:

 ```ts
 import { AI } from "@opencode/ai/promise"

 const ai = AI.make()
-const text = await ai.llm.generate({ model: openai.responses("gpt-4o-mini"), prompt: "Say hello." })
+const input = { model: openai.responses("gpt-4o-mini"), prompt: "Say hello." }
+const text = await ai.llm.generate(input)
 const generated = await ai.image.generate({ model: openai.image("gpt-image-2"), prompt: "A lighthouse" })
-for await (const event of ai.llm.stream({ model: openai.responses("gpt-4o-mini"), prompt: "Stream hello." })) {
+await ai.write(generated.image, "./lighthouse.png") // also ai.file(path), ai.bytes(asset), ai.base64(asset), ai.materialize(asset)
+for await (const event of ai.llm.stream(ai.llm.request(input))) {
  // LLMEvent
 }
 await ai.dispose()
@@ -322,10 +337,9 @@ and `moonshot/responses`; each exports `model(modelID, settings)`.
 MiniMax defaults to its Messages API and reads `MINIMAX_API_KEY` when `apiKey` is omitted:

 ```ts
-import { Effect, Layer } from "effect"
-import { LLM, LLMClient } from "@opencode/ai"
+import { Effect } from "effect"
+import { AIClient, LLM } from "@opencode/ai"
 import { MiniMax } from "@opencode/ai/providers"
-import { RequestExecutor } from "@opencode/ai/route"

 const minimax = MiniMax.configure({ apiKey: process.env.MINIMAX_API_KEY })
 const request = LLM.request({
@@ -335,8 +349,7 @@ const request = LLM.request({
  generation: { maxTokens: 1536 },
 })

-const layer = LLMClient.layer.pipe(Layer.provide(RequestExecutor.fetchLayer))
-const response = await Effect.runPromise(LLMClient.generate(request).pipe(Effect.provide(layer)))
+const response = await Effect.runPromise(LLM.generate(request).pipe(Effect.provide(AIClient.layer)))
 console.log(response.text)
 ```

@@ -405,14 +418,14 @@ Use `Image.generate` for one-off generation or editing:
 import { Image, Media } from "@opencode/ai"

 const generation = Image.generate({
-  model: meta("muse-image-1.0"),
+  model: meta.image("muse-image-1.0"),
  prompt: "A flat black square on a white background.",
  n: 1,
  providerOptions: { reasoningStrength: "low" },
 })

 const edit = Image.generate({
-  model: meta("muse-image-1.0"),
+  model: meta.image("muse-image-1.0"),
  prompt: "Make the square purple.",
  images: [Media.bytes(imageBytes, "image/webp")],
  format: "png",
@@ -459,6 +472,25 @@ const program = Effect.gen(function* () {
 })
 ```

+Common fields are portable in shape, not in support. Unsupported fields fail with a typed `AIError` before any network
+call rather than being dropped, so check this table before swapping only the `model`:
+
+| Provider              | `n` | `size`    | `aspectRatio` | `seed` | `format` | `images`                         | `mask`              |
+| --------------------- | --- | --------- | ------------- | ------ | -------- | -------------------------------- | ------------------- |
+| OpenAI                | ✓¹  | ✓         | ✗             | ✗      | ✓        | ✓                                | ✓                   |
+| Google (Gemini)       | 1   | ✗         | ✓             | ✓      | ✗        | ✓ (no public URLs)               | ✗                   |
+| xAI                   | ✓   | ✗         | ✓             | ✗      | ✗        | ✓                                | ✗                   |
+| Z.ai                  | ✗   | ✓         | ✗             | ✗      | ✗        | ✗                                | ✗                   |
+| Meta                  | ✓   | ✓ (hint)  | ✗             | ✗      | ✓        | ✓                                | ✗                   |
+| Black Forest Labs     | 1   | per model | per model     | ✓      | ✓        | per model (1–8)                  | `flux-pro-1.0-fill` |
+| fal                   | ✓   | per model | per model     | ✓      | ✓        | 1 (several on `/edit`, `/multi`) | ✓                   |
+| Replicate             | ✗   | ✗         | ✗             | ✗      | ✗        | ✗ (use `providerOptions`)        | ✗                   |
+| Stability `image`     | 1   | ✗         | ✓             | ✓      | ✓        | 1 (not on `core`)                | ✗                   |
+| Stability `upscale()` | ✗   | ✗         | ✗             | ✓      | ✓        | exactly 1 (required)             | ✗                   |
+
+✓ lowers natively; ✗ fails whenever the field is set (including `n: 1`); `1` means `n > 1` fails. ¹ `Image.stream` on OpenAI generates one image. fal
+rejects `size` and `aspectRatio` together; which one a fal or BFL model takes depends on the model.
+
 `Media.Asset` is the one asset type shared by image requests, image responses, LLM messages, and tool results.
 `asset.source` is the serializable `Media.Source` (`bytes`, `base64`, `url`, or `ref`); `asset.bytes()`,
 `asset.base64()`, and `asset.dataUrl()` decode or download lazily and cache; `asset.materialize()` pulls a `url`
@@ -468,9 +500,8 @@ asset into owned bytes before the provider URL expires. Construct assets with `M
 Pass ordered image inputs to the same method for editing, composition, or image-conditioned generation:

 ```ts
-const response =
-  yield *
-  Image.generate({
+const composed = Effect.gen(function* () {
+  const response = yield* Image.generate({
    model,
    prompt: "Combine these product photos into one studio scene",
    images: [
@@ -481,23 +512,25 @@ const response =
    providerOptions,
    http,
  })
+  return response.images
+})
 ```

 `Media.ref(provider, id)` represents provider file handles such as OpenAI file IDs or Gemini Files URIs; routes
-only forward refs that belong to their own provider. Raw strings are not accepted as image inputs, avoiding
-ambiguity between base64, URLs, and provider IDs. Empty or omitted `images` uses text-to-image generation; a
-non-empty array selects the provider's edit behavior without enforcing provider image-count limits locally. OpenAI
+only forward refs that belong to their own provider (OpenAI, xAI, and Gemini images accept them). No shipped route
+returns a ref yet, and `asset.bytes()` / `materialize()` on a ref fail by design. Raw strings are not accepted as
+image inputs, avoiding ambiguity between base64, URLs, and provider IDs. Empty or omitted `images` uses text-to-image generation; a
+non-empty array selects the provider's edit behavior (see the table above for routes that limit the count). OpenAI
 uses multipart for byte/data-URL edits and its JSON reference body for URL or file-ID edits. The common `mask`
 field selects inpainting; routes that cannot honor it fail with `UnsupportedOperation`:

 ```ts
-yield *
-  Image.generate({
-    model: openai.image("gpt-image-2"),
-    prompt,
-    images: [Media.bytes(sourceBytes, "image/png")],
-    mask: Media.bytes(maskBytes, "image/png"),
-  })
+const inpainted = Image.generate({
+  model: openai.image("gpt-image-2"),
+  prompt,
+  images: [Media.bytes(sourceBytes, "image/png")],
+  mask: Media.bytes(maskBytes, "image/png"),
+})
 ```

 On multipart requests, `http.body` can override option fields but not structural `model`, `prompt`, `image[]`,
@@ -508,31 +541,31 @@ not accept image inputs. These cases fail with a typed `AIError` before network
 Provider-native image options belong to each request. Raw `http.body` fields have final precedence over them:

 ```ts
-yield *
-  Image.generate({
-    model: openai.image("gpt-image-2"),
-    prompt,
-    providerOptions: { quality: "medium" },
-    http,
-  })
+const medium = Image.generate({
+  model: openai.image("gpt-image-2"),
+  prompt,
+  providerOptions: { quality: "medium" },
+  http,
+})
 ```

 xAI image models use the same request API with xAI-native controls:

 ```ts
-yield *
-  Image.generate({
-    model: XAI.configure({ apiKey })("any-model-id"),
-    prompt,
-    n: 2,
-    aspectRatio: "16:9",
-    providerOptions: {
-      resolution: "1k",
-      responseFormat: "b64_json",
-      future_option: true,
-    },
-    http,
-  })
+import { XAI } from "@opencode/ai/providers"
+
+const xai = Image.generate({
+  model: XAI.configure({ apiKey }).image("any-model-id"),
+  prompt,
+  n: 2,
+  aspectRatio: "16:9",
+  providerOptions: {
+    resolution: "1k",
+    responseFormat: "b64_json",
+    future_option: true,
+  },
+  http,
+})
 ```

 Google's current Gemini image models use the same direct API:
@@ -542,7 +575,7 @@ import { Google } from "@opencode/ai/providers"

 const googleProgram = Effect.gen(function* () {
  const response = yield* Image.generate({
-    model: Google.configure({ apiKey })("any-model-id"),
+    model: Google.configure({ apiKey }).image("any-model-id"),
    prompt: "A robot tending a rooftop garden",
    aspectRatio: "16:9",
    seed: 42,
@@ -567,23 +600,88 @@ their mapped aliases, and `http.body` is the final deep overlay. The selected mo
 Z.ai image models infer open Z.ai-native options from the selected model:

 ```ts
-yield *
-  Image.generate({
-    model: ZAI.configure({ apiKey })("any-model-id"),
-    prompt,
-    providerOptions: {
-      quality: "hd",
-      userID: "user-123",
-      future_option: true,
-    },
-    http,
-  })
+import { ZAI } from "@opencode/ai/providers"
+
+const zai = Image.generate({
+  model: ZAI.configure({ apiKey }).image("any-model-id"),
+  prompt,
+  providerOptions: {
+    quality: "hd",
+    userID: "user-123",
+    future_option: true,
+  },
+  http,
+})
 ```

 Z.ai does not include trustworthy MIME metadata for output URLs, so generated images use
 `application/octet-stream` until materialized. Output URLs expire after 30 days; call `asset.materialize()` and
 persist the bytes promptly if they must remain available.

+### Partial images
+
+OpenAI's GPT image models stream previews. `Image.stream` sends `stream: true` with `partialImages` (0–3, default 2)
+and emits `image-partial` events before each final `image`; `Image.generate` keeps the plain JSON request:
+
+```ts
+import { Stream } from "effect"
+import { ImageEvent } from "@opencode/ai"
+
+const previews = Image.stream({
+  model: openai.image("gpt-image-2"),
+  prompt: "A lighthouse at dusk",
+  providerOptions: { partialImages: 2 },
+}).pipe(Stream.runForEach((event) => (ImageEvent.is.imagePartial(event) ? showPreview(event.image) : Effect.void)))
+```
+
+The provider may send fewer previews than requested when the final image is ready first.
+
+### Queued image providers
+
+Black Forest Labs, fal, Replicate, and Stability's creative upscaler are submit-then-poll routes. `Image.generate`
+and `Image.stream` poll for you (pass `{ poll }` to tune the interval and timeout); `Image.start` returns a
+`Generation` whose `token` is serializable JSON for `Image.resume` in another process:
+
+```ts
+import { BlackForestLabs, Stability } from "@opencode/ai/providers"
+
+const bfl = BlackForestLabs.configure({ apiKey: process.env.BFL_API_KEY })
+
+const submit = Effect.gen(function* () {
+  const generation = yield* Image.start({ model: bfl.image("flux-2-pro"), prompt, size: "1024x768" })
+  persist({ provider: "black-forest-labs", modelID: "flux-2-pro", token: generation.token })
+})
+
+const finish = Effect.gen(function* () {
+  const saved = load()
+  const resumed = yield* Image.resume(bfl.image(saved.modelID), saved.token)
+  return yield* resumed.await({ poll: { interval: "2 seconds" } })
+})
+```
+
+The token carries no route identity, so persist the provider and model ID alongside it: `resume` needs the model.
+
+- **Black Forest Labs** — results are downloaded before returning, because `result.sample` expires in 10 minutes.
+- **Replicate** — inputs are model-defined, so only `prompt` lowers: sizing, count, seed, format, and files go in
+  `providerOptions` under the model's names, with files as `Media.Asset` (data URLs up to 256 KB, larger by URL).
+  Outputs are removed an hour after the prediction completes. `Prefer: wait=60` in `headers` or `http.headers` holds
+  the submission open so a fast prediction costs one result read.
+- **Stability** — `stability.image(id)` generates inline; `stability.upscale()` is the creative upscaler, queued:
+
+```ts
+const stability = Stability.configure({ apiKey: process.env.STABILITY_API_KEY })
+const upscaled = Effect.gen(function* () {
+  const small = yield* Media.file("./small.png")
+  return yield* Image.generate(
+    { model: stability.upscale(), prompt: "A lighthouse", images: [small] },
+    { poll: { interval: "5 seconds" } },
+  )
+})
+```
+
+Imagen is not available: Google shut it down on the Gemini API, and Vertex discontinued the Imagen 4 models on
+2026-06-30. `Google.image(...)` uses Gemini-native image models.
+
 Conversational image generation remains part of the LLM interaction. OpenAI Responses exposes it through its hosted image tool:

 ```ts
@@ -600,20 +698,257 @@ const program = Effect.gen(function* () {
 })
 ```

-The hosted result is represented as a provider-executed tool call and tool result, and the generated image is also emitted as a first-class `media` `LLMEvent` (`response.message` then carries a `media` part). Gemini image-capable models emit the same `media` event for inline image output. Retaining `response.message` preserves the generated image for continuation on both routes.
+The hosted result is represented as a provider-executed tool call and a tool result whose content carries the generated image as a file. Gemini image-capable models instead emit a first-class `media` `LLMEvent` for inline image output (`response.message` then carries a `media` part). Retaining `response.message` preserves the generated image for continuation on both routes.
+
+## Video generation
+
+Video mirrors `Image` with one difference: every provider is asynchronous, so the route is a submit-then-poll
+`Generation`. Models come from `.video(...)` selectors on the `Google` (Veo), `XAI`, `Fal`, and `Runway` facades.
+Common fields (`frames`, `references`, `video`, `durationSeconds`, `aspectRatio`, `resolution`, `audio`, `n`, `seed`,
+`negativePrompt`) lower natively or fail with a typed `AIError` before any network call; provider-native controls live
+under `providerOptions`, inferred from the selected model.
+
+```ts
+import { Video } from "@opencode/ai"
+import { Google, Runway } from "@opencode/ai/providers"
+
+const google = Google.configure({ apiKey: process.env.GOOGLE_GENERATIVE_AI_API_KEY })
+
+// Simple: submit and wait.
+const program = Effect.gen(function* () {
+  const response = yield* Video.generate(
+    {
+      model: google.video("veo-3.1-generate-preview"),
+      prompt: "Panning wide shot of a calico kitten sleeping in the sunshine",
+      aspectRatio: "16:9",
+      resolution: "1080p",
+      durationSeconds: 8,
+      providerOptions: { personGeneration: "allow_adult" },
+    },
+    { poll: { interval: "10 seconds", timeout: "10 minutes" } },
+  )
+  // Veo serves files for two days behind the API key. The asset knows the deadline (`expiresAt`) and carries the
+  // download credentials only on the live instance (`asset.headers`), never in `source` or JSON: materialize
+  // before persisting, or the persisted URL cannot be fetched again.
+  return yield* response.video.materialize()
+})
+
+// Explicit control: keep the handle, persist the token, resume elsewhere.
+const controlled = Effect.gen(function* () {
+  const generation = yield* Video.start({ model: google.video("veo-3.1-generate-preview"), prompt })
+  generation.id // provider operation / task / request id
+  generation.status // "queued" | "running" | "completed" | "failed" | "cancelled" | "expired"
+  generation.token // route-owned JSON: `{ operation }`, `{ requestID }`, `{ taskID }`, or fal's follow-up URLs
+  // The token carries no route identity: persist the provider and model ID alongside it, since `resume` needs the model.
+  const saved = JSON.stringify(generation.token)
+
+  const resumed = yield* Video.resume(google.video("veo-3.1-generate-preview"), JSON.parse(saved))
+  return yield* resumed.await({ poll: { interval: "10 seconds" } })
+})
+
+// Progress as a stream: generation-queued | generation-progress | video | finish.
+const events = Video.stream({ model: Runway.configure({ apiKey }).video("gen4.5"), prompt }, { poll })
+```
+
+Status polls, result fetches, cancels, and asset downloads all run through the same request executor with the route's
+auth. `Generation.await` and `Generation.events` fail with a
+`Timeout` reason when `poll.timeout` (default 10 minutes) elapses. Failed,
+cancelled, and expired generations fail typed with the provider's terminal document on `reason.body`; moderation
+outcomes (Veo `raiMediaFilteredReasons`, xAI `respect_moderation`, Runway `SAFETY.*` codes) surface as `notices` when
+a video is still returned and as a `ContentPolicy` reason when nothing is.
+
+Provider notes:
+
+- **Google Veo** takes inline bytes only (materialize `url` assets first); `frames.last` requires `frames.first`;
+  audio is always on, so `audio: false` fails typed; one video per request. Output URLs need the API key to
+  download, which the returned asset holds transiently (see above).
+- **xAI** sends a `video` input to `/videos/edits`, or `/videos/extensions` with `providerOptions.mode: "extend"`.
+  `seed` and `negativePrompt` are not supported.
+- **fal** endpoints are model-specific: `durationSeconds`, `references`, and `frames.last` fail typed and belong in
+  `providerOptions` under the model's own names (`duration: "8s"`, `end_image_url`, …). Auth is
+  `Authorization: Key <FAL_KEY>`.
+- **Runway** expects pixel ratios in `aspectRatio` for most models (`"1280:720"`), pins `X-Runway-Version`, reports
+  `usage: { type: "credits" }`, and its output URLs expire after 24–48 hours.
+
+The promise client exposes the same surface: `ai.video.start(...)` resolves to a handle with `await`, `events`,
+`result`, `refresh`, `cancel`, and `token`; `ai.video.generate`, `ai.video.resume(model, token)`, and
+`ai.video.stream` mirror the Effect API. The handle's `status` and `progress` are a snapshot from when it was
+created; `refresh()` resolves to a new handle.
+
+```ts
+import { ai } from "@opencode/ai/promise"
+
+const generation = await ai.video.start({ model, prompt })
+for await (const event of generation.events({ poll: { interval: 10_000 } })) console.log(event.type)
+const video = await generation.result({ signal })
+await ai.write(video.video, "./kite.mp4")
+```
+
+## Speech generation
+
+Speech (text-to-speech) is one request whose response is parsed incrementally, so every route supports both
+`Speech.generate` (the whole file) and `Speech.stream` (audio chunks as they arrive). Models come from `.speech(...)`
+selectors on the `OpenAI`, `Google` (Gemini TTS), `ElevenLabs`, `Cartesia`, and `Deepgram` facades. Common fields
+(`voice`, `format`, `speed`, `language`, `instructions`, `timestamps`) lower natively or fail with a typed `AIError`
+before any network call; provider-native controls live under `providerOptions`, inferred from the selected model.
+
+```ts
+import { Media, Speech, SpeechClient, SpeechEvent } from "@opencode/ai"
+import { ElevenLabs, OpenAI } from "@opencode/ai/providers"
+
+const openai = OpenAI.configure({ apiKey: process.env.OPENAI_API_KEY })
+
+// The whole file, written to disk.
+const program = Effect.gen(function* () {
+  const response = yield* Speech.generate({
+    model: openai.speech("gpt-4o-mini-tts"),
+    text: "Hello from OpenCode.",
+    voice: "coral",
+    format: "mp3",
+    instructions: "Warm and unhurried.",
+  })
+  response.audio // Media.Asset with bytes; headerless PCM carries info.encoding / sampleRate / channels
+  response.usage // undefined: OpenAI reports tokens only on SSE streams (Gemini: tokens; ElevenLabs: credits; Deepgram: characters)
+  yield* Media.write(response.audio, "hello.mp3")
+})
+
+// Chunks as they arrive: audio-delta* (interleaved with timestamps) then one finish carrying the assembled asset.
+const events = Speech.stream({
+  model: ElevenLabs.configure({ apiKey }).speech("eleven_flash_v2_5"),
+  text: "Hello from OpenCode.",
+  voice: "JBFqnCBsd6RMkjVDRZzb",
+  format: "pcm",
+  timestamps: true,
+}).pipe(
+  Stream.tap((event) => {
+    if (SpeechEvent.is.audioDelta(event)) return play(event.chunk)
+    if (SpeechEvent.is.timestamps(event)) return highlight(event.items) // { text, startSeconds, endSeconds }[]
+    return Effect.void
+  }),
+)
+```
+
+`voice` is the provider's own identifier — a name on OpenAI and Gemini (`"coral"`, `"Kore"`), a voice id on
+ElevenLabs and Cartesia. `{ id }` selects an OpenAI custom voice (`{ id: "voice_1234" }`) and means the same as the
+plain string elsewhere. There is no cross-provider voice catalog. `format` is the container-level word (`mp3`, `wav`,
+`pcm`, `opus`, `aac`, `flac`); sample rates and bitrates live under `providerOptions`, and a value the route cannot
+produce fails as `UnsupportedOperation`. Streams buffer every chunk so `finish` can carry the whole clip.
+
+Provider notes:
+
+- **OpenAI** streams over SSE (`stream_format: "sse"`), which is also the only place it reports token usage; `tts-1`
+  and `tts-1-hd` do not support SSE and stream the raw audio body instead. `pcm` is 24 kHz 16-bit mono. `language`
+  and `timestamps` are not supported.
+- **Gemini TTS** returns the provider's default output: WAV for Gemini 3.8 TTS `generate`, raw 16-bit PCM
+  (`audio/L16;codec=pcm;rate=24000`) otherwise. `pcm` is the only explicit `format` it accepts, and it fails typed on
+  Gemini 3.8 `generate`; the route never wraps PCM as WAV. Style is directed in the text, so `instructions` and
+  `speed` fail typed. Only `gemini-3.1-flash-tts-preview` and later support streaming. Two-speaker audio goes through
+  `providerOptions.speechConfig.multiSpeakerVoiceConfig`.
+- **ElevenLabs** requires `voice` (the path voice id) and authenticates with `xi-api-key`. `format` maps to the
+  `output_format` query parameter (`mp3_44100_128`, `pcm_24000`, `wav_24000`, `opus_48000_64`);
+  `providerOptions.outputFormat` sets the exact string. WAV is only available from `generate`. `timestamps: true`
+  selects the `with-timestamps` endpoints and yields character-level alignment. `instructions` is not supported.
+- **Cartesia** requires `voice` and pins `Cartesia-Version`. `generate` defaults to MP3 from `/tts/bytes`; streams
+  and `timestamps: true` (word-level) use `/tts/sse`, which only serves raw PCM. `providerOptions.sampleRate`,
+  `bitRate`, and `encoding` complete `output_format`. No usage is reported.
+- **Deepgram** Aura's voice is the model id (`aura-2-thalia-en`), so `voice` and `language` fail typed. `format`
+  and `providerOptions` lower to query parameters (`encoding`, `container`, `sample_rate`, `bit_rate`); `pcm` is
+  `linear16` without a container. Auth is `Authorization: Token <DEEPGRAM_API_KEY>`.
+
+The promise client mirrors the Effect API; `ai.speech.stream` is an `AsyncIterable`.
+
+```ts
+import { ai } from "@opencode/ai/promise"
+
+const response = await ai.speech.generate({ model, text: "Hello from OpenCode.", voice: "coral" })
+await ai.write(response.audio, "hello.mp3")
+
+for await (const event of ai.speech.stream({ model, text: "Hello from OpenCode.", voice: "coral" })) {
+  if (event.type === "audio-delta") player.write(event.chunk)
+}
+```
+
+## Transcription
+
+Transcription (speech-to-text) is the one modality whose providers use every route kind: OpenAI and Gemini stream,
+Deepgram answers inline, and AssemblyAI is queued. `Transcription.generate` and `Transcription.stream` work on all of
+them; `Transcription.start` / `resume` return a `Generation` on queued routes and fail with `UnsupportedOperation`
+elsewhere. Models come from `.transcription(...)` selectors on the `OpenAI`, `Google`, `Deepgram`, and `AssemblyAI`
+facades. Common fields (`language`, `prompt`, `timestamps: "none" | "segment" | "word"`, `diarize`, `speakers`) lower
+natively or fail with a typed `AIError` before any network call; a route may return more than asked.
+
+```ts
+import { Console, Effect, Stream } from "effect"
+import { Media, Transcription, TranscriptionEvent } from "@opencode/ai"
+import { AssemblyAI, Deepgram, OpenAI } from "@opencode/ai/providers"
+
+const openai = OpenAI.configure({ apiKey: process.env.OPENAI_API_KEY })
+
+const program = Effect.gen(function* () {
+  const audio = yield* Media.file("./call.mp3")
+
+  // Speaker-labelled segments; labels are provider-native strings ("A", "0", "spk:0").
+  const response = yield* Transcription.generate({
+    model: Deepgram.configure({ apiKey }).transcription("nova-3"),
+    audio,
+    diarize: true,
+    timestamps: "word",
+  })
+  response.text // "Hello from OpenCode."
+  response.segments // [{ text, startSeconds, endSeconds, speaker: "0" }]
+  response.words // [{ text, startSeconds, endSeconds, speaker, confidence }]
+  response.language // the provider's own value, lowercased ("en", "english", "en_us")
+
+  // Text deltas as the model transcribes, then one finish carrying the whole transcript.
+  yield* Transcription.stream({ model: openai.transcription("gpt-4o-mini-transcribe"), audio }).pipe(
+    Stream.tap((event) => (TranscriptionEvent.is.textDelta(event) ? Console.log(event.delta) : Effect.void)),
+    Stream.runDrain,
+  )
+
+  // Queued: persist the token with the provider and model ID (the token alone cannot pick the model), resume, and await.
+  const model = AssemblyAI.configure({ apiKey }).transcription("universal-3-5-pro")
+  const generation = yield* Transcription.start({ model, audio })
+  const resumed = yield* Transcription.resume(model, JSON.parse(JSON.stringify(generation.token)))
+  const transcript = yield* resumed.await({ poll: { interval: "3 seconds" } })
+})
+```
+
+Inline routes emit only `finish` from `stream` (no faked deltas); queued routes emit `generation-queued` /
+`generation-progress` before it.
+
+Provider notes:
+
+- **OpenAI** takes inline audio only; `diarize` needs `gpt-4o-transcribe-diarize`, timestamps need `whisper-1`, and `whisper-1` does not stream.
+- **Gemini** needs a transcribe model (`gemini-3.5-transcribe`); `prompt` and `speakers` fail typed.
+- **Deepgram** detects the language unless `language` is set; vocabulary goes in `providerOptions.keyterm`.
+- **AssemblyAI** uploads inline audio before submitting and is the only route that accepts `speakers`.
+
+The promise client mirrors the Effect API:
+
+```ts
+const audio = await ai.file("./call.mp3")
+const text = (await ai.transcription.generate({ model, audio })).text
+for await (const event of ai.transcription.stream({ model, audio })) if (event.type === "text-delta") write(event.delta)
+const generation = await ai.transcription.start({ model: assemblyai, audio })
+const transcript = await generation.await({ poll: { interval: 3_000 } })
+```

 ## Public API

 - **`LLM.request({...})`** — build a provider-neutral `LLMRequest`. Accepts ergonomic inputs (`system: string`, `prompt: string`) that normalize into the canonical Schema classes.
- **`LLM.generate` / `LLM.stream`** — re-exported from `LLMClient` for one-import use.
+- **`LLM.generate` / `LLM.stream`** — run direct input or an `LLMRequest` through `LLMClient` for one-import use.
 - **`Message.user(...)` / `Message.assistant(...)` / `Message.tool(...)`** — message constructors from the canonical schema model.
 - **`LanguageModel.make(...)` / `ToolCallPart.make(...)` / `ToolResultPart.make(...)` / `ToolDefinition.make(...)`** — model and tool-related constructors from the canonical schema model.
 - **`LLMEvent.is.*`** — typed guards (`is.textDelta`, `is.toolCall`, `is.finish`, …) for filtering streams.
- **`Image.request` / `Image.generate` / `Image.stream`** — generate images through a provider-neutral image request and response model.
+- **`Image.request` / `generate` / `stream` / `start` / `resume`** — images over inline, streaming (partial previews), and queued routes through a provider-neutral request and response model.
 - **`ImageClient`** — Effect service and layer for image execution, parallel to `LLMClient`.
 - **`Media`** — the shared asset type (`Media.Asset`, `Media.Source`) and constructors used by messages, tool results, and media requests.
 - **`Generation`** — provider-neutral handle for an in-flight media generation (`await`, `refresh`, `cancel`, `events`) used by queued media routes.
- **`@opencode/ai/promise`** — `AI.make({ layer? })` and a default `ai` client exposing `llm` and `image` as Promise / `AsyncIterable` APIs.
+- **`Video.request` / `generate` / `stream` / `start` / `resume`** — queued video generation through a provider-neutral request; `VideoClient` is its Effect service and layer.
+- **`Speech.request` / `Speech.generate` / `Speech.stream`** — text-to-speech through a provider-neutral request; `SpeechClient` is its Effect service and layer.
+- **`Transcription.request` / `generate` / `stream` / `start` / `resume`** — speech-to-text over inline, streaming, and queued routes; `TranscriptionClient` is its Effect service and layer.
+- **`AIClient.layer` / `AIClient.layerWith(executor)`** — every modality client plus the request executor in one layer.
+- **`@opencode/ai/promise`** — `AI.make({ layer? })` and a default `ai` client exposing `llm`, `image`, `video`, `speech`, and `transcription` as Promise / `AsyncIterable` APIs, plus `file`, `write`, `bytes`, `base64`, and `materialize` for assets.

 ## Testing

@@ -676,11 +1011,13 @@ This is different from prompt caching, server-side history storage, or truncatio
 Prefer this operation, where supported, when the application owns compaction policy and durable context updates.

 ```ts
-const result = yield * LLMClient.compact(request)
-const next = LLMRequest.update(request, {
-  messages: result.replacement,
+const compacted = Effect.gen(function* () {
+  const result = yield* LLMClient.compact(request)
+  const next = LLMRequest.update(request, {
+    messages: result.replacement,
+  })
+  return yield* LLMClient.generate(next)
 })
-const response = yield * LLMClient.generate(next)
 ```

 `replacement` replaces the complete input window. Do not append it to the original transcript or extract only the encrypted item: the provider may retain additional messages in its output. Retained user and assistant messages remain ordinary messages with typed text, media, or reasoning parts, in their original order. Provider-specific message IDs, status, and phase use `providerMetadata`, not a raw output array hidden in an assistant message. Unsupported returned item types fail explicitly.
@@ -696,16 +1033,16 @@ The input must still fit the model's context window. Explicit compaction is not
 OpenAI Responses also exposes a separate, explicitly selected mechanism:

 ```ts
-const result =
-  yield *
-  LLMClient.compact(request, {
+const checkpoint = Effect.gen(function* () {
+  const result = yield* LLMClient.compact(request, {
    mechanism: "trigger",
    webSocket, // Optional: without it, the request uses HTTP/SSE.
  })

-result.checkpoint // Successful encrypted CompactionPart.
-result.responseID
-result.usage
+  result.checkpoint // Successful encrypted CompactionPart.
+  result.responseID
+  result.usage
+})
 ```

 This appends a native `compaction_trigger` control item to the full input and sends a normal Responses request, with tools and instructions retained, `stream: true`, `store: false`, and parallel tool calls enabled. It removes normal-answer text/output-format controls, forced tool choices, output-token/tool-call limits, and automatic `context_management`. Body overlays cannot replace `input` or supply `previous_response_id`/`conversation`; the complete canonical history is required for safe stateless replay. Request metadata, auth, headers, query parameters, service tier, and supported prompt-cache settings are preserved.
@@ -719,9 +1056,11 @@ The supplied WebSocket executor can reuse a compatible append baseline for the c
 Trigger support is separate from endpoint support. Only the OpenAI Responses route advertises it; Azure, xAI, Chat, and compatible Responses routes do not inherit it. Untyped calls still fail before sending: missing route capabilities return `UnsupportedOperation`, while unknown mechanism names and invalid inputs return `InvalidRequest`. Dynamic callers must narrow for the selected mechanism:

 ```ts
-if (LLMClient.canCompact(request, { mechanism: "trigger" })) {
-  const result = yield * LLMClient.compact(request, { mechanism: "trigger" })
-}
+const narrowed = Effect.gen(function* () {
+  if (LLMClient.canCompact(request, { mechanism: "trigger" })) {
+    const result = yield* LLMClient.compact(request, { mechanism: "trigger" })
+  }
+})
 ```

 This capability describes protocol implementation, **not universal availability on OpenAI API deployments**. The host application owns subscription/deployment eligibility, OAuth, endpoint selection, and deployment-specific headers. Local protocol/socket tests do not establish live provider support.
@@ -730,9 +1069,10 @@ This capability describes protocol implementation, **not universal availability

 `providerOptions.contextManagement` lets the provider decide when to compact during an ordinary `generate` or `stream` call. This is an advanced option for callers that own persistence and recovery: persist the complete assistant message, including its checkpoint, before continuing. Enabling the option does not provide durable checkpoint storage, interruption recovery, or model-switch policy. Keep the prior context until a successful checkpoint has been persisted.

-Inside an `Effect.gen`, enable OpenAI compaction with typed provider options:
+Enable OpenAI compaction with typed provider options:

 ```ts
+import { Effect } from "effect"
 import { LLM, LLMClient, LLMRequest, Message } from "@opencode/ai"
 import { OpenAI } from "@opencode/ai/providers"

@@ -743,9 +1083,11 @@ const request = LLM.request({
    contextManagement: [{ type: "compaction", compactThreshold: 200_000 }],
  },
 })
-const response = yield * LLMClient.generate(request)
-const next = LLMRequest.update(request, {
-  messages: [...request.messages, response.message, Message.user("Continue")],
+const continued = Effect.gen(function* () {
+  const response = yield* LLMClient.generate(request)
+  return LLMRequest.update(request, {
+    messages: [...request.messages, response.message, Message.user("Continue")],
+  })
 })
 ```

@@ -973,7 +1315,7 @@ Compose a route with `Route.make({ protocol, endpoint, auth, framing, ... })`. T

 ## Effect

-This package is built on Effect. Public methods return `Effect` or `Stream`; provide `LLMClient.layer` for LLM dispatch and `ImageClient.layer` for image dispatch, then import the provider/protocol modules for the routes you use. The example at `example/tutorial.ts` is a runnable walkthrough.
+This package is built on Effect. Public methods return `Effect` or `Stream`; provide `AIClient.layer` (or `AIClient.layerWith(executor)`) for every modality, then import the provider/protocol modules for the routes you use. The example at `example/tutorial.ts` is a runnable walkthrough.

 ## See also

@@ -1,6 +1,7 @@
 # Media generation in `@opencode/ai` — public API direction

-Status: proposal. Branch `media-support`.
+Status: phases 1–4 implemented (through Image queued routes and partial images; ElevenLabs Scribe transcription
+pending); phase 5 proposal.

 ## Goal

@@ -11,11 +12,11 @@ The design below is derived from a survey of the raw provider APIs (OpenAI, Gemi
 ## What the survey forces

 1. **Three execution shapes, everywhere.** Inline sync (OpenAI images, all TTS, Gemini), async job with polling or webhook (every video provider, BFL, fal, Replicate, AssemblyAI), and bidirectional streams (ElevenLabs/Cartesia/Deepgram WS, realtime). Video has no sync provider at all.
-2. **Output is never just bytes.** base64, signed URLs with TTLs from 10 minutes (BFL) to 2 days (Veo), URLs that need auth plus redirect (Veo), separate download endpoints (Sora `/content?variant=`), raw bodies (Stability, TTS). Multi-output is the norm.
+2. **Output is never just bytes.** base64, signed URLs with TTLs from 10 minutes (BFL, so its route downloads before returning via `PollContext.materialize`) to 2 days (Veo), URLs that need auth plus redirect (Veo), separate download endpoints (Sora `/content?variant=`), raw bodies (Stability, TTS). Multi-output is the norm.
 3. **Inputs have roles.** First/last frame, mask, style/subject reference, source video for edit/extend, reference audio, prior generation id, provider-side file handles (`file_id`, `gs://`, `runway://`, `mm_file://`).
 4. **Partial streaming is modality-specific.** Images: a few whole partial frames. Audio: ordered chunks plus timestamp events. Jobs: status/progress/logs. Video: none.
 5. **Usage is a union**: tokens, seconds, characters (often only in headers), credits, compute time.
-6. **Moderation can be partial success** (Veo strips audio but returns video). Deprecations are constant (Sora API shuts down 2026-09-24, Imagen on Gemini API 2026-08-17).
+6. **Moderation can be partial success** (Veo strips audio but returns video). Deprecations are constant (Sora API shuts down 2026-09-24; Imagen is shut down on the Gemini API and past its 2026-06-30 discontinuation date on Vertex).

 ## Where existing SDKs are weak and we should not be

@@ -40,7 +41,7 @@ The design below is derived from a survey of the raw provider APIs (OpenAI, Gemi

 ### Model selection

-A model value is built as `OpenAI.configure({ apiKey }).responses("gpt-5")` or `.image("gpt-image-2")`: `configure` fixes credentials, endpoint, and defaults; the selector fixes which of the provider's APIs to hit and binds the typed `providerOptions` generic. Media follows the same shape with one selector per modality — `openai.image(id)` today, `.video(id)` / `.speech(id)` / `.transcription(id)` as those modalities land — mirroring `openai.responses(id)`. `Image.request` accepts `ImageModel` only, exactly as `LLM.request` accepts `LanguageModel`.
+A model value is built as `OpenAI.configure({ apiKey }).responses("gpt-5")` or `.image("gpt-image-2")`: `configure` fixes credentials, endpoint, and defaults; the selector fixes which of the provider's APIs to hit and binds the typed `providerOptions` generic. Media follows the same shape with one selector per modality — `.image(id)`, `.video(id)`, `.speech(id)`, `.transcription(id)` on the facades that offer each — mirroring `openai.responses(id)`. `Image.request` accepts `ImageModel` only, exactly as `LLM.request` accepts `LanguageModel`.

 ```ts
 import { OpenAI, Google } from "@opencode/ai/providers"
@@ -54,7 +55,7 @@ Speech.request({ model: openai.speech("gpt-4o-mini-tts"), text })
 Transcription.request({ model: openai.transcription("gpt-4o-transcribe"), audio })
 ```

-The request namespace and the selector share one word (`Image.request` + `.image(...)`). That redundancy is accepted: a callable facade returning a lazily resolved ref would be a second way to construct the same model, and the type machinery to infer `providerOptions` through it is not worth one word. Where a provider has two APIs for one modality, the selectors stay explicit (`openai.chat`, a future `google.imagen`), and one default per modality per provider is part of the facade definition (OpenAI image → Images API, Google image → Gemini-native since Imagen on the Gemini API shuts down 2026-08-17). Provider package entrypoints keep `model(modelID, settings)` per modality-specific path, e.g. `@opencode/ai/providers/openai/responses`.
+The request namespace and the selector share one word (`Image.request` + `.image(...)`). That redundancy is accepted: a callable facade returning a lazily resolved ref would be a second way to construct the same model, and the type machinery to infer `providerOptions` through it is not worth one word. Where a provider has two routes for one modality, the selectors stay explicit (`openai.chat`, `stability.image` inline vs `stability.upscale()` queued), and one default per modality per provider is part of the facade definition (OpenAI image → Images API, Google image → Gemini-native; Imagen is shut down, so there is no `google.imagen`). The facade selector (`openai.image(id)`) is the public path for media models. Modality-specific package entrypoints (`model(modelID, settings)` beside today's LLM paths such as `@opencode/ai/providers/openai/responses`) are deferred until Core has a modality-aware model resolver; Core's resolver accepts only `LanguageModel` today.

 ### `Media` — the asset type

@@ -66,7 +67,7 @@ import { Media } from "@opencode/ai"
 Media.Source =
  | { type: "bytes";  data: Uint8Array; mediaType: string }
  | { type: "base64"; data: string;     mediaType: string }
-  | { type: "url";    url: string; mediaType?: string; expiresAt?: number; headers?: Record<string, string> }
+  | { type: "url";    url: string; mediaType?: string; expiresAt?: number }
  | { type: "ref";    provider: ProviderID; id: string; mediaType?: string }   // file_id, gs://, runway://, prior generation

 class Media.Asset {
@@ -76,6 +77,7 @@ class Media.Asset {
  readonly info?: { width?; height?; durationSeconds?; sampleRate?; channels?; encoding?; format? }
  readonly expiresAt?: number
  readonly providerMetadata?: ProviderMetadata
+  readonly headers?: Record<string, string>     // transient download credentials (Veo); never in source/JSON

  bytes(): Effect<Uint8Array, AIError, RequestExecutor.Service>   // downloads/decodes lazily, cached
  base64(): Effect<string, AIError, RequestExecutor.Service>
@@ -85,10 +87,18 @@ class Media.Asset {

 Media.bytes(data, mediaType?)      Media.base64(data, mediaType?)
 Media.url(url, options?)           Media.ref(provider, id)
-Media.file(path)                   // Bun/Node: reads + sniffs; Effect FileSystem variant for layers
-Media.write(asset, path)           // convenience, uses FileSystem
+Media.file(path)                   // Effect<Asset, AIError, FileSystem>: reads + sniffs
+Media.write(asset, path)           // Effect<void, AIError, FileSystem | RequestExecutor.Service>
 ```

+`Media.file` and `Media.write` stay Effect-only: bring your platform's `FileSystem` layer. The Promise client owns the
+runtime path: `ai.file(path)` and `ai.write(asset, path)` read and write through `node:fs/promises` (loaded on first
+use) with the same media-type sniffing and `InvalidRequest` failures, and `ai.bytes`, `ai.base64`, and
+`ai.materialize` run the asset methods in its runtime.
+
+A `ref` source is accepted as input only by routes whose provider issues file handles. No shipped route produces one
+yet, so `bytes()` and `materialize()` on a ref fail by design until a producer exists.
+
 Raw-PCM outputs (Gemini TTS, Cartesia raw, Deepgram WS) carry `info.encoding/sampleRate/channels` because there is no container header.

 ### Modality namespaces
@@ -103,128 +113,253 @@ import { OpenAI, Google, ElevenLabs, Fal } from "@opencode/ai/providers"
 #### Image

 ```ts
-const request = Image.request({
-  model: openai.image("gpt-image-2"),
-  prompt: "A robot tending a rooftop garden",
-  images: [Media.file("./ref.png")],           // references / edit sources
-  mask: Media.file("./mask.png"),
-  n: 2,
-  size: "1536x1024",                            // or aspectRatio: "3:2"
-  seed: 7,
-  format: "webp",
-  providerOptions: { quality: "high", background: "transparent" },   // typed per model
+Effect.gen(function* () {
+  const request = Image.request({
+    model: openai.image("gpt-image-2"),
+    prompt: "A robot tending a rooftop garden",
+    images: [yield* Media.file("./ref.png")],     // references / edit sources
+    mask: yield* Media.file("./mask.png"),
+    n: 2,
+    size: "1536x1024",                            // OpenAI sizes by pixels; Gemini/xAI take aspectRatio instead
+    format: "webp",
+    providerOptions: { quality: "high", background: "transparent" },   // typed per model
+  })
+
+  const response = yield* Image.generate(request) // ImageResponse
+  response.image                                  // Media.Asset (first)
+  response.images                                 // Media.Asset[]
+  response.usage                                  // Usage union (see below)
+  response.notices                                // moderation / partial-result notices
+
+  Image.stream(request)                           // Stream<ImageEvent>
+  // ImageEvent: generation-queued | generation-progress | image-partial { index, image } | image { index, image } | finish { usage }
 })
-
-const response = yield* Image.generate(request)   // ImageResponse
-response.image                                    // Media.Asset (first)
-response.images                                   // Media.Asset[]
-response.usage                                    // Usage union (see below)
-response.notices                                  // moderation / partial-result notices
-
-yield* Image.stream(request)                      // Stream<ImageEvent>
-// ImageEvent: generation-queued | generation-progress | image-partial { index, image } | image { index, image } | finish { usage }
 ```

+`size` and `aspectRatio` are not interchangeable; each route rejects fields it cannot lower — see the portability table
+in the README's Image generation section.
+
 Editing is not a separate function; `images`/`mask` on the request select the edit path in the route (OpenAI `/images/edits`, Gemini multimodal parts, xAI `/images/edits`). Routes that cannot honor `mask` fail with `Unsupported`.

+`ImageRoute` is the inline | stream | queued union, dispatched on `route.kind`, like every modality route. `Image.stream` on a streaming route emits `image-partial` previews before each `image`; on a queued route it emits `generation-queued` / `generation-progress` observations, then the result's `image` and `finish` events.
+
 #### Video

+Shipped in phase 2 (`src/video.ts`, `src/video-client.ts`, protocols `google-video`, `xai-video`, `fal-video`, `runway-video`).
+
 ```ts
-const request = Video.request({
-  model: google.video("veo-3.1-generate-preview"),
-  prompt: "Panning wide shot of a calico kitten sleeping in the sunshine",
-  frames: { first: Media.file("./start.png"), last: Media.file("./end.png") },
-  references: [Media.url("https://…/style.png")],
-  video: Media.ref("openai", "video_123"),      // edit / extend / remix source
-  durationSeconds: 8,
-  aspectRatio: "16:9",
-  resolution: "1080p",
-  audio: true,
-  n: 1,
-  providerOptions: { personGeneration: "dont_allow", negativePrompt: "text, watermark" },
+Effect.gen(function* () {
+  const request = Video.request({
+    model: google.video("veo-3.1-generate-preview"),
+    prompt: "Panning wide shot of a calico kitten sleeping in the sunshine",
+    frames: { first: yield* Media.file("./start.png"), last: yield* Media.file("./end.png") },
+    references: [yield* Media.file("./style.png")],
+    video: Media.bytes(previous, "video/mp4"),    // edit / extend source
+    durationSeconds: 8,
+    aspectRatio: "16:9",
+    resolution: "1080p",
+    audio: true,
+    n: 1,
+    seed: 7,
+    negativePrompt: "text, watermark",            // common, not provider-native
+    providerOptions: { personGeneration: "allow_adult" },
+  })
+
+  // Simple: wait for it.
+  const response = yield* Video.generate(request, { poll: { interval: "10 seconds", timeout: "10 minutes" } })
+  response.video                                     // Media.Asset: url (expiresAt on Veo and Runway; transient `headers` for Veo downloads)
+  response.usage                                     // credits on Runway; the other three report none (xAI's usage.cost_in_usd_ticks is not decoded)
+  response.notices                                   // Veo raiMediaFilteredReasons → filtered, xAI respect_moderation → moderated
+  yield* response.video.materialize()                // pull bytes before the URL expires
+
+  // Explicit generation control.
+  const generation = yield* Video.start(request)     // Generation<VideoResponse>
+  generation.id; generation.status; generation.progress; generation.position; generation.token
+  yield* generation.await({ poll })                  // VideoResponse
+  yield* generation.cancel()                         // fal PUT cancel_url, Runway DELETE /tasks/{id}; Veo and xAI succeed without a request
+
+  // Resume from another process. The token is validated against the route's codec and refreshed once. It carries no
+  // route identity, so persist the provider and model ID alongside it: `resume` needs the model.
+  const resumed = yield* Video.resume(model, JSON.parse(saved))
+
+  // Progress as a stream.
+  Video.stream(request, { poll })                    // Stream<VideoEvent>: generation-queued { id, position } | generation-progress { id, progress } | video { index, video } | finish { usage, notices }
 })
-
-// Simple: wait for it.
-const response = yield* Video.generate(request, { poll: { interval: "10 seconds", timeout: "10 minutes" } })
-response.video                                     // Media.Asset (url with expiresAt, or bytes when the route downloads)
-yield* response.video.materialize()                // pull bytes before the URL expires
-
-// Explicit generation control.
-const generation = yield* Video.start(request)     // Generation<VideoResponse>
-generation.id; generation.status; generation.progress; generation.token   // token is serializable JSON
-yield* generation.await({ poll })                  // VideoResponse
-yield* generation.cancel()
-
-// Resume from another process.
-const resumed = yield* Video.resume(model, token)  // Generation<VideoResponse>
-
-// Progress as a stream.
-yield* Video.stream(request)                       // Stream<VideoEvent>: generation-queued { position } | generation-progress { progress, logs } | video { index, video } | finish
 ```

-Webhooks: `Video.complete(model, token, webhook)` finishes a generation from a webhook payload without polling. Token shape is route-owned and opaque (Veo operation name, fal `response_url`, Runway task id).
+Tokens are route-owned JSON: Veo `{ operation }`, xAI `{ requestID }`, Runway `{ taskID }`, fal
+`{ requestID, statusURL, responseURL, cancelURL }` (fal's follow-up URLs are authoritative and absolute). Common-field
+lowering per provider: Veo takes inline media only, rejects `audio: false` and `n > 1`, and requires `frames.first`
+when `frames.last` is set; xAI rejects `n`, `seed`, and `negativePrompt` and routes a `video` input to edits or
+(`providerOptions.mode: "extend"`) extensions; fal rejects `n`, plus `durationSeconds`, `references`, and `frames.last`
+because the field names and enums differ per model; Runway passes `aspectRatio` through as its pixel `ratio` and
+rejects `n`.
+
+Deferred: `Video.complete(model, token, webhook)` (finish from a webhook payload without polling) and provider poll
+hints (none of the four providers emit one). Later providers: Luma, Kling, MiniMax, Replicate.

 #### Speech (TTS)

+Shipped in phase 3 (`src/speech.ts`, `src/speech-client.ts`, protocols `openai-speech`, `google-speech`,
+`elevenlabs-speech`, `cartesia-speech`, `deepgram-speech`; new `ElevenLabs`, `Cartesia`, and `Deepgram` facades).
+
 ```ts
 const request = Speech.request({
-  model: elevenlabs.speech("eleven_v3"),
+  model: elevenlabs.speech("eleven_flash_v2_5"),
  text: "Hello from OpenCode.",
-  voice: "JBFqnCBsd6RMkjVDRZzb",                  // name, uuid, or { id } — provider-normalized
+  voice: "JBFqnCBsd6RMkjVDRZzb",                  // provider-native identifier, or { id }
  format: "mp3",                                   // mp3 | wav | pcm | opus | aac | flac | (string & {})
  speed: 1.0,
  language: "en",
-  instructions: "Warm, unhurried.",
-  providerOptions: { stability: 0.5 },
+  instructions: "Warm, unhurried.",                // only OpenAI; elsewhere fails typed
+  timestamps: true,                                // request alignment; routes without it fail typed
+  providerOptions: { voice_settings: { stability: 0.5 } },
 })

-const response = yield* Speech.generate(request)   // SpeechResponse: audio: Media.Asset, timestamps?, usage
-yield* Speech.stream(request)                      // Stream<SpeechEvent>: audio-delta { chunk } | timestamps { words } | finish
+const response = yield* Speech.generate(request)   // SpeechResponse: audio: Media.Asset, timestamps?, usage?, providerMetadata?
+yield* Speech.stream(request)                      // Stream<SpeechEvent>: generation-queued | generation-progress | audio-delta { chunk } | timestamps { items } | finish { audio, usage? }
 ```

-Streaming TTS is first-class on day one: OpenAI `stream_format: sse`, ElevenLabs `/stream`, Cartesia SSE, Deepgram chunked. Input-streaming TTS (WS, text arrives incrementally) is a later `Speech.session(...)` scoped resource, not part of `generate`.
+Execution is `MediaProtocol.stream` for every provider: one request whose body is framed and folded by a `step`
+state machine, with `generate` running the same stream and collecting it. The route submits the request with its
+`mode` (`"generate" | "stream"`), which lets one provider stay one protocol — OpenAI adds `stream_format: "sse"` (except `tts-1`/`tts-1-hd`, which stream raw bytes), ElevenLabs appends
+`/stream`, Cartesia switches `/tts/bytes` to `/tts/sse`, Gemini switches `generateContent` to
+`streamGenerateContent`. The terminal `finish` event carries the assembled asset (every provider's stream is
+concatenable chunks), so stream consumers also get the whole file and `generate` is just "take `finish`, gather
+`timestamps`". The cost is memory: a stream holds every chunk until `finish`, so even a consumer that only plays deltas
+keeps the whole clip in memory. That is bounded by the providers' input text limits (a few minutes of audio); a
+long-form or session API would need an opt-out.
+
+**Voice.** `voice?: string | { id: string }`. A string is passed through as the provider's native identifier — a
+name on OpenAI and Gemini, a voice id on ElevenLabs (path segment) and Cartesia. `{ id }` selects an OpenAI custom
+voice and is treated as the plain string on routes that do not distinguish custom from built-in. Deepgram's voice is
+the model id (`aura-2-thalia-en`), so `voice` is `unsupported` there. There is no cross-provider voice catalog or
+name→id resolution. Multi-speaker (Gemini `speechConfig.multiSpeakerVoiceConfig`) and per-voice settings
+(ElevenLabs `voice_settings`) go through `providerOptions`.
+
+**Format and PCM.** `format` is container-level; provider sample rates and bitrates live under `providerOptions`
+(ElevenLabs `outputFormat`, Cartesia `sampleRate`/`bitRate`/`encoding`, Deepgram `encoding`/`container`/`sampleRate`/
+`bitRate`). Each protocol maps `format` to its native value (ElevenLabs `mp3_44100_128`/`pcm_24000`/`wav_24000`/
+`opus_48000_64`, Cartesia `{ container, encoding, sample_rate }`, Deepgram `encoding`+`container`) and declares the
+asset's media type rather than sniffing, because headerless PCM can look like an MPEG frame sync. Headerless PCM
+always carries `info.encoding`, `info.sampleRate`, and `info.channels`; its media type is the provider's declaration
+(Gemini `audio/L16;codec=pcm;rate=24000`, Deepgram's `content-type`) or `audio/pcm`. Gemini's asset follows the
+provider's declared type: WAV for Gemini 3.8 TTS `generate`, headerless PCM otherwise. The route never wraps PCM as WAV,
+so `pcm` is the only explicit `format` it accepts, and not on Gemini 3.8 `generate`. Every `format` value a route cannot
+produce (unknown to it, a container on Cartesia SSE, WAV on an ElevenLabs stream, anything but `pcm` on Gemini, `pcm` on
+Gemini 3.8 `generate`) fails the same way as an unsupported field: `UnsupportedOperation` with
+`operation: "media.format"`.
+
+**Timestamps.** `timestamps: true` on the request asks for alignment. ElevenLabs selects the `with-timestamps`
+endpoints (character-level, NDJSON when streaming); Cartesia sets `add_timestamps` on `/tts/sse` (word-level; a
+`generate` with timestamps collects the SSE stream). OpenAI, Gemini, and Deepgram reject it.
+
+Common-field lowering per provider:
+
+| Provider | `voice` | `speed` | `language` | `instructions` | `timestamps` | Usage |
+|---|---|---|---|---|---|---|
+| OpenAI | `voice` (name or `{ id }`) | `speed` | unsupported | `instructions` | unsupported | `tokens` from SSE `speech.audio.done` only |
+| Gemini | `prebuiltVoiceConfig.voiceName` | unsupported | `speechConfig.languageCode` | unsupported (direct in text) | unsupported | `tokens` from `usageMetadata` |
+| ElevenLabs | path voice id (required) | `voice_settings.speed` | `language_code` | unsupported | `with-timestamps` | `credits` from `character-cost` header |
+| Cartesia | `voice` (required) | `generation_config.speed` | `language` | unsupported | `add_timestamps` | none |
+| Deepgram | unsupported (voice is the model) | `speed` query | unsupported | unsupported | unsupported | `characters` from `dg-char-count` header |
+
+Deferred: `Speech.session(...)` — input-streaming TTS where text arrives incrementally over a WebSocket (ElevenLabs
+`stream-input`, Cartesia WebSocket contexts, Deepgram WebSocket speak) — is a separate scoped resource, not part of
+`generate`/`stream`, and ships with the realtime work in phase 5.

 #### Transcription (STT)

+Shipped as the second half of phase 3 (`src/transcription.ts`, `src/transcription-client.ts`, protocols
+`openai-transcription`, `google-transcription`, `deepgram-transcription`, `assemblyai-transcription`; new `AssemblyAI`
+facade).
+
 ```ts
 const request = Transcription.request({
-  model: openai.transcription("gpt-4o-transcribe"),
-  audio: Media.file("./call.wav"),
-  language: "en",
-  prompt: "Names: Shoubhit, OpenCode.",
-  timestamps: "word",                              // none | segment | word
+  model: openai.transcription("gpt-4o-transcribe-diarize"),
+  audio: yield* Media.file("./call.wav"),
+  language: "en",                                  // provider-native passthrough
+  timestamps: "segment",                           // none | segment | word
  diarize: true,
-  providerOptions: { chunkingStrategy: "auto" },
+  speakers: 2,                                     // exact speaker count (AssemblyAI only)
+  providerOptions: { known_speaker_names: ["agent"] },
 })

 const response = yield* Transcription.generate(request)
-response.text; response.segments; response.words; response.language; response.durationSeconds
-yield* Transcription.stream(request)               // Stream<TranscriptionEvent>: text-delta | segment | finish
+response.text; response.segments; response.words; response.language; response.durationSeconds; response.usage
+yield* Transcription.stream(request)               // Stream<TranscriptionEvent>: generation-queued | generation-progress | text-delta | segment | finish
+const generation = yield* Transcription.start(request) // queued routes only
+yield* Transcription.resume(model, token)
 ```

-Realtime STT over WebSocket is the same future `session` shape as input-streaming TTS.
+Transcription is the first modality whose providers span all three protocol kinds, and it needed no fourth kind.
+Every `MediaRoute` now carries its `kind`; `TranscriptionRoute` is the union of the inline, stream, and queued routes;
+`TranscriptionModel.fromRoute` is overloaded per protocol kind (arity picks the overload: `<Options>`,
+`<Options, Frame, State>`, `<Options, Token>`) and composes through the shared `composeRoute` (`src/media-model.ts`),
+which picks `MediaRoute.inline` / `stream` / `queued`; and `TranscriptionClient`, like every modality client, is
+`MediaClient.make` (`src/media-client.ts`), which dispatches on `route.kind`. `generate` on a queued route is `start`
+then `await`; `stream` on an inline route is the response as a single `finish`, and on a queued route it is the status
+observations followed by `finish`. `start` / `resume` on a non-queued route fail with `UnsupportedOperation` (`transcription.start`). The
+`finish` event carries the whole transcript (text, segments, words, language, duration, usage), so the stream route's
+`collect` is just "take `finish`".
+
+The route layer gained a `binary` body with array-valued `query` (Deepgram) and `Queued.start.prepare` (AssemblyAI's
+upload); `packages/ai/AGENTS.md` (Media Routes) describes both.
+
+Settled rules:
+
+- **Timestamps.** A granularity the selected route or model cannot produce fails as `UnsupportedOperation`
+  (`media.timestamps`), following Speech; a route that returns more than asked (Deepgram and AssemblyAI always return
+  words) is not stripped. Segments always carry start and end times: Gemini times each transcription part from its
+  word offsets, so segment timestamps and diarization also request word offsets there.
+- **Diarization.** `diarize` means segments (and words, where the provider labels them) carry `speaker`. Labels are
+  provider-native strings — OpenAI `A` or a known speaker name, Deepgram `0`, Gemini `spk:0`, AssemblyAI `A` — with no
+  cross-provider speaker model. `speakers` is the exact number of speakers to label, which AssemblyAI (`speakers_expected`, the only route that
+  accepts it) treats as a constraint rather than a hint.
+- **Language** is passed through (`language`, OpenAI `gpt-transcribe` `languages[]`, Gemini `languageCodes`,
+  AssemblyAI `language_code`). `response.language` is the provider's own value, lowercased but not normalized: an
+  ISO code on most routes (AssemblyAI's detection returns `en`), `english` from whisper-1. Deepgram and AssemblyAI
+  assume English unless asked to detect, so a missing `language` enables their detection.
+- **Gemini** requires a transcribe model; other model ids fail with `UnsupportedOperation` before the call, because
+  general models ignore `audioTranscriptionConfig` and answer conversationally. Streamed chunks carry whole speaker
+  turns (one part per turn), which join with a space.
+- **Streaming inline providers** emit only `finish`; deltas are never faked.
+- **Units.** AssemblyAI milliseconds and Gemini protobuf durations (`"0.400s"`) are normalized to seconds at the
+  protocol boundary.
+
+| Provider | Kind | Audio input | `timestamps` | `diarize` | Unsupported | Usage |
+|---|---|---|---|---|---|---|
+| OpenAI | stream (`stream: true` in `stream` mode; `whisper-1` ignores `stream`, so it emits only `finish`) | multipart `file` (inline only) | `whisper-1` (`verbose_json`); diarize model: `segment` | `gpt-4o-transcribe-diarize` (`diarized_json`) | `speakers`; `prompt` on the diarize model | `tokens` or `seconds` |
+| Gemini | stream (`generateContent` / `streamGenerateContent`) | `inlineData` or Gemini Files `fileData` | `audioTranscriptionConfig.wordTimestamp` | `audioTranscriptionConfig.diarization` | `prompt`, `speakers` | `tokens` |
+| Deepgram | inline | raw body, or JSON `{ url }` | words always; `segment` → `utterances` | `diarize_model=latest` + `utterances` | `prompt`, `speakers` | `seconds` (`metadata.duration`) |
+| AssemblyAI | queued (upload → submit → poll) | `/v2/upload` then `audio_url`, or a URL | words always; `segment` → `speaker_labels` | `speaker_labels` | — | `seconds` (`audio_duration`) |
+
+Deferred: `Transcription.session(...)` — realtime STT over WebSocket (Deepgram live, AssemblyAI streaming, ElevenLabs
+realtime, OpenAI realtime transcription) — is the same future scoped `session` shape as input-streaming TTS and ships
+with the realtime work in phase 5. ElevenLabs Scribe is not implemented yet.

 ### `Generation` — shared async execution

 ```ts
 class Generation<Response> {
  readonly id: string
-  readonly model: MediaModel
+  readonly route: GenerationRoute<Response>        // token-free: { status, result, cancel?: Effect } closed over the decoded token
  readonly token: unknown                          // route-owned serializable JSON
  readonly status: "queued" | "running" | "completed" | "failed" | "cancelled" | "expired"
  readonly progress?: number                       // 0..1, normalized
  readonly position?: number
-  readonly expiresAt?: number
  refresh(): Effect<Generation<Response>, AIError>
-  await(options?: { poll?: Poll }): Effect<Response, AIError>
+  result(): Effect<Response, AIError>
+  await(options?: GenerationAwaitOptions): Effect<Response, AIError>
  cancel(): Effect<void, AIError>
-  events(options?): Stream<GenerationEvent, AIError>
+  events(options?: GenerationAwaitOptions): Stream<GenerationEvent, AIError>   // fails with Timeout past poll.timeout, checked per observation
 }

-Poll = { interval?: Duration; timeout?: Duration; schedule?: Schedule }   // route may override from provider hints (`openai-poll-after-ms`)
+GenerationAwaitOptions = { poll?: Poll }
+Poll = { interval?: Duration; timeout?: Duration }
 ```

-`Generation` is not video-specific. Image routes on BFL, fal, and Replicate are queued; `Image.start` exists for them. A route declares itself `inline` or `queued`; `generate` on a queued route is `start` then `await`.
+`Generation` is not video-specific. Image routes on BFL, fal, Replicate, and Stability `upscale()` are queued; `Image.start` exists for them. A route declares itself `inline` or `queued`; `generate` on a queued route is `start` then `await`.

 ### Usage

@@ -250,15 +385,18 @@ const ai = AI.make()                                // ManagedRuntime over Reque
 // AI.make({ layer }) to inject a custom executor / recorder / middleware

 const image = await ai.image.generate({ model, prompt })
-await image.image.bytes()
+await ai.bytes(image.image)                         // also ai.base64, ai.materialize, ai.write(asset, path)
+const reference = await ai.file("./ref.png")

 for await (const event of ai.speech.stream({ model, text, voice })) { … }

-const generation = await ai.video.start({ model, prompt })
+const generation = await ai.video.start({ model, prompt })    // snapshot handle; refresh() returns a new one
+for await (const event of generation.events({ poll: { interval: 10_000 } })) { … }
 const video = await generation.await({ poll: { interval: 10_000 }, signal })
-const resumed = ai.video.resume(model, JSON.parse(saved))
+const resumed = await ai.video.resume(model, JSON.parse(saved)) // persist provider + model ID with the token

-const text = await ai.llm.generate({ model, prompt })   // closes today's gap: LLM has no promise API either
+const request = ai.llm.request({ model, prompt })
+const text = await ai.llm.generate(request)
 for await (const event of ai.llm.stream(request)) { … }

 await ai.dispose()
@@ -268,30 +406,38 @@ Streams become `AsyncIterable` via `Stream.toAsyncIterable`. `AIError` is thrown

 ### Providers

-Existing facades gain per-modality selectors; the modality routes each facade provides:
+Existing facades gain per-modality selectors; the modality routes each facade provides (*italics* are not
+implemented):

 | Facade | llm | image | video | speech | transcription | other |
 |---|---|---|---|---|---|---|
-| `OpenAI` | responses (default), chat | Images API | Sora (deprecated 2026-09-24) | ✓ | ✓ | |
-| `Google` | Gemini | Gemini-native (default), `imagen` | Veo | Gemini TTS | Gemini transcribe | |
+| `OpenAI` | responses (default), chat | Images API (stream) | *Sora skipped (decision 8)* | ✓ | ✓ | |
+| `Google` | Gemini | Gemini-native | Veo | Gemini TTS | `gemini-3.5-transcribe` | |
 | `XAI` | ✓ | ✓ | ✓ | | | |
-| `ElevenLabs` | | | | ✓ | Scribe | soundEffect, music |
-| `Fal` | | ✓ | ✓ | | | |
-| `Replicate`, `Runway`, `Luma`, `Kling`, `MiniMax`, `Deepgram`, `Cartesia`, `AssemblyAI`, `BlackForestLabs`, `Stability` | | per provider | | | | |
+| `ElevenLabs` | | | | ✓ | *Scribe (pending)* | *soundEffect, music (phase 5)* |
+| `Cartesia` | | | | ✓ | | |
+| `Deepgram` | | | | Aura | ✓ | |
+| `Fal` | | ✓ (queued) | ✓ | | | |
+| `AssemblyAI` | | | | | ✓ (queued) | |
+| `BlackForestLabs` | | ✓ (queued) | | | | |
+| `Replicate` | | ✓ (queued) | | | | |
+| `Stability` | | `image` (inline), `upscale()` (queued) | | | | |
+| `Runway` | | | ✓ | | | |
+| `Luma`, `Kling`, `MiniMax` | | *deferred* | *deferred* | | | |

-New facades follow the existing one-file-per-provider rule. Package entrypoints are modality-specific, such as `@opencode/ai/providers/openai/images`, and return the concrete model.
+New facades follow the existing one-file-per-provider rule. The facade selector is the public path for media models; modality-specific package entrypoints (for example `@opencode/ai/providers/openai/images`) are deferred until Core has a modality-aware model resolver.

-`ImageModel<Options>` already gives typed `providerOptions` per model; `VideoModel`, `SpeechModel`, `TranscriptionModel` follow the same generic. A shared `MediaModel` union is what `Generation` and the promise client key on.
+`ImageModel<Options>` gives typed `providerOptions` per model; `VideoModel`, `SpeechModel`, and `TranscriptionModel` follow the same generic. As with `LanguageModel`, the route type does not carry `Options`, so `ImageModel<OpenAIImageOptions>` is an `ImageModel` and client methods take plain `ImageRequestFor`. They share an internal `MediaModel` base class (ids, route, `http` overlays) that is not part of the public exports; `Generation` and the promise client work with the concrete modality models.

 ### Routes and protocols

 Media does not fit the LLM four-axis route (SSE frames → event state machine) except for streaming TTS/STT. Reuse `Endpoint`, `Auth`, `Framing`, `RequestExecutor`, and add media protocol kinds:

 - `MediaProtocol.inline` — `body.from(request)` (JSON, multipart, or query), `response.decode(response)` (JSON, or binary body → `Media.Asset`).
- `MediaProtocol.queued` — `start`, `status`, `result`, `cancel`, optional `download`, `pollHint`, `token` schema.
- `MediaProtocol.stream` — framing + `step` state machine emitting modality events, same discipline as LLM protocols.
+- `MediaProtocol.queued` — `start` (body + decode to `{ token, snapshot }`), `status`, `result`, optional `cancel`, and a `token` codec. `result` is always a separate GET (against the status document for Veo/xAI/Runway, fal's `response_url` otherwise) so `await` after `start` and after `resume` share one path. `PollContext.auth` hands the auth headers the route sent to the protocol for output URLs that need them (Veo downloads); they become transient `Media.Asset.headers`, never part of `source`. There is no separate `download` step: `Media.Asset.bytes()` downloads through the executor with those headers. `MediaRoute.inline(...)` / `MediaRoute.queued(...)` compose each kind with endpoint and auth; the queued route decodes the token once and hands `Generation` a token-free `{ status, result, cancel? }`.
+- `MediaProtocol.stream` — `body.from(request)` over the request plus its `mode`, `frames` (a function that picks the framing for the call: `Framing.sse`, `lines`, `document`, or the raw bytes), fresh per-response `initial()` state, `step` emitting modality events, and `finish(state, context)` — with the observed response for header-only usage — emitting exactly one terminal event or failing as an incomplete stream. The route fills `reason.http` on stream errors. `MediaRoute.stream(...)` exposes `stream` and `generate` (the same stream folded by the modality's `collect`).

-`Route.make` for media composes one protocol kind with endpoint/auth. The existing `ImageRoute { generate(request, execute) }` is the ad-hoc version of `inline` and gets folded in.
+`MediaRoute.inline` / `MediaRoute.queued` / `MediaRoute.stream` compose one protocol kind with endpoint/auth and tag the route with its `kind`; `ImageModel`/`VideoModel`/`SpeechModel`/`TranscriptionModel` share the `MediaModel` base (`src/media-model.ts`).

 ### LLM integration

@@ -320,9 +466,9 @@ Foundation + Image ship together as the reference implementation, serially. Vide
 ## Phasing

 1. **Foundation** — per-modality selectors, `Media`, `Generation`, `Poll`, `Usage` union, `MediaProtocol` kinds, `@opencode/ai/promise` with `llm` + `image`. Port the five existing image protocols onto it. Unify `MediaPart` and add the `media` LLM event (fixes Gemini image output being dropped).
-2. **Video** — Veo, xAI, fal, Runway first. Then Luma, Kling, MiniMax, Replicate.
-3. **Speech + Transcription** — OpenAI, ElevenLabs, Gemini TTS, Deepgram, Cartesia, AssemblyAI. Streaming TTS from the start.
-4. **Image queued routes and partials** — BFL, fal, Replicate, Stability; OpenAI `partial_images` streaming.
+2. **Video** — ✅ Veo, xAI, fal, Runway shipped (`MediaProtocol.queued`, `Video.start/generate/resume/stream`, promise `ai.video`). Deferred: `Video.complete` (webhooks), Luma, Kling, MiniMax, Replicate.
+3. **Speech + Transcription** — ✅ Speech: OpenAI, Gemini TTS, ElevenLabs, Cartesia, Deepgram shipped (`MediaProtocol.stream`, `Speech.generate/stream`, promise `ai.speech`). ✅ Transcription: OpenAI, Gemini, Deepgram, AssemblyAI shipped across all three route kinds (`Transcription.generate/stream/start/resume`, promise `ai.transcription`). Pending: ElevenLabs Scribe. Deferred: `Speech.session` and `Transcription.session` (WebSocket streaming).
+4. **Image queued routes and partials** — ✅ BFL, fal, Replicate, and Stability creative upscale queued; Stability generate inline; OpenAI `partial_images` streaming (`image-partial` restored). Imagen dropped: shut down on the Gemini API and discontinued on Vertex (2026-06-30). Deferred: Stability's synchronous edit and fast/conservative upscale endpoints.
 5. **Later** — ElevenLabs music/SFX, Lyria, `Speech.session` / `Transcription.session`, realtime.

 Core adoption (session attachments beyond png/jpeg/gif/webp/pdf, image-generation tool, TUI rendering) comes after phase 1 and is a Core concern.
@@ -1,18 +1,7 @@
-import { Config, Effect, Formatter, Layer, Schema, Stream } from "effect"
+import { Config, Effect, Formatter, Schema, Stream } from "effect"
 import { NodeFileSystem } from "@effect/platform-node"
-import {
-  Image,
-  ImageClient,
-  LLM,
-  LLMClient,
-  LLMRequest,
-  Media,
-  Message,
-  ProviderID,
-  Tool,
-  ToolRuntime,
-} from "@opencode/ai"
-import { Route, Auth, Endpoint, Framing, Protocol, RequestExecutor } from "@opencode/ai/route"
+import { AIClient, Image, LLM, LLMRequest, Media, Message, ProviderID, Tool, ToolRuntime } from "@opencode/ai"
+import { Route, Auth, Endpoint, Framing, Protocol } from "@opencode/ai/route"
 import { OpenAI } from "@opencode/ai/providers"

 /**
@@ -243,12 +232,10 @@ const generateImage = Effect.gen(function* () {
  yield* Media.write(response.image, "tutorial-image.jpg").pipe(Effect.provide(NodeFileSystem.layer))
 })

-// Provide the LLM runtime and the HTTP request executor once. Keep one path
-// enabled at a time so the tutorial can demonstrate generate, stream, or
+// Provide every modality client and the HTTP request executor once with
+// `AIClient.layer` (`AIClient.layerWith(executor)` swaps the executor). Keep one
+// path enabled at a time so the tutorial can demonstrate generate, stream, or
 // tool-loop behavior without spending tokens on every example.
-const requestExecutorLayer = RequestExecutor.fetchLayer
-const llmClientLayer = LLMClient.layer.pipe(Layer.provide(requestExecutorLayer))
-const imageClientLayer = ImageClient.layer.pipe(Layer.provide(requestExecutorLayer))

 const program = Effect.gen(function* () {
  // yield* generateOnce
@@ -257,6 +244,6 @@ const program = Effect.gen(function* () {
  // yield* generateDynamicObject.pipe(Effect.andThen((response) => Effect.sync(() => console.log(response.object))))
  // yield* generateImage
  yield* streamWithTools
-}).pipe(Effect.provide(Layer.mergeAll(requestExecutorLayer, llmClientLayer, imageClientLayer)))
+}).pipe(Effect.provide(AIClient.layer))

 Effect.runPromise(program)
@@ -1,6 +1,6 @@
 {
  "$schema": "https://json.schemastore.org/package.json",
-  "version": "2.0.14",
+  "version": "2.0.16",
  "name": "@opencode/ai",
  "type": "module",
  "license": "MIT",
@@ -104,6 +104,111 @@ const PROVIDERS: ReadonlyArray<Provider> = [
    vars: [{ name: "XAI_API_KEY" }],
    validate: (env) => validateBearer("https://api.x.ai/v1/models", Redacted.make(env.XAI_API_KEY)),
  },
+  {
+    id: "fal",
+    label: "fal",
+    tier: "canary",
+    note: "fal queue image and video recorded tests",
+    vars: [{ name: "FAL_KEY" }],
+    // fal has no free authenticated list endpoint; a 404 for an unknown request id proves the key was accepted.
+    validate: (env) =>
+      Effect.gen(function* () {
+        const http = yield* HttpClient.HttpClient
+        const response = yield* http.execute(
+          HttpClientRequest.get(
+            "https://queue.fal.run/fal-ai/veo3.1/requests/00000000-0000-0000-0000-000000000000/status",
+          ).pipe(HttpClientRequest.setHeaders({ authorization: `Key ${Redacted.value(Redacted.make(env.FAL_KEY))}` })),
+        )
+        if (response.status === 404) return undefined
+        return yield* responseError(response)
+      }),
+  },
+  {
+    id: "black-forest-labs",
+    label: "Black Forest Labs",
+    tier: "canary",
+    note: "BFL FLUX queued image recorded tests",
+    vars: [{ name: "BFL_API_KEY" }],
+    validate: (env) =>
+      HttpClientRequest.get("https://api.bfl.ai/v1/credits").pipe(
+        HttpClientRequest.setHeader("x-key", Redacted.value(Redacted.make(env.BFL_API_KEY))),
+        executeRequest,
+      ),
+  },
+  {
+    id: "replicate",
+    label: "Replicate",
+    tier: "canary",
+    note: "Replicate prediction image recorded tests",
+    vars: [{ name: "REPLICATE_API_TOKEN" }],
+    validate: (env) => validateBearer("https://api.replicate.com/v1/account", Redacted.make(env.REPLICATE_API_TOKEN)),
+  },
+  {
+    id: "stability",
+    label: "Stability AI",
+    tier: "canary",
+    note: "Stability inline generate and queued upscale recorded tests",
+    vars: [{ name: "STABILITY_API_KEY" }],
+    validate: (env) => validateBearer("https://api.stability.ai/v1/user/balance", Redacted.make(env.STABILITY_API_KEY)),
+  },
+  {
+    id: "runway",
+    label: "Runway",
+    tier: "canary",
+    note: "Runway task video recorded tests",
+    vars: [{ name: "RUNWAYML_API_SECRET" }],
+    validate: (env) =>
+      validateBearer("https://api.dev.runwayml.com/v1/organization", Redacted.make(env.RUNWAYML_API_SECRET), {
+        "X-Runway-Version": "2024-11-06",
+      }),
+  },
+  {
+    id: "elevenlabs",
+    label: "ElevenLabs",
+    tier: "canary",
+    note: "ElevenLabs text-to-speech recorded tests",
+    vars: [{ name: "ELEVENLABS_API_KEY" }],
+    validate: (env) =>
+      HttpClientRequest.get("https://api.elevenlabs.io/v1/models").pipe(
+        HttpClientRequest.setHeader("xi-api-key", Redacted.value(Redacted.make(env.ELEVENLABS_API_KEY))),
+        executeRequest,
+      ),
+  },
+  {
+    id: "cartesia",
+    label: "Cartesia",
+    tier: "canary",
+    note: "Cartesia text-to-speech recorded tests",
+    vars: [{ name: "CARTESIA_API_KEY" }],
+    validate: (env) =>
+      validateBearer("https://api.cartesia.ai/voices?limit=1", Redacted.make(env.CARTESIA_API_KEY), {
+        "Cartesia-Version": "2026-08-14",
+      }),
+  },
+  {
+    id: "deepgram",
+    label: "Deepgram",
+    tier: "canary",
+    note: "Deepgram Aura text-to-speech and Nova transcription recorded tests",
+    vars: [{ name: "DEEPGRAM_API_KEY" }],
+    validate: (env) =>
+      HttpClientRequest.get("https://api.deepgram.com/v1/projects").pipe(
+        HttpClientRequest.setHeader("authorization", `Token ${Redacted.value(Redacted.make(env.DEEPGRAM_API_KEY))}`),
+        executeRequest,
+      ),
+  },
+  {
+    id: "assemblyai",
+    label: "AssemblyAI",
+    tier: "canary",
+    note: "AssemblyAI queued transcription recorded tests",
+    vars: [{ name: "ASSEMBLYAI_API_KEY" }],
+    validate: (env) =>
+      HttpClientRequest.get("https://api.assemblyai.com/v2/transcript?limit=1").pipe(
+        HttpClientRequest.setHeader("authorization", Redacted.value(Redacted.make(env.ASSEMBLYAI_API_KEY))),
+        executeRequest,
+      ),
+  },
  {
    id: "cloudflare-ai-gateway",
    label: "Cloudflare AI Gateway",
@@ -0,0 +1,24 @@
+import { Layer } from "effect"
+import { ImageClient } from "./image-client.js"
+import { LLMClient } from "./route/client.js"
+import { RequestExecutor } from "./route/executor.js"
+import { SpeechClient } from "./speech-client.js"
+import { TranscriptionClient } from "./transcription-client.js"
+import { VideoClient } from "./video-client.js"
+
+/** Every modality client over `executor`, which stays in the output so `asset.bytes()` and `Media.write` resolve. */
+export const layerWith = <E, R>(executor: Layer.Layer<RequestExecutor.Service, E, R>) =>
+  Layer.mergeAll(
+    LLMClient.layer,
+    ImageClient.layer,
+    VideoClient.layer,
+    SpeechClient.layer,
+    TranscriptionClient.layer,
+  ).pipe(Layer.provideMerge(executor))
+
+/** Every modality client plus the executor over `RequestExecutor.fetchLayer`: the one layer most programs need. */
+export const layer = layerWith(RequestExecutor.fetchLayer)
+
+export type Services = Layer.Success<typeof layer>
+
+export * as AIClient from "./ai-client.js"
@@ -86,7 +86,7 @@ export const layer: Layer.Layer<Service, never, RequestExecutor.Service> = Layer
    })
  }),
 )
-export const fetchLayer = layer.pipe(Layer.provide(RequestExecutor.fetchLayer))
+export const fetchLayer = layer.pipe(Layer.provideMerge(RequestExecutor.fetchLayer))

 export const EvaluationClient = {
  Service,
@@ -214,7 +214,7 @@ export function request(input: EvaluationRequest | EvaluationRequestInput) {
  return new EvaluationRequest({
    ...input,
    model: input.model as unknown as EvaluationModel,
-    http: input.http === undefined ? undefined : HttpOptions.make(input.http),
+    http: HttpOptions.make(input.http),
  })
 }

@@ -1,4 +1,4 @@
-import { Duration, Effect, Schedule, Schema, Stream } from "effect"
+import { Clock, Duration, Effect, Schedule, Schema, Stream } from "effect"
 import { AIError, TimeoutError } from "./schema/errors.js"

 export const Status = Schema.Literals(["queued", "running", "completed", "failed", "cancelled", "expired"])
@@ -11,47 +11,59 @@ export interface Snapshot {
  /** Normalized 0..1 when the provider reports progress. */
  readonly progress?: number
  readonly position?: number
-  readonly expiresAt?: number
 }

 /**
- * Route-owned generation operations. `token` is the route's serializable handle (operation name, task id, response URL)
- * so a generation can be resumed from another process; its shape is opaque to `Generation`.
+ * Route-owned generation operations for one generation. The media route decodes its serializable token once (from the
+ * submission response or a `resume` input) and closes over it, so `Generation` never sees the token's shape.
 */
 export interface Route<Response> {
-  readonly status: (token: unknown) => Effect.Effect<Snapshot, AIError>
-  readonly result: (token: unknown) => Effect.Effect<Response, AIError>
-  readonly cancel?: (token: unknown) => Effect.Effect<void, AIError>
-  /** Provider polling hint (e.g. `openai-poll-after-ms`) that overrides the default interval for the next poll. */
-  readonly pollHint?: (snapshot: Snapshot) => Duration.Duration | undefined
+  readonly status: Effect.Effect<Snapshot, AIError>
+  readonly result: Effect.Effect<Response, AIError>
+  readonly cancel?: Effect.Effect<void, AIError>
 }

 export interface Poll {
  readonly interval?: Duration.Input
  readonly timeout?: Duration.Input
-  /** Full override of the polling schedule; `interval` and `pollHint` are ignored when supplied. */
-  readonly schedule?: Schedule.Schedule<unknown, Snapshot>
+}
+
+export interface AwaitOptions {
+  readonly poll?: Poll
 }

 export const DEFAULT_POLL_INTERVAL = Duration.seconds(5)
 export const DEFAULT_POLL_TIMEOUT = Duration.minutes(10)

-export type Event =
-  | { readonly type: "generation-queued"; readonly id: string; readonly position?: number }
-  | { readonly type: "generation-progress"; readonly id: string; readonly progress?: number }
-  | { readonly type: "generation-finished"; readonly id: string; readonly status: Status }
+export const QueuedEvent = Schema.Struct({
+  type: Schema.tag("generation-queued"),
+  id: Schema.String,
+  position: Schema.optional(Schema.Number),
+}).annotate({ identifier: "Generation.Event.Queued" })
+
+export const ProgressEvent = Schema.Struct({
+  type: Schema.tag("generation-progress"),
+  id: Schema.String,
+  progress: Schema.optional(Schema.Number),
+}).annotate({ identifier: "Generation.Event.Progress" })
+
+export type Observation = Schema.Schema.Type<typeof QueuedEvent> | Schema.Schema.Type<typeof ProgressEvent>
+
+export type Event = Observation | { readonly type: "generation-finished"; readonly id: string; readonly status: Status }

 const TERMINAL: ReadonlySet<Status> = new Set(["completed", "failed", "cancelled", "expired"])

+export const isTerminal = (status: Status) => TERMINAL.has(status)
+
 export class Generation<Response> {
  readonly id: string
  readonly status: Status
  readonly progress?: number
  readonly position?: number
-  readonly expiresAt?: number

  constructor(
    readonly route: Route<Response>,
+    /** Route-owned serializable JSON; pass it to the modality's `resume` from another process. */
    readonly token: unknown,
    snapshot: Snapshot,
  ) {
@@ -59,7 +71,6 @@ export class Generation<Response> {
    this.status = snapshot.status
    this.progress = snapshot.progress
    this.position = snapshot.position
-    this.expiresAt = snapshot.expiresAt
  }

  get snapshot(): Snapshot {
@@ -68,70 +79,110 @@ export class Generation<Response> {
      status: this.status,
      progress: this.progress,
      position: this.position,
-      expiresAt: this.expiresAt,
    }
  }

  get terminal() {
-    return TERMINAL.has(this.status)
+    return isTerminal(this.status)
  }

  refresh(): Effect.Effect<Generation<Response>, AIError> {
-    return this.route.status(this.token).pipe(Effect.map((snapshot) => new Generation(this.route, this.token, snapshot)))
+    return this.route.status.pipe(Effect.map((snapshot) => new Generation(this.route, this.token, snapshot)))
+  }
+
+  /** Fetch the result without polling; non-completed terminal generations fail with the provider's terminal body. */
+  result(): Effect.Effect<Response, AIError> {
+    return this.route.result
  }

  /** Poll until the generation reaches a terminal status, then fetch the result. Fails with a `Timeout` reason on deadline. */
-  await(options?: { readonly poll?: Poll }): Effect.Effect<Response, AIError> {
+  await(options?: AwaitOptions): Effect.Effect<Response, AIError> {
    const timeout = Duration.fromInputUnsafe(options?.poll?.timeout ?? DEFAULT_POLL_TIMEOUT)
    const settled = this.terminal ? Effect.succeed(this) : this.poll(options?.poll)
    return settled.pipe(
      // Non-completed terminal states also go through `result` so the route can surface its provider failure body.
-      Effect.flatMap((generation) => generation.route.result(generation.token)),
-      Effect.timeoutOrElse({
-        duration: timeout,
-        orElse: () =>
-          new AIError({
-            reason: new TimeoutError({
-              message: `Generation ${this.id} did not finish within ${Duration.format(timeout)}`,
-              timeoutMs: Duration.toMillis(timeout),
-            }),
-          }),
-      }),
+      Effect.flatMap((generation) => generation.result()),
+      Effect.timeoutOrElse({ duration: timeout, orElse: () => this.timeoutError(timeout) }),
    )
  }

  cancel(): Effect.Effect<void, AIError> {
-    return this.route.cancel?.(this.token) ?? Effect.void
+    return this.route.cancel ?? Effect.void
  }

-  /** Status observations as a stream, ending after the first terminal observation. */
-  events(options?: { readonly poll?: Poll }): Stream.Stream<Event, AIError> {
-    const observations = this.terminal
-      ? Stream.make(this)
-      : Stream.fromEffectSchedule(this.refresh(), this.schedule(options?.poll)).pipe(
-          Stream.takeUntil((generation) => generation.terminal),
-        )
-    return observations.pipe(
-      Stream.map((generation): Event => {
-        if (generation.terminal) return { type: "generation-finished", id: generation.id, status: generation.status }
-        if (generation.status === "queued") return { type: "generation-queued", id: generation.id, position: generation.position }
-        return { type: "generation-progress", id: generation.id, progress: generation.progress }
-      }),
+  /**
+   * Status observations as a stream, ending after the first terminal observation. Each poll and each sleep between polls
+   * is bounded by the time remaining until `poll.timeout`, so a hung status request or a long interval fails the stream at
+   * the deadline instead of stalling it. (`Stream.interruptWhen` would express this directly but deadlocks under
+   * `TestClock` when the source completes while the timer sleeps.)
+   */
+  events(options?: AwaitOptions): Stream.Stream<Event, AIError> {
+    if (this.terminal) return Stream.make(this.event())
+    const timeout = Duration.fromInputUnsafe(options?.poll?.timeout ?? DEFAULT_POLL_TIMEOUT)
+    return Stream.unwrap(
+      Clock.currentTimeMillis.pipe(
+        Effect.map((start) => {
+          const deadline = start + Duration.toMillis(timeout)
+          // Fail before polling once the deadline has passed: a fast status request could otherwise win the zero-budget
+          // race and schedule another zero-delay poll.
+          const refresh = Clock.currentTimeMillis.pipe(
+            Effect.flatMap((now) =>
+              now >= deadline
+                ? this.timeoutError(timeout)
+                : this.refresh().pipe(
+                    Effect.timeoutOrElse({
+                      duration: Duration.millis(deadline - now),
+                      orElse: () => this.timeoutError(timeout),
+                    }),
+                  ),
+            ),
+          )
+          const schedule = this.schedule(options?.poll).pipe(
+            Schedule.modifyDelay((meta) =>
+              Effect.succeed(Duration.min(meta.duration, Duration.millis(Math.max(0, deadline - meta.now)))),
+            ),
+          )
+          return Stream.fromEffectSchedule(refresh, schedule).pipe(
+            Stream.takeUntil((generation) => generation.terminal),
+            Stream.map((generation) => generation.event()),
+          )
+        }),
+      ),
    )
  }

+  private event(): Event {
+    if (this.terminal) return { type: "generation-finished", id: this.id, status: this.status }
+    if (this.status === "queued") return { type: "generation-queued", id: this.id, position: this.position }
+    return { type: "generation-progress", id: this.id, progress: this.progress }
+  }
+
+  private timeoutError(timeout: Duration.Duration) {
+    return new AIError({
+      reason: new TimeoutError({
+        message: `Generation ${this.id} did not finish within ${Duration.format(timeout)}`,
+        timeoutMs: Duration.toMillis(timeout),
+      }),
+    })
+  }
+
  private poll(poll: Poll | undefined) {
-    return this.refresh().pipe(Effect.repeat({ schedule: this.schedule(poll), until: (generation) => generation.terminal }))
-  }
-
-  private schedule(poll: Poll | undefined): Schedule.Schedule<unknown, Generation<Response>> {
-    if (poll?.schedule) return poll.schedule.pipe(Schedule.setInputType<Generation<Response>>())
-    const interval = poll?.interval ?? DEFAULT_POLL_INTERVAL
-    const pollHint = this.route.pollHint
-    const spaced = Schedule.spaced(interval).pipe(Schedule.setInputType<Generation<Response>>())
-    if (!pollHint) return spaced
-    return spaced.pipe(
-      Schedule.modifyDelay((metadata) => Effect.succeed(pollHint(metadata.input.snapshot) ?? interval)),
+    return this.refresh().pipe(
+      Effect.repeat({ schedule: this.schedule(poll), until: (generation) => generation.terminal }),
    )
  }
+
+  private schedule(poll: Poll | undefined) {
+    return Schedule.spaced(poll?.interval ?? DEFAULT_POLL_INTERVAL)
+  }
 }
+
+export const resultEvents = <Response, A>(
+  generation: Generation<Response>,
+  expand: (response: Response) => ReadonlyArray<A>,
+  options?: AwaitOptions,
+): Stream.Stream<Observation | A, AIError> =>
+  generation.events(options).pipe(
+    Stream.filter((event): event is Observation => event.type !== "generation-finished"),
+    Stream.concat(Stream.fromIterableEffect(Effect.map(generation.result(), expand))),
+  )
@@ -1,61 +1,30 @@
-import { Context, Effect, Layer, Stream } from "effect"
-import { RequestExecutor } from "./route/executor.js"
-import type { AIError } from "./schema/index.js"
+import { Context } from "effect"
+import { MediaClient } from "./media-client.js"
 import {
-  responseEvents,
+  ImageOutputEvent,
+  ImageFinishEvent,
  type ImageEvent,
-  type ImageOptions,
  type ImageRequestFor,
  type ImageResponse,
 } from "./image.js"

-export interface Interface {
-  readonly generate: <Options extends ImageOptions>(
-    request: ImageRequestFor<Options>,
-  ) => Effect.Effect<ImageResponse, AIError>
-  readonly stream: <Options extends ImageOptions>(
-    request: ImageRequestFor<Options>,
-  ) => Stream.Stream<ImageEvent, AIError>
-}
+export type Interface = MediaClient.Interface<ImageRequestFor, ImageEvent, ImageResponse>

-export class Service extends Context.Service<Service, Interface>()("@opencode/ImageClient") {}
-
-export const generate = <Options extends ImageOptions>(
-  request: ImageRequestFor<Options>,
-): Effect.Effect<ImageResponse, AIError, Service> =>
-  Effect.gen(function* () {
-    const client = yield* Service
-    return yield* client.generate(request)
-  })
-
-export const stream = <Options extends ImageOptions>(
-  request: ImageRequestFor<Options>,
-): Stream.Stream<ImageEvent, AIError, Service> =>
-  Stream.unwrap(
-    Effect.gen(function* () {
-      const client = yield* Service
-      return client.stream(request)
-    }),
-  )
-
-export const layer: Layer.Layer<Service, never, RequestExecutor.Service> = Layer.effect(
-  Service,
-  Effect.gen(function* () {
-    const executor = yield* RequestExecutor.Service
-    const generate = <Options extends ImageOptions>(request: ImageRequestFor<Options>) =>
-      request.model.route.generate(request, executor.execute)
-    return Service.of({
-      generate,
-      // Inline routes have no partial frames yet; the stream is the completed response expanded into events.
-      stream: (request) =>
-        Stream.unwrap(generate(request).pipe(Effect.map((response) => Stream.fromIterable(responseEvents(response))))),
-    })
-  }),
-)
+export class ImageClientService extends Context.Service<ImageClientService, Interface>()("@opencode/ImageClient") {}
+export const Service = ImageClientService
+export type Service = ImageClientService

 export const ImageClient = {
  Service,
-  layer,
-  generate,
-  stream,
+  ...MediaClient.make(Service, {
+    modality: "image",
+    responseEvents: (response: ImageResponse) => [
+      ...response.images.map((image, index) => ImageOutputEvent.make({ index, image })),
+      ImageFinishEvent.make({
+        usage: response.usage,
+        notices: response.notices,
+        providerMetadata: response.providerMetadata,
+      }),
+    ],
+  }),
 } as const
@@ -1,98 +1,48 @@
 import { Effect, Schema, Stream } from "effect"
+import { Generation, ProgressEvent, QueuedEvent, type AwaitOptions } from "./generation.js"
 import { Media } from "./media.js"
-import { Endpoint } from "./route/endpoint.js"
+import { MediaModel, composeRoute, tryRequest } from "./media-model.js"
 import { MediaRoute } from "./route/media.js"
-import type { MediaProtocol } from "./route/media-protocol.js"
-import {
-  AIError,
-  HttpOptions,
-  InvalidRequestError,
-  MediaUsage,
-  ModelID,
-  ProviderID,
-  ProviderMetadata,
-} from "./schema/index.js"
+import { AIError, HttpOptions, MediaUsage, ProviderMetadata, type OpenString } from "./schema/index.js"
 import { ImageClient, Service } from "./image-client.js"

 // ---------------------------------------------------------------------------
 // Model
 // ---------------------------------------------------------------------------

-export type ImageOptions = Record<string, unknown>
+export type ImageOptions = MediaModel.Options

-export type ImageRoute<Options extends ImageOptions = ImageOptions> = MediaRoute.Route<
-  ImageRequestFor<Options>,
-  ImageResponse
->
+export type ImageRoute = MediaRoute.AnyRoute<ImageRequestFor, ImageEvent, ImageResponse>

-export class ImageModel<Options extends ImageOptions = ImageOptions> {
-  declare protected readonly _Options: (options: Options) => Options
-  readonly id: ModelID
-  readonly provider: ProviderID
-  readonly route: ImageRoute<Options>
-  readonly http?: HttpOptions
+export class ImageModel<Options extends ImageOptions = ImageOptions> extends MediaModel<ImageRoute, Options> {
+  declare protected readonly _ImageModel: void

-  constructor(input: ImageModel.Input<Options>) {
-    this.id = input.id
-    this.provider = input.provider
-    this.route = input.route
-    this.http = input.http
-  }
-
-  static make<Options extends ImageOptions = ImageOptions>(input: ImageModel.MakeInput<Options>) {
-    return new ImageModel<Options>({
-      id: ModelID.make(input.id),
-      provider: ProviderID.make(input.provider),
-      route: input.route,
-      http: input.http,
-    })
-  }
-
-  /** Compose an inline image protocol with its canonical path into a model for one deployment. */
-  static fromRoute<Options extends ImageOptions = ImageOptions>(
-    route: ImageModel.RouteInput<Options>,
+  /** The number of type arguments selects the kind: `<Options>`, `<Options, Frame, State>`, or `<Options, Token>`. */
+  static fromRoute<Options extends ImageOptions>(
+    route: MediaModel.InlineRouteInput<ImageRequestFor<Options>, ImageResponse>,
+    input: MediaRoute.ModelInput,
+  ): ImageModel<Options>
+  static fromRoute<Options extends ImageOptions, Frame, State>(
+    route: MediaModel.StreamRouteInput<ImageRequestFor<Options>, ImageEvent, Frame, State>,
+    input: MediaRoute.ModelInput,
+  ): ImageModel<Options>
+  static fromRoute<Options extends ImageOptions, Token>(
+    route: MediaModel.QueuedRouteInput<ImageRequestFor<Options>, ImageResponse, Token>,
+    input: MediaRoute.ModelInput,
+  ): ImageModel<Options>
+  static fromRoute<Options extends ImageOptions, Frame, State, Token>(
+    route: MediaModel.AnyRouteInput<ImageRequestFor<Options>, ImageEvent, ImageResponse, Frame, State, Token>,
    input: MediaRoute.ModelInput,
  ) {
-    return ImageModel.make<Options>({
+    return new ImageModel<Options>({
      id: input.id,
-      provider: route.provider,
+      provider: route.protocol.provider,
      http: input.http,
-      route: MediaRoute.make({
-        id: route.id,
-        provider: route.provider,
-        protocol: route.protocol,
-        endpoint: Endpoint.path(route.path, { baseURL: input.baseURL ?? route.baseURL }),
-        auth: input.auth,
-        headers: input.headers,
-      }),
+      route: composeRoute(route, input, collectResponse) as ImageRoute,
    })
  }
 }

-export namespace ImageModel {
-  export interface Input<Options extends ImageOptions = ImageOptions> {
-    readonly id: ModelID
-    readonly provider: ProviderID
-    readonly route: ImageRoute<Options>
-    readonly http?: HttpOptions
-  }
-
-  export interface MakeInput<Options extends ImageOptions = ImageOptions>
-    extends Omit<Input<Options>, "id" | "provider"> {
-    readonly id: string | ModelID
-    readonly provider: string | ProviderID
-  }
-
-  export interface RouteInput<Options extends ImageOptions = ImageOptions> {
-    readonly id: string
-    readonly provider: string | ProviderID
-    readonly protocol: MediaProtocol.Inline<ImageRequestFor<Options>, ImageResponse>
-    readonly path: Endpoint.EndpointPart<MediaProtocol.Body, ImageRequestFor<Options>>
-    /** Canonical base URL; `ModelInput.baseURL` overrides it per deployment. */
-    readonly baseURL?: string
-  }
-}
-
 export const ImageModelSchema = Schema.declare((value): value is ImageModel => value instanceof ImageModel, {
  expected: "Image.Model",
 })
@@ -107,13 +57,10 @@ export const ImageSize = Schema.declare<ImageSize>(
  { title: "ImageSize" },
 )

-export type ImageAspectRatio = `${number}:${number}`
-export const ImageAspectRatio = Schema.declare<ImageAspectRatio>(
-  (value): value is ImageAspectRatio => typeof value === "string" && /^\d+(?:\.\d+)?:\d+(?:\.\d+)?$/.test(value),
-  { title: "ImageAspectRatio" },
-)
+export type ImageAspectRatio = Media.AspectRatio
+export const ImageAspectRatio = Media.AspectRatio

-export type ImageFormat = "png" | "jpeg" | "webp" | (string & {})
+export type ImageFormat = OpenString<"png" | "jpeg" | "webp">

 export class ImageRequest extends Schema.Class<ImageRequest>("Image.Request")({
  model: ImageModelSchema,
@@ -168,6 +115,13 @@ export class ImageResponse extends Schema.Class<ImageResponse>("Image.Response")
  }
 }

+/** The `index`-th lower-fidelity preview of an image still generating; the `image` event supersedes it. */
+export const ImagePartialEvent = Schema.Struct({
+  type: Schema.tag("image-partial"),
+  index: Schema.Number,
+  image: Media.AssetSchema,
+}).annotate({ identifier: "Image.Event.Partial" })
+
 export const ImageOutputEvent = Schema.Struct({
  type: Schema.tag("image"),
  index: Schema.Number,
@@ -181,24 +135,37 @@ export const ImageFinishEvent = Schema.Struct({
  providerMetadata: Schema.optional(ProviderMetadata),
 }).annotate({ identifier: "Image.Event.Finish" })

-const imageEventTagged = Schema.Union([ImageOutputEvent, ImageFinishEvent]).pipe(Schema.toTaggedUnion("type"))
+const imageEventTagged = Schema.Union([
+  QueuedEvent,
+  ProgressEvent,
+  ImagePartialEvent,
+  ImageOutputEvent,
+  ImageFinishEvent,
+]).pipe(Schema.toTaggedUnion("type"))
 export const ImageEvent = Object.assign(imageEventTagged, {
  is: {
+    generationQueued: imageEventTagged.guards["generation-queued"],
+    generationProgress: imageEventTagged.guards["generation-progress"],
+    imagePartial: imageEventTagged.guards["image-partial"],
    image: imageEventTagged.guards.image,
    finish: imageEventTagged.guards.finish,
  },
 })
 export type ImageEvent = Schema.Schema.Type<typeof imageEventTagged>

-/** Inline routes produce every image at once; expand the response into the streaming event shape. */
-export const responseEvents = (response: ImageResponse): ReadonlyArray<ImageEvent> => [
-  ...response.images.map((image, index) => ImageOutputEvent.make({ index, image })),
-  ImageFinishEvent.make({
-    usage: response.usage,
-    notices: response.notices,
-    providerMetadata: response.providerMetadata,
-  }),
-]
+const collectResponse = (events: ReadonlyArray<ImageEvent>): Effect.Effect<ImageResponse> => {
+  const finish = events.find(ImageEvent.is.finish)
+  // Every image protocol's `finish` emits the terminal event or fails, so a completed stream always has one.
+  if (finish === undefined) return Effect.die(new Error("The image stream completed without a finish event"))
+  return Effect.succeed(
+    new ImageResponse({
+      images: events.filter(ImageEvent.is.image).map((event) => event.image),
+      usage: finish.usage,
+      notices: finish.notices,
+      providerMetadata: finish.providerMetadata,
+    }),
+  )
+}

 // ---------------------------------------------------------------------------
 // Request-shaped call API
@@ -212,40 +179,43 @@ export function request(input: ImageRequest | ImageRequestInput) {
  if (input instanceof ImageRequest) return input
  return new ImageRequest({
    ...input,
-    http: input.http === undefined ? undefined : HttpOptions.make(input.http),
+    http: HttpOptions.make(input.http),
  })
 }

-const requestEffect = (input: ImageRequest | ImageRequestInput) =>
-  Effect.try({
-    try: () => request(input),
-    catch: (error) =>
-      new AIError({
-        reason: new InvalidRequestError({
-          message: error instanceof Error ? error.message : String(error),
-          cause: error,
-        }),
-      }),
-  })
+const requestEffect = (input: ImageRequest | ImageRequestInput) => tryRequest(() => request(input))

 export function generate<const Model extends ImageModel>(
-  input: ImageRequestInput<Model>,
+  input: ImageRequest | ImageRequestInput<Model>,
+  options?: AwaitOptions,
 ): Effect.Effect<ImageResponse, AIError, Service>
-export function generate(input: ImageRequest): Effect.Effect<ImageResponse, AIError, Service>
-export function generate(input: ImageRequest | ImageRequestInput) {
-  return requestEffect(input).pipe(Effect.flatMap((request) => ImageClient.generate(request)))
+export function generate(input: ImageRequest | ImageRequestInput, options?: AwaitOptions) {
+  return requestEffect(input).pipe(Effect.flatMap((request) => ImageClient.generate(request, options)))
 }

 export function stream<const Model extends ImageModel>(
-  input: ImageRequestInput<Model>,
+  input: ImageRequest | ImageRequestInput<Model>,
+  options?: AwaitOptions,
 ): Stream.Stream<ImageEvent, AIError, Service>
-export function stream(input: ImageRequest): Stream.Stream<ImageEvent, AIError, Service>
-export function stream(input: ImageRequest | ImageRequestInput) {
-  return Stream.unwrap(requestEffect(input).pipe(Effect.map((request) => ImageClient.stream(request))))
+export function stream(input: ImageRequest | ImageRequestInput, options?: AwaitOptions) {
+  return Stream.unwrap(requestEffect(input).pipe(Effect.map((request) => ImageClient.stream(request, options))))
 }

+/** Inline and streaming routes fail with `UnsupportedOperation`. */
+export function start<const Model extends ImageModel>(
+  input: ImageRequest | ImageRequestInput<Model>,
+): Effect.Effect<Generation<ImageResponse>, AIError, Service>
+export function start(input: ImageRequest | ImageRequestInput) {
+  return requestEffect(input).pipe(Effect.flatMap((request) => ImageClient.start(request)))
+}
+
+export const resume = (model: ImageModel, token: unknown): Effect.Effect<Generation<ImageResponse>, AIError, Service> =>
+  ImageClient.resume(model, token)
+
 export const Image = {
  request,
  generate,
  stream,
+  start,
+  resume,
 } as const
@@ -1,3 +1,4 @@
+export { AIClient } from "./ai-client.js"
 export { LLMClient } from "./route/client.js"
 export { ImageClient } from "./image-client.js"
 export { Auth } from "./route/auth.js"
@@ -8,7 +9,7 @@ export type {
  RouteLanguageModelInput,
  RouteRoutedLanguageModelInput,
  Interface as LLMClientShape,
-  Service as LLMClientService,
+  LLMClientService,
 } from "./route/client.js"
 export * from "./schema/index.js"
 export {
@@ -29,9 +30,73 @@ export type {
  ImageRoute,
 } from "./image.js"
 export { Image } from "./image.js"
+export { VideoClient } from "./video-client.js"
+export {
+  VideoAspectRatio,
+  VideoEvent,
+  VideoFrames,
+  VideoModel,
+  VideoModelSchema,
+  VideoRequest,
+  VideoResponse,
+} from "./video.js"
+export type {
+  VideoModelOptions,
+  VideoOptions,
+  VideoRequestFor,
+  VideoRequestInput,
+  VideoResolution,
+  VideoRoute,
+} from "./video.js"
+export { Video } from "./video.js"
+export { SpeechClient } from "./speech-client.js"
+export {
+  SpeechEvent,
+  SpeechModel,
+  SpeechModelSchema,
+  SpeechRequest,
+  SpeechResponse,
+  SpeechTimestamp,
+  SpeechVoice,
+} from "./speech.js"
+export type {
+  SpeechFormat,
+  SpeechModelOptions,
+  SpeechOptions,
+  SpeechRequestFor,
+  SpeechRequestInput,
+  SpeechRoute,
+} from "./speech.js"
+export { Speech } from "./speech.js"
+export { TranscriptionClient } from "./transcription-client.js"
+export {
+  TranscriptionEvent,
+  TranscriptionModel,
+  TranscriptionModelSchema,
+  TranscriptionRequest,
+  TranscriptionResponse,
+  TranscriptionSegment,
+  TranscriptionTimestamps,
+  TranscriptionWord,
+} from "./transcription.js"
+export type {
+  TranscriptionModelOptions,
+  TranscriptionOptions,
+  TranscriptionRequestFor,
+  TranscriptionRequestInput,
+  TranscriptionRoute,
+} from "./transcription.js"
+export { Transcription } from "./transcription.js"
 export { Media } from "./media.js"
 export { Generation } from "./generation.js"
-export type { Event as GenerationEvent, Poll, Route as GenerationRoute, Snapshot as GenerationSnapshot, Status as GenerationStatus } from "./generation.js"
+export type {
+  AwaitOptions as GenerationAwaitOptions,
+  Event as GenerationEvent,
+  Poll,
+  Route as GenerationRoute,
+  Snapshot as GenerationSnapshot,
+  Status as GenerationStatus,
+} from "./generation.js"
 export { Tool, ToolFailure, toDefinitions } from "./tool.js"
 export { ToolRuntime } from "./tool-runtime.js"
 export type { DispatchResult as ToolDispatchResult, ToolSettlement } from "./tool-runtime.js"
@@ -1,5 +1,6 @@
-import { Effect, JsonSchema, Schema } from "effect"
-import { LLMClient, Service } from "./route/client.js"
+import { Effect, JsonSchema, Schema, Stream } from "effect"
+import { tryRequest } from "./media-model.js"
+import { LLMClient, Service, type StreamOptions } from "./route/client.js"
 import {
  GenerationOptions,
  HttpOptions,
@@ -35,9 +36,26 @@ export type RequestInput<SelectedLanguageModel extends LanguageModel = LanguageM
  readonly http?: HttpOptions.Input
 }

-export const generate = LLMClient.generate
+export function generate<const Model extends LanguageModel>(
+  input: RequestInput<Model>,
+  options?: StreamOptions,
+): Effect.Effect<LLMResponse, AIError, Service>
+export function generate(input: LLMRequest, options?: StreamOptions): Effect.Effect<LLMResponse, AIError, Service>
+export function generate(input: RequestInput | LLMRequest, options?: StreamOptions) {
+  return requestEffect(input).pipe(Effect.flatMap((request) => LLMClient.generate(request, options)))
+}

-export const stream = LLMClient.stream
+export function stream<const Model extends LanguageModel>(
+  input: RequestInput<Model>,
+  options?: StreamOptions,
+): Stream.Stream<LLMEvent, AIError, Service>
+export function stream(input: LLMRequest, options?: StreamOptions): Stream.Stream<LLMEvent, AIError, Service>
+export function stream(input: RequestInput | LLMRequest, options?: StreamOptions) {
+  return Stream.unwrap(requestEffect(input).pipe(Effect.map((request) => LLMClient.stream(request, options))))
+}
+
+const requestEffect = (input: RequestInput | LLMRequest) =>
+  input instanceof LLMRequest ? Effect.succeed(input) : tryRequest(() => request(input))

 export const request = <const SelectedLanguageModel extends LanguageModel>(
  input: RequestInput<SelectedLanguageModel>,
@@ -61,7 +79,7 @@ export const request = <const SelectedLanguageModel extends LanguageModel>(
    toolChoice: requestToolChoice ? ToolChoice.make(requestToolChoice) : undefined,
    generation: requestGeneration === undefined ? undefined : GenerationOptions.make(requestGeneration),
    providerOptions: requestProviderOptions,
-    http: requestHttp === undefined ? undefined : HttpOptions.make(requestHttp),
+    http: HttpOptions.make(requestHttp),
  })
 }

@@ -0,0 +1,77 @@
+import { type Context, Effect, Layer, Stream } from "effect"
+import { resultEvents, type AwaitOptions, type Generation, type Observation } from "./generation.js"
+import { RequestExecutor } from "./route/executor.js"
+import type { MediaRoute } from "./route/media.js"
+import { AIError, UnsupportedOperationError } from "./schema/index.js"
+
+/** A media request whose model carries the route that executes it. */
+export interface RoutedRequest<Self extends MediaRoute.MediaRequest, Event, Response> extends MediaRoute.MediaRequest {
+  readonly model: MediaRoute.MediaRequest["model"] & { readonly route: MediaRoute.AnyRoute<Self, Event, Response> }
+}
+
+/** `start` and `resume` fail with `UnsupportedOperation` on inline and stream routes. */
+export interface Interface<Req extends RoutedRequest<Req, Event, Response>, Event, Response> {
+  readonly generate: (request: Req, options?: AwaitOptions) => Effect.Effect<Response, AIError>
+  readonly stream: (request: Req, options?: AwaitOptions) => Stream.Stream<Event | Observation, AIError>
+  readonly start: (request: Req) => Effect.Effect<Generation<Response>, AIError>
+  readonly resume: (model: Req["model"], token: unknown) => Effect.Effect<Generation<Response>, AIError>
+}
+
+/** One modality's layer and service accessors, dispatching each request on its route's `kind`. */
+export const make = <Self, Req extends RoutedRequest<Req, Event, Response>, Event, Response>(
+  service: Context.Service<Self, Interface<Req, Event, Response>>,
+  input: {
+    readonly modality: string
+    /** A completed response expanded into the streaming event shape. */
+    readonly responseEvents: (response: Response) => ReadonlyArray<Event>
+  },
+) => ({
+  layer: Layer.effect(
+    service,
+    Effect.gen(function* () {
+      const executor = yield* RequestExecutor.Service
+      const notQueued = (route: MediaRoute.AnyRoute<Req, Event, Response>, operation: string) =>
+        new AIError({
+          reason: new UnsupportedOperationError({
+            operation: `${input.modality}.${operation}`,
+            provider: route.provider,
+            route: route.id,
+            message: `${route.provider}/${route.id} is not a queued route; use generate or stream`,
+          }),
+        })
+      const start = (request: Req) => {
+        const route = request.model.route
+        if (route.kind !== "queued") return Effect.fail(notQueued(route, "start"))
+        return route.start(request, executor.execute)
+      }
+      return service.of({
+        start,
+        resume: (model, token) => {
+          if (model.route.kind !== "queued") return Effect.fail(notQueued(model.route, "resume"))
+          return model.route.resume(model, token, executor.execute)
+        },
+        generate: (request, options) => {
+          const route = request.model.route
+          if (route.kind !== "queued") return route.generate(request, executor.execute)
+          return start(request).pipe(Effect.flatMap((generation) => generation.await(options)))
+        },
+        stream: (request, options) => {
+          const route = request.model.route
+          if (route.kind === "stream") return route.stream(request, executor.execute)
+          if (route.kind === "queued")
+            return Stream.unwrap(
+              start(request).pipe(Effect.map((generation) => resultEvents(generation, input.responseEvents, options))),
+            )
+          return Stream.fromIterableEffect(Effect.map(route.generate(request, executor.execute), input.responseEvents))
+        },
+      })
+    }),
+  ),
+  generate: (request: Req, options?: AwaitOptions) => service.use((client) => client.generate(request, options)),
+  stream: (request: Req, options?: AwaitOptions) =>
+    Stream.unwrap(service.useSync((client) => client.stream(request, options))),
+  start: (request: Req) => service.use((client) => client.start(request)),
+  resume: (model: Req["model"], token: unknown) => service.use((client) => client.resume(model, token)),
+})
+
+export * as MediaClient from "./media-client.js"
@@ -0,0 +1,110 @@
+import { Effect } from "effect"
+import { Endpoint } from "./route/endpoint.js"
+import { MediaRoute } from "./route/media.js"
+import type { MediaProtocol } from "./route/media-protocol.js"
+import { AIError, HttpOptions, InvalidRequestError, ModelID, ProviderID } from "./schema/index.js"
+
+/**
+ * What every media model carries: ids, the configured route, and deployment `http` overlays. Modality classes
+ * (`ImageModel`, `VideoModel`, `SpeechModel`, `TranscriptionModel`) extend it with their route type and a nominal
+ * marker so one cannot stand in for the other in requests.
+ */
+export class MediaModel<Route, Options> {
+  // As with `LanguageModel`, the route type is erased over `Options`; `fromRoute` and the constructor trust that the
+  // route accepts every request this model's `Options` admit.
+  declare protected readonly _Options: Options
+  readonly id: ModelID
+  readonly provider: ProviderID
+  readonly route: Route
+  readonly http?: HttpOptions
+
+  constructor(input: MediaModel.Input<Route>) {
+    this.id = ModelID.make(input.id)
+    this.provider = ProviderID.make(input.provider)
+    this.route = input.route
+    this.http = input.http
+  }
+}
+
+export namespace MediaModel {
+  export type Options = Record<string, unknown>
+
+  export interface Input<Route> {
+    readonly id: string | ModelID
+    readonly provider: string | ProviderID
+    readonly route: Route
+    readonly http?: HttpOptions
+  }
+
+  /** A protocol plus its canonical start path; `ModelInput.baseURL` overrides `baseURL` per deployment. */
+  export interface RouteInput<Request extends MediaRoute.MediaRequest, Protocol> {
+    readonly protocol: Protocol
+    readonly path: Endpoint.EndpointPart<MediaProtocol.Body, Request>
+    readonly baseURL?: string
+    /** Headers the protocol requires on every call, such as a pinned API version; deployment headers win. */
+    readonly headers?: Record<string, string>
+  }
+
+  export type InlineRouteInput<Request extends MediaRoute.MediaRequest, Response> = RouteInput<
+    Request,
+    MediaProtocol.Inline<Request, Response>
+  >
+
+  export type StreamRouteInput<Request extends MediaRoute.MediaRequest, Event, Frame, State> = RouteInput<
+    MediaProtocol.Addressed<Request>,
+    MediaProtocol.Streamed<Request, Event, Frame, State>
+  >
+
+  export type QueuedRouteInput<Request extends MediaRoute.MediaRequest, Response, Token> = RouteInput<
+    Request,
+    MediaProtocol.Queued<Request, Response, Token>
+  >
+
+  export type AnyRouteInput<Request extends MediaRoute.MediaRequest, Event, Response, Frame, State, Token> =
+    | InlineRouteInput<Request, Response>
+    | StreamRouteInput<Request, Event, Frame, State>
+    | QueuedRouteInput<Request, Response, Token>
+}
+
+/** Compose a protocol route input with one deployment through `MediaRoute.inline`, `queued`, or `stream`. */
+export const composeRoute = <Request extends MediaRoute.MediaRequest, Event, Response, Frame, State, Token>(
+  route: MediaModel.AnyRouteInput<Request, Event, Response, Frame, State, Token>,
+  input: MediaRoute.ModelInput,
+  collect: (events: ReadonlyArray<Event>) => Effect.Effect<Response, AIError>,
+): MediaRoute.AnyRoute<Request, Event, Response> => {
+  if (isStreamInput(route)) return MediaRoute.stream({ ...composition(route, input), collect })
+  if (isQueuedInput(route)) return MediaRoute.queued(composition(route, input))
+  return MediaRoute.inline(composition(route, input))
+}
+
+const composition = <Request extends MediaRoute.MediaRequest, Protocol>(
+  route: MediaModel.RouteInput<Request, Protocol>,
+  input: MediaRoute.ModelInput,
+): MediaRoute.Composition<Request> & { readonly protocol: Protocol } => ({
+  protocol: route.protocol,
+  endpoint: Endpoint.path(route.path, { baseURL: input.baseURL ?? route.baseURL }),
+  auth: input.auth,
+  headers:
+    route.headers === undefined && input.headers === undefined ? undefined : { ...route.headers, ...input.headers },
+})
+
+const isStreamInput = <Request extends MediaRoute.MediaRequest, Event, Response, Frame, State, Token>(
+  route: MediaModel.AnyRouteInput<Request, Event, Response, Frame, State, Token>,
+): route is MediaModel.StreamRouteInput<Request, Event, Frame, State> => route.protocol.kind === "stream"
+
+const isQueuedInput = <Request extends MediaRoute.MediaRequest, Event, Response, Frame, State, Token>(
+  route: MediaModel.AnyRouteInput<Request, Event, Response, Frame, State, Token>,
+): route is MediaModel.QueuedRouteInput<Request, Response, Token> => route.protocol.kind === "queued"
+
+/** Lift a synchronous Schema-class constructor into a typed `InvalidRequest` failure. */
+export const tryRequest = <A>(make: () => A): Effect.Effect<A, AIError> =>
+  Effect.try({
+    try: make,
+    catch: (error) =>
+      new AIError({
+        reason: new InvalidRequestError({
+          message: error instanceof Error ? error.message : String(error),
+          cause: error,
+        }),
+      }),
+  })
@@ -6,7 +6,7 @@ import { ProviderID } from "./schema/ids.js"
 import { AIError, HttpContext, InvalidProviderOutputError, InvalidRequestError } from "./schema/errors.js"
 import { ProviderMetadata } from "./schema/options.js"
 import { Service } from "./route/executor-service.js"
-import { detectMediaType, extensionMediaType } from "./utils/media-type.js"
+import { detectMediaType, fileMediaType } from "./utils/media-type.js"

 export { detectMediaType } from "./utils/media-type.js"

@@ -34,8 +34,6 @@ const UrlSource = Schema.Struct({
  mediaType: Schema.optional(Schema.String),
  /** Epoch milliseconds after which the provider no longer serves the URL. */
  expiresAt: Schema.optional(Schema.Number),
-  /** Headers required to fetch the URL, such as provider auth for Veo downloads. */
-  headers: Schema.optional(Schema.Record(Schema.String, Schema.String)),
 })

 /** A provider-side handle: OpenAI `file_id`, Gemini file URI, `gs://`, `runway://`, or a prior generation id. */
@@ -55,6 +53,12 @@ export type Source = Schema.Schema.Type<typeof Source>
 // Kind, Info, Notice
 // ---------------------------------------------------------------------------

+export type AspectRatio = `${number}:${number}`
+export const AspectRatio = Schema.declare<AspectRatio>(
+  (value): value is AspectRatio => typeof value === "string" && /^\d+(?:\.\d+)?:\d+(?:\.\d+)?$/.test(value),
+  { title: "Media.AspectRatio" },
+)
+
 export const Kind = Schema.Literals(["image", "video", "audio", "document", "other"])
 export type Kind = Schema.Schema.Type<typeof Kind>

@@ -110,6 +114,8 @@ export class Asset {
  /** Epoch milliseconds after which a `url` source stops resolving. */
  readonly expiresAt?: number
  readonly providerMetadata?: ProviderMetadata
+  /** Transient download credentials for `url` sources; see `Asset.Input.headers`. */
+  readonly headers?: Record<string, string>

  // Derived payload forms are cached on the instance because every protocol lowering re-reads the same payload. The
  // cache is check-then-set (concurrent first reads of a `url` source may both download) and is never observable
@@ -127,6 +133,7 @@ export class Asset {
    this.info = input.info
    this.expiresAt = input.source.type === "url" ? input.source.expiresAt : undefined
    this.providerMetadata = input.providerMetadata
+    this.headers = input.source.type === "url" ? input.headers : undefined
  }

  /** Inline payload without effects, for protocols that embed base64 or data URLs directly. */
@@ -151,7 +158,7 @@ export class Asset {
          ? Effect.fromResult(Encoding.decodeBase64(source.data)).pipe(
              Effect.mapError((cause) => invalid(`Media asset contains invalid base64 data`, cause)),
            )
-          : download(source)
+          : download(source, this.headers)
      return decoded.pipe(Effect.tap((data) => Effect.sync(() => (this.#bytes = data))))
    })
  }
@@ -198,6 +205,12 @@ export namespace Asset {
    readonly source: Source
    readonly info?: Info
    readonly providerMetadata?: ProviderMetadata
+    /**
+     * Headers required to download a `url` source, such as the provider API key Veo demands for its file URIs.
+     * They are runtime-only: never part of `source`, `toJSON()`, or `AssetSchema`, so a persisted asset cannot leak
+     * credentials and cannot be downloaded again after a round-trip. Call `materialize()` before persisting.
+     */
+    readonly headers?: Record<string, string>
  }
 }

@@ -226,10 +239,13 @@ export const AssetSchema = AssetEncoded.pipe(
  }),
 )

-const download = Effect.fn("Media.download")(function* (source: Extract<Source, { readonly type: "url" }>) {
+const download = Effect.fn("Media.download")(function* (
+  source: Extract<Source, { readonly type: "url" }>,
+  headers: Record<string, string> | undefined,
+) {
  const executor = yield* Service
  const response = yield* executor.execute(
-    HttpClientRequest.get(source.url).pipe(HttpClientRequest.setHeaders(source.headers ?? {})),
+    HttpClientRequest.get(source.url).pipe(HttpClientRequest.setHeaders(headers ?? {})),
  )
  const buffer = yield* response.arrayBuffer.pipe(
    Effect.mapError(
@@ -264,8 +280,8 @@ export const url = (
  value: string,
  options?: AssetOptions & Omit<Extract<Source, { readonly type: "url" }>, "type" | "url">,
 ) => {
-  const { mediaType, expiresAt, headers, ...rest } = options ?? {}
-  return from({ type: "url", url: value, mediaType, expiresAt, headers }, rest)
+  const { mediaType, expiresAt, ...rest } = options ?? {}
+  return from({ type: "url", url: value, mediaType, expiresAt }, rest)
 }

 export const ref = (provider: string | ProviderID, id: string, mediaType?: string, options?: AssetOptions) =>
@@ -293,7 +309,7 @@ export const file = (path: string, options?: AssetOptions): Effect.Effect<Asset,
    const data = yield* fs
      .readFile(path)
      .pipe(Effect.mapError((cause) => invalid(`Failed to read media file ${path}`, cause)))
-    return bytes(data, detectMediaType(data) ?? extensionMediaType(path), options)
+    return bytes(data, fileMediaType(data, path), options)
  })

 /** Materialize an asset and write its bytes through `FileSystem`. */
@@ -1,18 +1,30 @@
 import { Effect, Layer, ManagedRuntime, Stream } from "effect"
-import { Image, ImageModel, ImageRequest, type ImageRequestInput } from "./image.js"
-import { ImageClient } from "./image-client.js"
+import { AIClient } from "./ai-client.js"
+import type { AwaitOptions, Event, Generation, Snapshot } from "./generation.js"
+import { Image, type ImageModel, type ImageRequest, type ImageRequestInput } from "./image.js"
 import { LLM } from "./index.js"
-import { LLMClient } from "./route/client.js"
+import { Media } from "./media.js"
+import { tryRequest } from "./media-model.js"
 import { RequestExecutor } from "./route/executor.js"
-import { LanguageModel, LLMRequest } from "./schema/index.js"
+import { AIError, InvalidRequestError, LanguageModel, LLMRequest } from "./schema/index.js"
 import type { RequestInput } from "./llm.js"
+import { Speech, type SpeechModel, type SpeechRequest, type SpeechRequestInput } from "./speech.js"
+import {
+  Transcription,
+  type TranscriptionModel,
+  type TranscriptionRequest,
+  type TranscriptionRequestInput,
+} from "./transcription.js"
+import { fileMediaType } from "./utils/media-type.js"
+import { Video, type VideoModel, type VideoRequest, type VideoRequestInput } from "./video.js"

 /**
- * Promise-first entrypoint for scripts and non-Effect callers. One `ManagedRuntime` hosts the LLM and image clients
- * over a request executor; every method runs the corresponding Effect API and rethrows `AIError` unchanged.
+ * Promise-first entrypoint for scripts and non-Effect callers. One `ManagedRuntime` hosts the LLM, image, video, speech,
+ * and transcription clients over a request executor; every method runs the corresponding Effect API and rethrows
+ * `AIError` unchanged. `file` and `write` load `node:fs/promises` on first use, so importing this module does not.
 */
 export interface Options {
-  /** Executor layer; defaults to `RequestExecutor.fetchLayer`. Inject a recorder or middleware here. */
+  /** Executor layer; defaults to `RequestExecutor.fetchLayer`. Inject a recorder or `RequestExecutor.middleware(fn)` here. */
  readonly layer?: Layer.Layer<RequestExecutor.Service>
 }

@@ -20,10 +32,23 @@ export interface RunOptions {
  readonly signal?: AbortSignal
 }

-export type Services =
-  | Layer.Success<typeof LLMClient.layer>
-  | Layer.Success<typeof ImageClient.layer>
-  | RequestExecutor.Service
+export type Services = AIClient.Services
+
+/**
+ * Promise view of a `Generation`. Its fields are a snapshot taken when the handle was created; `refresh()` resolves to a
+ * new handle rather than updating this one.
+ */
+export type GenerationHandle<Response> = Snapshot & {
+  /** Serializable JSON; pass it back to `resume` from another process. */
+  readonly token: unknown
+  readonly await: (options?: AwaitOptions & RunOptions) => Promise<Response>
+  /** Status observations until the first terminal one, polling like `await`; abort ends iteration without throwing. */
+  readonly events: (options?: AwaitOptions & RunOptions) => AsyncIterable<Event>
+  /** The result without polling; fails when the generation has not completed. */
+  readonly result: (options?: RunOptions) => Promise<Response>
+  readonly refresh: (options?: RunOptions) => Promise<GenerationHandle<Response>>
+  readonly cancel: (options?: RunOptions) => Promise<void>
+}

 const abortEffect = (signal: AbortSignal | undefined) =>
  signal === undefined
@@ -39,13 +64,9 @@ const abortEffect = (signal: AbortSignal | undefined) =>
      })

 export const make = (options: Options = {}) => {
-  const runtime = ManagedRuntime.make(
-    Layer.mergeAll(LLMClient.layer, ImageClient.layer).pipe(
-      Layer.provideMerge(options.layer ?? RequestExecutor.fetchLayer),
-    ),
-  )
+  const runtime = ManagedRuntime.make(AIClient.layerWith(options.layer ?? RequestExecutor.fetchLayer))

-  /** Run any package Effect (for example `asset.bytes()`) inside this runtime. */
+  /** Run any package Effect (for example `LLMClient.compact(...)`) inside this runtime. */
  const run = <A, E>(effect: Effect.Effect<A, E, Services>, options?: RunOptions) =>
    runtime.runPromise(effect, { signal: options?.signal })

@@ -61,32 +82,121 @@ export const make = (options: Options = {}) => {
      ),
    )

-  // The typed `generate`/`stream` overloads take a concrete input or a request, not the union; normalize once here.
-  const llmRequest = (input: RequestInput | LLMRequest) => (input instanceof LLMRequest ? input : LLM.request(input))
-  const imageRequest = (input: ImageRequestInput | ImageRequest) =>
-    input instanceof ImageRequest ? input : Image.request(input)
+  const handle = <Response>(generation: Generation<Response>): GenerationHandle<Response> => ({
+    ...generation.snapshot,
+    token: generation.token,
+    await: (options) => run(generation.await({ poll: options?.poll }), options),
+    events: (options) => iterate(generation.events({ poll: options?.poll }), options),
+    result: (options) => run(generation.result(), options),
+    refresh: (options) => run(generation.refresh(), options).then(handle),
+    cancel: (options) => run(generation.cancel(), options),
+  })
+
+  const llmRequest = (input: RequestInput | LLMRequest) =>
+    input instanceof LLMRequest ? Effect.succeed(input) : tryRequest(() => LLM.request(input))

  return {
    run,
+    /** Decoded asset bytes, downloading `url` sources through the executor. */
+    bytes: (asset: Media.Asset, options?: RunOptions) => run(asset.bytes(), options),
+    base64: (asset: Media.Asset, options?: RunOptions) => run(asset.base64(), options),
+    /** Pull a `url` asset into owned bytes before the provider URL expires. */
+    materialize: (asset: Media.Asset, options?: RunOptions) => run(asset.materialize(), options),
+    /** Read a file into an asset like `Media.file`: sniffed media type, then the extension's. */
+    file: async (path: string, options?: Media.AssetOptions & RunOptions) => {
+      const { readFile } = await import("node:fs/promises")
+      return run(
+        Effect.tryPromise({
+          try: (signal) => readFile(path, { signal }),
+          catch: (cause) => fileError(`Failed to read media file ${path}`, cause),
+        }).pipe(
+          Effect.map((buffer) => {
+            const data = new Uint8Array(buffer)
+            return Media.bytes(data, fileMediaType(data, path), options)
+          }),
+        ),
+        options,
+      )
+    },
+    /** Write an asset's bytes like `Media.write`, downloading `url` sources through the executor. */
+    write: async (asset: Media.Asset, path: string, options?: RunOptions) => {
+      const { writeFile } = await import("node:fs/promises")
+      return run(
+        asset.bytes().pipe(
+          Effect.flatMap((data) =>
+            Effect.tryPromise({
+              try: (signal) => writeFile(path, data, { signal }),
+              catch: (cause) => fileError(`Failed to write media file ${path}`, cause),
+            }),
+          ),
+        ),
+        options,
+      )
+    },
    llm: {
      request: LLM.request,
-      generate: <const Model extends LanguageModel>(
-        input: RequestInput<Model> | LLMRequest,
-        options?: RunOptions,
-      ) => run(LLM.generate(llmRequest(input)), options),
+      generate: <const Model extends LanguageModel>(input: RequestInput<Model> | LLMRequest, options?: RunOptions) =>
+        run(Effect.flatMap(llmRequest(input), LLM.generate), options),
      stream: <const Model extends LanguageModel>(input: RequestInput<Model> | LLMRequest, options?: RunOptions) =>
-        iterate(LLM.stream(llmRequest(input)), options),
+        iterate(Stream.unwrap(Effect.map(llmRequest(input), LLM.stream)), options),
    },
    image: {
      request: Image.request,
      generate: <const Model extends ImageModel>(
        input: ImageRequestInput<Model> | ImageRequest,
-        options?: RunOptions,
-      ) => run(Image.generate(imageRequest(input)), options),
+        options?: AwaitOptions & RunOptions,
+      ) => run(Image.generate(input, { poll: options?.poll }), options),
      stream: <const Model extends ImageModel>(
        input: ImageRequestInput<Model> | ImageRequest,
+        options?: AwaitOptions & RunOptions,
+      ) => iterate(Image.stream(input, { poll: options?.poll }), options),
+      start: <const Model extends ImageModel>(input: ImageRequestInput<Model> | ImageRequest, options?: RunOptions) =>
+        run(Image.start(input), options).then(handle),
+      resume: (model: ImageModel, token: unknown, options?: RunOptions) =>
+        run(Image.resume(model, token), options).then(handle),
+    },
+    video: {
+      request: Video.request,
+      start: <const Model extends VideoModel>(input: VideoRequestInput<Model> | VideoRequest, options?: RunOptions) =>
+        run(Video.start(input), options).then(handle),
+      generate: <const Model extends VideoModel>(
+        input: VideoRequestInput<Model> | VideoRequest,
+        options?: AwaitOptions & RunOptions,
+      ) => run(Video.generate(input, { poll: options?.poll }), options),
+      resume: (model: VideoModel, token: unknown, options?: RunOptions) =>
+        run(Video.resume(model, token), options).then(handle),
+      stream: <const Model extends VideoModel>(
+        input: VideoRequestInput<Model> | VideoRequest,
+        options?: AwaitOptions & RunOptions,
+      ) => iterate(Video.stream(input, { poll: options?.poll }), options),
+    },
+    speech: {
+      request: Speech.request,
+      generate: <const Model extends SpeechModel>(
+        input: SpeechRequestInput<Model> | SpeechRequest,
        options?: RunOptions,
-      ) => iterate(Image.stream(imageRequest(input)), options),
+      ) => run(Speech.generate(input), options),
+      stream: <const Model extends SpeechModel>(
+        input: SpeechRequestInput<Model> | SpeechRequest,
+        options?: RunOptions,
+      ) => iterate(Speech.stream(input), options),
+    },
+    transcription: {
+      request: Transcription.request,
+      generate: <const Model extends TranscriptionModel>(
+        input: TranscriptionRequestInput<Model> | TranscriptionRequest,
+        options?: AwaitOptions & RunOptions,
+      ) => run(Transcription.generate(input, { poll: options?.poll }), options),
+      stream: <const Model extends TranscriptionModel>(
+        input: TranscriptionRequestInput<Model> | TranscriptionRequest,
+        options?: AwaitOptions & RunOptions,
+      ) => iterate(Transcription.stream(input, { poll: options?.poll }), options),
+      start: <const Model extends TranscriptionModel>(
+        input: TranscriptionRequestInput<Model> | TranscriptionRequest,
+        options?: RunOptions,
+      ) => run(Transcription.start(input), options).then(handle),
+      resume: (model: TranscriptionModel, token: unknown, options?: RunOptions) =>
+        run(Transcription.resume(model, token), options).then(handle),
    },
    dispose: () => runtime.dispose(),
  }
@@ -94,6 +204,9 @@ export const make = (options: Options = {}) => {

 export type Client = ReturnType<typeof make>

+const fileError = (message: string, cause: unknown) =>
+  new AIError({ reason: new InvalidRequestError({ message, cause }) })
+
 /** Default client over `RequestExecutor.fetchLayer` for scripts; the runtime builds its layer on first use. */
 export const ai = make()

@@ -70,7 +70,11 @@ export const protocol = Protocol.make({
      return {
        ...(yield* OpenAIChat.protocol.body.from(req)),
        enable_thinking: opts.enableThinking,
-        thinking_budget: opts.thinkingBudget,
+        // Alibaba also rejects an explicit budget that is not below `max_completion_tokens`.
+        thinking_budget:
+          opts.thinkingBudget === undefined
+            ? undefined
+            : ProviderShared.fitThinkingBudget(opts.thinkingBudget, req.generation?.maxTokens),
        preserve_thinking: opts.preserveThinking,
        clear_thinking: opts.clearThinking,
        thinking: opts.thinking,
@@ -26,18 +26,21 @@ export const protocol = Protocol.make({
    from: Effect.fn("AlibabaMessages.fromRequest")(function* (req) {
      const opts = yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(Options))(req.providerOptions ?? {})
      // Model Studio accepts enabled thinking without Anthropic's mandatory token budget.
+      const body = yield* AnthropicMessages.protocol.body.from(
+        LLMRequest.update(req, {
+          providerOptions: { ...req.providerOptions, thinking: undefined },
+        }),
+      )
+      const budget = opts.thinking?.budgetTokens ?? opts.thinking?.budget_tokens
      return {
-        ...(yield* AnthropicMessages.protocol.body.from(
-          LLMRequest.update(req, {
-            providerOptions: { ...req.providerOptions, thinking: undefined },
-          }),
-        )),
+        ...body,
        thinking:
          opts.thinking === undefined
            ? undefined
            : {
                type: opts.thinking.type,
-                budget_tokens: opts.thinking.budgetTokens ?? opts.thinking.budget_tokens,
+                budget_tokens:
+                  budget === undefined ? undefined : ProviderShared.fitThinkingBudget(budget, body.max_tokens),
              },
      }
    }),
@@ -18,7 +18,6 @@ import {
  type CacheHint,
  type FinishReasonDetails,
  type FinishReason,
-  type JsonSchema,
  type MediaPart,
  type ProviderMetadata,
  type ProviderOptions,
@@ -31,13 +30,13 @@ import { classifyProviderFailure } from "../provider-error.js"
 import { effortUpdate, resolveEffortUpdates } from "../effort-updates.js"
 import * as Cache from "./utils/cache.js"
 import { Lifecycle } from "./utils/lifecycle.js"
-import { ToolSchemaProjection } from "./utils/tool-schema.js"
 import { ToolStream } from "./utils/tool-stream.js"

 const ADAPTER = "anthropic-messages"
 export const DEFAULT_BASE_URL = "https://api.anthropic.com/v1"
 export const PATH = "/messages"
 export const DEFAULT_MAX_TOKENS = 32_000
+const MIN_THINKING_BUDGET = 1_024
 const DEFAULT_EFFORT = "high"

 const SSE_EVENTS = new Set([
@@ -524,10 +523,10 @@ const redactedDataFromMetadata = (metadata: ProviderMetadata | undefined, key: s
  return typeof provider.redactedData === "string" ? provider.redactedData : undefined
 }

-const lowerTool = (breakpoints: Cache.Breakpoints, tool: ToolDefinition, inputSchema: JsonSchema): AnthropicTool => ({
+const lowerTool = (breakpoints: Cache.Breakpoints, tool: ToolDefinition): AnthropicTool => ({
  name: tool.name,
  description: tool.description,
-  input_schema: inputSchema,
+  input_schema: tool.inputSchema,
  cache_control: cacheControl(breakpoints, tool.cache),
 })

@@ -1027,6 +1026,15 @@ const applyThinkingBindingDefault = (model: LLMRequest["model"], thinking: Anthr
  }
 }

+// Anthropic also requires an explicit thinking budget below `max_tokens` and at or above its minimum.
+const fitThinking = (thinking: AnthropicThinking | undefined, maxTokens: number) =>
+  thinking?.type === "enabled"
+    ? {
+        ...thinking,
+        budget_tokens: ProviderShared.fitThinkingBudget(thinking.budget_tokens, maxTokens, MIN_THINKING_BUDGET),
+      }
+    : thinking
+
 const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (request: LLMRequest) {
  const options = yield* decodeOptions(request.providerOptions ?? {})
  const management = options.contextManagement
@@ -1034,22 +1042,12 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
  const format = outputConfig?.format ?? undefined
  const updates = resolveEffortUpdates(request, options.effort ?? outputConfig?.effort ?? undefined)
  const generation = request.generation
-  const toolSchemaCompatibility = request.model.compatibility?.toolSchema
  // Allocate the 4-breakpoint budget in invalidation order: tools → system →
  // messages. Tools live highest in the cache hierarchy, so when callers
  // over-mark we keep their tool hints and shed the message-tail ones first.
  const breakpoints = Cache.newBreakpoints(ANTHROPIC_BREAKPOINT_CAP)
  const flattened = ProviderShared.flattenToolRequest(updates.request)
-  const tools =
-    flattened.tools.length === 0
-      ? undefined
-      : flattened.tools.map((tool) =>
-          lowerTool(
-            breakpoints,
-            tool,
-            ToolSchemaProjection.modelCompatibility(tool.inputSchema, toolSchemaCompatibility),
-          ),
-        )
+  const tools = flattened.tools.length === 0 ? undefined : flattened.tools.map((tool) => lowerTool(breakpoints, tool))
  // Anthropic rejects tool_choice when tools are absent; "none" is only meaningful with tools present.
  const toolChoice = tools === undefined || !request.toolChoice ? undefined : yield* lowerToolChoice(request.toolChoice)
  const systemParts = request.system.filter((part) => part.text.length > 0)
@@ -1069,6 +1067,7 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
  }
  const output_config =
    updates.effort === undefined && format === undefined ? undefined : { effort: updates.effort, format }
+  const maxTokens = generation?.maxTokens ?? DEFAULT_MAX_TOKENS
  const body = {
    model: request.model.id,
    system,
@@ -1076,12 +1075,12 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
    tools,
    tool_choice: toolChoice,
    stream: true as const,
-    max_tokens: generation?.maxTokens ?? DEFAULT_MAX_TOKENS,
+    max_tokens: maxTokens,
    temperature: generation?.temperature,
    top_p: generation?.topP,
    top_k: generation?.topK,
    stop_sequences: generation?.stop,
-    thinking: applyThinkingBindingDefault(request.model, options.thinking),
+    thinking: applyThinkingBindingDefault(request.model, fitThinking(options.thinking, maxTokens)),
    output_config,
    // top-level passthrough per SDK MessageCreateParamsBase:4638,4643,4649,4654,4670
    cache_control: options.cache_control ?? options.cacheControl,
@@ -0,0 +1,209 @@
+import { Effect, Schema } from "effect"
+import type { HttpClientResponse } from "effect/unstable/http"
+import type { Status } from "../generation.js"
+import { Media } from "../media.js"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { mergeJsonRecords, type OpenString } from "../schema/index.js"
+import { TranscriptionModel, TranscriptionResponse, type TranscriptionRequestFor } from "../transcription.js"
+import { ProviderShared, optionalNull } from "./shared.js"
+import { MediaInput } from "./utils/media-input.js"
+
+const route = MediaProtocol.identity({ id: "assemblyai-transcription", name: "AssemblyAI", provider: "assemblyai" })
+export const DEFAULT_BASE_URL = "https://api.assemblyai.com"
+export const PATH = "/v2/transcript"
+export const UPLOAD_PATH = "/v2/upload"
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+export type AssemblyAITranscriptionOptions = {
+  readonly keyterms_prompt?: ReadonlyArray<string>
+  readonly punctuate?: boolean
+  readonly format_text?: boolean
+  readonly disfluencies?: boolean
+  readonly filter_profanity?: boolean
+  readonly temperature?: number
+  readonly speaker_options?: { readonly min_speakers_expected?: number; readonly max_speakers_expected?: number }
+  readonly language_detection_options?: {
+    readonly expected_languages?: ReadonlyArray<string>
+    readonly fallback_language?: string
+    readonly code_switching?: boolean
+  }
+  readonly speech_models?: ReadonlyArray<OpenString<"universal-3-5-pro" | "universal-2">>
+} & Record<string, unknown>
+
+export type Request = TranscriptionRequestFor<AssemblyAITranscriptionOptions>
+
+// ---------------------------------------------------------------------------
+// 2. Token and response schemas
+// ---------------------------------------------------------------------------
+
+export const Token = Schema.Struct({ transcriptID: Schema.String })
+export type Token = Schema.Schema.Type<typeof Token>
+
+const Upload = Schema.Struct({ upload_url: Schema.String })
+
+/** Word and utterance times are milliseconds. */
+const Transcript = Schema.Struct({
+  id: Schema.String,
+  status: Schema.String,
+  text: optionalNull(Schema.String),
+  words: optionalNull(
+    Schema.Array(
+      Schema.Struct({
+        text: Schema.String,
+        start: Schema.Number,
+        end: Schema.Number,
+        confidence: optionalNull(Schema.Number),
+        speaker: optionalNull(Schema.String),
+      }),
+    ),
+  ),
+  utterances: optionalNull(
+    Schema.Array(
+      Schema.Struct({
+        text: Schema.String,
+        start: Schema.Number,
+        end: Schema.Number,
+        speaker: optionalNull(Schema.String),
+      }),
+    ),
+  ),
+  language_code: optionalNull(Schema.String),
+  audio_duration: optionalNull(Schema.Number),
+  speech_model_used: optionalNull(Schema.String),
+  error: optionalNull(Schema.String),
+})
+
+const STATUS = {
+  queued: "queued",
+  processing: "running",
+  completed: "completed",
+  error: "failed",
+} as const satisfies Record<string, Status>
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+const decodeUpload = route.decodeJson(Upload)
+
+/** `/v2/transcript` only takes a URL, so inline audio is uploaded to `/v2/upload` first. */
+const prepare = Effect.fn("AssemblyAITranscription.prepare")(function* (request: Request, send: MediaProtocol.Send) {
+  if (request.audio.source.type !== "bytes" && request.audio.source.type !== "base64") return request
+  const audio = yield* MediaInput.inlineBytes(route.id, request.audio)
+  const uploaded = yield* send(UPLOAD_PATH, MediaProtocol.binary(audio, "application/octet-stream")).pipe(
+    Effect.flatMap(decodeUpload),
+  )
+  return { ...request, audio: Media.url(uploaded.value.upload_url, { mediaType: request.audio.mediaType }) }
+})
+
+const fromRequest = Effect.fn("AssemblyAITranscription.fromRequest")(function* (request: Request) {
+  const audio = yield* ProviderShared.mediaReference(request.audio, route.provider, route.name)
+  return MediaProtocol.json(
+    mergeJsonRecords(
+      {
+        audio_url: audio.value,
+        speech_models: [request.model.id],
+        language_code: request.language,
+        language_detection: request.language === undefined ? true : undefined,
+        prompt: request.prompt,
+        // Turn-level `utterances`, the only segments AssemblyAI returns, and `speakers_expected` require speaker labels.
+        speaker_labels:
+          request.diarize === true || request.timestamps === "segment" || request.speakers !== undefined
+            ? true
+            : undefined,
+        speakers_expected: request.speakers,
+      },
+      request.providerOptions,
+      request.http?.body,
+    ) ?? {},
+  )
+})
+
+// ---------------------------------------------------------------------------
+// 6. Response decoding
+// ---------------------------------------------------------------------------
+
+const decodeTranscript = route.decodeJson(Transcript)
+
+const decodeStart = Effect.fn("AssemblyAITranscription.decodeStart")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+) {
+  const output = yield* decodeTranscript(response)
+  const status = yield* MediaProtocol.status(STATUS, output.value.status, output)
+  return { token: { transcriptID: output.value.id }, snapshot: { id: output.value.id, status } }
+})
+
+const decodeStatus = Effect.fn("AssemblyAITranscription.decodeStatus")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.PollContext<Token>,
+) {
+  const output = yield* decodeTranscript(response)
+  const status = yield* MediaProtocol.status(STATUS, output.value.status, output)
+  return { id: context.token.transcriptID, status }
+})
+
+const seconds = (milliseconds: number) => milliseconds / 1000
+
+const decodeResult = Effect.fn("AssemblyAITranscription.decodeResult")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.PollContext<Token>,
+) {
+  const output = yield* decodeTranscript(response)
+  const transcript = output.value
+  const status = yield* MediaProtocol.status(STATUS, transcript.status, output)
+  const error = transcript.error ?? undefined
+  if (status === "failed")
+    return yield* output.ended("failed", `${route.name} transcription failed${error === undefined ? "" : `: ${error}`}`)
+  if (status !== "completed") return yield* output.pending(context.token.transcriptID)
+  const duration = transcript.audio_duration ?? undefined
+  return new TranscriptionResponse({
+    text: transcript.text ?? "",
+    segments: transcript.utterances?.map((utterance) => ({
+      text: utterance.text,
+      startSeconds: seconds(utterance.start),
+      endSeconds: seconds(utterance.end),
+      speaker: utterance.speaker ?? undefined,
+    })),
+    words: transcript.words?.map((word) => ({
+      text: word.text,
+      startSeconds: seconds(word.start),
+      endSeconds: seconds(word.end),
+      speaker: word.speaker ?? undefined,
+      confidence: word.confidence ?? undefined,
+    })),
+    language: transcript.language_code?.toLowerCase(),
+    durationSeconds: duration,
+    usage: duration === undefined ? undefined : { type: "seconds", seconds: duration },
+    providerMetadata: {
+      assemblyai: { transcriptId: transcript.id, speechModel: transcript.speech_model_used ?? undefined },
+    },
+  })
+})
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+const transcriptPath = (token: Token) => `${PATH}/${token.transcriptID}`
+
+export const protocol = MediaProtocol.queued<Request, TranscriptionResponse, Token>(route, {
+  token: Token,
+  start: { prepare, body: { from: fromRequest }, decode: decodeStart },
+  status: { path: transcriptPath, decode: decodeStatus },
+  result: { path: transcriptPath, decode: decodeResult },
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  TranscriptionModel.fromRoute<AssemblyAITranscriptionOptions, Token>(
+    { protocol, baseURL: DEFAULT_BASE_URL, path: PATH },
+    input,
+  )
+
+export const AssemblyAITranscription = {
+  protocol,
+  model,
+} as const
@@ -9,9 +9,8 @@ import {
  type CacheHint,
  type FinishReason,
  type FinishReasonDetails,
-  type JsonSchema,
  type LLMRequest,
-  type LanguageModelToolSchemaCompatibility,
+  type LanguageModel,
  type ProviderMetadata,
  type ReasoningPart,
  type ToolCallPart,
@@ -26,8 +25,8 @@ import { BedrockCache } from "./utils/bedrock-cache.js"
 import { BedrockMedia } from "./utils/bedrock-media.js"
 import { Lifecycle } from "./utils/lifecycle.js"
 import { MistralToolID } from "./utils/mistral-tool-id.js"
-import { ToolSchemaProjection } from "./utils/tool-schema.js"
 import { ToolStream } from "./utils/tool-stream.js"
+import { concatBytes } from "../utils/bytes.js"

 const ADAPTER = "bedrock-converse"

@@ -220,22 +219,18 @@ type BedrockEvent = Schema.Schema.Type<typeof BedrockEvent>
 // =============================================================================
 // Request Lowering
 // =============================================================================
-const lowerToolSpec = (tool: ToolDefinition, inputSchema: JsonSchema): BedrockToolSpec => ({
+const lowerToolSpec = (tool: ToolDefinition): BedrockToolSpec => ({
  toolSpec: {
    name: tool.name,
    ...(tool.description.trim().length > 0 ? { description: tool.description } : {}),
-    inputSchema: { json: inputSchema },
+    inputSchema: { json: tool.inputSchema },
  },
 })

-const lowerTools = (
-  compatibility: LanguageModelToolSchemaCompatibility | undefined,
-  breakpoints: BedrockCache.Breakpoints,
-  tools: ReadonlyArray<ToolDefinition>,
-): BedrockTool[] => {
+const lowerTools = (breakpoints: BedrockCache.Breakpoints, tools: ReadonlyArray<ToolDefinition>): BedrockTool[] => {
  const result: BedrockTool[] = []
  for (const tool of tools) {
-    result.push(lowerToolSpec(tool, ToolSchemaProjection.modelCompatibility(tool.inputSchema, compatibility)))
+    result.push(lowerToolSpec(tool))
    const cachePoint = BedrockCache.block(breakpoints, tool.cache)
    if (cachePoint) result.push(cachePoint)
  }
@@ -429,17 +424,50 @@ const lowerSystem = (breakpoints: BedrockCache.Breakpoints, system: ReadonlyArra
  return content.length === 0 ? undefined : content
 }

+// Nova 2 rejects `maxTokens` at high reasoning effort, where its output can exceed the field's maximum. Other models
+// that take `reasoningConfig`, such as Grok on Bedrock, accept it.
+const isNova2 = (model: LanguageModel) => /\bamazon\.nova-2-/.test(model.id)
+const isHighReasoningEffort = Schema.is(
+  Schema.Struct({
+    additionalModelRequestFields: Schema.Struct({
+      reasoningConfig: Schema.Struct({ maxReasoningEffort: Schema.Literal("high") }),
+    }),
+  }),
+)
+
+const Options = Schema.Struct({
+  thinking: Schema.optional(Schema.Struct({ type: Schema.Literal("enabled"), budgetTokens: Schema.Number })),
+})
+export type OptionsInput = typeof Options.Type
+const decodeOptions = ProviderShared.validateWith(Schema.decodeUnknownEffect(Options))
+// Claude on Bedrock requires the thinking budget below `maxTokens`, with a minimum of 1,024.
+const MIN_THINKING_BUDGET = 1_024
+
 const fromRequest = Effect.fn("BedrockConverse.fromRequest")(function* (request: LLMRequest) {
  const toolChoice = request.toolChoice ? yield* lowerToolChoice(request.toolChoice) : undefined
  const flattened = ProviderShared.flattenToolRequest(request)
  const generation = request.generation
+  const options = yield* decodeOptions(request.providerOptions ?? {})
+  const maxTokens =
+    isNova2(request.model) && isHighReasoningEffort(request.http?.body) ? undefined : generation?.maxTokens
+  const thinking =
+    options.thinking === undefined
+      ? undefined
+      : {
+          type: "enabled",
+          budget_tokens: ProviderShared.fitThinkingBudget(
+            options.thinking.budgetTokens,
+            maxTokens,
+            MIN_THINKING_BUDGET,
+          ),
+        }
  // Bedrock-Claude shares Anthropic's 4-breakpoint cap. Spend the budget in
  // tools → system → messages order to favour the highest-impact prefixes.
  const breakpoints = BedrockCache.breakpoints(request.model.id)
  const toolConfig = (() => {
    if (flattened.tools.length === 0) return undefined
    return {
-      tools: lowerTools(request.model.compatibility?.toolSchema, breakpoints, flattened.tools),
+      tools: lowerTools(breakpoints, flattened.tools),
      // Converse has no native "none". Keep definitions stable for prompt
      // caching and omit only the unsupported choice.
      toolChoice,
@@ -454,14 +482,14 @@ const fromRequest = Effect.fn("BedrockConverse.fromRequest")(function* (request:
  }
  const inferenceConfig = (() => {
    if (
-      generation?.maxTokens === undefined &&
+      maxTokens === undefined &&
      generation?.temperature === undefined &&
      generation?.topP === undefined &&
      (generation?.stop === undefined || generation.stop.length === 0)
    )
      return undefined
    return {
-      maxTokens: generation?.maxTokens,
+      maxTokens,
      temperature: generation?.temperature,
      topP: generation?.topP,
      stopSequences: generation?.stop,
@@ -473,9 +501,15 @@ const fromRequest = Effect.fn("BedrockConverse.fromRequest")(function* (request:
    system,
    inferenceConfig,
    toolConfig,
-    // Converse's base inferenceConfig has no topK; Anthropic/Nova accept it
-    // as a model-specific field, so it goes through additionalModelRequestFields.
-    additionalModelRequestFields: generation?.topK === undefined ? undefined : { top_k: generation.topK },
+    // Converse's base inferenceConfig has no topK or thinking; Anthropic/Nova accept them
+    // as model-specific fields, so they go through additionalModelRequestFields.
+    additionalModelRequestFields:
+      generation?.topK === undefined && thinking === undefined
+        ? undefined
+        : {
+            ...(generation?.topK === undefined ? {} : { top_k: generation.topK }),
+            ...(thinking === undefined ? {} : { thinking }),
+          },
  }
 })

@@ -524,14 +558,7 @@ interface ParserState {
  readonly reasoningRedactedContent: Readonly<Record<number, ReadonlyArray<Uint8Array>>>
 }

-const encodeRedactedContent = (chunks: ReadonlyArray<Uint8Array>) => {
-  const bytes = new Uint8Array(chunks.reduce((total, chunk) => total + chunk.length, 0))
-  chunks.reduce((offset, chunk) => {
-    bytes.set(chunk, offset)
-    return offset + chunk.length
-  }, 0)
-  return Encoding.encodeBase64(bytes)
-}
+const encodeRedactedContent = (chunks: ReadonlyArray<Uint8Array>) => Encoding.encodeBase64(concatBytes(chunks))

 const step = (state: ParserState, event: BedrockEvent) =>
  Effect.gen(function* () {
@@ -0,0 +1,220 @@
+import { Effect, Schema } from "effect"
+import type { HttpClientResponse } from "effect/unstable/http"
+import type { Status } from "../generation.js"
+import { ImageModel, ImageResponse, type ImageRequestFor } from "../image.js"
+import { Media } from "../media.js"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { mergeJsonRecords } from "../schema/index.js"
+import { ProviderShared, optionalNull } from "./shared.js"
+import { MediaInput } from "./utils/media-input.js"
+
+const route = MediaProtocol.identity({ id: "bfl-images", name: "Black Forest Labs", provider: "black-forest-labs" })
+export const DEFAULT_BASE_URL = "https://api.bfl.ai"
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+export type BlackForestLabsImageOptions = {
+  readonly safety_tolerance?: number
+  readonly prompt_upsampling?: boolean
+  readonly disable_pup?: boolean
+  readonly raw?: boolean
+  readonly guidance?: number
+  readonly steps?: number
+} & Record<string, unknown>
+
+export type Request = ImageRequestFor<BlackForestLabsImageOptions>
+
+// ---------------------------------------------------------------------------
+// 2. Token and response schemas
+// ---------------------------------------------------------------------------
+
+/**
+ * Regional clusters answer on different hosts, so the returned `polling_url` is followed verbatim. BFL reports the
+ * credit cost on submit, so it rides on the token; it is optional so tokens persisted before it existed still decode.
+ */
+export const Token = Schema.Struct({
+  id: Schema.String,
+  pollingURL: Schema.String,
+  cost: Schema.optionalKey(Schema.Number),
+})
+export type Token = Schema.Schema.Type<typeof Token>
+
+const StartResponse = Schema.Struct({
+  id: Schema.String,
+  polling_url: Schema.String,
+  cost: optionalNull(Schema.Number),
+})
+
+const Result = Schema.Struct({
+  id: Schema.String,
+  status: Schema.String,
+  result: optionalNull(
+    Schema.StructWithRest(
+      Schema.Struct({ sample: Schema.String, seed: optionalNull(Schema.Number), prompt: optionalNull(Schema.String) }),
+      [Schema.Record(Schema.String, Schema.Unknown)],
+    ),
+  ),
+  cost: optionalNull(Schema.Number),
+})
+
+const STATUS = {
+  Pending: "running",
+  Reasoning: "running",
+  Generating: "running",
+  Ready: "completed",
+  Error: "failed",
+  // Moderation is terminal; `decodeResult` reports it as a content-policy failure.
+  "Content Moderated": "failed",
+  "Request Moderated": "failed",
+  "Task not found": "expired",
+} as const satisfies Record<string, Status>
+
+const isModerated = (status: string) => status === "Content Moderated" || status === "Request Moderated"
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+interface Capabilities {
+  readonly sizing: "dimensions" | "aspectRatio" | "none"
+  /** `input_image` numbers extra references `input_image_2`…; the other fields take one image. */
+  readonly imageField: "input_image" | "image_prompt" | "image"
+  readonly maxImages: number
+  readonly mask: boolean
+}
+
+const capabilities = (model: string): Capabilities => {
+  if (model.startsWith("flux-pro-1.0-fill")) return { sizing: "none", imageField: "image", maxImages: 1, mask: true }
+  if (model.startsWith("flux-pro-1.0-expand")) return { sizing: "none", imageField: "image", maxImages: 1, mask: false }
+  if (model.startsWith("flux-kontext"))
+    return { sizing: "aspectRatio", imageField: "input_image", maxImages: 4, mask: false }
+  if (model.startsWith("flux-pro-1.1-ultra"))
+    return { sizing: "aspectRatio", imageField: "image_prompt", maxImages: 1, mask: false }
+  if (model.startsWith("flux-pro-1.1") || model.startsWith("flux-dev"))
+    return { sizing: "dimensions", imageField: "image_prompt", maxImages: 1, mask: false }
+  if (model.startsWith("flux-2-klein"))
+    return { sizing: "dimensions", imageField: "input_image", maxImages: 4, mask: false }
+  return { sizing: "dimensions", imageField: "input_image", maxImages: 8, mask: false }
+}
+
+const validate = (request: Request, model: Capabilities) => {
+  const id = request.model.id
+  const images = request.images?.length ?? 0
+  if (request.n !== undefined && request.n > 1)
+    return Effect.fail(route.unsupported("media.n", `${id} generates one image per request; call it once per image`))
+  if (request.size !== undefined && model.sizing !== "dimensions")
+    return Effect.fail(route.unsupported("media.size", `${id} does not take size (width and height)`))
+  if (request.aspectRatio !== undefined && model.sizing !== "aspectRatio")
+    return Effect.fail(route.unsupported("media.aspectRatio", `${id} does not take aspectRatio`))
+  if (images > model.maxImages)
+    return Effect.fail(route.unsupported("media.images", `${id} takes at most ${model.maxImages} images`))
+  if (request.mask !== undefined && !model.mask)
+    return Effect.fail(route.unsupported("media.mask", `${id} does not inpaint; use flux-pro-1.0-fill`))
+  return Effect.void
+}
+
+const imageInput = (asset: Media.Asset) => {
+  const value = asset.inline()?.base64 ?? ProviderShared.mediaUrl(asset)
+  if (value === undefined)
+    return Effect.fail(ProviderShared.invalidRequest(`${route.name} accepts inline images or https URLs`))
+  return Effect.succeed(value)
+}
+
+const fromRequest = Effect.fn("BlackForestLabsImages.fromRequest")(function* (request: Request) {
+  const model = capabilities(request.model.id)
+  yield* validate(request, model)
+  const images = yield* Effect.forEach(request.images ?? [], imageInput)
+  const fields = images.map((image, index) => [
+    index === 0 ? model.imageField : `${model.imageField}_${index + 1}`,
+    image,
+  ])
+  return MediaProtocol.json(
+    mergeJsonRecords(
+      {
+        prompt: request.prompt,
+        ...(request.size === undefined ? {} : MediaInput.dimensions(request.size)),
+        aspect_ratio: request.aspectRatio,
+        seed: request.seed,
+        output_format: request.format,
+        mask: request.mask === undefined ? undefined : yield* imageInput(request.mask),
+        ...Object.fromEntries(fields),
+      },
+      request.providerOptions,
+      request.http?.body,
+    ) ?? {},
+  )
+})
+
+// ---------------------------------------------------------------------------
+// 6. Response decoding
+// ---------------------------------------------------------------------------
+
+const decodeStart = route.decodeStarted(StartResponse, (value) => ({
+  token: {
+    id: value.id,
+    pollingURL: value.polling_url,
+    ...(value.cost === undefined || value.cost === null ? {} : { cost: value.cost }),
+  },
+  snapshot: { id: value.id, status: "queued" },
+}))
+
+const decodeDocument = route.decodeJson(Result)
+
+const decodeStatus = Effect.fn("BlackForestLabsImages.decodeStatus")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.PollContext<Token>,
+) {
+  const output = yield* decodeDocument(response)
+  return { id: context.token.id, status: yield* MediaProtocol.status(STATUS, output.value.status, output) }
+})
+
+const decodeResult = Effect.fn("BlackForestLabsImages.decodeResult")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.PollContext<Token>,
+) {
+  const output = yield* decodeDocument(response)
+  const document = output.value
+  const status = yield* MediaProtocol.status(STATUS, document.status, output)
+  if (isModerated(document.status)) return yield* output.contentPolicy(`${route.name} moderated the generation`)
+  if (status === "failed" || status === "expired")
+    return yield* output.ended(status, `${route.name} generation ${context.token.id} ended with ${document.status}`)
+  if (status !== "completed") return yield* output.pending(context.token.id)
+  if (document.result === undefined || document.result === null)
+    return yield* output.invalid(`${route.name} generation ${context.token.id} has no result`)
+  const { sample, seed, prompt, ...rest } = document.result
+  // A settled `cost` on the result supersedes the submit-time cost carried on the token.
+  const cost = document.cost ?? context.token.cost
+  return new ImageResponse({
+    // `sample` is a signed URL that expires 10 minutes after the result is ready, so it is downloaded now.
+    images: [yield* context.materialize(Media.url(sample))],
+    usage: cost === undefined ? undefined : { type: "credits", credits: cost },
+    providerMetadata: {
+      bfl: { id: context.token.id, seed: seed ?? undefined, prompt: prompt ?? undefined, ...rest },
+    },
+  })
+})
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+export const protocol = MediaProtocol.queued<Request, ImageResponse, Token>(route, {
+  token: Token,
+  start: { body: { from: fromRequest }, decode: decodeStart },
+  status: { path: (token) => token.pollingURL, decode: decodeStatus },
+  result: { path: (token) => token.pollingURL, decode: decodeResult },
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  ImageModel.fromRoute<BlackForestLabsImageOptions, Token>(
+    { protocol, baseURL: DEFAULT_BASE_URL, path: ({ request }) => `/v1/${request.model.id}` },
+    input,
+  )
+
+export const BlackForestLabsImages = {
+  protocol,
+  model,
+} as const
@@ -0,0 +1,185 @@
+import { Effect, Schema } from "effect"
+import { classifyProviderFailure } from "../provider-error.js"
+import { Framing } from "../route/framing.js"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { AIError, mergeJsonRecords, type OpenString } from "../schema/index.js"
+import { SpeechModel, type SpeechEvent, type SpeechRequestFor } from "../speech.js"
+import { ProviderShared, optionalNull } from "./shared.js"
+import { SpeechStream } from "./utils/speech-stream.js"
+
+const route = MediaProtocol.identity({ id: "cartesia-speech", name: "Cartesia", provider: "cartesia" })
+export const DEFAULT_BASE_URL = "https://api.cartesia.ai"
+export const API_VERSION = "2026-08-14"
+export const BYTES_PATH = "/tts/bytes"
+export const SSE_PATH = "/tts/sse"
+const DEFAULT_SAMPLE_RATE = 44100
+const DEFAULT_BIT_RATE = 128000
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+export type CartesiaEncoding = SpeechStream.PcmEncoding
+
+export type CartesiaSpeechOptions = {
+  readonly sampleRate?: 8000 | 16000 | 22050 | 24000 | 44100 | 48000
+  readonly bitRate?: 32000 | 64000 | 96000 | 128000 | 192000
+  readonly encoding?: CartesiaEncoding
+  readonly generation_config?: {
+    readonly volume?: number
+    readonly emotion?: OpenString<"neutral" | "calm" | "angry" | "content" | "sad" | "scared">
+  }
+  readonly pronunciation_dict_id?: string
+} & Record<string, unknown>
+
+export type Request = SpeechRequestFor<CartesiaSpeechOptions>
+
+// ---------------------------------------------------------------------------
+// 3. Streaming event schema
+// ---------------------------------------------------------------------------
+
+/** `phoneme_timestamps` and future record types are ignored. */
+const SseEvent = Schema.Struct({
+  type: Schema.String,
+  data: Schema.optional(Schema.Uint8ArrayFromBase64),
+  word_timestamps: Schema.optional(
+    Schema.Struct({
+      words: Schema.Array(Schema.String),
+      start: Schema.Array(Schema.Number),
+      end: Schema.Array(Schema.Number),
+    }),
+  ),
+  status_code: Schema.optional(Schema.Number),
+  title: Schema.optional(Schema.String),
+  message: Schema.optional(Schema.String),
+  error_code: optionalNull(Schema.String),
+})
+
+const decodeEvent = route.decodeFrame(SseEvent)
+
+// ---------------------------------------------------------------------------
+// 4. Parser state
+// ---------------------------------------------------------------------------
+
+interface State extends SpeechStream.Audio {
+  readonly done: boolean
+}
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+/** Timestamps exist only on the SSE endpoint, so a `generate` that asks for them collects an SSE stream. */
+const usesSse = (request: MediaProtocol.Addressed<Request>) => request.mode === "stream" || request.timestamps === true
+
+const CONTAINERS: Readonly<Record<string, "raw" | "wav" | "mp3">> = { pcm: "raw", wav: "wav", mp3: "mp3" }
+
+const outputFormat = Effect.fn("CartesiaSpeech.outputFormat")(function* (request: MediaProtocol.Addressed<Request>) {
+  const sse = usesSse(request)
+  const format = request.format ?? (sse ? "pcm" : "mp3")
+  const container = CONTAINERS[format]
+  if (container === undefined)
+    return yield* route.unsupported(
+      "media.format",
+      `${route.name} supports the pcm, wav, and mp3 formats, not "${format}"`,
+    )
+  if (sse && container !== "raw")
+    return yield* route.unsupported(
+      "media.format",
+      `${route.name} streams and timestamps only raw PCM; request format "pcm" instead of "${format}"`,
+    )
+  const sampleRate = request.providerOptions?.sampleRate ?? DEFAULT_SAMPLE_RATE
+  if (container === "mp3")
+    return { container, sample_rate: sampleRate, bit_rate: request.providerOptions?.bitRate ?? DEFAULT_BIT_RATE }
+  return { container, encoding: request.providerOptions?.encoding ?? "pcm_s16le", sample_rate: sampleRate }
+})
+
+const fromRequest = Effect.fn("CartesiaSpeech.fromRequest")(function* (request: MediaProtocol.Addressed<Request>) {
+  const voice = SpeechStream.voiceID(request.voice)
+  if (voice === undefined)
+    return yield* ProviderShared.invalidRequest(`${route.name} requires a voice id; pass it as \`voice\``)
+  const { sampleRate: _sampleRate, bitRate: _bitRate, encoding: _encoding, ...native } = request.providerOptions ?? {}
+  return MediaProtocol.json(
+    mergeJsonRecords(
+      {
+        model_id: request.model.id,
+        transcript: request.text,
+        voice,
+        output_format: yield* outputFormat(request),
+        language: request.language,
+        generation_config: request.speed === undefined ? undefined : { speed: request.speed },
+        add_timestamps: request.timestamps === true ? true : undefined,
+      },
+      native,
+      request.http?.body,
+    ) ?? {},
+  )
+})
+
+// ---------------------------------------------------------------------------
+// 6. Stream parsing
+// ---------------------------------------------------------------------------
+
+const onEvent = Effect.fn("CartesiaSpeech.onEvent")(function* (state: State, frame: string) {
+  const event = yield* decodeEvent(frame)
+  if (event.type === "chunk" && event.data !== undefined) return SpeechStream.delta(state, event.data)
+  if (event.type === "timestamps" && event.word_timestamps !== undefined) {
+    const words = event.word_timestamps
+    return [state, SpeechStream.timestamps(words.words, words.start, words.end)] as const
+  }
+  if (event.type === "done") return [{ ...state, done: true }, []] as const
+  if (event.type === "error")
+    return yield* new AIError({
+      reason: classifyProviderFailure({
+        message: `${route.name} stream failed${event.title === undefined ? "" : ` (${event.title})`}: ${event.message ?? "unknown error"}`,
+        status: event.status_code,
+        rawBody: frame,
+      }),
+    })
+  return [state, []] as const
+})
+
+const finish = Effect.fn("CartesiaSpeech.finish")(function* (
+  state: State,
+  context: MediaProtocol.ResponseContext<Request>,
+) {
+  if (usesSse(context.request) && !state.done) return yield* route.incomplete()
+  const format = yield* outputFormat(context.request)
+  return yield* SpeechStream.finish(
+    route,
+    state,
+    format.container === "raw"
+      ? SpeechStream.pcm(format.encoding, format.sample_rate)
+      : SpeechStream.container(format.container, format.sample_rate),
+  )
+})
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+export const protocol = MediaProtocol.stream<Request, SpeechEvent, string | Uint8Array, State>(route, {
+  unsupported: ["instructions"],
+  body: { from: fromRequest },
+  frames: (bytes, context) => (usesSse(context.request) ? Framing.sse.frame(bytes) : bytes),
+  initial: () => ({ chunks: [], done: false }),
+  step: SpeechStream.step(onEvent),
+  finish,
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  SpeechModel.fromRoute<CartesiaSpeechOptions, string | Uint8Array, State>(
+    {
+      protocol,
+      baseURL: DEFAULT_BASE_URL,
+      headers: { "Cartesia-Version": API_VERSION },
+      path: ({ request }) => (usesSse(request) ? SSE_PATH : BYTES_PATH),
+    },
+    input,
+  )
+
+export const CartesiaSpeech = {
+  protocol,
+  model,
+} as const
@@ -0,0 +1,140 @@
+import { Effect } from "effect"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { mergeJsonRecords, type OpenString } from "../schema/index.js"
+import { SpeechModel, type SpeechEvent, type SpeechRequestFor } from "../speech.js"
+import { MediaInput } from "./utils/media-input.js"
+import { SpeechStream } from "./utils/speech-stream.js"
+
+const route = MediaProtocol.identity({ id: "deepgram-speech", name: "Deepgram", provider: "deepgram" })
+export const DEFAULT_BASE_URL = "https://api.deepgram.com"
+export const PATH = "/v1/speak"
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+export type DeepgramEncoding = OpenString<"linear16" | "mulaw" | "alaw" | "mp3" | "opus" | "flac" | "aac">
+
+export type DeepgramSpeechOptions = {
+  readonly encoding?: DeepgramEncoding
+  readonly container?: OpenString<"wav" | "ogg" | "none">
+  readonly sampleRate?: number
+  readonly bitRate?: number
+  readonly mip_opt_out?: boolean
+  readonly tag?: string
+} & Record<string, unknown>
+
+export type Request = SpeechRequestFor<DeepgramSpeechOptions>
+
+// ---------------------------------------------------------------------------
+// 4. Parser state
+// ---------------------------------------------------------------------------
+
+type State = SpeechStream.Audio
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+const FORMATS: Readonly<Record<string, { readonly encoding: string; readonly container?: string }>> = {
+  mp3: { encoding: "mp3" },
+  wav: { encoding: "linear16", container: "wav" },
+  pcm: { encoding: "linear16", container: "none" },
+  opus: { encoding: "opus" },
+  flac: { encoding: "flac" },
+  aac: { encoding: "aac" },
+}
+
+const audioFormat = (request: Request) => {
+  const format = request.format === undefined ? undefined : FORMATS[request.format]
+  return {
+    encoding: request.providerOptions?.encoding ?? format?.encoding,
+    container: request.providerOptions?.container ?? format?.container,
+  }
+}
+
+const queryParameters = (request: Request) => {
+  const { encoding: _encoding, container: _container, sampleRate, bitRate, ...native } = request.providerOptions ?? {}
+  return MediaInput.query(route.id, {
+    ...native,
+    model: request.model.id,
+    ...audioFormat(request),
+    sample_rate: sampleRate,
+    bit_rate: bitRate,
+    speed: request.speed,
+  })
+}
+
+const fromRequest = Effect.fn("DeepgramSpeech.fromRequest")(function* (request: Request) {
+  // Not in `unsupported`: that list would also reject `timestamps: false`, which asks for nothing.
+  if (request.timestamps === true)
+    return yield* route.unsupported("media.timestamps", `${route.name} does not return timestamps`)
+  if (
+    request.format !== undefined &&
+    FORMATS[request.format] === undefined &&
+    request.providerOptions?.encoding === undefined
+  )
+    return yield* route.unsupported(
+      "media.format",
+      `${route.name} has no encoding for format "${request.format}"; pass providerOptions.encoding`,
+    )
+  return MediaProtocol.json(
+    mergeJsonRecords({ text: request.text }, request.http?.body) ?? {},
+    yield* queryParameters(request),
+  )
+})
+
+// ---------------------------------------------------------------------------
+// 6. Stream parsing
+// ---------------------------------------------------------------------------
+
+const HEADERLESS_ENCODINGS: Readonly<Record<string, SpeechStream.PcmEncoding>> = {
+  linear16: "pcm_s16le",
+  mulaw: "pcm_mulaw",
+  alaw: "pcm_alaw",
+}
+
+const finish = (state: State, context: MediaProtocol.ResponseContext<Request>) => {
+  const headers = context.http.headers
+  const mediaType = headers["content-type"]
+  const format = audioFormat(context.request)
+  const encoding = HEADERLESS_ENCODINGS[format.encoding ?? ""]
+  const requestID = headers["dg-request-id"]
+  const modelName = headers["dg-model-name"]
+  return SpeechStream.finish(route, state, {
+    ...(format.container === "none" && encoding !== undefined
+      ? SpeechStream.pcm(encoding, SpeechStream.sampleRate(mediaType), mediaType)
+      : // Deepgram's default encoding is MP3; WAV is a container around any encoding.
+        { mediaType, info: { format: format.container === "wav" ? "wav" : (format.encoding ?? "mp3") } }),
+    usage: SpeechStream.headerUsage("characters", headers["dg-char-count"]),
+    providerMetadata:
+      requestID === undefined && modelName === undefined
+        ? undefined
+        : { deepgram: { requestId: requestID, modelName } },
+  })
+}
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+export const protocol = MediaProtocol.stream<Request, SpeechEvent, Uint8Array, State>(route, {
+  unsupported: ["voice", "language", "instructions"],
+  body: { from: fromRequest },
+  frames: (bytes) => bytes,
+  initial: () => ({ chunks: [] }),
+  step: (state, frame) => Effect.succeed(SpeechStream.delta(state, frame)),
+  finish,
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  SpeechModel.fromRoute<DeepgramSpeechOptions, Uint8Array, State>(
+    { protocol, baseURL: DEFAULT_BASE_URL, path: PATH },
+    input,
+  )
+
+export const DeepgramSpeech = {
+  protocol,
+  model,
+} as const
@@ -0,0 +1,186 @@
+import { Effect, Schema } from "effect"
+import type { HttpClientResponse } from "effect/unstable/http"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { mergeJsonRecords, type OpenString } from "../schema/index.js"
+import { TranscriptionModel, TranscriptionResponse, type TranscriptionRequestFor } from "../transcription.js"
+import { ProviderShared } from "./shared.js"
+import { MediaInput } from "./utils/media-input.js"
+
+const route = MediaProtocol.identity({ id: "deepgram-transcription", name: "Deepgram", provider: "deepgram" })
+export const DEFAULT_BASE_URL = "https://api.deepgram.com"
+export const PATH = "/v1/listen"
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+export type DeepgramTranscriptionOptions = {
+  readonly smart_format?: boolean
+  readonly punctuate?: boolean
+  readonly paragraphs?: boolean
+  readonly utterances?: boolean
+  readonly detect_language?: boolean | ReadonlyArray<string>
+  readonly keyterm?: ReadonlyArray<string>
+  readonly diarize_model?: OpenString<"latest" | "v1" | "v2">
+  readonly filler_words?: boolean
+  readonly numerals?: boolean
+  readonly mip_opt_out?: boolean
+  readonly tag?: string | ReadonlyArray<string>
+} & Record<string, unknown>
+
+export type Request = TranscriptionRequestFor<DeepgramTranscriptionOptions>
+
+// ---------------------------------------------------------------------------
+// 2. Response schema
+// ---------------------------------------------------------------------------
+
+const Word = Schema.Struct({
+  word: Schema.String,
+  start: Schema.Number,
+  end: Schema.Number,
+  confidence: Schema.optional(Schema.Number),
+  speaker: Schema.optional(Schema.Number),
+  punctuated_word: Schema.optional(Schema.String),
+})
+
+const ListenResponse = Schema.Struct({
+  metadata: Schema.optional(
+    Schema.Struct({ request_id: Schema.optional(Schema.String), duration: Schema.optional(Schema.Number) }),
+  ),
+  results: Schema.Struct({
+    channels: Schema.Array(
+      Schema.Struct({
+        alternatives: Schema.optional(
+          Schema.Array(Schema.Struct({ transcript: Schema.String, words: Schema.optional(Schema.Array(Word)) })),
+        ),
+        detected_language: Schema.optional(Schema.String),
+      }),
+    ),
+    utterances: Schema.optional(
+      Schema.Array(
+        Schema.Struct({
+          start: Schema.Number,
+          end: Schema.Number,
+          transcript: Schema.String,
+          speaker: Schema.optional(Schema.Number),
+          words: Schema.optional(Schema.Array(Word)),
+        }),
+      ),
+    ),
+  }),
+})
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+const query = (request: Request) =>
+  MediaInput.query(
+    route.id,
+    mergeJsonRecords(
+      {
+        model: request.model.id,
+        smart_format: true,
+        language: request.language,
+        // Deepgram assumes English unless asked to detect, unlike the other routes' auto-detection.
+        detect_language: request.language === undefined ? true : undefined,
+        // `diarize=true` is deprecated in favor of choosing a diarization model.
+        diarize_model: request.diarize === true ? "latest" : undefined,
+        utterances: request.diarize === true || request.timestamps === "segment" ? true : undefined,
+      },
+      request.providerOptions,
+    ) ?? {},
+  )
+
+const fromRequest = Effect.fn("DeepgramTranscription.fromRequest")(function* (request: Request) {
+  const url = ProviderShared.mediaUrl(request.audio)
+  if (url !== undefined)
+    return MediaProtocol.json(mergeJsonRecords({ url }, request.http?.body) ?? {}, yield* query(request))
+  if (request.http?.body !== undefined)
+    return yield* ProviderShared.invalidRequest(
+      `${route.name} sends inline audio as the raw body, so http.body cannot apply`,
+    )
+  const audio = yield* MediaInput.inlineBytes(route.id, request.audio)
+  return MediaProtocol.binary(audio, request.audio.mediaType, yield* query(request))
+})
+
+// ---------------------------------------------------------------------------
+// 6. Response decoding
+// ---------------------------------------------------------------------------
+
+const decodeListen = route.decodeJson(ListenResponse)
+
+const speaker = (value: number | undefined) => (value === undefined ? undefined : String(value))
+
+const wordText = (word: typeof Word.Type) => word.punctuated_word ?? word.word
+
+// Utterances split on pauses, not speakers: the v2 diarizer labels a whole utterance with one speaker even when its
+// words change speaker, so segments split each utterance at speaker changes.
+const speakerTurns = (words: ReadonlyArray<typeof Word.Type>) =>
+  words.reduce<Array<Array<typeof Word.Type>>>((turns, word) => {
+    const last = turns.at(-1)
+    if (last === undefined || last[0].speaker !== word.speaker) return [...turns, [word]]
+    last.push(word)
+    return turns
+  }, [])
+
+const decodeResponse = Effect.fn("DeepgramTranscription.decodeResponse")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+) {
+  const output = yield* decodeListen(response)
+  const channel = output.value.results.channels[0]
+  const alternative = channel?.alternatives?.[0]
+  if (alternative === undefined) return yield* output.invalid(`${route.name} returned no transcript`)
+  const duration = output.value.metadata?.duration
+  const requestID = output.value.metadata?.request_id
+  return new TranscriptionResponse({
+    text: alternative.transcript,
+    segments: output.value.results.utterances?.flatMap((utterance) =>
+      utterance.words === undefined || utterance.words.length === 0
+        ? [
+            {
+              text: utterance.transcript,
+              startSeconds: utterance.start,
+              endSeconds: utterance.end,
+              speaker: speaker(utterance.speaker),
+            },
+          ]
+        : speakerTurns(utterance.words).map((turn) => ({
+            text: turn.map(wordText).join(" "),
+            startSeconds: turn[0].start,
+            endSeconds: turn[turn.length - 1].end,
+            speaker: speaker(turn[0].speaker),
+          })),
+    ),
+    words: alternative.words?.map((word) => ({
+      text: wordText(word),
+      startSeconds: word.start,
+      endSeconds: word.end,
+      speaker: speaker(word.speaker),
+      confidence: word.confidence,
+    })),
+    language: channel?.detected_language?.toLowerCase(),
+    durationSeconds: duration,
+    usage: duration === undefined ? undefined : { type: "seconds", seconds: duration },
+    providerMetadata: requestID === undefined ? undefined : { deepgram: { requestId: requestID } },
+  })
+})
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+export const protocol = MediaProtocol.inline<Request, TranscriptionResponse>(route, {
+  unsupported: ["prompt", "speakers"],
+  body: { from: fromRequest },
+  response: { decode: decodeResponse },
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  TranscriptionModel.fromRoute<DeepgramTranscriptionOptions>({ protocol, baseURL: DEFAULT_BASE_URL, path: PATH }, input)
+
+export const DeepgramTranscription = {
+  protocol,
+  model,
+} as const
@@ -0,0 +1,210 @@
+import { Effect, Schema } from "effect"
+import { Framing } from "../route/framing.js"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { mergeJsonRecords, type OpenString } from "../schema/index.js"
+import { SpeechModel, type SpeechEvent, type SpeechRequestFor } from "../speech.js"
+import { ProviderShared, optionalNull } from "./shared.js"
+import { SpeechStream } from "./utils/speech-stream.js"
+
+const route = MediaProtocol.identity({ id: "elevenlabs-speech", name: "ElevenLabs", provider: "elevenlabs" })
+export const DEFAULT_BASE_URL = "https://api.elevenlabs.io"
+export const PATH = "/v1/text-to-speech"
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+export type ElevenLabsOutputFormat = OpenString<
+  | "mp3_22050_32"
+  | "mp3_24000_48"
+  | "mp3_44100_32"
+  | "mp3_44100_64"
+  | "mp3_44100_96"
+  | "mp3_44100_128"
+  | "mp3_44100_192"
+  | "pcm_8000"
+  | "pcm_16000"
+  | "pcm_22050"
+  | "pcm_24000"
+  | "pcm_32000"
+  | "pcm_44100"
+  | "pcm_48000"
+  | "wav_8000"
+  | "wav_16000"
+  | "wav_22050"
+  | "wav_24000"
+  | "wav_32000"
+  | "wav_44100"
+  | "wav_48000"
+  | "ulaw_8000"
+  | "alaw_8000"
+  | "opus_48000_32"
+  | "opus_48000_64"
+  | "opus_48000_96"
+  | "opus_48000_128"
+  | "opus_48000_192"
+>
+
+export type ElevenLabsSpeechOptions = {
+  readonly outputFormat?: ElevenLabsOutputFormat
+  readonly voice_settings?: {
+    readonly stability?: number
+    readonly similarity_boost?: number
+    readonly style?: number
+    readonly use_speaker_boost?: boolean
+  }
+  readonly seed?: number
+  readonly apply_text_normalization?: OpenString<"auto" | "on" | "off">
+} & Record<string, unknown>
+
+export type Request = SpeechRequestFor<ElevenLabsSpeechOptions>
+
+// ---------------------------------------------------------------------------
+// 3. Streaming event schema
+// ---------------------------------------------------------------------------
+
+const Alignment = Schema.Struct({
+  characters: Schema.Array(Schema.String),
+  character_start_times_seconds: Schema.Array(Schema.Number),
+  character_end_times_seconds: Schema.Array(Schema.Number),
+})
+
+const TimestampedAudio = Schema.Struct({
+  audio_base64: Schema.Uint8ArrayFromBase64,
+  alignment: optionalNull(Alignment),
+})
+
+const decodeRecord = route.decodeFrame(TimestampedAudio)
+
+// ---------------------------------------------------------------------------
+// 4. Parser state
+// ---------------------------------------------------------------------------
+
+type State = SpeechStream.Audio
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+const OUTPUT_FORMATS: Readonly<Record<string, string>> = {
+  mp3: "mp3_44100_128",
+  pcm: "pcm_24000",
+  wav: "wav_24000",
+  opus: "opus_48000_64",
+}
+
+/** WAV is served only by the non-streaming endpoints. */
+const outputFormat = Effect.fn("ElevenLabsSpeech.outputFormat")(function* (request: MediaProtocol.Addressed<Request>) {
+  const format = request.providerOptions?.outputFormat ?? OUTPUT_FORMATS[request.format ?? "mp3"]
+  if (format === undefined)
+    return yield* route.unsupported(
+      "media.format",
+      `${route.name} has no default output format for "${request.format}"; pass providerOptions.outputFormat`,
+    )
+  if (request.mode === "stream" && format.startsWith("wav_"))
+    return yield* route.unsupported(
+      "media.format",
+      `${route.name} streams mp3, pcm, opus, ulaw, and alaw but not "${format}"; use generate for WAV`,
+    )
+  return format
+})
+
+const fromRequest = Effect.fn("ElevenLabsSpeech.fromRequest")(function* (request: MediaProtocol.Addressed<Request>) {
+  if (request.voice === undefined)
+    return yield* ProviderShared.invalidRequest(`${route.name} requires a voice id; pass it as \`voice\``)
+  const { outputFormat: _outputFormat, ...native } = request.providerOptions ?? {}
+  return MediaProtocol.json(
+    mergeJsonRecords(
+      {
+        text: request.text,
+        model_id: request.model.id,
+        language_code: request.language,
+        voice_settings: request.speed === undefined ? undefined : { speed: request.speed },
+      },
+      native,
+      request.http?.body,
+    ) ?? {},
+    { output_format: yield* outputFormat(request) },
+  )
+})
+
+const path = (request: MediaProtocol.Addressed<Request>) =>
+  `${PATH}/${encodeURIComponent(SpeechStream.voiceID(request.voice) ?? "")}${request.mode === "stream" ? "/stream" : ""}${
+    request.timestamps === true ? "/with-timestamps" : ""
+  }`
+
+// ---------------------------------------------------------------------------
+// 6. Stream parsing
+// ---------------------------------------------------------------------------
+
+const onRecord = Effect.fn("ElevenLabsSpeech.onRecord")(function* (state: State, frame: string) {
+  const record = yield* decodeRecord(frame)
+  const [next, events] = SpeechStream.delta(state, record.audio_base64)
+  const alignment = record.alignment
+  if (!alignment) return [next, events] as const
+  return [
+    next,
+    [
+      ...events,
+      ...SpeechStream.timestamps(
+        alignment.characters,
+        alignment.character_start_times_seconds,
+        alignment.character_end_times_seconds,
+      ),
+    ],
+  ] as const
+})
+
+const PCM_CODECS: Readonly<Record<string, SpeechStream.PcmEncoding>> = {
+  pcm: "pcm_s16le",
+  ulaw: "pcm_mulaw",
+  alaw: "pcm_alaw",
+}
+
+const describeOutput = (format: string) => {
+  const [codec = format, rate] = format.split("_")
+  const sampleRate = rate === undefined ? undefined : Number(rate)
+  const encoding = PCM_CODECS[codec]
+  return encoding === undefined ? SpeechStream.container(codec, sampleRate) : SpeechStream.pcm(encoding, sampleRate)
+}
+
+const finish = Effect.fn("ElevenLabsSpeech.finish")(function* (
+  state: State,
+  context: MediaProtocol.ResponseContext<Request>,
+) {
+  const requestID = context.http.headers["request-id"]
+  return yield* SpeechStream.finish(route, state, {
+    ...describeOutput(yield* outputFormat(context.request)),
+    // `character-cost` is billed credits, not a character count (3 for 20 characters on `eleven_flash_v2_5`).
+    usage: SpeechStream.headerUsage("credits", context.http.headers["character-cost"]),
+    providerMetadata: requestID === undefined ? undefined : { elevenlabs: { requestId: requestID } },
+  })
+})
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+export const protocol = MediaProtocol.stream<Request, SpeechEvent, string | Uint8Array, State>(route, {
+  unsupported: ["instructions"],
+  body: { from: fromRequest },
+  frames: (bytes, context) => {
+    if (context.request.timestamps !== true) return bytes
+    return context.request.mode === "stream" ? Framing.lines.frame(bytes) : Framing.document.frame(bytes)
+  },
+  initial: () => ({ chunks: [] }),
+  step: SpeechStream.step(onRecord),
+  finish,
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  SpeechModel.fromRoute<ElevenLabsSpeechOptions, string | Uint8Array, State>(
+    { protocol, baseURL: DEFAULT_BASE_URL, path: ({ request }) => path(request) },
+    input,
+  )
+
+export const ElevenLabsSpeech = {
+  protocol,
+  model,
+} as const
@@ -0,0 +1,156 @@
+import { Effect, Schema } from "effect"
+import type { HttpClientResponse } from "effect/unstable/http"
+import { ImageModel, ImageResponse, type ImageRequestFor } from "../image.js"
+import { Media } from "../media.js"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { mergeJsonRecords, type OpenString } from "../schema/index.js"
+import { ProviderShared, optionalNull } from "./shared.js"
+import { FalQueue } from "./utils/fal-queue.js"
+import { MediaInput } from "./utils/media-input.js"
+
+const route = MediaProtocol.identity({ id: "fal-images", name: "fal Images", provider: "fal" })
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+export type FalImageOptions = {
+  readonly image_size?: OpenString<
+    "square_hd" | "square" | "portrait_4_3" | "portrait_16_9" | "landscape_4_3" | "landscape_16_9"
+  >
+  readonly enable_safety_checker?: boolean
+} & Record<string, unknown>
+
+export type Request = ImageRequestFor<FalImageOptions>
+
+// ---------------------------------------------------------------------------
+// 2. Response schema
+// ---------------------------------------------------------------------------
+
+const QueueResult = Schema.StructWithRest(
+  Schema.Struct({
+    images: Schema.Array(
+      Schema.Struct({
+        url: Schema.String,
+        width: optionalNull(Schema.Number),
+        height: optionalNull(Schema.Number),
+        content_type: optionalNull(Schema.String),
+      }),
+    ),
+    seed: optionalNull(Schema.Number),
+    has_nsfw_concepts: optionalNull(Schema.Array(Schema.Boolean)),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+const sizing = (model: string) => {
+  if (/^fal-ai\/(nano-banana|flux-pro\/(v1\.1-ultra|kontext))/.test(model)) return "aspect_ratio"
+  if (model.startsWith("fal-ai/flux")) return "image_size"
+  return undefined
+}
+
+const validate = (request: Request) => {
+  const id = request.model.id
+  const field = sizing(id)
+  if (request.size !== undefined && request.aspectRatio !== undefined)
+    return Effect.fail(ProviderShared.invalidRequest(`${route.name} accepts either size or aspectRatio, not both`))
+  if (request.size !== undefined && field === "aspect_ratio")
+    return Effect.fail(route.unsupported("media.size", `${id} sizes by aspectRatio`))
+  if (request.aspectRatio !== undefined && field === "image_size")
+    return Effect.fail(route.unsupported("media.aspectRatio", `${id} sizes by size (image_size)`))
+  if ((request.images?.length ?? 0) > 1 && !takesImageList(id))
+    return Effect.fail(
+      route.unsupported(
+        "media.images",
+        `${id} takes one image_url; use an /edit or /multi endpoint for several images`,
+      ),
+    )
+  return Effect.void
+}
+
+// `/edit` and `/multi` (Kontext) endpoints take an `image_urls` list; image-to-image, fill, and Ultra take one
+// `image_url` (beside `mask_url`).
+const takesImageList = (model: string) => model.endsWith("/edit") || model.endsWith("/multi")
+
+const fromRequest = Effect.fn("FalImages.fromRequest")(function* (request: Request) {
+  yield* validate(request)
+  const images = yield* Effect.forEach(request.images ?? [], (image) => FalQueue.mediaUrl(image, route.name))
+  const list = takesImageList(request.model.id)
+  return MediaProtocol.json(
+    mergeJsonRecords(
+      {
+        prompt: request.prompt,
+        num_images: request.n,
+        seed: request.seed,
+        image_size: request.size === undefined ? undefined : MediaInput.dimensions(request.size),
+        aspect_ratio: request.aspectRatio,
+        output_format: request.format,
+        image_urls: list && images.length > 0 ? images : undefined,
+        image_url: list ? undefined : images[0],
+        mask_url: request.mask === undefined ? undefined : yield* FalQueue.mediaUrl(request.mask, route.name),
+      },
+      request.providerOptions,
+      request.http?.body,
+    ) ?? {},
+  )
+})
+
+// ---------------------------------------------------------------------------
+// 6. Response decoding
+// ---------------------------------------------------------------------------
+
+const decodeQueueResult = route.decodeJson(QueueResult)
+
+const decodeResult = Effect.fn("FalImages.decodeResult")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.PollContext<FalQueue.Token>,
+) {
+  const output = yield* decodeQueueResult(response)
+  const { images, seed, has_nsfw_concepts, ...rest } = output.value
+  if (images.length === 0) return yield* output.invalid(`${route.name} returned no images`)
+  // With the safety checker on, flagged images come back blacked out rather than omitted.
+  const flagged = (has_nsfw_concepts ?? []).flatMap((value, index) => (value ? [index] : []))
+  return new ImageResponse({
+    images: images.map((image) => {
+      const info = { width: image.width ?? undefined, height: image.height ?? undefined }
+      // `sync_mode: true` returns data URIs instead of hosted URLs.
+      return (
+        Media.parseDataUrl(image.url, { info }) ??
+        Media.url(image.url, { mediaType: image.content_type ?? undefined, info })
+      )
+    }),
+    notices:
+      flagged.length === 0
+        ? undefined
+        : flagged.map((index) => ({
+            type: "moderated" as const,
+            message: `${route.name} flagged image ${index} as NSFW`,
+          })),
+    providerMetadata: { fal: { requestId: context.token.requestID, seed: seed ?? undefined, ...rest } },
+  })
+})
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+export const protocol = FalQueue.protocol<Request, ImageResponse>(route, {
+  from: fromRequest,
+  decodeResult,
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  ImageModel.fromRoute<FalImageOptions, FalQueue.Token>(
+    { protocol, baseURL: FalQueue.DEFAULT_BASE_URL, path: ({ request }) => `/${request.model.id}` },
+    input,
+  )
+
+export const FalImages = {
+  protocol,
+  model,
+} as const
@@ -0,0 +1,121 @@
+import { Effect, Schema } from "effect"
+import type { HttpClientResponse } from "effect/unstable/http"
+import { Media } from "../media.js"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { mergeJsonRecords, type OpenString } from "../schema/index.js"
+import { VideoModel, VideoResponse, type VideoRequestFor } from "../video.js"
+import { optionalNull } from "./shared.js"
+import { FalQueue } from "./utils/fal-queue.js"
+
+const route = MediaProtocol.identity({ id: "fal-video", name: "fal Video", provider: "fal" })
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+/**
+ * Provider-native input. fal video endpoints are model-specific: `duration` is a string enum whose values differ per
+ * model (`"8s"` for Veo, `"5"` for Kling), and last-frame fields are named per model (`end_image_url`,
+ * `last_frame_url`, `tail_image_url`), so those pass through here instead of lowering from common fields.
+ */
+export type FalVideoOptions = {
+  readonly duration?: OpenString<"4s" | "6s" | "8s" | "5" | "10">
+} & Record<string, unknown>
+
+export type Request = VideoRequestFor<FalVideoOptions>
+
+// ---------------------------------------------------------------------------
+// 2. Response schema
+// ---------------------------------------------------------------------------
+
+const QueueResult = Schema.StructWithRest(
+  Schema.Struct({
+    video: Schema.Struct({
+      url: Schema.String,
+      content_type: optionalNull(Schema.String),
+      file_name: optionalNull(Schema.String),
+      file_size: optionalNull(Schema.Number),
+    }),
+    seed: optionalNull(Schema.Number),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+const fromRequest = Effect.fn("FalVideo.fromRequest")(function* (request: Request) {
+  if (request.frames?.last !== undefined)
+    return yield* route.unsupported(
+      "video.frames.last",
+      `${route.name} names the last frame per model; pass it through providerOptions (e.g. end_image_url) instead of frames.last`,
+    )
+  const imageUrl =
+    request.frames?.first === undefined ? undefined : yield* FalQueue.mediaUrl(request.frames.first, route.name)
+  const videoUrl = request.video === undefined ? undefined : yield* FalQueue.mediaUrl(request.video, route.name)
+  return MediaProtocol.json(
+    mergeJsonRecords(
+      {
+        prompt: request.prompt,
+        negative_prompt: request.negativePrompt,
+        seed: request.seed,
+        aspect_ratio: request.aspectRatio,
+        resolution: request.resolution,
+        generate_audio: request.audio,
+        image_url: imageUrl,
+        video_url: videoUrl,
+      },
+      request.providerOptions,
+      request.http?.body,
+    ) ?? {},
+  )
+})
+
+// ---------------------------------------------------------------------------
+// 6. Response decoding
+// ---------------------------------------------------------------------------
+
+const decodeQueueResult = route.decodeJson(QueueResult)
+
+const decodeResult = Effect.fn("FalVideo.decodeResult")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.PollContext<FalQueue.Token>,
+) {
+  const output = yield* decodeQueueResult(response)
+  const { video, seed, ...rest } = output.value
+  return new VideoResponse({
+    videos: [Media.url(video.url, { mediaType: video.content_type ?? "video/mp4" })],
+    providerMetadata: {
+      fal: {
+        requestId: context.token.requestID,
+        seed: seed ?? undefined,
+        fileName: video.file_name ?? undefined,
+        fileSize: video.file_size ?? undefined,
+        ...rest,
+      },
+    },
+  })
+})
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+export const protocol = FalQueue.protocol<Request, VideoResponse>(route, {
+  unsupported: ["n", "durationSeconds", "references"],
+  from: fromRequest,
+  decodeResult,
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  VideoModel.fromRoute<FalVideoOptions, FalQueue.Token>(
+    { protocol, baseURL: FalQueue.DEFAULT_BASE_URL, path: ({ request }) => `/${request.model.id}` },
+    input,
+  )
+
+export const FalVideo = {
+  protocol,
+  model,
+} as const
@@ -10,7 +10,6 @@ import {
  LLMEvent,
  Usage,
  type FinishReason,
-  type JsonSchema,
  type LLMRequest,
  type MediaPart,
  type ProviderMetadata,
@@ -22,13 +21,14 @@ import {
 import { classifyProviderFailure } from "../provider-error.js"
 import { Media } from "../media.js"
 import { JsonObject, knownString, lenient, optionalArray, optionalNull, ProviderShared } from "./shared.js"
-import { GeminiToolSchema } from "./utils/gemini-tool-schema.js"
+import { GeminiGenerateContent } from "./utils/gemini-generate-content.js"
 import { Lifecycle } from "./utils/lifecycle.js"
-import { ToolSchemaProjection } from "./utils/tool-schema.js"

 const ADAPTER = "gemini"
 // Google documents this sentinel for replaying Gemini 3 function calls after their original signature was lost.
 const SKIP_THOUGHT_SIGNATURE_VALIDATOR = "skip_thought_signature_validator"
+// Gemini 2.5 rejects a budget under the model's minimum: 512 on Flash-Lite, the highest, and 128 on Pro.
+const MIN_THINKING_BUDGET = 512
 export const DEFAULT_BASE_URL = "https://generativelanguage.googleapis.com/v1beta"

 // Gemini 3 rejects replayed function calls without a thought signature. Google's SDKs avoid that in normal chats by
@@ -132,7 +132,7 @@ const GeminiSystemInstruction = Schema.Struct({
 const GeminiFunctionDeclaration = Schema.Struct({
  name: Schema.String,
  description: Schema.String,
-  parameters: Schema.optional(JsonObject),
+  parametersJsonSchema: JsonObject,
 })

 const GeminiTool = Schema.Struct({
@@ -265,36 +265,14 @@ interface ParserState {
  readonly seenCallIds?: ReadonlySet<string>
 }

-// =============================================================================
-// Tool Schema Conversion
-// =============================================================================
-// Tool-schema conversion has two distinct concerns:
-//
-// 1. Sanitize — fix common authoring mistakes Gemini rejects: integer/number
-//    enums (must be strings), `required` entries that don't match a property,
-//    untyped arrays (`items` must be present), and `properties`/`required`
-//    keys on non-object scalars. Mirrors OpenCode's historical Gemini rules.
-//
-// 2. Project — lossy mapping from JSON Schema to Gemini's schema dialect:
-//    drop empty root parameter schemas while preserving nested empty objects,
-//    expand type arrays into `anyOf`, derive `nullable: true` from null members,
-//    coerce `const` to `[const]` enum, recurse properties/items, and propagate
-//    only an allowlisted set of keys (description, required, format, type,
-//    nullable, enum, properties, items, allOf, anyOf, oneOf, minLength).
-//    Anything outside the allowlist (e.g. `additionalProperties`, `$ref`) is
-//    silently dropped.
-//
-// Sanitize runs first, then project. The implementation lives in
-// `utils/gemini-tool-schema` so this protocol keeps the same shape as the other
-// provider protocols.
-
 // =============================================================================
 // Request Lowering
 // =============================================================================
-const lowerTool = (tool: ToolDefinition, inputSchema: JsonSchema) => ({
+// Tool schemas go in `parametersJsonSchema`, which accepts standard JSON Schema.
+const lowerTool = (tool: ToolDefinition) => ({
  name: tool.name,
  description: tool.description,
-  parameters: GeminiToolSchema.convert(inputSchema),
+  parametersJsonSchema: tool.inputSchema,
 })

 const lowerToolConfig = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
@@ -305,14 +283,9 @@ const lowerToolConfig = (toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
    tool: (name) => ({ functionCallingConfig: { mode: "ANY" as const, allowedFunctionNames: [name] } }),
  })

-// Gemini does not fetch public URLs; inline payloads and Gemini Files references are the accepted inputs.
 const lowerContentPart = Effect.fn("Gemini.lowerContentPart")(function* (part: TextPart | MediaPart) {
  if (part.type === "text") return { text: part.text }
-  const source = part.media.source
-  if (source.type === "ref" && source.provider === "google")
-    return { fileData: { mimeType: part.media.mediaType, fileUri: source.id } }
-  const media = yield* ProviderShared.requireInlineMedia("Gemini", part.media)
-  return { inlineData: { mimeType: media.mime, data: media.base64 } }
+  return yield* GeminiGenerateContent.mediaPart("Gemini", part.media)
 })

 const providerMetadata = (key: string, metadata: Record<string, unknown>): ProviderMetadata => ({ [key]: metadata })
@@ -469,7 +442,6 @@ const fromRequest = Effect.fn("Gemini.fromRequest")(function* (request: LLMReque
  const hasTools = flattened.tools.length > 0
  const generation = request.generation
  const options = yield* decodeOptions(request.providerOptions ?? {})
-  const toolSchemaCompatibility = request.model.compatibility?.toolSchema
  const generationConfig = {
    maxOutputTokens: generation?.maxTokens,
    temperature: generation?.temperature,
@@ -479,10 +451,22 @@ const fromRequest = Effect.fn("Gemini.fromRequest")(function* (request: LLMReque
    presencePenalty: generation?.presencePenalty,
    seed: generation?.seed,
    stopSequences: generation?.stop,
+    // Gemini accepts a budget above `maxOutputTokens`, but thinking then leaves the answer empty.
    thinkingConfig:
      options.thinkingConfig === undefined
        ? undefined
-        : { ...options.thinkingConfig, includeThoughts: options.thinkingConfig.includeThoughts ?? true },
+        : {
+            ...options.thinkingConfig,
+            includeThoughts: options.thinkingConfig.includeThoughts ?? true,
+            thinkingBudget:
+              options.thinkingConfig.thinkingBudget === undefined
+                ? undefined
+                : ProviderShared.fitThinkingBudget(
+                    options.thinkingConfig.thinkingBudget,
+                    generation?.maxTokens,
+                    MIN_THINKING_BUDGET,
+                  ),
+          },
  }

  return {
@@ -495,9 +479,7 @@ const fromRequest = Effect.fn("Gemini.fromRequest")(function* (request: LLMReque
    tools: hasTools
      ? [
          {
-            functionDeclarations: flattened.tools.map((tool) =>
-              lowerTool(tool, ToolSchemaProjection.modelCompatibility(tool.inputSchema, toolSchemaCompatibility)),
-            ),
+            functionDeclarations: flattened.tools.map(lowerTool),
          },
        ]
      : undefined,
@@ -833,6 +815,8 @@ export const protocol = Protocol.make({
    schema: GeminiBody,
    from: fromRequest,
  },
+  // Gemini's schema rules are this API's default, including for tuned endpoints whose IDs do not name Gemini.
+  sanitizer: "gemini",
  stream: {
    event: Protocol.jsonEvent(GeminiEvent),
    initial: (request) => ({
@@ -1,28 +1,24 @@
 import { Effect, Schema } from "effect"
 import type { HttpClientResponse } from "effect/unstable/http"
 import { ImageModel, ImageResponse, type ImageRequestFor } from "../image.js"
-import { Media } from "../media.js"
 import { MediaProtocol } from "../route/media-protocol.js"
 import { MediaRoute } from "../route/media.js"
-import { ProviderID, mergeJsonRecords, type AIError } from "../schema/index.js"
+import { mergeJsonRecords, type OpenString } from "../schema/index.js"
 import { ProviderShared } from "./shared.js"
+import { GeminiGenerateContent } from "./utils/gemini-generate-content.js"
 import { MediaInput } from "./utils/media-input.js"

-const ADAPTER = "google-images"
-const NAME = "Google Images"
-const PROVIDER = ProviderID.make("google")
+const route = MediaProtocol.identity({ id: "google-images", name: "Google Images", provider: "google" })
 export const DEFAULT_BASE_URL = "https://generativelanguage.googleapis.com/v1beta"

 // ---------------------------------------------------------------------------
 // 1. Public model input
 // ---------------------------------------------------------------------------

-export type GoogleImageString<Known extends string> = Known | (string & {})
-
 /** Provider-native options. Common fields (`aspectRatio`, `seed`, `images`) live on the request. */
 export type GoogleImageOptions = {
-  readonly imageSize?: GoogleImageString<"1K" | "2K" | "4K">
-  readonly thinkingLevel?: GoogleImageString<"MINIMAL" | "LOW" | "MEDIUM" | "HIGH">
+  readonly imageSize?: OpenString<"1K" | "2K" | "4K">
+  readonly thinkingLevel?: OpenString<"MINIMAL" | "LOW" | "MEDIUM" | "HIGH">
  readonly includeThoughts?: boolean
 } & Record<string, unknown>

@@ -102,34 +98,15 @@ const generationConfig = (request: Request) => {
  )
 }

-// Gemini does not fetch public URLs; inline payloads or Gemini Files references are the only accepted inputs.
-const imagePart = (asset: Media.Asset): Effect.Effect<Record<string, unknown>, AIError> => {
-  const inline = asset.inline()
-  if (inline) return Effect.succeed({ inlineData: { mimeType: inline.mime, data: inline.base64 } })
-  const id = MediaInput.refID(asset, PROVIDER)
-  if (id) return Effect.succeed({ fileData: { mimeType: asset.mediaType, fileUri: id } })
-  if (asset.source.type === "ref")
-    return Effect.fail(
-      ProviderShared.invalidRequest(
-        "Google generateContent requires Gemini file references rather than other providers' file IDs",
-      ),
-    )
-  return Effect.fail(
-    ProviderShared.invalidRequest(
-      "Google generateContent does not fetch public image URLs; use bytes, a data URL, or a Gemini file reference",
-    ),
-  )
-}
-
 const fromRequest = Effect.fn("GoogleImages.fromRequest")(function* (request: Request) {
  if (request.n !== undefined && request.n > 1)
-    return yield* ProviderShared.unsupportedOperation({
-      operation: "image.n",
-      provider: PROVIDER,
-      route: ADAPTER,
-      message: `${NAME} generates one image per request; call it once per image instead of n=${request.n}`,
-    })
-  const parts = yield* Effect.forEach(request.images ?? [], imagePart)
+    return yield* route.unsupported(
+      "media.n",
+      `${route.name} generates one image per request; call it once per image instead of n=${request.n}`,
+    )
+  const parts = yield* Effect.forEach(request.images ?? [], (image) =>
+    GeminiGenerateContent.mediaPart(route.name, image),
+  )
  return MediaProtocol.json(
    mergeJsonRecords(
      {
@@ -145,10 +122,12 @@ const fromRequest = Effect.fn("GoogleImages.fromRequest")(function* (request: Re
 // 6. Response decoding
 // ---------------------------------------------------------------------------

+const decodeDocument = route.decodeJson(GoogleImageResponse)
+
 const decodeResponse = Effect.fn("GoogleImages.decodeResponse")(function* (
  response: HttpClientResponse.HttpClientResponse,
 ) {
-  const output = yield* MediaProtocol.decodeJson(ADAPTER, NAME, GoogleImageResponse)(response)
+  const output = yield* decodeDocument(response)
  const decoded = output.value
  const candidates = decoded.candidates ?? []
  const candidateMetadata = candidates.map((candidate, candidateIndex) => ({
@@ -188,7 +167,7 @@ const decodeResponse = Effect.fn("GoogleImages.decodeResponse")(function* (
  const images = yield* Effect.forEach(encoded, (item) =>
    MediaInput.decodedAsset(
      output.invalid,
-      `${NAME} candidate ${item.candidateIndex} part ${item.partIndex}`,
+      `${route.name} candidate ${item.candidateIndex} part ${item.partIndex}`,
      item.inlineData.data,
      item.inlineData.mimeType,
      {
@@ -211,7 +190,7 @@ const decodeResponse = Effect.fn("GoogleImages.decodeResponse")(function* (
      candidate.finishReason === undefined ? [] : [candidate.finishReason],
    )
    return yield* output.invalid(
-      `${NAME} returned no final images${
+      `${route.name} returned no final images${
        finishReasons.length === 0 ? "" : ` (finish reasons: ${finishReasons.join(", ")})`
      }; inspect body for prompt feedback and candidate details`,
    )
@@ -223,7 +202,7 @@ const decodeResponse = Effect.fn("GoogleImages.decodeResponse")(function* (
      : [
          {
            type: "filtered" as const,
-            message: `${NAME} reported prompt feedback`,
+            message: `${route.name} reported prompt feedback`,
            providerMetadata: { google: { promptFeedback: decoded.promptFeedback } },
          },
        ]),
@@ -233,7 +212,7 @@ const decodeResponse = Effect.fn("GoogleImages.decodeResponse")(function* (
        : [
            {
              type: "filtered" as const,
-              message: `${NAME} candidate ${candidate.index ?? index} finished with ${candidate.finishReason}${
+              message: `${route.name} candidate ${candidate.index ?? index} finished with ${candidate.finishReason}${
                candidate.finishMessage === undefined ? "" : `: ${candidate.finishMessage}`
              }`,
              providerMetadata: {
@@ -283,9 +262,7 @@ const decodeResponse = Effect.fn("GoogleImages.decodeResponse")(function* (
 // 7. Protocol and route
 // ---------------------------------------------------------------------------

-export const protocol = MediaProtocol.inline<Request, ImageResponse>({
-  id: ADAPTER,
-  name: NAME,
+export const protocol = MediaProtocol.inline<Request, ImageResponse>(route, {
  unsupported: ["mask", "size", "format"],
  body: { from: fromRequest },
  response: { decode: decodeResponse },
@@ -294,8 +271,6 @@ export const protocol = MediaProtocol.inline<Request, ImageResponse>({
 export const model = (input: MediaRoute.ModelInput) =>
  ImageModel.fromRoute<GoogleImageOptions>(
    {
-      id: ADAPTER,
-      provider: PROVIDER,
      protocol,
      baseURL: DEFAULT_BASE_URL,
      path: ({ request }) => `/models/${request.model.id}:generateContent`,
@@ -0,0 +1,153 @@
+import { Effect, Schema } from "effect"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { mergeJsonRecords } from "../schema/index.js"
+import { SpeechModel, type SpeechEvent, type SpeechRequestFor } from "../speech.js"
+import { GeminiGenerateContent } from "./utils/gemini-generate-content.js"
+import { SpeechStream } from "./utils/speech-stream.js"
+
+const route = MediaProtocol.identity({ id: "google-speech", name: "Google Speech", provider: "google" })
+export const DEFAULT_BASE_URL = "https://generativelanguage.googleapis.com/v1beta"
+const DEFAULT_SAMPLE_RATE = 24000
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+/** Style is directed in the text itself, and `speechConfig.multiSpeakerVoiceConfig` excludes `voice`. */
+export type GoogleSpeechOptions = {
+  readonly temperature?: number
+  readonly seed?: number
+  readonly speechConfig?: {
+    readonly multiSpeakerVoiceConfig?: {
+      readonly speakerVoiceConfigs: ReadonlyArray<{
+        readonly speaker: string
+        readonly voiceConfig: { readonly prebuiltVoiceConfig: { readonly voiceName: string } }
+      }>
+    }
+  }
+} & Record<string, unknown>
+
+export type Request = SpeechRequestFor<GoogleSpeechOptions>
+
+// ---------------------------------------------------------------------------
+// 3. Streaming event schema
+// ---------------------------------------------------------------------------
+
+const GenerateContentChunk = GeminiGenerateContent.chunk(
+  Schema.Struct({
+    text: Schema.optional(Schema.String),
+    inlineData: Schema.optional(Schema.Struct({ mimeType: Schema.String, data: Schema.Uint8ArrayFromBase64 })),
+  }),
+)
+
+const decodeChunk = route.decodeFrame(GenerateContentChunk)
+
+// ---------------------------------------------------------------------------
+// 4. Parser state
+// ---------------------------------------------------------------------------
+
+interface State extends SpeechStream.Audio, GeminiGenerateContent.Metadata {
+  readonly mimeType?: string
+}
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+const fromRequest = Effect.fn("GoogleSpeech.fromRequest")(function* (request: MediaProtocol.Addressed<Request>) {
+  // Not in `unsupported`: that list would also reject `timestamps: false`, which asks for nothing.
+  if (request.timestamps === true)
+    return yield* route.unsupported("media.timestamps", `${route.name} does not return timestamps`)
+  if (request.format === "pcm" && request.mode === "generate" && /^gemini-3\.8-.*-tts(?:-|$)/.test(request.model.id))
+    return yield* route.unsupported(
+      "media.format",
+      `${route.name} returns WAV by default for Gemini 3.8 TTS unary requests; omit the format to accept it`,
+    )
+  if (request.format !== undefined && request.format !== "pcm")
+    return yield* route.unsupported(
+      "media.format",
+      `${route.name} only accepts raw PCM as an explicit format; omit it to accept the provider's default output`,
+    )
+  const voiceName = SpeechStream.voiceID(request.voice)
+  return MediaProtocol.json(
+    mergeJsonRecords(
+      {
+        contents: [{ role: "user", parts: [{ text: request.text }] }],
+        generationConfig: mergeJsonRecords(
+          {
+            responseModalities: ["AUDIO"],
+            speechConfig: {
+              voiceConfig: voiceName === undefined ? undefined : { prebuiltVoiceConfig: { voiceName } },
+              languageCode: request.language,
+            },
+          },
+          request.providerOptions,
+        ),
+      },
+      request.http?.body,
+    ) ?? {},
+  )
+})
+
+// ---------------------------------------------------------------------------
+// 6. Stream parsing
+// ---------------------------------------------------------------------------
+
+const step = Effect.fn("GoogleSpeech.step")(function* (state: State, frame: string) {
+  const chunk = yield* decodeChunk(frame)
+  const blocked = GeminiGenerateContent.blocked(route.name, chunk, frame)
+  if (blocked !== undefined) return yield* blocked
+  const audio = (chunk.candidates?.[0]?.content?.parts ?? []).flatMap((part) =>
+    part.inlineData === undefined ? [] : [part.inlineData],
+  )
+  const next: State = { ...GeminiGenerateContent.track(state, chunk), mimeType: state.mimeType ?? audio[0]?.mimeType }
+  return [next, audio.flatMap((part) => SpeechStream.delta(next, part.data)[1])] as const
+})
+
+const finish = (state: State, context: MediaProtocol.ResponseContext<Request>) => {
+  const sampleRate = SpeechStream.sampleRate(state.mimeType) ?? DEFAULT_SAMPLE_RATE
+  const output =
+    state.mimeType?.split(";")[0]?.toLowerCase() === "audio/wav"
+      ? SpeechStream.container("wav", sampleRate)
+      : SpeechStream.pcm("pcm_s16le", sampleRate, state.mimeType ?? `audio/L16;codec=pcm;rate=${sampleRate}`)
+  if (context.request.format === "pcm" && output.info.format !== "pcm")
+    return Effect.fail(
+      route.frameError(`Google Speech returned ${output.info.format} instead of the requested raw PCM`),
+    )
+  return SpeechStream.finish(route, state, {
+    ...output,
+    usage: GeminiGenerateContent.usage(state.usage),
+    providerMetadata: GeminiGenerateContent.providerMetadata(state),
+    detail: state.finishReason === undefined ? undefined : `finish reason: ${state.finishReason}`,
+  })
+}
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+export const protocol = MediaProtocol.stream<Request, SpeechEvent, string, State>(route, {
+  unsupported: ["instructions", "speed"],
+  body: { from: fromRequest },
+  frames: (bytes, context) => GeminiGenerateContent.frames(bytes, context.request.mode),
+  initial: () => ({ chunks: [] }),
+  step,
+  finish,
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  SpeechModel.fromRoute<GoogleSpeechOptions, string, State>(
+    {
+      protocol,
+      baseURL: DEFAULT_BASE_URL,
+      // Only `gemini-3.1-flash-tts-preview` and later stream; earlier TTS models reject `streamGenerateContent`.
+      path: ({ request }) => GeminiGenerateContent.path(request.model.id, request.mode),
+    },
+    input,
+  )
+
+export const GoogleSpeech = {
+  protocol,
+  model,
+} as const
@@ -0,0 +1,203 @@
+import { Effect, Schema, SchemaGetter } from "effect"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { mergeJsonRecords, type OpenString } from "../schema/index.js"
+import {
+  TranscriptionFinishEvent,
+  TranscriptionModel,
+  TranscriptionSegmentEvent,
+  TranscriptionTextDeltaEvent,
+  type TranscriptionRequestFor,
+  type TranscriptionSegment,
+  type TranscriptionWord,
+  type TranscriptionEvent,
+} from "../transcription.js"
+import { GeminiGenerateContent } from "./utils/gemini-generate-content.js"
+
+const route = MediaProtocol.identity({ id: "google-transcription", name: "Google Transcription", provider: "google" })
+export const DEFAULT_BASE_URL = "https://generativelanguage.googleapis.com/v1beta"
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+/**
+ * Merged into `generationConfig`. The API rejects `customVocabulary` and `mode: "SMART"` alongside diarization or word
+ * timestamps.
+ */
+export type GoogleTranscriptionOptions = {
+  readonly audioTranscriptionConfig?: {
+    readonly mode?: OpenString<"VERBATIM" | "SMART">
+    readonly customVocabulary?: ReadonlyArray<string>
+    readonly languageCodes?: ReadonlyArray<string>
+  }
+} & Record<string, unknown>
+
+export type Request = TranscriptionRequestFor<GoogleTranscriptionOptions>
+
+// ---------------------------------------------------------------------------
+// 3. Streaming event schema
+// ---------------------------------------------------------------------------
+
+const Seconds = Schema.String.check(Schema.isPattern(/^\d+(\.\d+)?s$/)).pipe(
+  Schema.decodeTo(Schema.Number, {
+    decode: SchemaGetter.transform((value) => Number.parseFloat(value)),
+    encode: SchemaGetter.transform((value) => `${value}s`),
+  }),
+)
+
+const AudioTranscription = Schema.Struct({
+  text: Schema.String,
+  speakerLabel: Schema.optional(Schema.String),
+  words: Schema.optional(
+    Schema.Array(
+      Schema.Struct({
+        word: Schema.String,
+        startOffset: Seconds,
+        endOffset: Seconds,
+        speakerLabel: Schema.optional(Schema.String),
+      }),
+    ),
+  ),
+})
+
+const decodeChunk = route.decodeFrame(
+  GeminiGenerateContent.chunk(Schema.Struct({ audioTranscription: Schema.optional(AudioTranscription) })),
+)
+
+// ---------------------------------------------------------------------------
+// 4. Parser state
+// ---------------------------------------------------------------------------
+
+interface State extends GeminiGenerateContent.Metadata {
+  readonly text: string
+  readonly segments: Array<TranscriptionSegment>
+  readonly words: Array<TranscriptionWord>
+}
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+const fromRequest = Effect.fn("GoogleTranscription.fromRequest")(function* (request: MediaProtocol.Addressed<Request>) {
+  // General Gemini models ignore `audioTranscriptionConfig` and answer the audio conversationally.
+  if (!request.model.id.includes("transcribe"))
+    return yield* route.unsupported(
+      "transcription.model",
+      `${request.model.id} is not a transcription model; use a transcribe model such as gemini-3.5-transcribe`,
+    )
+  return MediaProtocol.json(
+    mergeJsonRecords(
+      {
+        contents: [{ role: "user", parts: [yield* GeminiGenerateContent.mediaPart(route.id, request.audio)] }],
+        generationConfig: mergeJsonRecords(
+          {
+            audioTranscriptionConfig: {
+              languageCodes: request.language === undefined ? undefined : [request.language],
+              // Parts carry no offsets of their own, so segment times come from word offsets.
+              wordTimestamp:
+                request.diarize === true || request.timestamps === "word" || request.timestamps === "segment"
+                  ? true
+                  : undefined,
+              diarization: request.diarize === true ? true : undefined,
+            },
+          },
+          request.providerOptions,
+        ),
+      },
+      request.http?.body,
+    ) ?? {},
+  )
+})
+
+// ---------------------------------------------------------------------------
+// 6. Stream parsing
+// ---------------------------------------------------------------------------
+
+const turn = (part: Schema.Schema.Type<typeof AudioTranscription>) => {
+  const words = (part.words ?? []).map((word) => ({
+    text: word.word,
+    startSeconds: word.startOffset,
+    endSeconds: word.endOffset,
+    speaker: word.speakerLabel ?? part.speakerLabel,
+  }))
+  const first = words[0]
+  const last = words.at(-1)
+  return {
+    text: part.text,
+    words,
+    segment:
+      first === undefined || last === undefined
+        ? undefined
+        : {
+            text: part.text,
+            startSeconds: first.startSeconds,
+            endSeconds: last.endSeconds,
+            speaker: part.speakerLabel,
+          },
+  }
+}
+
+const step = Effect.fn("GoogleTranscription.step")(function* (state: State, frame: string) {
+  const chunk = yield* decodeChunk(frame)
+  const blocked = GeminiGenerateContent.blocked(route.name, chunk, frame)
+  if (blocked !== undefined) return yield* blocked
+  const turns = (chunk.candidates?.[0]?.content?.parts ?? []).flatMap((part) =>
+    part.audioTranscription === undefined ? [] : [turn(part.audioTranscription)],
+  )
+  const segments = turns.flatMap((item) => (item.segment === undefined ? [] : [item.segment]))
+  state.words.push(...turns.flatMap((item) => item.words))
+  state.segments.push(...segments)
+  // Each part is one whole speaker turn without surrounding whitespace, so turns join with a space.
+  const text = turns
+    .map((item) => item.text)
+    .filter((item) => item.length > 0)
+    .join(" ")
+  const delta = text.length === 0 || state.text.length === 0 ? text : ` ${text}`
+  const events: ReadonlyArray<TranscriptionEvent> = [
+    ...(delta.length === 0 ? [] : [TranscriptionTextDeltaEvent.make({ delta })]),
+    ...segments.map((segment) => TranscriptionSegmentEvent.make({ segment })),
+  ]
+  return [{ ...GeminiGenerateContent.track(state, chunk), text: state.text + delta }, events] as const
+})
+
+const finish = (state: State) => {
+  if (state.finishReason === undefined) return Effect.fail(route.incomplete())
+  return Effect.succeed([
+    TranscriptionFinishEvent.make({
+      text: state.text,
+      segments: state.segments.length === 0 ? undefined : state.segments,
+      words: state.words.length === 0 ? undefined : state.words,
+      usage: GeminiGenerateContent.usage(state.usage),
+      providerMetadata: GeminiGenerateContent.providerMetadata(state),
+    }),
+  ])
+}
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+export const protocol = MediaProtocol.stream<Request, TranscriptionEvent, string, State>(route, {
+  unsupported: ["prompt", "speakers"],
+  body: { from: fromRequest },
+  frames: (bytes, context) => GeminiGenerateContent.frames(bytes, context.request.mode),
+  initial: () => ({ text: "", segments: [], words: [] }),
+  step,
+  finish,
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  TranscriptionModel.fromRoute<GoogleTranscriptionOptions, string, State>(
+    {
+      protocol,
+      baseURL: DEFAULT_BASE_URL,
+      path: ({ request }) => GeminiGenerateContent.path(request.model.id, request.mode),
+    },
+    input,
+  )
+
+export const GoogleTranscription = {
+  protocol,
+  model,
+} as const
@@ -0,0 +1,220 @@
+import { Duration, Effect, Schema } from "effect"
+import type { HttpClientResponse } from "effect/unstable/http"
+import type { Status } from "../generation.js"
+import { Media } from "../media.js"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { mergeJsonRecords, type OpenString } from "../schema/index.js"
+import { VideoModel, VideoResponse, type VideoRequestFor } from "../video.js"
+import { ProviderShared, optionalArray } from "./shared.js"
+
+const route = MediaProtocol.identity({ id: "google-video", name: "Google Veo", provider: "google" })
+export const DEFAULT_BASE_URL = "https://generativelanguage.googleapis.com/v1beta"
+/** Veo keeps generated files for two days; the asset carries that deadline so callers materialize in time. */
+const FILE_RETENTION = Duration.days(2)
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+/** Provider-native `parameters`. Common fields (`aspectRatio`, `resolution`, `durationSeconds`, `seed`) live on the request. */
+export type GoogleVideoOptions = {
+  readonly personGeneration?: OpenString<"allow_all" | "allow_adult" | "dont_allow">
+} & Record<string, unknown>
+
+export type Request = VideoRequestFor<GoogleVideoOptions>
+
+// ---------------------------------------------------------------------------
+// 2. Token and response schemas
+// ---------------------------------------------------------------------------
+
+/** The long-running operation name, e.g. `models/veo-3.1-generate-preview/operations/abc123`. */
+export const Token = Schema.Struct({ operation: Schema.String })
+export type Token = Schema.Schema.Type<typeof Token>
+
+const StartResponse = Schema.Struct({ name: Schema.String })
+
+const Operation = Schema.Struct({
+  done: Schema.optional(Schema.Boolean),
+  error: Schema.optional(Schema.Struct({ message: Schema.optional(Schema.String) })),
+  response: Schema.optional(
+    Schema.Struct({
+      generateVideoResponse: Schema.optional(
+        Schema.Struct({
+          generatedSamples: optionalArray(
+            Schema.Struct({
+              video: Schema.optional(
+                Schema.Struct({
+                  uri: Schema.optional(Schema.String),
+                  mimeType: Schema.optional(Schema.String),
+                }),
+              ),
+            }),
+          ),
+          raiMediaFilteredCount: Schema.optional(Schema.Number),
+          raiMediaFilteredReasons: optionalArray(Schema.String),
+        }),
+      ),
+    }),
+  ),
+  metadata: Schema.optional(Schema.Unknown),
+})
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+// Veo takes inline media only; a prior Veo output is `Media.url` with transient auth, so materialize it first.
+const inlineMedia = (asset: Media.Asset) =>
+  ProviderShared.requireInlineMedia(route.name, asset).pipe(
+    Effect.map((inline) => ({ inlineData: { mimeType: inline.mime, data: inline.base64 } })),
+  )
+
+const fromRequest = Effect.fn("GoogleVideo.fromRequest")(function* (request: Request) {
+  if (request.n !== undefined && request.n > 1)
+    return yield* route.unsupported(
+      "video.n",
+      `${route.name} generates one video per request; call it once per video instead of n=${request.n}`,
+    )
+  if (request.audio === false)
+    return yield* route.unsupported(
+      "video.audio",
+      `${route.name} always generates audio; audio: false cannot be honored`,
+    )
+  if (request.frames?.last !== undefined && request.frames.first === undefined)
+    return yield* ProviderShared.invalidRequest(`${route.name} requires frames.first when frames.last is set`)
+  const image = request.frames?.first === undefined ? undefined : yield* inlineMedia(request.frames.first)
+  const lastFrame = request.frames?.last === undefined ? undefined : yield* inlineMedia(request.frames.last)
+  const video = request.video === undefined ? undefined : yield* inlineMedia(request.video)
+  const referenceImages = yield* Effect.forEach(request.references ?? [], (asset) =>
+    inlineMedia(asset).pipe(Effect.map((image) => ({ image, referenceType: "asset" }))),
+  )
+  return MediaProtocol.json(
+    mergeJsonRecords(
+      {
+        instances: [
+          {
+            prompt: request.prompt,
+            image,
+            lastFrame,
+            referenceImages: referenceImages.length === 0 ? undefined : referenceImages,
+            video,
+          },
+        ],
+        parameters: mergeJsonRecords(
+          {
+            aspectRatio: request.aspectRatio,
+            resolution: request.resolution,
+            durationSeconds: request.durationSeconds,
+            negativePrompt: request.negativePrompt,
+            seed: request.seed,
+          },
+          request.providerOptions,
+        ),
+      },
+      request.http?.body,
+    ) ?? {},
+  )
+})
+
+// ---------------------------------------------------------------------------
+// 6. Response decoding
+// ---------------------------------------------------------------------------
+
+const decodeStart = route.decodeStarted(StartResponse, (value) => ({
+  token: { operation: value.name },
+  snapshot: { id: value.name, status: "running" },
+}))
+
+// Operations carry no status string: not done is running, done with `error` failed, otherwise completed.
+const statusOf = (operation: typeof Operation.Type): Status => {
+  if (operation.done !== true) return "running"
+  return operation.error === undefined ? "completed" : "failed"
+}
+
+const decodeOperation = route.decodeJson(Operation)
+
+const decodeStatus = Effect.fn("GoogleVideo.decodeStatus")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.PollContext<Token>,
+) {
+  const output = yield* decodeOperation(response)
+  return { id: context.token.operation, status: statusOf(output.value) }
+})
+
+const decodeResult = Effect.fn("GoogleVideo.decodeResult")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.PollContext<Token>,
+) {
+  const output = yield* decodeOperation(response)
+  const operation = output.value
+  const status = statusOf(operation)
+  if (status === "running") return yield* output.pending(context.token.operation)
+  if (status === "failed")
+    return yield* output.ended(
+      "failed",
+      `${route.name} operation failed${operation.error?.message === undefined ? "" : `: ${operation.error.message}`}`,
+    )
+  const generated = operation.response?.generateVideoResponse
+  // Downloads require the same API key as the poll; the asset carries it transiently and follows the redirect.
+  const videos = yield* Effect.forEach(
+    (generated?.generatedSamples ?? []).flatMap((sample) =>
+      sample.video?.uri === undefined ? [] : [{ uri: sample.video.uri, mimeType: sample.video.mimeType }],
+    ),
+    (video) =>
+      MediaProtocol.expiringUrl(video.uri, FILE_RETENTION, {
+        mediaType: video.mimeType ?? "video/mp4",
+        headers: context.auth,
+      }),
+  )
+  const reasons = generated?.raiMediaFilteredReasons ?? []
+  const notices = reasons.map((reason) => ({
+    type: "filtered" as const,
+    message: `${route.name} filtered media: ${reason}`,
+    providerMetadata: { google: { raiMediaFilteredReason: reason } },
+  }))
+  if (videos.length === 0 && (reasons.length > 0 || (generated?.raiMediaFilteredCount ?? 0) > 0))
+    return yield* output.contentPolicy(
+      `${route.name} filtered every video${reasons.length === 0 ? "" : `: ${reasons.join("; ")}`}`,
+    )
+  if (videos.length === 0) return yield* output.invalid(`${route.name} operation completed without any video`)
+  return new VideoResponse({
+    videos,
+    notices: notices.length === 0 ? undefined : notices,
+    providerMetadata: {
+      google: {
+        operation: context.token.operation,
+        raiMediaFilteredCount: generated?.raiMediaFilteredCount,
+        metadata: operation.metadata,
+      },
+    },
+  })
+})
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+const operationPath = (token: Token) => `/${token.operation}`
+
+export const protocol = MediaProtocol.queued<Request, VideoResponse, Token>(route, {
+  token: Token,
+  start: { body: { from: fromRequest }, decode: decodeStart },
+  status: { path: operationPath, decode: decodeStatus },
+  result: { path: operationPath, decode: decodeResult },
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  VideoModel.fromRoute<GoogleVideoOptions, Token>(
+    {
+      protocol,
+      baseURL: DEFAULT_BASE_URL,
+      path: ({ request }) => `/models/${request.model.id}:predictLongRunning`,
+    },
+    input,
+  )
+
+export const GoogleVideo = {
+  protocol,
+  model,
+} as const
@@ -4,20 +4,17 @@ import { ImageModel, ImageResponse, type ImageRequestFor } from "../image.js"
 import { Media } from "../media.js"
 import { MediaProtocol } from "../route/media-protocol.js"
 import { MediaRoute } from "../route/media.js"
-import { ProviderID, mergeJsonRecords, type AIError } from "../schema/index.js"
+import { mergeJsonRecords, type OpenString } from "../schema/index.js"
 import { JsonObject, ProviderShared, optionalNull } from "./shared.js"
 import { MediaInput } from "./utils/media-input.js"

-const ADAPTER = "meta-images"
-const NAME = "Meta Images"
-const PROVIDER = ProviderID.make("meta")
+const route = MediaProtocol.identity({ id: "meta-images", name: "Meta Images", provider: "meta" })
+export const DEFAULT_BASE_URL = "https://api.meta.ai/v1"

 // ---------------------------------------------------------------------------
 // 1. Public model input
 // ---------------------------------------------------------------------------

-type OpenString<Known extends string> = Known | (string & {})
-
 /** Provider-native options. Common fields (`n`, `size`, `format`, `images`) live on the request. */
 export type ImageOptions = {
  readonly responseFormat?: OpenString<"b64_json" | "url">
@@ -70,13 +67,9 @@ const Response = Schema.Struct({

 const isEdit = (request: Request) => (request.images?.length ?? 0) > 0

-const reference = (asset: Media.Asset): Effect.Effect<Record<string, unknown>, AIError> => {
-  const inline = asset.inline()
-  if (inline) return Effect.succeed({ image_url: inline.dataUrl })
-  const url = ProviderShared.mediaUrl(asset)
-  if (url) return Effect.succeed({ image_url: url })
-  return Effect.fail(ProviderShared.invalidRequest(`${NAME} accepts image bytes and URLs`))
-}
+// Meta has no file handles: refs are rejected even when they name this provider.
+const reference = (asset: Media.Asset) =>
+  ProviderShared.mediaReference(asset, undefined, route.name).pipe(Effect.map((item) => ({ image_url: item.value })))

 const fromRequest = Effect.fn("MetaImages.fromRequest")(function* (request: Request) {
  const images = yield* Effect.forEach(request.images ?? [], reference)
@@ -105,24 +98,23 @@ const fromRequest = Effect.fn("MetaImages.fromRequest")(function* (request: Requ
 // 6. Response decoding
 // ---------------------------------------------------------------------------

+const decodeDocument = route.decodeJson(Response)
+
 const decodeResponse = Effect.fn("MetaImages.decodeResponse")(function* (
  response: HttpClientResponse.HttpClientResponse,
  context: MediaProtocol.DecodeContext<Request>,
 ) {
-  const output = yield* MediaProtocol.decodeJson(ADAPTER, NAME, Response)(response)
+  const output = yield* decodeDocument(response)
  const decoded = output.value
  const requested = context.body.type === "json" ? context.body.value.output_format : undefined
  const format = decoded.output_format ?? (typeof requested === "string" ? requested : "webp")
  const mediaType = `image/${format}`
-  const images = yield* Effect.forEach(decoded.data, (item, index) => {
-    if (item.b64_json)
-      return MediaInput.decodedAsset(output.invalid, `${NAME} result ${index}`, item.b64_json, mediaType, {
-        info: { format },
-      })
-    if (item.url) return Effect.succeed(Media.url(item.url, { mediaType, info: { format } }))
-    return Effect.fail(output.invalid(`${NAME} result ${index} has neither image data nor a URL`))
-  })
-  if (images.length === 0) return yield* output.invalid(`${NAME} returned no images`)
+  const images = yield* Effect.forEach(decoded.data, (item, index) =>
+    MediaInput.imageOutput(output.invalid, `${route.name} result ${index}`, item, mediaType, {
+      info: { format },
+    }),
+  )
+  if (images.length === 0) return yield* output.invalid(`${route.name} returned no images`)
  return new ImageResponse({
    images,
    usage:
@@ -143,20 +135,17 @@ const decodeResponse = Effect.fn("MetaImages.decodeResponse")(function* (
 // 7. Protocol and route
 // ---------------------------------------------------------------------------

-export const protocol = MediaProtocol.inline<Request, ImageResponse>({
-  id: ADAPTER,
-  name: NAME,
+export const protocol = MediaProtocol.inline<Request, ImageResponse>(route, {
  unsupported: ["mask", "aspectRatio", "seed"],
  body: { from: fromRequest },
  response: { decode: decodeResponse },
 })

-export const model = (input: MediaRoute.ModelInput & { readonly baseURL: string }) =>
+export const model = (input: MediaRoute.ModelInput) =>
  ImageModel.fromRoute<ImageOptions>(
    {
-      id: ADAPTER,
-      provider: PROVIDER,
      protocol,
+      baseURL: DEFAULT_BASE_URL,
      path: ({ request }) => `/images/${isEdit(request) ? "edits" : "generations"}`,
    },
    input,
@@ -5,8 +5,7 @@ import { LLMEvent, LLMRequest, Message, ToolResultPart } from "../schema/index.j
 import { OpenResponses } from "./open-responses.js"
 import { JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared.js"
 import { ResponsesHostedTools } from "./utils/responses-hosted-tools.js"
-import { ToolSchemaProjection } from "./utils/tool-schema.js"
-import { MetaImage } from "./utils/meta-image.js"
+import { detectMediaType } from "../utils/media-type.js"

 const ADAPTER = "meta-responses"
 const NAME = "Meta Responses"
@@ -103,12 +102,7 @@ const fromRequest = Effect.fn("MetaResponses.fromRequest")(function* (request: L
        ? undefined
        : yield* Effect.forEach(projected.tools, (tool) =>
            Effect.gen(function* () {
-              if (tool.native === undefined)
-                return yield* OpenResponses.lowerTool(
-                  NAME,
-                  tool,
-                  ToolSchemaProjection.modelCompatibility(tool.inputSchema, request.model.compatibility?.toolSchema),
-                )
+              if (tool.native === undefined) return yield* OpenResponses.lowerTool(NAME, tool)
              return yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(NativeTool))(tool.native.meta)
            }),
          ),
@@ -151,7 +145,11 @@ const HOSTED_TOOLS = {
          ),
        ),
      )
-      const mime = MetaImage.mediaType(data, item.output_format)
+      // Responses image items can omit output_format, including when PNG/JPEG was requested.
+      const mime =
+        item.output_format === undefined
+          ? (detectMediaType(data) ?? "application/octet-stream")
+          : `image/${item.output_format}`
      return {
        type: "content" as const,
        value: [{ type: "file" as const, uri: `data:${mime};base64,${item.result}`, mime }],
@@ -8,7 +8,6 @@ import {
  ProviderInternalError,
  Usage,
  type FinishReason,
-  type JsonSchema,
  type LLMRequest,
  type MediaPart,
  type ProviderMetadata,
@@ -24,7 +23,6 @@ import { classifyProviderFailure } from "../provider-error.js"
 import { effortUpdate } from "../effort-updates.js"
 import { OpenResponsesOptions } from "./utils/open-responses-options.js"
 import { Lifecycle } from "./utils/lifecycle.js"
-import { ToolSchemaProjection } from "./utils/tool-schema.js"
 import { ToolStream } from "./utils/tool-stream.js"

 const ADAPTER = "open-responses"
@@ -443,23 +441,24 @@ interface ReasoningStreamItem {
 // =============================================================================
 // Request Lowering
 // =============================================================================
-export const lowerTool = Effect.fn("OpenResponses.lowerTool")(function* (
-  protocolName: string,
-  tool: ToolDefinition,
-  inputSchema: JsonSchema,
-) {
+export const lowerTool = Effect.fn("OpenResponses.lowerTool")(function* (protocolName: string, tool: ToolDefinition) {
  if (tool.native !== undefined)
    return yield* ProviderShared.invalidRequest(`${protocolName} does not support provider-native tool ${tool.name}`)
  return {
    type: "function" as const,
    name: tool.name,
    description: tool.description,
-    parameters: inputSchema,
+    parameters: tool.inputSchema,
    // The common tool definition does not currently express Responses strict-schema policy.
    strict: false,
  }
 })

+export const lowerTools = (tools: ReadonlyArray<ToolDefinition>, adapter: ProviderAdapter) =>
+  Effect.forEach(tools, (tool) =>
+    tool.native !== undefined && adapter.nativeTool ? adapter.nativeTool(tool.native) : lowerTool(adapter.name, tool),
+  )
+
 export const lowerToolChoice = (protocolName: string, toolChoice: NonNullable<LLMRequest["toolChoice"]>) =>
  ProviderShared.matchToolChoice(protocolName, toolChoice, {
    auto: () => "auto" as const,
@@ -818,22 +817,10 @@ export const fromRequestWithAdapter = Effect.fn("OpenResponses.fromRequestWithAd
  adapter: ProviderAdapter,
 ) {
  const projected = ProviderShared.flattenToolRequest(request)
-  const toolSchemaCompatibility = request.model.compatibility?.toolSchema
  return {
    ...(yield* lowerConversation(projected.request, adapter)),
    ...lowerGeneration(request),
-    tools:
-      projected.tools.length === 0
-        ? undefined
-        : yield* Effect.forEach(projected.tools, (tool) =>
-            tool.native !== undefined && adapter.nativeTool
-              ? adapter.nativeTool(tool.native)
-              : lowerTool(
-                  adapter.name,
-                  tool,
-                  ToolSchemaProjection.modelCompatibility(tool.inputSchema, toolSchemaCompatibility),
-                ),
-          ),
+    tools: projected.tools.length === 0 ? undefined : yield* lowerTools(projected.tools, adapter),
    tool_choice:
      allowedToolChoice(request) ??
      (request.toolChoice ? yield* lowerToolChoice(adapter.name, request.toolChoice) : undefined),
@@ -1,4 +1,4 @@
-import { Effect, Schema } from "effect"
+import { Effect, Option, Schema } from "effect"
 import { Tool } from "@opencode/schema/tool"
 import { Route } from "../route/client.js"
 import { Auth } from "../route/auth.js"
@@ -17,7 +17,6 @@ import {
  type FinishReason,
  type FinishReasonDetails,
  type CacheHint,
-  type JsonSchema,
  type LLMRequest,
  type MediaPart,
  type ReasoningPart,
@@ -29,7 +28,6 @@ import { classifyProviderFailure } from "../provider-error.js"
 import { isRecord, JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared.js"
 import { OpenAIOptions } from "./utils/openai-options.js"
 import { Lifecycle } from "./utils/lifecycle.js"
-import { ToolSchemaProjection } from "./utils/tool-schema.js"
 import { ToolStream } from "./utils/tool-stream.js"

 const ADAPTER = "openai-chat"
@@ -76,6 +74,44 @@ const OpenAIChatAssistantToolCall = Schema.Struct({
 })
 type OpenAIChatAssistantToolCall = Schema.Schema.Type<typeof OpenAIChatAssistantToolCall>

+// `reasoning_details` carries two dialects. OpenRouter's `reasoning.*` entries
+// must be replayed unmodified (`index` included), so they keep every field they
+// arrived with. Kimi's OpenAI-compatible surface streams preserved thinking as
+// bare `summary` / `encrypted` entries keyed by a stream-only `index`; Kimi does
+// not document this publicly, so the handling follows Kimi Code (Kimi's own
+// client): merge summary deltas by `index`, replay without `index`, and always
+// send `reasoning_content` alongside. Anything else is dropped at the boundary.
+const OpenRouterDetailFields = {
+  id: Schema.optional(Schema.NullOr(Schema.String)),
+  format: Schema.optional(Schema.String),
+  index: Schema.optional(Schema.Number),
+  signature: Schema.optional(Schema.NullOr(Schema.String)),
+}
+const ReasoningDetail = Schema.Union([
+  Schema.StructWithRest(
+    Schema.Struct({ type: Schema.Literal("reasoning.text"), text: Schema.optional(Schema.String), ...OpenRouterDetailFields }),
+    [Schema.Record(Schema.String, Schema.Unknown)],
+  ),
+  Schema.StructWithRest(
+    Schema.Struct({
+      type: Schema.Literal("reasoning.summary"),
+      summary: Schema.optional(Schema.String),
+      ...OpenRouterDetailFields,
+    }),
+    [Schema.Record(Schema.String, Schema.Unknown)],
+  ),
+  Schema.StructWithRest(
+    Schema.Struct({ type: Schema.Literal("reasoning.encrypted"), data: Schema.String, ...OpenRouterDetailFields }),
+    [Schema.Record(Schema.String, Schema.Unknown)],
+  ),
+  Schema.Struct({ type: Schema.Literal("summary"), summary: Schema.String, index: Schema.optional(Schema.Number) }),
+  Schema.Struct({ type: Schema.Literal("encrypted"), encrypted: Schema.String, index: Schema.optional(Schema.Number) }),
+])
+type ReasoningDetail = Schema.Schema.Type<typeof ReasoningDetail>
+const decodeReasoningDetail = Schema.decodeUnknownOption(ReasoningDetail)
+const knownReasoningDetails = (details: ReadonlyArray<unknown>) =>
+  details.flatMap((detail) => Option.toArray(decodeReasoningDetail(detail)))
+
 // Intentionally omit Gemini's provider-specific `extra_content.google.thought_signature`
 // extension until direct Google OpenAI-compatible routing is supported here:
 // https://github.com/vercel/ai/issues/11590
@@ -92,6 +128,10 @@ const OpenAIChatUserContent = Schema.Union([
    type: Schema.Literal("image_url"),
    image_url: Schema.Struct({ url: Schema.String }),
  }),
+  Schema.Struct({
+    type: Schema.Literal("file"),
+    file: Schema.Struct({ filename: Schema.String, file_data: Schema.String }),
+  }),
 ])

 const OpenAIChatMessage = Schema.Union([
@@ -265,7 +305,9 @@ export interface ParserState {
  readonly finishReason?: FinishReasonDetails
  readonly lifecycle: Lifecycle.State
  readonly reasoningField?: string
-  readonly reasoningDetails: Array<unknown>
+  /** A scalar reasoning field (`reasoning_content`, ...) has carried text in this stream. */
+  readonly reasoningTextObserved: boolean
+  readonly reasoningDetails: Array<ReasoningDetail>
  readonly reasoningDetailsObserved: boolean
  readonly reasoningEmitted: boolean
  readonly latestToolIndex?: number
@@ -286,17 +328,12 @@ interface LoweringOptions {
  readonly toolCallID?: (id: string) => string
 }

-const lowerTool = (
-  tool: ToolDefinition,
-  inputSchema: JsonSchema,
-  options: LoweringOptions,
-  supportsStrictMode: boolean,
-): OpenAIChatTool => ({
+const lowerTool = (tool: ToolDefinition, options: LoweringOptions, supportsStrictMode: boolean): OpenAIChatTool => ({
  type: "function",
  function: {
    name: tool.name,
    description: tool.description,
-    parameters: inputSchema,
+    parameters: tool.inputSchema,
    ...(supportsStrictMode ? { strict: false } : {}),
  },
  cache_control: options.cacheControl?.(tool.cache),
@@ -320,6 +357,15 @@ const lowerToolCall = (part: ToolCallPart, options: LoweringOptions): OpenAIChat
 })

 const lowerMedia = Effect.fn("OpenAIChat.lowerMedia")(function* (part: MediaPart) {
+  // Chat Completions accepts PDFs, and no other documents, as inline `file` parts; file URLs are not supported.
+  if (part.media.mediaType.toLowerCase() === "application/pdf")
+    return {
+      type: "file" as const,
+      file: {
+        filename: part.filename ?? "document.pdf",
+        file_data: (yield* ProviderShared.requireInlineMedia("OpenAI Chat", part.media)).dataUrl,
+      },
+    }
  if (part.media.kind !== "image")
    return yield* ProviderShared.invalidRequest(`OpenAI Chat does not support media type ${part.media.mediaType}`)
  const url =
@@ -341,10 +387,21 @@ const reasoningDetails = (parts: ReadonlyArray<ReasoningPart>, native: unknown,
    return Array.isArray(details) ? details : []
  })
  if (parts.some((part) => Array.isArray(part.providerMetadata?.[providerMetadataKey]?.reasoningDetails)))
-    return observed
-  if (isRecord(native) && Array.isArray(native.reasoning_details)) return native.reasoning_details
+    return knownReasoningDetails(observed).map(lowerReasoningDetail)
+  if (isRecord(native) && Array.isArray(native.reasoning_details))
+    return knownReasoningDetails(native.reasoning_details).map(lowerReasoningDetail)
 }

+// Kimi rejects its stream-only `index` on requests
+// ("the reasoning_details ... must not contain streaming index").
+const lowerReasoningDetail = (detail: ReasoningDetail) => {
+  if (detail.type === "summary") return { type: detail.type, summary: detail.summary }
+  if (detail.type === "encrypted") return { type: detail.type, encrypted: detail.encrypted }
+  return detail
+}
+
+const isKimiDetail = (detail: { readonly type: string }) => detail.type === "summary" || detail.type === "encrypted"
+
 const lowerUserMessage = Effect.fn("OpenAIChat.lowerUserMessage")(function* (
  message: OpenAIChatRequestMessage,
  options: LoweringOptions,
@@ -410,6 +467,9 @@ const lowerAssistantMessage = Effect.fn("OpenAIChat.lowerAssistantMessage")(func
    if (observedField !== undefined) return observedField
    if (nativeReasoning !== undefined) return "reasoning_content"
    if (!fullyStructured || requireReasoning) return "reasoning_content"
+    // Kimi always expects `reasoning_content` on replayed assistant messages,
+    // even when thinking arrived only through structured details.
+    if (details?.some(isKimiDetail)) return "reasoning_content"
  })()
  const reasoningText = (() => {
    if (configuredField !== undefined)
@@ -435,7 +495,7 @@ const lowerToolMessages = Effect.fn("OpenAIChat.lowerToolMessages")(function* (
  options: LoweringOptions,
 ) {
  const messages: OpenAIChatMessage[] = []
-  const images: Array<Schema.Schema.Type<typeof OpenAIChatUserContent>> = []
+  const attachments: Array<Schema.Schema.Type<typeof OpenAIChatUserContent>> = []
  for (const part of message.content) {
    if (!ProviderShared.supportsContent(part, ["tool-result"]))
      return yield* ProviderShared.unsupportedContent("OpenAI Chat", "tool", ["tool-result"])
@@ -457,9 +517,9 @@ const lowerToolMessages = Effect.fn("OpenAIChat.lowerToolMessages")(function* (
      cache_control: options.cacheControl?.(part.cache),
    })
    const files = content.filter((item) => item.type === "file")
-    images.push(...(yield* Effect.forEach(files, (item) => lowerMedia(ProviderShared.toolFileMedia(item)))))
+    attachments.push(...(yield* Effect.forEach(files, (item) => lowerMedia(ProviderShared.toolFileMedia(item)))))
  }
-  return { messages, images }
+  return { messages, attachments }
 })

 const lowerMessage = Effect.fn("OpenAIChat.lowerMessage")(function* (
@@ -520,21 +580,21 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
    if (requireAssistantAfterTool && messages.at(-1)?.role === "tool")
      messages.push({ role: "assistant", content: "Done." })
  }
-  const pendingImages: Array<Schema.Schema.Type<typeof OpenAIChatUserContent>> = []
-  const flushImages = () => {
-    if (pendingImages.length === 0) return
+  const pendingAttachments: Array<Schema.Schema.Type<typeof OpenAIChatUserContent>> = []
+  const flushAttachments = () => {
+    if (pendingAttachments.length === 0) return
    bridgeTools()
-    messages.push({ role: "user", content: pendingImages.splice(0) })
+    messages.push({ role: "user", content: pendingAttachments.splice(0) })
  }
  for (const message of request.messages) {
    if (message.role === "user") bridgeTools()
    if (message.role === "system") {
      const part = yield* ProviderShared.wrappedSystemUpdate("OpenAI Chat", message)
-      if (pendingImages.length > 0) {
+      if (pendingAttachments.length > 0) {
        messages.push({
          role: "user",
          content: [
-            ...pendingImages.splice(0),
+            ...pendingAttachments.splice(0),
            { type: "text", text: part.text, cache_control: options.cacheControl?.(part.cache) },
          ],
        })
@@ -578,13 +638,13 @@ const lowerMessages = Effect.fn("OpenAIChat.lowerMessages")(function* (request:
    if (message.role === "tool") {
      const lowered = yield* lowerToolMessages(message, lowering)
      messages.push(...lowered.messages)
-      pendingImages.push(...lowered.images)
+      pendingAttachments.push(...lowered.attachments)
      continue
    }
-    flushImages()
+    flushAttachments()
    messages.push(...(yield* lowerMessage(message, reasoningField, requireReasoning, lowering)))
  }
-  flushImages()
+  flushAttachments()
  return messages
 })

@@ -736,7 +796,6 @@ export const fromRequest = Effect.fn("OpenAIChat.fromRequest")(function* (
      `OpenAI Chat reasoning field conflicts with reserved field ${reasoningField}`,
    )
  const generation = request.generation
-  const toolSchemaCompatibility = request.model.compatibility?.toolSchema
  const flattened = ProviderShared.flattenToolRequest(request)
  const provider = String(request.model.provider)
  const baseURL = request.model.route.endpoint.baseURL
@@ -759,14 +818,7 @@ export const fromRequest = Effect.fn("OpenAIChat.fromRequest")(function* (
        ? hasHistory
          ? []
          : undefined
-        : flattened.tools.map((tool) =>
-            lowerTool(
-              tool,
-              ToolSchemaProjection.modelCompatibility(tool.inputSchema, toolSchemaCompatibility),
-              options,
-              supportsStrictMode,
-            ),
-          ),
+        : flattened.tools.map((tool) => lowerTool(tool, options, supportsStrictMode)),
    tool_choice: hasActiveTools && request.toolChoice ? yield* lowerToolChoice(request.toolChoice) : undefined,
    stream: true as const,
    ...(supportsUsageInStreaming ? { stream_options: { include_usage: true } } : {}),
@@ -881,44 +933,74 @@ const reasoningDelta = (
  return undefined
 }

-const detailText = (details: ReadonlyArray<unknown>) => {
+const detailText = (details: ReadonlyArray<ReasoningDetail>, hideKimiSummary: boolean) => {
  const text = details.flatMap((detail) => {
-    if (!isRecord(detail)) return []
-    if (detail.type === "reasoning.text" && typeof detail.text === "string" && detail.text) return [detail.text]
-    if (detail.type === "reasoning.summary" && typeof detail.summary === "string" && detail.summary)
-      return [detail.summary]
+    if (detail.type === "reasoning.text") return detail.text ? [detail.text] : []
+    if (detail.type === "reasoning.summary") return detail.summary ? [detail.summary] : []
+    // Kimi streams the full thinking through `reasoning_content` and a separate
+    // summary through details; show the summary only when nothing else does.
+    if (detail.type === "summary") return detail.summary && !hideKimiSummary ? [detail.summary] : []
    return []
  })
  if (text.length > 0) return text.join("")
 }

-const appendReasoningDetails = (result: Array<unknown>, details: ReadonlyArray<unknown>) => {
+const appendReasoningDetails = (result: Array<ReasoningDetail>, details: ReadonlyArray<ReasoningDetail>) => {
  for (const detail of details) {
    const previous = result.at(-1)
-    if (
-      !isRecord(previous) ||
-      previous.type !== "reasoning.text" ||
-      !isRecord(detail) ||
-      detail.type !== "reasoning.text" ||
-      conflictingReasoningTextDetails(previous, detail)
-    ) {
+    const merged = previous === undefined ? undefined : mergeReasoningDetails(previous, detail)
+    if (merged === undefined) {
      result.push(detail)
      continue
    }
-    result[result.length - 1] = {
-      ...previous,
-      ...Object.fromEntries(Object.entries(detail).filter((entry) => entry[1] !== undefined)),
-      text: `${typeof previous.text === "string" ? previous.text : ""}${typeof detail.text === "string" ? detail.text : ""}`,
-      signature: mergeDetailValue(previous.signature, detail.signature),
-      format: mergeDetailValue(previous.format, detail.format),
-    }
+    result[result.length - 1] = merged
  }
 }

-const mergeDetailValue = (previous: unknown, current: unknown) =>
+// Consecutive text or summary deltas of the same kind accumulate into one
+// entry; encrypted entries are opaque and never merge.
+const mergeReasoningDetails = (previous: ReasoningDetail, detail: ReasoningDetail): ReasoningDetail | undefined => {
+  if (conflictingReasoningDetails(previous, detail)) return undefined
+  if (previous.type === "reasoning.text" && detail.type === "reasoning.text")
+    return {
+      ...previous,
+      ...detail,
+      text: `${previous.text ?? ""}${detail.text ?? ""}`,
+      ...mergeDetailIdentity(previous, detail),
+    }
+  if (previous.type === "reasoning.summary" && detail.type === "reasoning.summary")
+    return {
+      ...previous,
+      ...detail,
+      summary: `${previous.summary ?? ""}${detail.summary ?? ""}`,
+      ...mergeDetailIdentity(previous, detail),
+    }
+  if (previous.type === "summary" && detail.type === "summary")
+    return { ...previous, ...detail, summary: previous.summary + detail.summary }
+}
+
+type DetailIdentity = {
+  readonly id?: string | null
+  readonly index?: number
+  readonly format?: string
+  readonly signature?: string | null
+}
+
+// The first non-empty signature and format win; a later delta may carry the
+// signature for text that streamed earlier.
+const mergeDetailIdentity = (previous: DetailIdentity, current: DetailIdentity) => {
+  const signature = mergeDetailValue(previous.signature, current.signature)
+  const format = mergeDetailValue(previous.format, current.format)
+  return {
+    ...(signature === undefined ? {} : { signature }),
+    ...(format === undefined ? {} : { format }),
+  }
+}
+
+const mergeDetailValue = <T>(previous: T | undefined, current: T | undefined) =>
  previous || current || (previous !== undefined ? previous : current)

-const conflictingReasoningTextDetails = (previous: Record<string, unknown>, current: Record<string, unknown>) =>
+const conflictingReasoningDetails = (previous: DetailIdentity, current: DetailIdentity) =>
  conflictingDetailValue(previous.id, current.id) ||
  conflictingDetailValue(previous.index, current.index) ||
  conflictingDetailValue(previous.format, current.format) ||
@@ -930,7 +1012,7 @@ const conflictingDetailValue = (previous: unknown, current: unknown) =>
 const reasoningMetadata = (
  providerMetadataKey: string,
  field: ParserState["reasoningField"],
-  details?: ReadonlyArray<unknown>,
+  details?: ReadonlyArray<ReasoningDetail>,
 ) => ({
  [providerMetadataKey]: {
    ...(field ? { reasoningField: field } : {}),
@@ -993,11 +1075,16 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
    }

    const reasoningField = state.reasoningField ?? reasoning?.field
-    const detailDelta = Array.isArray(delta?.reasoning_details) ? delta.reasoning_details : undefined
+    const reasoningTextObserved = state.reasoningTextObserved || reasoning !== undefined
+    const detailDelta = Array.isArray(delta?.reasoning_details)
+      ? knownReasoningDetails(delta.reasoning_details)
+      : undefined
    if (detailDelta !== undefined) appendReasoningDetails(state.reasoningDetails, detailDelta)
    const reasoningDetailsObserved = state.reasoningDetailsObserved || detailDelta !== undefined
    const deltaMetadata = reasoningMetadata(state.providerMetadataKey, reasoningField)
-    const text = detailDelta?.length ? (detailText(detailDelta) ?? reasoning?.text) : reasoning?.text
+    const text = detailDelta?.length
+      ? (detailText(detailDelta, reasoningTextObserved) ?? reasoning?.text)
+      : reasoning?.text
    if (text !== undefined) lifecycle = Lifecycle.reasoningDelta(lifecycle, events, "reasoning-0", text, deltaMetadata)
    else if (
      reasoningDetailsObserved &&
@@ -1093,6 +1180,7 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
        finishReason,
        lifecycle,
        reasoningField,
+        reasoningTextObserved,
        reasoningDetails: state.reasoningDetails,
        reasoningDetailsObserved,
        reasoningEmitted,
@@ -1173,6 +1261,7 @@ export const protocol = Protocol.make({
      toolCallEvents: [],
      lifecycle: Lifecycle.initial(),
      reasoningField: request.model.compatibility?.reasoningField,
+      reasoningTextObserved: false,
      reasoningDetails: [],
      reasoningDetailsObserved: false,
      reasoningEmitted: false,
@@ -1,16 +1,21 @@
-import { Effect, Schema } from "effect"
-import type { HttpClientResponse } from "effect/unstable/http"
-import { ImageModel, ImageResponse, type ImageRequestFor } from "../image.js"
+import { Effect, Schema, Stream } from "effect"
+import {
+  ImageFinishEvent,
+  ImageModel,
+  ImageOutputEvent,
+  ImagePartialEvent,
+  type ImageEvent,
+  type ImageRequestFor,
+} from "../image.js"
 import { Media } from "../media.js"
+import { Framing } from "../route/framing.js"
 import { MediaProtocol } from "../route/media-protocol.js"
 import { MediaRoute } from "../route/media.js"
-import { ProviderID, mergeJsonRecords, type AIError } from "../schema/index.js"
+import { mergeJsonRecords, type MediaUsage, type OpenString } from "../schema/index.js"
 import { ProviderShared } from "./shared.js"
 import { MediaInput } from "./utils/media-input.js"

-const ADAPTER = "openai-images"
-const NAME = "OpenAI Images"
-const PROVIDER = ProviderID.make("openai")
+const route = MediaProtocol.identity({ id: "openai-images", name: "OpenAI Images", provider: "openai" })
 export const DEFAULT_BASE_URL = "https://api.openai.com/v1"
 export const PATH = "/images/generations"
 export const EDIT_PATH = "/images/edits"
@@ -19,14 +24,14 @@ export const EDIT_PATH = "/images/edits"
 // 1. Public model input
 // ---------------------------------------------------------------------------

-export type OpenAIImageString<Known extends string> = Known | (string & {})
-
 /** Provider-native options. Common fields (`n`, `size`, `format`, `images`, `mask`) live on the request. */
 export type OpenAIImageOptions = {
-  readonly quality?: OpenAIImageString<"auto" | "low" | "medium" | "high" | "standard" | "hd">
-  readonly background?: OpenAIImageString<"auto" | "opaque" | "transparent">
-  readonly moderation?: OpenAIImageString<"auto" | "low">
+  readonly quality?: OpenString<"auto" | "low" | "medium" | "high" | "standard" | "hd">
+  readonly background?: OpenString<"auto" | "opaque" | "transparent">
+  readonly moderation?: OpenString<"auto" | "low">
  readonly outputCompression?: number
+  /** Previews sent before the final image when streaming (default 2); ignored by `Image.generate`. */
+  readonly partialImages?: number
 } & Record<string, unknown>

 export type Request = ImageRequestFor<OpenAIImageOptions>
@@ -35,26 +40,68 @@ export type Request = ImageRequestFor<OpenAIImageOptions>
 // 2. Response schema
 // ---------------------------------------------------------------------------

-const OpenAIImageResponse = Schema.Struct({
-  data: Schema.Array(
-    Schema.Struct({
-      b64_json: Schema.optional(Schema.String),
-      url: Schema.optional(Schema.String),
-      revised_prompt: Schema.optional(Schema.String),
-    }),
-  ),
-  output_format: Schema.optional(Schema.String),
-  usage: Schema.optional(
-    Schema.Struct({
-      input_tokens: Schema.optional(Schema.Number),
-      output_tokens: Schema.optional(Schema.Number),
-      total_tokens: Schema.optional(Schema.Number),
-      input_tokens_details: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
-      output_tokens_details: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
-    }),
-  ),
+const Usage = Schema.Struct({
+  input_tokens: Schema.optional(Schema.Number),
+  output_tokens: Schema.optional(Schema.Number),
+  total_tokens: Schema.optional(Schema.Number),
+  input_tokens_details: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
+  output_tokens_details: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
 })

+/** What the provider actually rendered; it can differ from the request when `auto` or a default applied. */
+const Settings = {
+  output_format: Schema.optional(Schema.String),
+  size: Schema.optional(Schema.String),
+  quality: Schema.optional(Schema.String),
+  background: Schema.optional(Schema.String),
+}
+
+const OpenAIImageResponse = Schema.Struct({
+  data: Schema.Array(Schema.Struct({ b64_json: Schema.String })),
+  ...Settings,
+  usage: Schema.optional(Usage),
+})
+
+// ---------------------------------------------------------------------------
+// 3. Streaming event schema
+// ---------------------------------------------------------------------------
+
+const StreamEvent = Schema.Union([
+  Schema.Struct({
+    type: Schema.Literals(["image_generation.partial_image", "image_edit.partial_image"]),
+    b64_json: Schema.String,
+    partial_image_index: Schema.Number,
+    ...Settings,
+    output_format: Schema.String,
+  }),
+  Schema.Struct({
+    type: Schema.Literals(["image_generation.completed", "image_edit.completed"]),
+    b64_json: Schema.String,
+    ...Settings,
+    output_format: Schema.String,
+    usage: Schema.optional(Usage),
+  }),
+])
+
+const decodeEvent = route.decodeFrame(StreamEvent)
+const decodeDocument = Schema.decodeUnknownEffect(Schema.fromJsonString(OpenAIImageResponse))
+
+/** `generate` reads the whole JSON response as one frame, with the requested format for responses that omit it. */
+type Frame = string | { readonly document: string; readonly requested: string | undefined }
+
+// ---------------------------------------------------------------------------
+// 4. Parser state
+// ---------------------------------------------------------------------------
+
+interface State {
+  readonly completed: number
+  readonly format?: string
+  readonly size?: string
+  readonly quality?: string
+  readonly background?: string
+  readonly usage?: MediaUsage
+}
+
 // ---------------------------------------------------------------------------
 // 5. Request body construction
 // ---------------------------------------------------------------------------
@@ -64,39 +111,35 @@ const RESERVED_FORM_FIELDS = new Set(["model", "prompt", "image", "image[]", "im

 const nativeOptions = (options: OpenAIImageOptions | undefined) => {
  if (!options) return undefined
-  const { outputCompression, ...native } = options
+  const { outputCompression, partialImages: _, ...native } = options
  return { output_compression: outputCompression, ...native }
 }

+const streamOptions = (request: MediaProtocol.Addressed<Request>) => {
+  if (request.mode !== "stream") return Effect.succeed(undefined)
+  if (request.n !== undefined && request.n > 1)
+    return Effect.fail(
+      route.unsupported("media.n", `${route.name} streams one image; use Image.generate for n=${request.n}`),
+    )
+  return Effect.succeed({ stream: true, partial_images: request.providerOptions?.partialImages ?? 2 })
+}
+
 const isEdit = (request: Request) => (request.images?.length ?? 0) > 0

-const isInline = (asset: Media.Asset) => asset.inline() !== undefined
+const isInline = (asset: Media.Asset) => asset.source.type === "bytes" || asset.source.type === "base64"

-const blob = (data: Uint8Array, mediaType: string) => {
-  const buffer = new ArrayBuffer(data.byteLength)
-  new Uint8Array(buffer).set(data)
-  return new Blob([buffer], { type: mediaType })
-}
-
-const reference = (asset: Media.Asset): Effect.Effect<Record<string, unknown>, AIError> => {
-  const inline = asset.inline()
-  if (inline) return Effect.succeed({ image_url: inline.dataUrl })
-  const url = ProviderShared.mediaUrl(asset)
-  if (url) return Effect.succeed({ image_url: url })
-  const id = MediaInput.refID(asset, PROVIDER)
-  if (id) return Effect.succeed({ file_id: id })
-  return Effect.fail(
-    ProviderShared.invalidRequest("OpenAI Images accepts image URLs, data URLs, bytes, and OpenAI file IDs"),
+const reference = (asset: Media.Asset) =>
+  ProviderShared.mediaReference(asset, route.provider, route.name).pipe(
+    Effect.map((item) => (item.type === "ref" ? { file_id: item.value } : { image_url: item.value })),
  )
-}

-const fromRequest = Effect.fn("OpenAIImages.fromRequest")(function* (request: Request) {
+const fromRequest = Effect.fn("OpenAIImages.fromRequest")(function* (request: MediaProtocol.Addressed<Request>) {
  const images = request.images ?? []
  const mask = request.mask
  if (mask !== undefined && images.length === 0)
    return yield* ProviderShared.invalidRequest("An OpenAI image mask requires at least one input image")
  const fields = mergeJsonRecords(
-    { n: request.n, size: request.size, output_format: request.format },
+    { n: request.n, size: request.size, output_format: request.format, ...(yield* streamOptions(request)) },
    nativeOptions(request.providerOptions),
    request.http?.body,
  )
@@ -104,16 +147,17 @@ const fromRequest = Effect.fn("OpenAIImages.fromRequest")(function* (request: Re
  // Owned bytes go through multipart edits; remote URLs and file IDs use the JSON edits body instead.
  if (images.length > 0 && images.every(isInline) && (mask === undefined || isInline(mask))) {
    const form = new FormData()
-    form.append("model", request.model.id)
-    form.append("prompt", request.prompt)
-    Object.entries(fields ?? {}).forEach(([key, value]) => {
-      if (RESERVED_FORM_FIELDS.has(key)) return
-      form.append(key, typeof value === "string" ? value : ProviderShared.encodeJson(value))
-    })
-    const uploads = yield* Effect.forEach(images, (image) => MediaInput.inlineBytes(ADAPTER, image))
-    uploads.forEach((data, index) => form.append("image[]", blob(data, images[index].mediaType), `image-${index}`))
+    MediaInput.appendFields(
+      form,
+      { model: request.model.id, prompt: request.prompt },
+      { overlay: fields, reserved: RESERVED_FORM_FIELDS },
+    )
+    const uploads = yield* Effect.forEach(images, (image) => MediaInput.inlineBytes(route.id, image))
+    uploads.forEach((data, index) =>
+      form.append("image[]", MediaInput.blob(data, images[index].mediaType), `image-${index}`),
+    )
    if (mask !== undefined)
-      form.append("mask", blob(yield* MediaInput.inlineBytes(ADAPTER, mask), mask.mediaType), "mask")
+      form.append("mask", MediaInput.blob(yield* MediaInput.inlineBytes(route.id, mask), mask.mediaType), "mask")
    return MediaProtocol.multipart(form)
  }

@@ -133,71 +177,121 @@ const fromRequest = Effect.fn("OpenAIImages.fromRequest")(function* (request: Re
 })

 // ---------------------------------------------------------------------------
-// 6. Response decoding
+// 6. Stream parsing
 // ---------------------------------------------------------------------------

 const requestedFormat = (body: MediaProtocol.Body) => {
+  if (body.type === "binary") return undefined
  const value = body.type === "json" ? body.value.output_format : body.value.get("output_format")
  return typeof value === "string" ? value : undefined
 }

-const decodeResponse = Effect.fn("OpenAIImages.decodeResponse")(function* (
-  response: HttpClientResponse.HttpClientResponse,
-  context: MediaProtocol.DecodeContext<Request>,
-) {
-  const output = yield* MediaProtocol.decodeJson(ADAPTER, NAME, OpenAIImageResponse)(response)
-  const decoded = output.value
-  const format = decoded.output_format ?? requestedFormat(context.body) ?? "png"
-  const mediaType = `image/${format}`
-  const images = yield* Effect.forEach(decoded.data, (item, index) => {
-    const providerMetadata =
-      item.revised_prompt === undefined ? undefined : { openai: { revisedPrompt: item.revised_prompt } }
-    if (item.b64_json)
-      return MediaInput.decodedAsset(output.invalid, `${NAME} result ${index}`, item.b64_json, mediaType, {
-        info: { format },
-        providerMetadata,
-      })
-    if (item.url) return Effect.succeed(Media.url(item.url, { mediaType, info: { format }, providerMetadata }))
-    return Effect.fail(output.invalid(`${NAME} result ${index} has neither image data nor a URL`))
-  })
-  if (images.length === 0) return yield* output.invalid(`${NAME} returned no images`)
-  return new ImageResponse({
-    images,
-    usage:
-      decoded.usage === undefined
-        ? undefined
-        : {
-            type: "tokens",
-            input: decoded.usage.input_tokens,
-            output: decoded.usage.output_tokens,
-            total: decoded.usage.total_tokens,
-            details: { openai: decoded.usage },
-          },
-    providerMetadata: { openai: { outputFormat: format } },
+const usage = (value: Schema.Schema.Type<typeof Usage> | undefined): MediaUsage | undefined =>
+  value === undefined
+    ? undefined
+    : {
+        type: "tokens",
+        input: value.input_tokens,
+        output: value.output_tokens,
+        total: value.total_tokens,
+        details: { openai: value },
+      }
+
+/** `size` echoes the rendered `WIDTHxHEIGHT`; `auto` or any other value leaves the dimensions unknown. */
+const info = (format: string, size: string | undefined): Media.Info => {
+  const match = size?.match(/^(\d+)x(\d+)$/)
+  return match ? { format, width: Number(match[1]), height: Number(match[2]) } : { format }
+}
+
+const eventImage = (frame: string, label: string, data: string, format: string, size: string | undefined) =>
+  MediaInput.decodedAsset((message, cause) => route.frameError(message, frame, cause), label, data, `image/${format}`, {
+    info: info(format, size),
  })
+
+const onEvent = Effect.fn("OpenAIImages.onEvent")(function* (state: State, frame: string) {
+  const event = yield* decodeEvent(frame)
+  const format = event.output_format
+  if ("partial_image_index" in event) {
+    const image = yield* eventImage(frame, `${route.name} partial image`, event.b64_json, format, event.size)
+    return [state, [ImagePartialEvent.make({ index: event.partial_image_index, image })]] as const
+  }
+  const image = yield* eventImage(frame, `${route.name} result ${state.completed}`, event.b64_json, format, event.size)
+  return [
+    {
+      completed: state.completed + 1,
+      format,
+      size: event.size,
+      quality: event.quality,
+      background: event.background,
+      usage: usage(event.usage),
+    },
+    [ImageOutputEvent.make({ index: state.completed, image })],
+  ] as const
 })

+const onDocument = Effect.fn("OpenAIImages.onDocument")(function* (frame: Exclude<Frame, string>) {
+  const invalid = (message: string, cause?: unknown) => route.frameError(message, frame.document, cause)
+  const decoded = yield* decodeDocument(frame.document).pipe(
+    Effect.mapError((cause) => invalid(`${route.name} returned an invalid response`, cause)),
+  )
+  const format = decoded.output_format ?? frame.requested ?? "png"
+  const images = yield* Effect.forEach(decoded.data, (item, index) =>
+    MediaInput.decodedAsset(invalid, `${route.name} result ${index}`, item.b64_json, `image/${format}`, {
+      info: info(format, decoded.size),
+    }),
+  )
+  if (images.length === 0) return yield* invalid(`${route.name} returned no images`)
+  const state: State = {
+    completed: images.length,
+    format,
+    size: decoded.size,
+    quality: decoded.quality,
+    background: decoded.background,
+    usage: usage(decoded.usage),
+  }
+  return [state, images.map((image, index) => ImageOutputEvent.make({ index, image }))] as const
+})
+
+const step = (state: State, frame: Frame) => (typeof frame === "string" ? onEvent(state, frame) : onDocument(frame))
+
+const finish = (state: State) => {
+  if (state.completed === 0) return Effect.fail(route.incomplete())
+  return Effect.succeed([
+    ImageFinishEvent.make({
+      usage: state.usage,
+      providerMetadata: {
+        openai: {
+          outputFormat: state.format,
+          size: state.size,
+          quality: state.quality,
+          background: state.background,
+        },
+      },
+    }),
+  ])
+}
+
 // ---------------------------------------------------------------------------
 // 7. Protocol and route
 // ---------------------------------------------------------------------------

-export const protocol = MediaProtocol.inline<Request, ImageResponse>({
-  id: ADAPTER,
-  name: NAME,
+export const protocol = MediaProtocol.stream<Request, ImageEvent, Frame, State>(route, {
  unsupported: ["aspectRatio", "seed"],
  body: { from: fromRequest },
-  response: { decode: decodeResponse },
+  frames: (bytes, context) =>
+    context.request.mode === "stream"
+      ? Framing.sse.frame(bytes)
+      : Framing.document
+          .frame(bytes)
+          .pipe(Stream.map((document) => ({ document, requested: requestedFormat(context.body) }))),
+  initial: () => ({ completed: 0 }),
+  step,
+  finish,
 })

 export const model = (input: MediaRoute.ModelInput) =>
-  ImageModel.fromRoute<OpenAIImageOptions>(
-    {
-      id: ADAPTER,
-      provider: PROVIDER,
-      protocol,
-      baseURL: DEFAULT_BASE_URL,
-      path: ({ request }) => (isEdit(request) ? EDIT_PATH : PATH),
-    },
+  ImageModel.fromRoute<OpenAIImageOptions, Frame, State>(
+    { protocol, baseURL: DEFAULT_BASE_URL, path: ({ request }) => (isEdit(request) ? EDIT_PATH : PATH) },
    input,
  )

@@ -5,14 +5,12 @@ import { Auth } from "../route/auth.js"
 import { Endpoint } from "../route/endpoint.js"
 import { Protocol } from "../route/protocol.js"
 import { HttpTransport } from "../route/transport/index.js"
-import { LLMRequest, mergeJsonRecords, type JsonSchema, type ToolDefinition, type ToolEntry } from "../schema/index.js"
+import { LLMRequest, type ToolDefinition, type ToolEntry } from "../schema/index.js"
 import { resolveEffortUpdates } from "../effort-updates.js"
 import { OpenResponses } from "./open-responses.js"
 import { OpenResponsesOptions } from "./utils/open-responses-options.js"
 import { JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared.js"
-import { OpenAIImage } from "./utils/openai-image.js"
 import { ResponsesHostedTools } from "./utils/responses-hosted-tools.js"
-import { ToolSchemaProjection } from "./utils/tool-schema.js"
 import { OpenResponsesChannel } from "./open-responses-channel.js"
 import { ResponsesCompaction } from "./utils/responses-compaction.js"
 import { ResponsesCheckpoint } from "./utils/responses-checkpoint.js"
@@ -41,7 +39,16 @@ const OpenAIResponsesImageGenerationTool = Schema.Struct({
  output_format: Schema.optional(Schema.Literals(["png", "jpeg", "webp"])),
  partial_images: Schema.optional(Schema.Int.check(Schema.isGreaterThanOrEqualTo(0))),
  quality: Schema.optional(Schema.Literals(["auto", "low", "medium", "high"])),
-  size: Schema.optional(OpenAIImage.Size),
+  size: Schema.optional(
+    Schema.String.check(
+      Schema.makeFilter((value) => {
+        if (value === "auto") return undefined
+        const match = /^(\d+)x(\d+)$/.exec(value)
+        if (!match) return "image size must be `auto` or `{width}x{height}`"
+        return Number(match[1]) > 0 && Number(match[2]) > 0 ? undefined : "image dimensions must be positive integers"
+      }),
+    ),
+  ),
 })

 const OpenAIResponsesHostedToolItem = Schema.Union([
@@ -128,11 +135,6 @@ export const CompactionTrigger = Schema.Struct({ type: Schema.Literal("compactio
 const CheckpointBody = Schema.Struct({
  ...OpenAIResponsesBody.fields,
  input: Schema.Array(Schema.Union([OpenAIResponsesInputItem, CompactionTrigger])),
-  store: Schema.Literal(false),
-  prompt_cache_retention: optionalNull(Schema.String),
-  prompt_cache_options: optionalNull(
-    Schema.Struct({ mode: Schema.optional(Schema.String), ttl: Schema.optional(Schema.String) }),
-  ),
 })

 const adapter = {
@@ -141,12 +143,14 @@ const adapter = {
  restoreHostedToolItem: (item: unknown) => (Schema.is(OpenAIResponsesHostedToolItem)(item) ? item : undefined),
 } satisfies OpenResponses.ProviderAdapter

-// Only GPT-6 Astra accepts `configuration_update`, and never alongside automatic `context_management` compaction.
+// GPT-6 Astra, Sol, and Luna accept `configuration_update` only in standard mode (not `reasoning.mode: "pro"` or
+// `-pro` slugs), and never alongside automatic `context_management` compaction.
 const supportsEffortUpdates = (request: LLMRequest) => {
  if (request.providerOptions?.contextManagement !== undefined) return false
+  if (Schema.is(Schema.Struct({ mode: Schema.Literal("pro") }))(request.http?.body?.reasoning)) return false
  const override = request.model.compatibility?.supportsEffortUpdates
  if (override !== undefined) return override
-  return /(?:^|\/)gpt-6-astra$/i.test(request.model.id)
+  return /(?:^|\/)gpt-6-(?:astra|sol|luna)$/i.test(request.model.id)
 }

 const nativeImageToolInput = (tool: ToolDefinition) => {
@@ -159,23 +163,19 @@ const nativeImageTool = (tool: ToolDefinition) => {
  return Schema.is(OpenAIResponsesImageGenerationTool)(native) ? native : undefined
 }

-const lowerTool = Effect.fn("OpenAIResponses.lowerTool")(function* (tool: ToolDefinition, inputSchema: JsonSchema) {
+const lowerTool = Effect.fn("OpenAIResponses.lowerTool")(function* (tool: ToolDefinition) {
  const native = nativeImageToolInput(tool)
  if (native !== undefined) {
    if (Schema.is(OpenAIResponsesImageGenerationTool)(native)) return native
    return yield* ProviderShared.invalidRequest("OpenAI Responses image generation tool options are invalid")
  }
-  return yield* OpenResponses.lowerTool(NAME, tool, inputSchema)
+  return yield* OpenResponses.lowerTool(NAME, tool)
 })

 // Native namespaces hold only function tools, so deeper levels flatten into
 // the leaf names the same way non-native protocols flatten the whole tree.
-const lowerToolEntry = Effect.fn("OpenAIResponses.lowerToolEntry")(function* (
-  tool: ToolEntry,
-  compatibility: Parameters<typeof ToolSchemaProjection.modelCompatibility>[1],
-) {
-  if (tool.type === "tool")
-    return yield* lowerTool(tool, ToolSchemaProjection.modelCompatibility(tool.inputSchema, compatibility))
+const lowerToolEntry = Effect.fn("OpenAIResponses.lowerToolEntry")(function* (tool: ToolEntry) {
+  if (tool.type === "tool") return yield* lowerTool(tool)
  // OpenAI requires a namespace description; fall back to a generic one so a
  // missing description never blocks the request.
  return {
@@ -183,11 +183,13 @@ const lowerToolEntry = Effect.fn("OpenAIResponses.lowerToolEntry")(function* (
    name: tool.name,
    description: tool.description ?? `Tools in the ${tool.name} namespace.`,
    tools: yield* Effect.forEach(ProviderShared.flattenTools(tool.tools), (leaf) =>
-      OpenResponses.lowerTool(NAME, leaf, ToolSchemaProjection.modelCompatibility(leaf.inputSchema, compatibility)),
+      OpenResponses.lowerTool(NAME, leaf),
    ),
  }
 })

+const lowerTools = (request: LLMRequest) => Effect.forEach(request.tools, lowerToolEntry)
+
 const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>, tools: ReadonlyArray<ToolEntry>) =>
  ProviderShared.matchToolChoice(NAME, toolChoice, {
    auto: () => "auto" as const,
@@ -207,15 +209,11 @@ const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request:
  )(request.providerOptions?.contextManagement)
  const options = OpenResponsesOptions.resolve(request)
  const updates = resolveEffortUpdates(request, options.reasoningEffort)
-  const toolSchemaCompatibility = request.model.compatibility?.toolSchema
  return yield* decodeBody({
    ...(yield* OpenResponses.lowerConversation(updates.request, adapter)),
    ...OpenResponses.lowerGeneration(request, { ...options, reasoningEffort: updates.effort }),
    context_management: management?.map((edit) => ({ type: edit.type, compact_threshold: edit.compactThreshold })),
-    tools:
-      request.tools.length === 0
-        ? undefined
-        : yield* Effect.forEach(request.tools, (tool) => lowerToolEntry(tool, toolSchemaCompatibility)),
+    tools: request.tools.length === 0 ? undefined : yield* lowerTools(request),
    tool_choice:
      request.tools.length === 0
        ? undefined
@@ -227,7 +225,6 @@ const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request:
 const checkpointBody = {
  schema: CheckpointBody,
  from: Effect.fn("OpenAIResponses.checkpointBody")(function* (request: LLMRequest) {
-    const native = yield* fromRequest(LLMRequest.update(request, { toolChoice: undefined }))
    const overlay = request.http?.body
    // Complete history is required for stateless replay and SSE recovery. Raw input overrides bypass that contract.
    if (
@@ -238,18 +235,13 @@ const checkpointBody = {
      return yield* ProviderShared.invalidRequest(
        "Trigger compaction requires complete canonical history, not an input or continuation override",
      )
-    return yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(CheckpointBody))({
-      ...mergeJsonRecords(native, overlay),
-      input: [...native.input, { type: "compaction_trigger" }],
-      stream: true,
-      store: false,
-      parallel_tool_calls: true,
-      tool_choice: undefined,
-      context_management: undefined,
-      text: undefined,
-      max_output_tokens: undefined,
-      max_tool_calls: undefined,
-    })
+    if (overlay?.stream !== undefined && overlay.stream !== true)
+      return yield* ProviderShared.invalidRequest("Trigger compaction requires a streamed response")
+    const native = yield* fromRequest(request)
+    return {
+      ...native,
+      input: [...native.input, { type: "compaction_trigger" as const }],
+    }
  }),
 }

@@ -331,7 +323,10 @@ export const transport = channelTransport({
 })

 export const route = Route.make({
-  compact: { endpoint: ResponsesCompaction.make(adapter), trigger: ResponsesCheckpoint.make(checkpointBody) },
+  compact: {
+    endpoint: ResponsesCompaction.make(adapter, lowerTools),
+    trigger: ResponsesCheckpoint.make(checkpointBody),
+  },
  id: ADAPTER,
  provider: "openai",
  providerMetadataKey: "openai",
@@ -0,0 +1,144 @@
+import { Effect, Schema } from "effect"
+import { Framing } from "../route/framing.js"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { mergeJsonRecords, type MediaUsage } from "../schema/index.js"
+import { SpeechModel, type SpeechEvent, type SpeechRequestFor } from "../speech.js"
+import { SpeechStream } from "./utils/speech-stream.js"
+
+const route = MediaProtocol.identity({ id: "openai-speech", name: "OpenAI Speech", provider: "openai" })
+export const DEFAULT_BASE_URL = "https://api.openai.com/v1"
+export const PATH = "/audio/speech"
+/** `pcm` is raw 24 kHz, 16-bit signed little-endian mono samples without a header. */
+const PCM_SAMPLE_RATE = 24000
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+/** `voice`, `instructions`, `speed`, and `format` are common request fields; other native body fields pass through. */
+export type OpenAISpeechOptions = {
+  /** Defaults to `"sse"` in `stream` mode on models that support it; the merged value selects the response framing. */
+  readonly stream_format?: "sse" | "audio"
+} & Record<string, unknown>
+
+export type Request = SpeechRequestFor<OpenAISpeechOptions>
+
+// ---------------------------------------------------------------------------
+// 3. Streaming event schema
+// ---------------------------------------------------------------------------
+
+const SpeechStreamEvent = Schema.Union([
+  Schema.Struct({ type: Schema.Literal("speech.audio.delta"), audio: Schema.Uint8ArrayFromBase64 }),
+  Schema.Struct({
+    type: Schema.Literal("speech.audio.done"),
+    usage: Schema.optional(
+      Schema.Struct({
+        input_tokens: Schema.optional(Schema.Number),
+        output_tokens: Schema.optional(Schema.Number),
+        total_tokens: Schema.optional(Schema.Number),
+      }),
+    ),
+  }),
+])
+
+const decodeEvent = route.decodeFrame(SpeechStreamEvent)
+
+// ---------------------------------------------------------------------------
+// 4. Parser state
+// ---------------------------------------------------------------------------
+
+interface State extends SpeechStream.Audio {
+  readonly done: boolean
+  readonly usage?: MediaUsage
+}
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+// `sse` is not supported for `tts-1` or `tts-1-hd`; those models stream the raw audio body instead.
+const supportsSse = (model: string) => !/^tts-1(-hd)?(-|$)/.test(model)
+
+const fromRequest = Effect.fn("OpenAISpeech.fromRequest")(function* (request: MediaProtocol.Addressed<Request>) {
+  // Not in `unsupported`: that list would also reject `timestamps: false`, which asks for nothing.
+  if (request.timestamps === true)
+    return yield* route.unsupported("media.timestamps", `${route.name} does not return timestamps`)
+  return MediaProtocol.json(
+    mergeJsonRecords(
+      {
+        model: request.model.id,
+        input: request.text,
+        voice: request.voice,
+        instructions: request.instructions,
+        response_format: request.format,
+        speed: request.speed,
+        stream_format: request.mode === "stream" && supportsSse(request.model.id) ? "sse" : undefined,
+      },
+      request.providerOptions,
+      request.http?.body,
+    ) ?? {},
+  )
+})
+
+// ---------------------------------------------------------------------------
+// 6. Stream parsing
+// ---------------------------------------------------------------------------
+
+const isSse = (body: MediaProtocol.Body) => body.type === "json" && body.value.stream_format === "sse"
+
+const onEvent = Effect.fn("OpenAISpeech.onEvent")(function* (state: State, frame: string) {
+  const event = yield* decodeEvent(frame)
+  if (event.type === "speech.audio.delta") return SpeechStream.delta(state, event.audio)
+  const usage = event.usage
+  return [
+    {
+      ...state,
+      done: true,
+      usage:
+        usage === undefined
+          ? undefined
+          : {
+              type: "tokens" as const,
+              input: usage.input_tokens,
+              output: usage.output_tokens,
+              total: usage.total_tokens,
+              details: { openai: usage },
+            },
+    },
+    [],
+  ] as const
+})
+
+const finish = (state: State, context: MediaProtocol.ResponseContext<Request>) => {
+  if (isSse(context.body) && !state.done) return Effect.fail(route.incomplete())
+  const format = context.request.format ?? "mp3"
+  return SpeechStream.finish(route, state, {
+    ...(format === "pcm" ? SpeechStream.pcm("pcm_s16le", PCM_SAMPLE_RATE) : SpeechStream.container(format)),
+    usage: state.usage,
+  })
+}
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+export const protocol = MediaProtocol.stream<Request, SpeechEvent, string | Uint8Array, State>(route, {
+  unsupported: ["language"],
+  body: { from: fromRequest },
+  frames: (bytes, context) => (isSse(context.body) ? Framing.sse.frame(bytes) : bytes),
+  initial: () => ({ chunks: [], done: false }),
+  step: SpeechStream.step(onEvent),
+  finish,
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  SpeechModel.fromRoute<OpenAISpeechOptions, string | Uint8Array, State>(
+    { protocol, baseURL: DEFAULT_BASE_URL, path: PATH },
+    input,
+  )
+
+export const OpenAISpeech = {
+  protocol,
+  model,
+} as const
@@ -0,0 +1,289 @@
+import { Effect, Schema, Stream } from "effect"
+import { classifyProviderFailure } from "../provider-error.js"
+import { Framing } from "../route/framing.js"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { AIError, mergeJsonRecords, type MediaUsage } from "../schema/index.js"
+import {
+  TranscriptionFinishEvent,
+  TranscriptionModel,
+  TranscriptionSegmentEvent,
+  TranscriptionTextDeltaEvent,
+  type TranscriptionEvent,
+  type TranscriptionRequestFor,
+  type TranscriptionSegment,
+} from "../transcription.js"
+import { mediaTypeExtension } from "../utils/media-type.js"
+import { ProviderShared } from "./shared.js"
+import { MediaInput } from "./utils/media-input.js"
+
+const route = MediaProtocol.identity({ id: "openai-transcription", name: "OpenAI Transcription", provider: "openai" })
+export const DEFAULT_BASE_URL = "https://api.openai.com/v1"
+export const PATH = "/audio/transcriptions"
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+export type OpenAITranscriptionOptions = {
+  readonly chunking_strategy?:
+    | "auto"
+    | {
+        readonly type: "server_vad"
+        readonly prefix_padding_ms?: number
+        readonly silence_duration_ms?: number
+        readonly threshold?: number
+      }
+  readonly include?: ReadonlyArray<"logprobs">
+  readonly keywords?: ReadonlyArray<string>
+  readonly known_speaker_names?: ReadonlyArray<string>
+  readonly known_speaker_references?: ReadonlyArray<string>
+  readonly temperature?: number
+} & Record<string, unknown>
+
+export type Request = TranscriptionRequestFor<OpenAITranscriptionOptions>
+
+// ---------------------------------------------------------------------------
+// 3. Streaming event schema
+// ---------------------------------------------------------------------------
+
+const Segment = Schema.Struct({
+  text: Schema.String,
+  start: Schema.Number,
+  end: Schema.Number,
+  speaker: Schema.optional(Schema.String),
+})
+
+const Usage = Schema.Union([
+  Schema.Struct({
+    type: Schema.Literal("tokens"),
+    input_tokens: Schema.optional(Schema.Number),
+    output_tokens: Schema.optional(Schema.Number),
+    total_tokens: Schema.optional(Schema.Number),
+    input_token_details: Schema.optional(
+      Schema.Struct({ audio_tokens: Schema.optional(Schema.Number), text_tokens: Schema.optional(Schema.Number) }),
+    ),
+  }),
+  Schema.Struct({ type: Schema.Literal("duration"), seconds: Schema.Number }),
+])
+
+const transcriptFields = {
+  text: Schema.String,
+  language: Schema.optional(Schema.String),
+  languages: Schema.optional(Schema.Array(Schema.Struct({ code: Schema.String }))),
+  duration: Schema.optional(Schema.Number),
+  segments: Schema.optional(Schema.Array(Segment)),
+  words: Schema.optional(
+    Schema.Array(Schema.Struct({ word: Schema.String, start: Schema.Number, end: Schema.Number })),
+  ),
+  usage: Schema.optional(Usage),
+}
+
+/** OpenAI may add stream event types; frames outside `EVENT_TYPES` are ignored. */
+const EventType = Schema.Struct({ type: Schema.String })
+const Event = Schema.Union([
+  Schema.Struct({ type: Schema.Literal("transcript.text.delta"), delta: Schema.String }),
+  Schema.Struct({ type: Schema.Literal("transcript.text.segment"), ...Segment.fields }),
+  Schema.Struct({ type: Schema.Literal("transcript.text.done"), ...transcriptFields }),
+  Schema.Struct({
+    type: Schema.Literal("error"),
+    message: Schema.optional(Schema.String),
+    error: Schema.optional(Schema.Struct({ message: Schema.optional(Schema.String) })),
+  }),
+])
+const EVENT_TYPES = new Set(["transcript.text.delta", "transcript.text.segment", "transcript.text.done", "error"])
+const Transcript = Schema.Struct(transcriptFields)
+type Transcript = Schema.Schema.Type<typeof Transcript>
+
+const decodeEventType = route.decodeFrame(EventType)
+const decodeEvent = route.decodeFrame(Event)
+const decodeTranscript = route.decodeFrame(Transcript)
+
+type Frame = string | { readonly document: string }
+
+// ---------------------------------------------------------------------------
+// 4. Parser state
+// ---------------------------------------------------------------------------
+
+interface State {
+  readonly segments: Array<TranscriptionSegment>
+  readonly transcript?: Transcript
+}
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+interface Capabilities {
+  readonly stream: boolean
+  readonly timestamps: ReadonlyArray<"segment" | "word">
+  readonly diarize: boolean
+  readonly languageField: "language" | "languages"
+}
+
+const TRANSCRIBE: Capabilities = { stream: true, timestamps: [], diarize: false, languageField: "language" }
+
+const capabilities = (model: string): Capabilities => {
+  if (model.startsWith("whisper")) return { ...TRANSCRIBE, stream: false, timestamps: ["segment", "word"] }
+  if (model.includes("diarize")) return { ...TRANSCRIBE, timestamps: ["segment"], diarize: true }
+  // `gpt-transcribe` replaces `language` with `languages[]` and rejects both together.
+  if (model.startsWith("gpt-transcribe")) return { ...TRANSCRIBE, languageField: "languages" }
+  return TRANSCRIBE
+}
+
+/** whisper-1 ignores `stream`, so its `stream` mode sends a plain request and emits only `finish`. */
+const streamsEvents = (request: MediaProtocol.Addressed<Request>) =>
+  request.mode === "stream" && capabilities(request.model.id).stream
+
+const validate = (request: MediaProtocol.Addressed<Request>, model: Capabilities) => {
+  const id = request.model.id
+  if (request.diarize === true && !model.diarize)
+    return Effect.fail(route.unsupported("media.diarize", `${id} does not diarize; use gpt-4o-transcribe-diarize`))
+  if (request.prompt !== undefined && model.diarize)
+    return Effect.fail(route.unsupported("media.prompt", `${id} does not accept a prompt`))
+  if (
+    request.timestamps === undefined ||
+    request.timestamps === "none" ||
+    model.timestamps.includes(request.timestamps)
+  )
+    return Effect.void
+  return Effect.fail(route.unsupported("media.timestamps", `${id} does not return ${request.timestamps} timestamps`))
+}
+
+const RESERVED_FORM_FIELDS = new Set([
+  "file",
+  "model",
+  "prompt",
+  "language",
+  "response_format",
+  "timestamp_granularities",
+  "stream",
+])
+
+const fromRequest = Effect.fn("OpenAITranscription.fromRequest")(function* (request: MediaProtocol.Addressed<Request>) {
+  const model = capabilities(request.model.id)
+  yield* validate(request, model)
+  // The API detects the audio format from the upload's filename extension.
+  const extension = mediaTypeExtension(request.audio.mediaType)
+  if (extension === undefined)
+    return yield* ProviderShared.invalidRequest(
+      `${route.name} cannot name a ${request.audio.mediaType} upload; send mp3, mp4, m4a, wav, webm, ogg, or flac audio`,
+    )
+  const audio = yield* MediaInput.inlineBytes(route.id, request.audio)
+  const responseFormat = model.diarize
+    ? "diarized_json"
+    : request.timestamps === undefined || request.timestamps === "none"
+      ? undefined
+      : "verbose_json"
+  const form = new FormData()
+  form.append("file", MediaInput.blob(audio, request.audio.mediaType), `audio.${extension}`)
+  MediaInput.appendFields(
+    form,
+    {
+      model: request.model.id,
+      language: model.languageField === "language" ? request.language : undefined,
+      languages: model.languageField === "languages" && request.language !== undefined ? [request.language] : undefined,
+      prompt: request.prompt,
+      response_format: responseFormat,
+      timestamp_granularities: responseFormat === "verbose_json" ? [request.timestamps] : undefined,
+      // Diarizing audio longer than 30 seconds requires a chunking strategy.
+      chunking_strategy: model.diarize ? "auto" : undefined,
+      stream: streamsEvents(request) ? true : undefined,
+    },
+    {
+      overlay: mergeJsonRecords(request.providerOptions, request.http?.body),
+      reserved: RESERVED_FORM_FIELDS,
+      repeatArrays: true,
+    },
+  )
+  return MediaProtocol.multipart(form)
+})
+
+// ---------------------------------------------------------------------------
+// 6. Stream parsing
+// ---------------------------------------------------------------------------
+
+const segment = (value: Schema.Schema.Type<typeof Segment>): TranscriptionSegment => ({
+  text: value.text.trim(),
+  startSeconds: value.start,
+  endSeconds: value.end,
+  speaker: value.speaker,
+})
+
+const onEvent = Effect.fn("OpenAITranscription.onEvent")(function* (state: State, frame: string) {
+  if (!EVENT_TYPES.has((yield* decodeEventType(frame)).type)) return [state, []] as const
+  const event = yield* decodeEvent(frame)
+  if (event.type === "error")
+    return yield* new AIError({
+      reason: classifyProviderFailure({
+        message: `${route.name} stream failed: ${event.message ?? event.error?.message ?? "unknown error"}`,
+        rawBody: frame,
+      }),
+    })
+  if (event.type === "transcript.text.done") return [{ ...state, transcript: event }, []] as const
+  if (event.type === "transcript.text.delta")
+    return [state, event.delta.length === 0 ? [] : [TranscriptionTextDeltaEvent.make({ delta: event.delta })]] as const
+  const next = segment(event)
+  state.segments.push(next)
+  return [state, [TranscriptionSegmentEvent.make({ segment: next })]] as const
+})
+
+const step = (state: State, frame: Frame) =>
+  typeof frame === "string"
+    ? onEvent(state, frame)
+    : decodeTranscript(frame.document).pipe(Effect.map((transcript) => [{ ...state, transcript }, []] as const))
+
+const usage = (value: Transcript["usage"]): MediaUsage | undefined => {
+  if (value === undefined) return undefined
+  if (value.type === "duration") return { type: "seconds", seconds: value.seconds }
+  return {
+    type: "tokens",
+    input: value.input_tokens,
+    output: value.output_tokens,
+    total: value.total_tokens,
+    details: { openai: value },
+  }
+}
+
+const finish = (state: State) => {
+  const transcript = state.transcript
+  if (transcript === undefined) return Effect.fail(route.incomplete())
+  const segments = transcript.segments?.map(segment) ?? state.segments
+  return Effect.succeed([
+    TranscriptionFinishEvent.make({
+      text: transcript.text,
+      segments: segments.length === 0 ? undefined : segments,
+      words: transcript.words?.map((word) => ({ text: word.word, startSeconds: word.start, endSeconds: word.end })),
+      language: (transcript.language ?? transcript.languages?.[0]?.code)?.toLowerCase(),
+      durationSeconds: transcript.duration,
+      usage: usage(transcript.usage),
+    }),
+  ])
+}
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+export const protocol = MediaProtocol.stream<Request, TranscriptionEvent, Frame, State>(route, {
+  unsupported: ["speakers"],
+  body: { from: fromRequest },
+  frames: (bytes, context) =>
+    streamsEvents(context.request)
+      ? Framing.sse.frame(bytes)
+      : Framing.document.frame(bytes).pipe(Stream.map((document) => ({ document }))),
+  initial: () => ({ segments: [] }),
+  step,
+  finish,
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  TranscriptionModel.fromRoute<OpenAITranscriptionOptions, Frame, State>(
+    { protocol, baseURL: DEFAULT_BASE_URL, path: PATH },
+    input,
+  )
+
+export const OpenAITranscription = {
+  protocol,
+  model,
+} as const
@@ -0,0 +1,180 @@
+import { Duration, Effect, Schema } from "effect"
+import type { HttpClientResponse } from "effect/unstable/http"
+import type { Status } from "../generation.js"
+import { ImageModel, ImageResponse, type ImageRequestFor } from "../image.js"
+import { Media } from "../media.js"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { mergeJsonRecords, type AIError } from "../schema/index.js"
+import { ProviderShared, optionalNull } from "./shared.js"
+
+const route = MediaProtocol.identity({ id: "replicate-images", name: "Replicate", provider: "replicate" })
+export const DEFAULT_BASE_URL = "https://api.replicate.com"
+const OUTPUT_RETENTION = Duration.hours(1)
+const MAX_DATA_URL_BYTES = 256 * 1024
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+/** Model-defined `input` fields under the model's own names; `Media.Asset` values lower to URLs or data URLs. */
+export type ReplicateImageOptions = Record<string, unknown>
+
+export type Request = ImageRequestFor<ReplicateImageOptions>
+
+// ---------------------------------------------------------------------------
+// 2. Token and response schemas
+// ---------------------------------------------------------------------------
+
+export const Token = Schema.Struct({ id: Schema.String, getURL: Schema.String, cancelURL: Schema.String })
+export type Token = Schema.Schema.Type<typeof Token>
+
+const Prediction = Schema.Struct({
+  id: Schema.String,
+  status: Schema.String,
+  output: optionalNull(Schema.Unknown),
+  error: optionalNull(Schema.Unknown),
+  data_removed: optionalNull(Schema.Boolean),
+  completed_at: optionalNull(Schema.String),
+  metrics: optionalNull(Schema.Struct({ predict_time: optionalNull(Schema.Number) })),
+  urls: Schema.Struct({ get: Schema.String, cancel: Schema.String }),
+})
+
+const Output = Schema.Union([Schema.String, Schema.Array(Schema.String)])
+const isOutput = Schema.is(Output)
+
+const STATUS = {
+  starting: "queued",
+  processing: "running",
+  succeeded: "completed",
+  failed: "failed",
+  canceled: "cancelled",
+  // The prediction hit its `Cancel-After` deadline before it started running.
+  aborted: "expired",
+} as const satisfies Record<string, Status>
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+// Official models run by `owner/name`; anything else (`owner/name:version` or a bare version id) is a pinned version.
+const isOfficial = (model: string) => /^[^/:]+\/[^/:]+$/.test(model)
+
+const inlineSize = (source: Media.Source) => {
+  if (source.type === "bytes") return source.data.byteLength
+  if (source.type === "base64") return source.data.length * 0.75
+  return 0
+}
+
+const fileInput = (asset: Media.Asset) => {
+  if (inlineSize(asset.source) > MAX_DATA_URL_BYTES)
+    return Effect.fail(
+      ProviderShared.invalidRequest(
+        `${route.name} data URL inputs are limited to 256 KB; pass a larger file by https URL`,
+      ),
+    )
+  return ProviderShared.mediaReference(asset, undefined, route.name).pipe(Effect.map((reference) => reference.value))
+}
+
+const inputValue = (value: unknown): Effect.Effect<unknown, AIError> => {
+  if (value instanceof Media.Asset) return fileInput(value)
+  if (Array.isArray(value) && value.some((item) => item instanceof Media.Asset))
+    return Effect.forEach(value, (item) => (item instanceof Media.Asset ? fileInput(item) : Effect.succeed(item)))
+  return Effect.succeed(value)
+}
+
+const fromRequest = Effect.fn("ReplicateImages.fromRequest")(function* (request: Request) {
+  const native = yield* Effect.forEach(Object.entries(request.providerOptions ?? {}), ([key, value]) =>
+    inputValue(value).pipe(Effect.map((lowered) => [key, lowered] as const)),
+  )
+  const input = mergeJsonRecords({ prompt: request.prompt }, Object.fromEntries(native))
+  const model = request.model.id
+  return MediaProtocol.json(
+    mergeJsonRecords(isOfficial(model) ? { input } : { version: model, input }, request.http?.body) ?? {},
+  )
+})
+
+// ---------------------------------------------------------------------------
+// 6. Response decoding
+// ---------------------------------------------------------------------------
+
+const decodePrediction = route.decodeJson(Prediction)
+
+const decodeStart = Effect.fn("ReplicateImages.decodeStart")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+) {
+  const output = yield* decodePrediction(response)
+  const prediction = output.value
+  return {
+    token: { id: prediction.id, getURL: prediction.urls.get, cancelURL: prediction.urls.cancel },
+    // `Prefer: wait` can return an already-finished prediction, so `await` skips straight to the result.
+    snapshot: { id: prediction.id, status: yield* MediaProtocol.status(STATUS, prediction.status, output) },
+  }
+})
+
+const decodeStatus = Effect.fn("ReplicateImages.decodeStatus")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.PollContext<Token>,
+) {
+  const output = yield* decodePrediction(response)
+  return { id: context.token.id, status: yield* MediaProtocol.status(STATUS, output.value.status, output) }
+})
+
+const decodeResult = Effect.fn("ReplicateImages.decodeResult")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.PollContext<Token>,
+) {
+  const output = yield* decodePrediction(response)
+  const prediction = output.value
+  const status = yield* MediaProtocol.status(STATUS, prediction.status, output)
+  if (status === "failed" || status === "cancelled" || status === "expired")
+    return yield* output.ended(
+      status,
+      `${route.name} prediction ${context.token.id} ${prediction.status}${typeof prediction.error === "string" ? `: ${prediction.error}` : ""}`,
+    )
+  if (status !== "completed") return yield* output.pending(context.token.id)
+  if (prediction.data_removed === true)
+    return yield* output.ended("expired", `${route.name} removed the output of prediction ${context.token.id}`)
+  if (!isOutput(prediction.output))
+    return yield* output.invalid(`${route.name} prediction ${context.token.id} returned output that is not image URLs`)
+  const urls = typeof prediction.output === "string" ? [prediction.output] : prediction.output
+  if (urls.length === 0) return yield* output.invalid(`${route.name} prediction ${context.token.id} returned no images`)
+  const predictTime = prediction.metrics?.predict_time ?? undefined
+  const completedAt = prediction.completed_at ?? undefined
+  const expiresAt =
+    completedAt === undefined ? undefined : Date.parse(completedAt) + Duration.toMillis(OUTPUT_RETENTION)
+  return new ImageResponse({
+    images: urls.map((url) => Media.url(url, { expiresAt })),
+    usage: predictTime === undefined ? undefined : { type: "compute", seconds: predictTime },
+    providerMetadata: { replicate: { predictionId: prediction.id } },
+  })
+})
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+export const protocol = MediaProtocol.queued<Request, ImageResponse, Token>(route, {
+  token: Token,
+  unsupported: ["images", "mask", "n", "size", "aspectRatio", "seed", "format"],
+  start: { body: { from: fromRequest }, decode: decodeStart },
+  status: { path: (token) => token.getURL, decode: decodeStatus },
+  result: { path: (token) => token.getURL, decode: decodeResult },
+  cancel: { method: "POST", path: (token) => token.cancelURL },
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  ImageModel.fromRoute<ReplicateImageOptions, Token>(
+    {
+      protocol,
+      baseURL: DEFAULT_BASE_URL,
+      path: ({ request }) =>
+        isOfficial(request.model.id) ? `/v1/models/${request.model.id}/predictions` : "/v1/predictions",
+    },
+    input,
+  )
+
+export const ReplicateImages = {
+  protocol,
+  model,
+} as const
@@ -0,0 +1,196 @@
+import { Duration, Effect, Schema } from "effect"
+import type { HttpClientResponse } from "effect/unstable/http"
+import type { Status } from "../generation.js"
+import { Media } from "../media.js"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { mergeJsonRecords, type OpenString } from "../schema/index.js"
+import { VideoModel, VideoResponse, type VideoRequestFor } from "../video.js"
+import { ProviderShared, optionalArray, optionalNull } from "./shared.js"
+
+const route = MediaProtocol.identity({ id: "runway-video", name: "Runway", provider: "runway" })
+export const DEFAULT_BASE_URL = "https://api.dev.runwayml.com/v1"
+/** Every Runway request must pin the API version. */
+export const API_VERSION = "2024-11-06"
+export const TEXT_TO_VIDEO_PATH = "/text_to_video"
+export const IMAGE_TO_VIDEO_PATH = "/image_to_video"
+export const VIDEO_TO_VIDEO_PATH = "/video_to_video"
+export const TASKS_PATH = "/tasks"
+/** Output URLs are valid for 24–48 hours; the asset carries the conservative bound. */
+const OUTPUT_RETENTION = Duration.hours(24)
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+/**
+ * Provider-native options. Common fields lower to Runway's names: `aspectRatio` → `ratio` (Runway expects pixel
+ * ratios such as `1280:720` for most models), `durationSeconds` → `duration`, `audio`, `negativePrompt`,
+ * `resolution`, `references`, and `frames` → `promptImage`.
+ */
+export type RunwayVideoOptions = {
+  readonly contentModeration?: { readonly publicFigureThreshold?: OpenString<"auto" | "low"> }
+  readonly outputFormat?: OpenString<"mp4" | "prores" | "png_sequence">
+} & Record<string, unknown>
+
+export type Request = VideoRequestFor<RunwayVideoOptions>
+
+// ---------------------------------------------------------------------------
+// 2. Token and response schemas
+// ---------------------------------------------------------------------------
+
+export const Token = Schema.Struct({ taskID: Schema.String })
+export type Token = Schema.Schema.Type<typeof Token>
+
+const Cost = Schema.Struct({ credits: Schema.Number })
+
+const StartResponse = Schema.Struct({ id: Schema.String })
+
+const Task = Schema.Struct({
+  status: Schema.String,
+  progress: optionalNull(Schema.Number),
+  output: optionalArray(Schema.String),
+  failure: optionalNull(Schema.String),
+  failureCode: optionalNull(Schema.String),
+  cost: Schema.optional(Cost),
+  estimatedCost: Schema.optional(Cost),
+})
+
+const STATUS = {
+  PENDING: "queued",
+  THROTTLED: "queued",
+  RUNNING: "running",
+  SUCCEEDED: "completed",
+  FAILED: "failed",
+  CANCELLED: "cancelled",
+} as const satisfies Record<string, Status>
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+// Runway accepts HTTPS URLs, `runway://` upload URIs, and data URIs, all as one string.
+const mediaUri = (asset: Media.Asset) =>
+  ProviderShared.mediaReference(asset, route.provider, route.name).pipe(Effect.map((reference) => reference.value))
+
+const fromRequest = Effect.fn("RunwayVideo.fromRequest")(function* (request: Request) {
+  const first = request.frames?.first === undefined ? undefined : yield* mediaUri(request.frames.first)
+  const last = request.frames?.last === undefined ? undefined : yield* mediaUri(request.frames.last)
+  const promptImage = [
+    ...(first === undefined ? [] : [{ uri: first, position: "first" }]),
+    ...(last === undefined ? [] : [{ uri: last, position: "last" }]),
+  ]
+  const videoUri = request.video === undefined ? undefined : yield* mediaUri(request.video)
+  const references = yield* Effect.forEach(request.references ?? [], (asset) =>
+    mediaUri(asset).pipe(Effect.map((uri) => ({ uri }))),
+  )
+  return MediaProtocol.json(
+    mergeJsonRecords(
+      {
+        model: request.model.id,
+        promptText: request.prompt,
+        promptImage: promptImage.length === 0 ? undefined : promptImage,
+        videoUri,
+        references: references.length === 0 ? undefined : references,
+        ratio: request.aspectRatio,
+        duration: request.durationSeconds,
+        resolution: request.resolution,
+        audio: request.audio,
+        negativePrompt: request.negativePrompt,
+        seed: request.seed,
+      },
+      request.providerOptions,
+      request.http?.body,
+    ) ?? {},
+  )
+})
+
+// ---------------------------------------------------------------------------
+// 6. Response decoding
+// ---------------------------------------------------------------------------
+
+const decodeStart = route.decodeStarted(StartResponse, (value) => ({
+  token: { taskID: value.id },
+  snapshot: { id: value.id, status: "queued" },
+}))
+
+const decodeTask = route.decodeJson(Task)
+
+const decodeStatus = Effect.fn("RunwayVideo.decodeStatus")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.PollContext<Token>,
+) {
+  const output = yield* decodeTask(response)
+  const status = yield* MediaProtocol.status(STATUS, output.value.status, output)
+  return { id: context.token.taskID, status, progress: output.value.progress ?? undefined }
+})
+
+const decodeResult = Effect.fn("RunwayVideo.decodeResult")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.PollContext<Token>,
+) {
+  const output = yield* decodeTask(response)
+  const task = output.value
+  const status = yield* MediaProtocol.status(STATUS, task.status, output)
+  if (status === "failed") {
+    const code = task.failureCode ?? undefined
+    const message = `${route.name} task failed${code === undefined ? "" : ` (${code})`}${task.failure ? `: ${task.failure}` : ""}`
+    // Runway failure codes are dotted paths; every moderation outcome carries a SAFETY segment.
+    if (code !== undefined && /(^|\.)SAFETY(\.|$)/.test(code)) return yield* output.contentPolicy(message)
+    return yield* output.ended("failed", message)
+  }
+  if (status === "cancelled")
+    return yield* output.ended("cancelled", `${route.name} task ${context.token.taskID} was cancelled`)
+  if (status !== "completed") return yield* output.pending(context.token.taskID)
+  const urls = task.output ?? []
+  if (urls.length === 0) return yield* output.invalid(`${route.name} task succeeded without any output`)
+  return new VideoResponse({
+    videos: yield* Effect.forEach(urls, (url) =>
+      MediaProtocol.expiringUrl(url, OUTPUT_RETENTION, { mediaType: "video/mp4" }),
+    ),
+    usage: task.cost === undefined ? undefined : { type: "credits", credits: task.cost.credits },
+    providerMetadata: {
+      runway: {
+        taskId: context.token.taskID,
+        estimatedCredits: task.estimatedCost?.credits,
+      },
+    },
+  })
+})
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+const taskPath = (token: Token) => `${TASKS_PATH}/${token.taskID}`
+
+export const protocol = MediaProtocol.queued<Request, VideoResponse, Token>(route, {
+  token: Token,
+  unsupported: ["n"],
+  start: { body: { from: fromRequest }, decode: decodeStart },
+  status: { path: taskPath, decode: decodeStatus },
+  result: { path: taskPath, decode: decodeResult },
+  cancel: { method: "DELETE", path: taskPath, activeOnly: true },
+})
+
+const startPath = (request: Request) => {
+  if (request.video !== undefined) return VIDEO_TO_VIDEO_PATH
+  if (request.frames?.first !== undefined || request.frames?.last !== undefined) return IMAGE_TO_VIDEO_PATH
+  return TEXT_TO_VIDEO_PATH
+}
+
+export const model = (input: MediaRoute.ModelInput) =>
+  VideoModel.fromRoute<RunwayVideoOptions, Token>(
+    {
+      protocol,
+      baseURL: DEFAULT_BASE_URL,
+      headers: { "X-Runway-Version": API_VERSION },
+      path: ({ request }) => startPath(request),
+    },
+    input,
+  )
+
+export const RunwayVideo = {
+  protocol,
+  model,
+} as const
@@ -1,6 +1,5 @@
 import { Tool } from "@opencode/schema/tool"
-import { Effect, Option, Schema, Stream } from "effect"
-import * as Sse from "effect/unstable/encoding/Sse"
+import { Effect, Option, Schema } from "effect"
 import { Headers, HttpClientRequest } from "effect/unstable/http"
 import { Media } from "../media.js"
 import {
@@ -13,17 +12,16 @@ import {
  ToolDefinition,
  type ContentPart,
  type MediaPart,
+  type OpenString,
  type ProviderID,
  type TextPart,
  type ToolEntry,
  type ToolResultPart,
 } from "../schema/index.js"
+import { Json, decodeJson, encodeJson } from "../utils/json.js"
 import { isRecord } from "../utils/record.js"
-export { isRecord }
+export { Json, decodeJson, encodeJson, isRecord }

-export const Json = Schema.fromJsonString(Schema.Unknown)
-export const decodeJson = Schema.decodeUnknownSync(Json)
-export const encodeJson = Schema.encodeSync(Json)
 const isJson = Schema.is(Schema.Json)
 export const JsonObject = Schema.Record(Schema.String, Schema.Unknown)
 export const optionalArray = <const S extends Schema.Top>(schema: S) => Schema.optional(Schema.Array(schema))
@@ -35,7 +33,7 @@ export const lenient = <const S extends Schema.Top>(schema: S) =>
  )
 /** Provider-defined string enum: known values for autocomplete, any string accepted at runtime. */
 export const knownString = <Known extends string>() =>
-  Schema.declare<Known | (string & {})>((value): value is Known | (string & {}) => typeof value === "string", {
+  Schema.declare<OpenString<Known>>((value): value is OpenString<Known> => typeof value === "string", {
    expected: "string",
  })

@@ -112,6 +110,14 @@ export const sumTokens = (...values: ReadonlyArray<number | undefined>): number
  return values.reduce((acc: number, value) => acc + (value ?? 0), 0)
 }

+/**
+ * Caps an explicit thinking budget at half the output limit. Thinking counts against the output limit, so a budget
+ * near it leaves the answer, a tool call, or a summary without room. Smaller budgets, special values such as `-1` and
+ * `0`, and requests without an output limit pass through unchanged.
+ */
+export const fitThinkingBudget = (budget: number, maxTokens: number | undefined, minimum = 1) =>
+  maxTokens === undefined || budget <= maxTokens / 2 ? budget : Math.max(minimum, Math.floor(maxTokens / 2))
+
 export const eventError = (route: string, message: string, body?: string, cause?: unknown) =>
  new AIError({
    reason: new InvalidProviderOutputError({ route, message, body, cause }),
@@ -192,6 +198,29 @@ export const inlineRequired = (route: string, asset: Media.Asset) =>
 /** The remote URL of a `url` asset, for protocols that accept `http(s)` references natively. */
 export const mediaUrl = (asset: Media.Asset) => (asset.source.type === "url" ? asset.source.url : undefined)

+export type MediaReference = { readonly type: "dataUrl" | "url" | "ref"; readonly value: string }
+
+/**
+ * The one string a provider can address an asset by: inline payloads as a data URL, `url` sources as their URL, and
+ * this provider's own `ref` as its id. Other providers' refs are never forwarded and fail typed; omit `provider` for
+ * APIs with no file handles at all.
+ */
+export const mediaReference = (
+  asset: Media.Asset,
+  provider: ProviderID | undefined,
+  label: string,
+): Effect.Effect<MediaReference, AIError> => {
+  const inline = asset.inline()
+  if (inline) return Effect.succeed({ type: "dataUrl", value: inline.dataUrl })
+  const url = mediaUrl(asset)
+  if (url) return Effect.succeed({ type: "url", value: url })
+  if (provider !== undefined && asset.source.type === "ref" && asset.source.provider === provider)
+    return Effect.succeed({ type: "ref", value: asset.source.id })
+  const accepted = provider === undefined ? "" : `, and ${provider} references`
+  const got = asset.source.type === "ref" ? `; got ${asset.source.provider}:${asset.source.id}` : ""
+  return Effect.fail(invalidRequest(`${label} accepts inline bytes, data URLs, http(s) URLs${accepted}${got}`))
+}
+
 /**
 * Lift a tool-result file into a `MediaPart`. Tool files carry either a data URL, an `http(s)` URL, or raw base64 in
 * `uri`; the declared `mime` wins over any data-URL prefix so tool authors control the type the model sees.
@@ -206,8 +235,6 @@ export const toolFileMedia = (item: Tool.FileContent): MediaPart => {
  return Message.media(asset, { filename: item.name })
 }

-export const trimBaseUrl = (value: string) => value.replace(/\/+$/, "")
-
 export const toolResultText = (part: ToolResultPart) => {
  if (part.result.type === "text") return String(part.result.value)
  if (part.result.type === "error") {
@@ -229,47 +256,6 @@ export const errorText = (error: unknown) => {
  return "Unknown stream error"
 }

-/**
- * `framing` step for Server-Sent Events. Decodes UTF-8, runs the SSE channel
- * decoder, optionally filters named events, and drops empty events. `[DONE]`
- * is dropped by default or retained for protocols that use it as their stream
- * boundary. Retry control events are ignored without interrupting the stream.
- * Decoder failures become provider output errors so the public error channel
- * stays `AIError`.
- */
-export const sseFraming = (
-  bytes: Stream.Stream<Uint8Array, AIError>,
-  events?: ReadonlySet<string>,
-  includeDone = false,
-): Stream.Stream<string, AIError> =>
-  bytes.pipe(
-    Stream.decodeText(),
-    Stream.mapAccumEffect(
-      () => {
-        const output: Sse.Event[] = []
-        return {
-          output,
-          parser: Sse.makeParser((event) => {
-            if (event._tag === "Event") output.push(event)
-          }),
-        }
-      },
-      (state, chunk) =>
-        Effect.gen(function* () {
-          const error = state.parser.feed(chunk)
-          if (error) return yield* eventError("sse", error.message, chunk, error)
-          return [state, state.output.splice(0)] as const
-        }),
-    ),
-    Stream.filter(
-      (event) =>
-        (events === undefined || events.has(event.event)) &&
-        event.data.length > 0 &&
-        (event.data !== "[DONE]" || includeDone || (events !== undefined && event.event !== "message")),
-    ),
-    Stream.map((event) => event.data),
-  )
-
 /**
 * Canonical invalid-request constructor shared by protocol lowering.
 */
@@ -0,0 +1,229 @@
+import { Effect, Schema } from "effect"
+import type { HttpClientResponse } from "effect/unstable/http"
+import { ImageModel, ImageResponse, type ImageRequestFor } from "../image.js"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { mergeJsonRecords, type OpenString } from "../schema/index.js"
+import { ProviderShared } from "./shared.js"
+import { MediaInput } from "./utils/media-input.js"
+
+const route = MediaProtocol.identity({ id: "stability-images", name: "Stability AI", provider: "stability" })
+const upscaleRoute = MediaProtocol.identity({ id: "stability-upscale", name: "Stability AI", provider: "stability" })
+export const DEFAULT_BASE_URL = "https://api.stability.ai"
+const RESULTS_PATH = "/v2beta/results"
+const UPSCALE_MODEL = "creative"
+/** Base64 JSON instead of raw bytes, so the seed and finish reason arrive with the image. */
+const HEADERS = { accept: "application/json" }
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+export type StabilityStylePreset = OpenString<
+  | "enhance"
+  | "anime"
+  | "photographic"
+  | "digital-art"
+  | "comic-book"
+  | "fantasy-art"
+  | "line-art"
+  | "analog-film"
+  | "neon-punk"
+  | "isometric"
+  | "low-poly"
+  | "origami"
+  | "modeling-compound"
+  | "cinematic"
+  | "3d-model"
+  | "pixel-art"
+  | "tile-texture"
+>
+
+export type StabilityImageOptions = {
+  readonly negative_prompt?: string
+  readonly style_preset?: StabilityStylePreset
+  readonly strength?: number
+  readonly cfg_scale?: number
+} & Record<string, unknown>
+
+/** Creative upscale options; the one `images` source is the image to upscale. */
+export type StabilityUpscaleOptions = {
+  readonly negative_prompt?: string
+  readonly style_preset?: StabilityStylePreset
+  readonly creativity?: number
+} & Record<string, unknown>
+
+export type Request = ImageRequestFor<StabilityImageOptions>
+export type UpscaleRequest = ImageRequestFor<StabilityUpscaleOptions>
+
+// ---------------------------------------------------------------------------
+// 2. Token and response schemas
+// ---------------------------------------------------------------------------
+
+export const Token = Schema.Struct({ id: Schema.String })
+export type Token = Schema.Schema.Type<typeof Token>
+
+// Generate names the base64 image `image`; `/v2beta/results/{id}` names it `result`.
+const ImageDocument = Schema.Struct({
+  image: Schema.optional(Schema.String),
+  result: Schema.optional(Schema.String),
+  seed: Schema.optional(Schema.Number),
+  finish_reason: Schema.optional(Schema.String),
+})
+
+const Started = Schema.Struct({ id: Schema.String })
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+/** `sd3.5-*` ids share the `sd3` endpoint and select the model with a form field. */
+const endpoint = (model: string) => (model.startsWith("sd3") ? "sd3" : model)
+
+const RESERVED_FORM_FIELDS = new Set(["image", "prompt", "mode", "model"])
+
+const form = Effect.fn("StabilityImages.form")(function* (
+  identity: MediaProtocol.Identity,
+  fields: Record<string, unknown>,
+  native: Record<string, unknown> | undefined,
+  source: Request["images"],
+) {
+  if ((source?.length ?? 0) > 1)
+    return yield* identity.unsupported("media.images", `${identity.name} takes one source image`)
+  const body = new FormData()
+  MediaInput.appendFields(body, fields, { overlay: native, reserved: RESERVED_FORM_FIELDS })
+  const image = source?.[0]
+  if (image !== undefined)
+    body.append("image", MediaInput.blob(yield* MediaInput.inlineBytes(identity.id, image), image.mediaType), "image")
+  return MediaProtocol.multipart(body)
+})
+
+const fromRequest = Effect.fn("StabilityImages.fromRequest")(function* (request: Request) {
+  if (request.n !== undefined && request.n > 1)
+    return yield* route.unsupported("media.n", `${route.name} generates one image per request; call it once per image`)
+  const target = endpoint(request.model.id)
+  const edit = (request.images?.length ?? 0) > 0
+  if (edit && target === "core")
+    return yield* route.unsupported("media.images", `${route.name} core is text-to-image only; use ultra or sd3.5-*`)
+  return yield* form(
+    route,
+    {
+      prompt: request.prompt,
+      aspect_ratio: request.aspectRatio,
+      seed: request.seed,
+      output_format: request.format,
+      model: target === "sd3" ? request.model.id : undefined,
+      mode: target === "sd3" && edit ? "image-to-image" : undefined,
+    },
+    mergeJsonRecords(request.providerOptions, request.http?.body),
+    request.images,
+  )
+})
+
+const fromUpscaleRequest = Effect.fn("StabilityImages.fromUpscaleRequest")(function* (request: UpscaleRequest) {
+  if ((request.images?.length ?? 0) === 0)
+    return yield* ProviderShared.invalidRequest(`${upscaleRoute.name} upscale requires the source image in images`)
+  return yield* form(
+    upscaleRoute,
+    { prompt: request.prompt, seed: request.seed, output_format: request.format },
+    mergeJsonRecords(request.providerOptions, request.http?.body),
+    request.images,
+  )
+})
+
+// ---------------------------------------------------------------------------
+// 6. Response decoding
+// ---------------------------------------------------------------------------
+
+const decodeImageDocument = (identity: MediaProtocol.Identity) => {
+  const decode = identity.decodeJson(ImageDocument)
+  return Effect.fn("StabilityImages.decodeImage")(function* (response: HttpClientResponse.HttpClientResponse) {
+    const output = yield* decode(response)
+    const document = output.value
+    const data = document.image ?? document.result
+    if (data === undefined) return yield* output.invalid(`${identity.name} returned no image`)
+    const image = yield* MediaInput.decodedAsset(output.invalid, `${identity.name} result`, data, undefined)
+    return new ImageResponse({
+      images: [image],
+      notices:
+        document.finish_reason === "CONTENT_FILTERED"
+          ? [{ type: "moderated", message: `${identity.name} blurred the image for violating its content policy` }]
+          : undefined,
+      providerMetadata: { stability: { seed: document.seed, finishReason: document.finish_reason } },
+    })
+  })
+}
+
+const decodeResponse = decodeImageDocument(route)
+const decodeUpscaleImage = decodeImageDocument(upscaleRoute)
+
+const decodeStart = upscaleRoute.decodeStarted(Started, (value) => ({
+  token: { id: value.id },
+  snapshot: { id: value.id, status: "queued" },
+}))
+
+// `/v2beta/results/{id}` answers 202 while in progress and 200 with the finished image document.
+const decodeStatus = (response: HttpClientResponse.HttpClientResponse, context: MediaProtocol.PollContext<Token>) =>
+  Effect.succeed({
+    id: context.token.id,
+    status: response.status === 202 ? ("running" as const) : ("completed" as const),
+  })
+
+const decodeUpscaleResult = Effect.fn("StabilityImages.decodeUpscaleResult")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.PollContext<Token>,
+) {
+  if (response.status === 202) {
+    const output = yield* upscaleRoute.text(response)
+    return yield* output.pending(context.token.id)
+  }
+  return yield* decodeUpscaleImage(response)
+})
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+export const protocol = MediaProtocol.inline<Request, ImageResponse>(route, {
+  unsupported: ["size", "mask"],
+  body: { from: fromRequest },
+  response: { decode: decodeResponse },
+})
+
+export const upscaleProtocol = MediaProtocol.queued<UpscaleRequest, ImageResponse, Token>(upscaleRoute, {
+  token: Token,
+  unsupported: ["n", "size", "aspectRatio", "mask"],
+  start: { body: { from: fromUpscaleRequest }, decode: decodeStart },
+  status: { path: (token) => `${RESULTS_PATH}/${token.id}`, decode: decodeStatus },
+  result: { path: (token) => `${RESULTS_PATH}/${token.id}`, decode: decodeUpscaleResult },
+})
+
+export const model = (input: MediaRoute.ModelInput) =>
+  ImageModel.fromRoute<StabilityImageOptions>(
+    {
+      protocol,
+      baseURL: DEFAULT_BASE_URL,
+      headers: HEADERS,
+      path: ({ request }) => `/v2beta/stable-image/generate/${endpoint(request.model.id)}`,
+    },
+    input,
+  )
+
+export const upscaleModel = (input: Omit<MediaRoute.ModelInput, "id">) =>
+  ImageModel.fromRoute<StabilityUpscaleOptions, Token>(
+    {
+      protocol: upscaleProtocol,
+      baseURL: DEFAULT_BASE_URL,
+      headers: HEADERS,
+      // Only the creative upscaler is asynchronous; fast and conservative answer inline.
+      path: `/v2beta/stable-image/upscale/${UPSCALE_MODEL}`,
+    },
+    { ...input, id: UPSCALE_MODEL },
+  )
+
+export const StabilityImages = {
+  protocol,
+  upscaleProtocol,
+  model,
+  upscaleModel,
+} as const
@@ -0,0 +1,91 @@
+import { Effect, Schema } from "effect"
+import type { HttpClientResponse } from "effect/unstable/http"
+import type { Status } from "../../generation.js"
+import type { Media } from "../../media.js"
+import { MediaProtocol } from "../../route/media-protocol.js"
+import type { AIError } from "../../schema/index.js"
+import { ProviderShared, optionalNull } from "../shared.js"
+
+export const DEFAULT_BASE_URL = "https://queue.fal.run"
+
+/** fal hands back absolute follow-up URLs on submit; they are authoritative for status, result, and cancel. */
+export const Token = Schema.Struct({
+  requestID: Schema.String,
+  statusURL: Schema.String,
+  responseURL: Schema.String,
+  cancelURL: Schema.String,
+})
+export type Token = Schema.Schema.Type<typeof Token>
+
+const StartResponse = Schema.Struct({
+  request_id: Schema.String,
+  status_url: Schema.String,
+  response_url: Schema.String,
+  cancel_url: Schema.String,
+  queue_position: optionalNull(Schema.Number),
+})
+
+const QueueStatus = Schema.Struct({
+  status: Schema.String,
+  queue_position: optionalNull(Schema.Number),
+  error: optionalNull(Schema.Unknown),
+})
+
+const STATUS = {
+  IN_QUEUE: "queued",
+  IN_PROGRESS: "running",
+  COMPLETED: "completed",
+} as const satisfies Record<string, Status>
+
+// fal accepts public URLs and data URIs; there is no provider file handle to forward.
+export const mediaUrl = (asset: Media.Asset, name: string) =>
+  ProviderShared.mediaReference(asset, undefined, name).pipe(Effect.map((reference) => reference.value))
+
+export const protocol = <Request, Response>(
+  route: MediaProtocol.Identity,
+  input: {
+    readonly unsupported?: ReadonlyArray<keyof Request & string>
+    readonly from: (request: Request) => Effect.Effect<MediaProtocol.Body, AIError>
+    readonly decodeResult: (
+      response: HttpClientResponse.HttpClientResponse,
+      context: MediaProtocol.PollContext<Token>,
+    ) => Effect.Effect<Response, AIError>
+  },
+) => {
+  const decodeQueueStatus = route.decodeJson(QueueStatus)
+  return MediaProtocol.queued<Request, Response, Token>(route, {
+    token: Token,
+    unsupported: input.unsupported,
+    start: {
+      body: { from: input.from },
+      decode: route.decodeStarted(StartResponse, (value) => ({
+        token: {
+          requestID: value.request_id,
+          statusURL: value.status_url,
+          responseURL: value.response_url,
+          cancelURL: value.cancel_url,
+        },
+        snapshot: { id: value.request_id, status: "queued", position: value.queue_position ?? undefined },
+      })),
+    },
+    status: {
+      path: (token) => token.statusURL,
+      decode: Effect.fn("FalQueue.decodeStatus")(function* (response, context) {
+        const output = yield* decodeQueueStatus(response)
+        const decoded = output.value
+        const status = yield* MediaProtocol.status(STATUS, decoded.status, output)
+        // fal reports request failures as COMPLETED with an `error`; the response endpoint carries the details.
+        const failed = status === "completed" && decoded.error !== undefined && decoded.error !== null
+        return {
+          id: context.token.requestID,
+          status: failed ? "failed" : status,
+          position: status === "queued" ? (decoded.queue_position ?? undefined) : undefined,
+        }
+      }),
+    },
+    result: { path: (token) => token.responseURL, decode: input.decodeResult },
+    cancel: { method: "PUT", path: (token) => token.cancelURL },
+  })
+}
+
+export * as FalQueue from "./fal-queue.js"
@@ -0,0 +1,113 @@
+import { Effect, Schema, type Stream } from "effect"
+import type { Media } from "../../media.js"
+import { Framing } from "../../route/framing.js"
+import type { MediaProtocol } from "../../route/media-protocol.js"
+import { AIError, ContentPolicyError, ProviderID, type MediaUsage, type ProviderMetadata } from "../../schema/index.js"
+import { ProviderShared } from "../shared.js"
+import { MediaInput } from "./media-input.js"
+
+const PROVIDER = ProviderID.make("google")
+
+const UsageMetadata = Schema.Struct({
+  promptTokenCount: Schema.optional(Schema.Number),
+  candidatesTokenCount: Schema.optional(Schema.Number),
+  totalTokenCount: Schema.optional(Schema.Number),
+})
+type UsageMetadata = Schema.Schema.Type<typeof UsageMetadata>
+
+export const chunk = <const Part extends Schema.Top>(part: Part) =>
+  Schema.Struct({
+    candidates: Schema.optional(
+      Schema.Array(
+        Schema.Struct({
+          content: Schema.optional(Schema.Struct({ parts: Schema.optional(Schema.Array(part)) })),
+          finishReason: Schema.optional(Schema.String),
+        }),
+      ),
+    ),
+    promptFeedback: Schema.optional(
+      Schema.Struct({
+        blockReason: Schema.optional(Schema.String),
+        blockReasonMessage: Schema.optional(Schema.String),
+      }),
+    ),
+    usageMetadata: Schema.optional(UsageMetadata),
+    modelVersion: Schema.optional(Schema.String),
+    responseId: Schema.optional(Schema.String),
+  })
+
+interface Chunk {
+  readonly candidates?: ReadonlyArray<{ readonly finishReason?: string }>
+  readonly promptFeedback?: { readonly blockReason?: string; readonly blockReasonMessage?: string }
+  readonly usageMetadata?: UsageMetadata
+  readonly modelVersion?: string
+  readonly responseId?: string
+}
+
+export interface Metadata {
+  readonly usage?: UsageMetadata
+  readonly finishReason?: string
+  readonly modelVersion?: string
+  readonly responseId?: string
+}
+
+export const track = <State extends Metadata>(state: State, chunk: Chunk): State => ({
+  ...state,
+  usage: chunk.usageMetadata ?? state.usage,
+  finishReason: chunk.candidates?.[0]?.finishReason ?? state.finishReason,
+  modelVersion: chunk.modelVersion ?? state.modelVersion,
+  responseId: chunk.responseId ?? state.responseId,
+})
+
+export const blocked = (name: string, chunk: Chunk, frame: string) => {
+  const feedback = chunk.promptFeedback
+  if (feedback?.blockReason === undefined) return undefined
+  return new AIError({
+    reason: new ContentPolicyError({
+      message: `${name} blocked the request (${feedback.blockReason})${
+        feedback.blockReasonMessage === undefined ? "" : `: ${feedback.blockReasonMessage}`
+      }`,
+      body: frame,
+    }),
+  })
+}
+
+export const usage = (usage: UsageMetadata | undefined): MediaUsage | undefined =>
+  usage === undefined
+    ? undefined
+    : {
+        type: "tokens",
+        input: usage.promptTokenCount,
+        output: usage.candidatesTokenCount,
+        total: ProviderShared.totalTokens(usage.promptTokenCount, usage.candidatesTokenCount, usage.totalTokenCount),
+        details: { google: usage },
+      }
+
+export const providerMetadata = (state: Metadata): ProviderMetadata => ({
+  google: { finishReason: state.finishReason, modelVersion: state.modelVersion, responseId: state.responseId },
+})
+
+export const path = (model: string, mode: MediaProtocol.Mode) =>
+  mode === "stream" ? `/models/${model}:streamGenerateContent?alt=sse` : `/models/${model}:generateContent`
+
+// `generateContent` answers with one document shaped exactly like a streamed chunk, so it is a single frame.
+export const frames = (bytes: Stream.Stream<Uint8Array, AIError>, mode: MediaProtocol.Mode) =>
+  mode === "stream" ? Framing.sse.frame(bytes) : Framing.document.frame(bytes)
+
+// Gemini does not fetch public URLs; inline payloads and Gemini Files references are the accepted inputs.
+export const mediaPart = (
+  route: string,
+  asset: Media.Asset,
+): Effect.Effect<
+  | { readonly fileData: { readonly mimeType: string; readonly fileUri: string } }
+  | { readonly inlineData: { readonly mimeType: string; readonly data: string } },
+  AIError
+> => {
+  const fileUri = MediaInput.refID(asset, PROVIDER)
+  if (fileUri !== undefined) return Effect.succeed({ fileData: { mimeType: asset.mediaType, fileUri } })
+  return ProviderShared.requireInlineMedia(route, asset).pipe(
+    Effect.map((media) => ({ inlineData: { mimeType: media.mime, data: media.base64 } })),
+  )
+}
+
+export * as GeminiGenerateContent from "./gemini-generate-content.js"
@@ -0,0 +1,77 @@
+import type { JsonSchema } from "../../schema/index.js"
+import { isRecord } from "../../utils/record.js"
+
+// Gemini's `parametersJsonSchema` accepts standard JSON Schema, but rejects a few shapes that
+// published tool schemas commonly contain. Rewrite only those and send everything else unchanged.
+const SCHEMA_MAPS = new Set([
+  "properties",
+  "patternProperties",
+  "$defs",
+  "definitions",
+  "dependentSchemas",
+  "dependencies",
+])
+const VALUES = new Set(["const", "default", "enum", "examples", "dependentRequired"])
+
+const mapValues = (record: Record<string, unknown>, map: (value: unknown, key: string) => unknown) =>
+  Object.fromEntries(Object.entries(record).map(([key, value]) => [key, map(value, key)]))
+
+const normalizeNode = (schema: unknown): unknown => {
+  if (Array.isArray(schema)) return schema.map(normalizeNode)
+  if (!isRecord(schema)) return schema
+  const properties = isRecord(schema.properties) ? schema.properties : undefined
+  return Object.fromEntries(
+    Object.entries(schema).flatMap(([key, value]) => {
+      if (VALUES.has(key)) return [[key, value]]
+      if (SCHEMA_MAPS.has(key) && isRecord(value)) return [[key, mapValues(value, normalizeNode)]]
+      // `required` may only name declared properties.
+      if (key === "required" && properties && Array.isArray(value))
+        return [[key, value.filter((name) => typeof name === "string" && Object.hasOwn(properties, name))]]
+      // Draft-04 boolean exclusive bounds become the numeric form.
+      if (key === "exclusiveMinimum" && typeof value === "boolean")
+        return value && typeof schema.minimum === "number" ? [[key, schema.minimum]] : []
+      if (key === "exclusiveMaximum" && typeof value === "boolean")
+        return value && typeof schema.maximum === "number" ? [[key, schema.maximum]] : []
+      if (key === "minimum" && schema.exclusiveMinimum === true) return []
+      if (key === "maximum" && schema.exclusiveMaximum === true) return []
+      // Draft-07 tuples (`items` array plus `additionalItems`) are `prefixItems` plus `items` in 2020-12.
+      if (key === "items" && Array.isArray(value)) return [["prefixItems", value.map(normalizeNode)]]
+      if (key === "additionalItems" && Array.isArray(schema.items)) return [["items", normalizeNode(value)]]
+      return [[key, normalizeNode(value)]]
+    }),
+  )
+}
+
+// Gemini accepts a recursive `$ref` only when the loop passes through an optional property or
+// potentially empty array `items`. Replace other self-references with an unconstrained schema.
+const cutLoops = (schema: unknown, target: string, safe: boolean): unknown => {
+  if (Array.isArray(schema)) return schema.map((item) => cutLoops(item, target, safe))
+  if (!isRecord(schema)) return schema
+  if (schema.$ref === target && !safe) return {}
+  const required = Array.isArray(schema.required) ? schema.required : []
+  return mapValues(schema, (value, key) => {
+    if (VALUES.has(key)) return value
+    if (key === "items") return cutLoops(value, target, safe || !(Number(schema.minItems) > 0))
+    if (key === "properties" && isRecord(value))
+      return mapValues(value, (child, name) => cutLoops(child, target, safe || !required.includes(name)))
+    if (SCHEMA_MAPS.has(key) && isRecord(value)) return mapValues(value, (child) => cutLoops(child, target, safe))
+    return cutLoops(value, target, safe)
+  })
+}
+
+export const normalize = (schema: JsonSchema): JsonSchema => {
+  const normalized = normalizeNode(schema)
+  if (!isRecord(normalized)) return {}
+  const result = cutLoops(
+    mapValues(normalized, (value, key) =>
+      (key === "$defs" || key === "definitions") && isRecord(value)
+        ? mapValues(value, (def, name) => cutLoops(def, `#/${key}/${name}`, false))
+        : value,
+    ),
+    "#",
+    false,
+  )
+  return isRecord(result) ? result : {}
+}
+
+export * as GeminiJsonSchema from "./gemini-json-schema.js"
@@ -1,119 +0,0 @@
-import { isRecord } from "../../utils/record.js"
-
-// Gemini accepts a JSON Schema-like dialect for tool parameters, but rejects a
-// handful of common JSON Schema shapes. Keep this projection isolated so the
-// Gemini protocol file still reads like the other protocol modules.
-const SCHEMA_INTENT_KEYS = [
-  "type",
-  "properties",
-  "items",
-  "prefixItems",
-  "enum",
-  "const",
-  "$ref",
-  "additionalProperties",
-  "patternProperties",
-  "required",
-  "not",
-  "if",
-  "then",
-  "else",
-]
-
-const hasCombiner = (schema: unknown) =>
-  isRecord(schema) && (Array.isArray(schema.anyOf) || Array.isArray(schema.oneOf) || Array.isArray(schema.allOf))
-
-const hasSchemaIntent = (schema: unknown) =>
-  isRecord(schema) && (hasCombiner(schema) || SCHEMA_INTENT_KEYS.some((key) => key in schema))
-
-const sanitizeNode = (schema: unknown): unknown => {
-  if (!isRecord(schema)) return Array.isArray(schema) ? schema.map(sanitizeNode) : schema
-
-  const result: Record<string, unknown> = Object.fromEntries(
-    Object.entries(schema).map(([key, value]) => [
-      key,
-      key === "enum" && Array.isArray(value) ? value.map(String) : sanitizeNode(value),
-    ]),
-  )
-
-  if (Array.isArray(result.enum) && (result.type === "integer" || result.type === "number")) result.type = "string"
-
-  const properties = result.properties
-  if (result.type === "object" && isRecord(properties) && Array.isArray(result.required)) {
-    result.required = result.required.filter((field) => typeof field === "string" && field in properties)
-  }
-
-  if (result.type === "array" && !hasCombiner(result)) {
-    result.items = result.items ?? {}
-    if (isRecord(result.items) && !hasSchemaIntent(result.items)) result.items = { ...result.items, type: "string" }
-  }
-
-  if (typeof result.type === "string" && result.type !== "object" && !hasCombiner(result)) {
-    delete result.properties
-    delete result.required
-  }
-
-  return result
-}
-
-const emptyObjectSchema = (schema: Record<string, unknown>) =>
-  schema.type === "object" &&
-  (!isRecord(schema.properties) || Object.keys(schema.properties).length === 0) &&
-  !schema.additionalProperties
-
-const projectNode = (schema: unknown, nested = false): Record<string, unknown> | undefined => {
-  if (!isRecord(schema)) return undefined
-  if (!nested && emptyObjectSchema(schema)) return undefined
-  const types = Array.isArray(schema.type) ? schema.type.filter((type) => type !== "null") : undefined
-  const anyOf = Array.isArray(schema.anyOf) ? schema.anyOf : undefined
-  const hasNullAnyOf = anyOf?.some((item) => isRecord(item) && item.type === "null") ?? false
-  const anyOfTypes = hasNullAnyOf ? anyOf?.filter((item) => !isRecord(item) || item.type !== "null") : anyOf
-  const flattenedAnyOf = hasNullAnyOf && anyOfTypes?.length === 1 ? projectNode(anyOfTypes[0], true) : undefined
-  const result = Object.fromEntries(
-    [
-      ["description", schema.description],
-      ["required", schema.required],
-      ["format", schema.format],
-      ["type", types ? (types.length === 0 ? "null" : undefined) : schema.type],
-      [
-        "nullable",
-        (Array.isArray(schema.type) && schema.type.includes("null") && types && types.length > 0) || hasNullAnyOf
-          ? true
-          : undefined,
-      ],
-      ["enum", schema.const !== undefined ? [schema.const] : schema.enum],
-      [
-        "properties",
-        isRecord(schema.properties)
-          ? Object.fromEntries(Object.entries(schema.properties).map(([key, value]) => [key, projectNode(value, true)]))
-          : undefined,
-      ],
-      [
-        "items",
-        Array.isArray(schema.items)
-          ? schema.items.map((item) => projectNode(item, true))
-          : schema.items === undefined
-            ? undefined
-            : projectNode(schema.items, true),
-      ],
-      ["allOf", Array.isArray(schema.allOf) ? schema.allOf.map((item) => projectNode(item, true)) : undefined],
-      [
-        "anyOf",
-        anyOfTypes
-          ? hasNullAnyOf && anyOfTypes.length === 1
-            ? undefined
-            : anyOfTypes.map((item) => projectNode(item, true))
-          : types && types.length > 0
-            ? types.map((type) => ({ type }))
-            : undefined,
-      ],
-      ["oneOf", Array.isArray(schema.oneOf) ? schema.oneOf.map((item) => projectNode(item, true)) : undefined],
-      ["minLength", schema.minLength],
-    ].filter((entry) => entry[1] !== undefined),
-  )
-  return flattenedAnyOf ? { ...result, ...flattenedAnyOf } : result
-}
-
-export const convert = (schema: unknown) => projectNode(sanitizeNode(schema))
-
-export * as GeminiToolSchema from "./gemini-tool-schema.js"
@@ -1,6 +1,8 @@
 import { Effect, Encoding } from "effect"
 import { Media } from "../../media.js"
-import type { AIError, ProviderID } from "../../schema/index.js"
+import type { MediaProtocol } from "../../route/media-protocol.js"
+import { mergeJsonRecords, type AIError, type ProviderID } from "../../schema/index.js"
+import { encodeJson } from "../../utils/json.js"
 import { ProviderShared } from "../shared.js"

 /** Owned bytes for multipart uploads; decodes `base64` sources and rejects remote sources. */
@@ -13,6 +15,31 @@ export const inlineBytes = (route: string, asset: Media.Asset): Effect.Effect<Ui
  )
 }

+/** Copied because `BlobPart` requires a plain `ArrayBuffer`. */
+export const blob = (data: Uint8Array, mediaType: string) => {
+  const buffer = new ArrayBuffer(data.byteLength)
+  new Uint8Array(buffer).set(data)
+  return new Blob([buffer], { type: mediaType })
+}
+
+const isScalar = (value: unknown): value is string | number | boolean =>
+  typeof value === "string" || typeof value === "number" || typeof value === "boolean"
+
+export const query = (route: string, values: Record<string, unknown>): Effect.Effect<MediaProtocol.Query, AIError> => {
+  const entries = Object.entries(values).filter(([, value]) => value !== undefined)
+  const invalid = entries.find(([, value]) => !isScalar(value) && !(Array.isArray(value) && value.every(isScalar)))
+  if (invalid !== undefined)
+    return Effect.fail(ProviderShared.invalidRequest(`${route} cannot send "${invalid[0]}" as a query parameter`))
+  return Effect.succeed(
+    Object.fromEntries(entries.map(([key, value]) => [key, Array.isArray(value) ? value.map(String) : String(value)])),
+  )
+}
+
+export const dimensions = (size: string) => {
+  const [width, height] = size.split("x").map(Number)
+  return { width, height }
+}
+
 /** Provider file handle when the ref belongs to this provider; refs from other providers are never forwarded. */
 export const refID = (asset: Media.Asset, provider: ProviderID) =>
  asset.source.type === "ref" && asset.source.provider === provider ? asset.source.id : undefined
@@ -30,4 +57,38 @@ export const decodedAsset = (
    Effect.map((bytes) => Media.bytes(bytes, mediaType, options)),
  )

+/** One image of an OpenAI-shaped `data` array, which carries either `b64_json` or a `url`. */
+export const imageOutput = (
+  invalid: (message: string, cause?: unknown) => AIError,
+  label: string,
+  item: { readonly b64_json?: string | null; readonly url?: string | null },
+  mediaType: string | undefined,
+  options?: Media.AssetOptions,
+) => {
+  if (item.b64_json) return decodedAsset(invalid, label, item.b64_json, mediaType, options)
+  if (item.url) return Effect.succeed(Media.url(item.url, { ...options, mediaType }))
+  return Effect.fail(invalid(`${label} has neither image data nor a URL`))
+}
+
+/**
+ * Append multipart text fields: strings as-is, other values as JSON, or arrays as repeated `key[]` parts with
+ * `repeatArrays`. `overlay` keys in `reserved` are dropped so `http.body` cannot replace route-owned fields.
+ */
+export const appendFields = (
+  form: FormData,
+  fields: Record<string, unknown>,
+  options: {
+    readonly overlay?: Record<string, unknown>
+    readonly reserved: ReadonlySet<string>
+    readonly repeatArrays?: true
+  },
+) => {
+  const overlay = Object.entries(options.overlay ?? {}).filter(([key]) => !options.reserved.has(key))
+  Object.entries(mergeJsonRecords(fields, Object.fromEntries(overlay)) ?? {}).forEach(([key, value]) => {
+    if (Array.isArray(value) && options.repeatArrays)
+      return value.forEach((item) => form.append(`${key}[]`, String(item)))
+    form.append(key, typeof value === "string" ? value : encodeJson(value))
+  })
+}
+
 export * as MediaInput from "./media-input.js"
@@ -1,11 +0,0 @@
-// Responses image items can omit output_format, including when PNG/JPEG was requested.
-export const mediaType = (data: Uint8Array, format?: string) => {
-  if (format !== undefined) return `image/${format}`
-  if (data[0] === 137 && data[1] === 80 && data[2] === 78 && data[3] === 71) return "image/png"
-  if (data[0] === 255 && data[1] === 216 && data[2] === 255) return "image/jpeg"
-  if (new TextDecoder().decode(data.slice(0, 4)) === "RIFF" && new TextDecoder().decode(data.slice(8, 12)) === "WEBP")
-    return "image/webp"
-  return "application/octet-stream"
-}
-
-export * as MetaImage from "./meta-image.js"
@@ -1,20 +0,0 @@
-import { Schema } from "effect"
-
-const dimensions = (value: string) => {
-  const match = /^(\d+)x(\d+)$/.exec(value)
-  if (!match) return undefined
-  return { width: Number(match[1]), height: Number(match[2]) }
-}
-
-export const Size = Schema.String.check(
-  Schema.makeFilter((value) => {
-    if (value === "auto") return undefined
-    const parsed = dimensions(value)
-    if (!parsed) return "image size must be `auto` or `{width}x{height}`"
-    return parsed.width > 0 && parsed.height > 0 ? undefined : "image dimensions must be positive integers"
-  }),
-)
-
-export const OpenAIImage = {
-  Size,
-} as const
@@ -1,7 +1,7 @@
 import { Effect, Schema, Stream } from "effect"
 import { Route, type RouteBody, type TriggerCompactOperation } from "../../route/client.js"
 import { Protocol } from "../../route/protocol.js"
-import { CompactionCheckpointResponse, HttpOptions, LLMEvent, LLMRequest } from "../../schema/index.js"
+import { CompactionCheckpointResponse, LLMEvent, LLMRequest } from "../../schema/index.js"
 import { OpenResponses } from "../open-responses.js"
 import { ProviderShared } from "../shared.js"

@@ -109,12 +109,8 @@ export const make = <Body>(body: RouteBody<Body>): TriggerCompactOperation =>
      transport: source.transport,
    })
    const native = yield* body.from(request)
-    // The body builder already applied and validated overlays. Do not let transport reapply them.
-    const preparedRequest = LLMRequest.update(request, {
-      http: request.http === undefined ? undefined : new HttpOptions({ ...request.http, body: undefined }),
-    })
-    const prepared = yield* route.prepareTransport(native, preparedRequest, options)
-    yield* route.streamPrepared(prepared, preparedRequest, { http: executor }, options).pipe(Stream.runDrain)
+    const prepared = yield* route.prepareTransport(native, request, options)
+    yield* route.streamPrepared(prepared, request, { http: executor }, options).pipe(Stream.runDrain)
    if (!result) return yield* ProviderShared.eventError(source.id, "Compaction response ended without a checkpoint")
    return result
  })
@@ -19,12 +19,18 @@ import { OpenResponses } from "../open-responses.js"
 import { JsonObject, optionalNull, ProviderShared } from "../shared.js"
 import { Media } from "../../media.js"

+// /compact has a smaller wire contract than /responses; keep the request controls it accepts.
 const Body = Schema.Struct({
  model: Schema.String,
  input: Schema.Array(Schema.Unknown),
  instructions: optionalNull(Schema.String),
  previous_response_id: optionalNull(Schema.String),
  service_tier: optionalNull(Schema.String),
+  reasoning: Schema.optional(JsonObject),
+  text: Schema.optional(JsonObject),
+  include: OpenResponses.coreFields.include,
+  parallel_tool_calls: OpenResponses.coreFields.parallel_tool_calls,
+  tools: Schema.optional(Schema.Array(JsonObject)),
  prompt_cache_key: optionalNull(Schema.String),
  prompt_cache_retention: optionalNull(Schema.String),
  prompt_cache_options: optionalNull(
@@ -74,17 +80,27 @@ const Response = Schema.Struct({
  usage: Schema.optional(Schema.StructWithRest(OpenResponses.OpenResponsesUsage, [JsonObject])),
 })

-export const make = (adapter: OpenResponses.ProviderAdapter): CompactOperation =>
+export const make = (
+  adapter: OpenResponses.ProviderAdapter,
+  lowerTools: (request: LLMRequest) => Effect.Effect<ReadonlyArray<Record<string, unknown>>, AIError>,
+): CompactOperation =>
  Effect.fn("ResponsesCompaction.execute")(function* (request, executor, options) {
    const route = request.model.route
    // The standalone compaction endpoint rejects histories containing configuration updates.
    const native = yield* OpenResponses.lowerConversation(stripEffortUpdates(request), adapter)
+    const generation = OpenResponses.lowerGeneration(request)
+    const tools = request.tools.length === 0 ? undefined : yield* lowerTools(request)
    const body = yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(Body))(
      mergeJsonRecords(
        {
          ...native,
-          service_tier: request.providerOptions?.serviceTier,
-          prompt_cache_key: ProviderShared.promptCacheKey(request),
+          service_tier: generation.service_tier,
+          reasoning: generation.reasoning,
+          text: generation.text,
+          include: generation.include,
+          parallel_tool_calls: generation.parallel_tool_calls,
+          tools,
+          prompt_cache_key: generation.prompt_cache_key,
        },
        request.http?.body,
      ),
@@ -0,0 +1,111 @@
+import { Effect } from "effect"
+import { Media } from "../../media.js"
+import type { MediaProtocol } from "../../route/media-protocol.js"
+import type { AIError, MediaUsage, ProviderMetadata } from "../../schema/index.js"
+import {
+  SpeechAudioDeltaEvent,
+  SpeechFinishEvent,
+  SpeechTimestampsEvent,
+  type SpeechEvent,
+  type SpeechVoice,
+} from "../../speech.js"
+import { concatBytes } from "../../utils/bytes.js"
+
+export interface Audio {
+  /** Appended in place: the route creates fresh state for each response through `initial`. */
+  readonly chunks: Array<Uint8Array>
+}
+
+export type StepResult<State> = readonly [State, ReadonlyArray<SpeechEvent>]
+
+/** Empty chunks (keep-alive records) emit nothing. */
+export const delta = <State extends Audio>(state: State, chunk: Uint8Array): StepResult<State> => {
+  if (chunk.length === 0) return [state, []]
+  state.chunks.push(chunk)
+  return [state, [SpeechAudioDeltaEvent.make({ chunk })]]
+}
+
+export const step =
+  <State extends Audio>(onRecord: (state: State, frame: string) => Effect.Effect<StepResult<State>, AIError>) =>
+  (state: State, frame: string | Uint8Array) =>
+    typeof frame === "string" ? onRecord(state, frame) : Effect.succeed(delta(state, frame))
+
+export const timestamps = (
+  texts: ReadonlyArray<string>,
+  starts: ReadonlyArray<number>,
+  ends: ReadonlyArray<number>,
+): ReadonlyArray<SpeechEvent> =>
+  texts.length === 0
+    ? []
+    : [
+        SpeechTimestampsEvent.make({
+          items: texts.map((text, index) => ({ text, startSeconds: starts[index] ?? 0, endSeconds: ends[index] ?? 0 })),
+        }),
+      ]
+
+export const voiceID = (voice: SpeechVoice | undefined) => (typeof voice === "object" ? voice.id : voice)
+
+const CONTAINER_MEDIA_TYPES: Readonly<Record<string, string>> = {
+  mp3: "audio/mpeg",
+  wav: "audio/wav",
+  opus: "audio/ogg",
+  aac: "audio/aac",
+  flac: "audio/flac",
+}
+
+export const container = (format: string, sampleRate?: number) => ({
+  mediaType: CONTAINER_MEDIA_TYPES[format],
+  info: { format, sampleRate },
+})
+
+const PCM_MEDIA_TYPES = {
+  pcm_s16le: "audio/pcm",
+  pcm_f32le: "audio/pcm",
+  pcm_mulaw: "audio/mulaw",
+  pcm_alaw: "audio/alaw",
+} as const
+
+export type PcmEncoding = keyof typeof PCM_MEDIA_TYPES
+
+export const pcm = (encoding: PcmEncoding, sampleRate: number | undefined, mediaType?: string) => ({
+  mediaType: mediaType ?? PCM_MEDIA_TYPES[encoding],
+  info: { format: "pcm", encoding, sampleRate, channels: 1 },
+})
+
+export const sampleRate = (mediaType: string | undefined) => {
+  const rate = /rate=(\d+)/i.exec(mediaType ?? "")?.[1]
+  return rate === undefined ? undefined : Number(rate)
+}
+
+/** A declared `mediaType` wins over sniffing: headerless PCM can start with bytes that look like an MPEG frame sync. */
+export const finish = (
+  route: MediaProtocol.Identity,
+  state: Audio,
+  output: {
+    readonly mediaType: string | undefined
+    readonly info?: Media.Info
+    readonly usage?: MediaUsage
+    readonly providerMetadata?: ProviderMetadata
+    readonly detail?: string
+  },
+): Effect.Effect<ReadonlyArray<SpeechEvent>, AIError> => {
+  if (state.chunks.length === 0)
+    return Effect.fail(
+      route.frameError(`The provider returned no audio${output.detail === undefined ? "" : ` (${output.detail})`}`),
+    )
+  return Effect.succeed([
+    SpeechFinishEvent.make({
+      audio: Media.bytes(concatBytes(state.chunks), output.mediaType, { info: output.info }),
+      usage: output.usage,
+      providerMetadata: output.providerMetadata,
+    }),
+  ])
+}
+
+export const headerUsage = (type: "characters" | "credits", value: string | undefined): MediaUsage | undefined => {
+  const amount = Number(value)
+  if (!Number.isFinite(amount)) return undefined
+  return type === "credits" ? { type, credits: amount } : { type, characters: amount }
+}
+
+export * as SpeechStream from "./speech-stream.js"
@@ -1,6 +1,6 @@
-import type { JsonSchema, LanguageModelToolSchemaCompatibility } from "../../schema/index.js"
+import { ToolDefinition, type JsonSchema, type LanguageModel, type LLMRequest } from "../../schema/index.js"
 import { isRecord } from "../../utils/record.js"
-import { GeminiToolSchema } from "./gemini-tool-schema.js"
+import { GeminiJsonSchema } from "./gemini-json-schema.js"

 const tupleItemsSchema = (items: ReadonlyArray<unknown>) => {
  const projected = items.map(moonshotNode)
@@ -9,12 +9,24 @@ const tupleItemsSchema = (items: ReadonlyArray<unknown>) => {
  return { anyOf: projected }
 }

+// Moonshot rejects an `enum` without a sibling `type`. Its `type` may be one type, or one type plus "null".
+const enumType = (values: ReadonlyArray<unknown>) => {
+  const types = [
+    ...new Set(values.map((value) => (value === null ? "null" : Array.isArray(value) ? "array" : typeof value))),
+  ]
+  if (types.length === 1) return types[0]
+  if (types.length === 2 && types.includes("null")) return [...types.filter((type) => type !== "null"), "null"]
+  return undefined
+}
+
 const moonshotNode = (schema: unknown): unknown => {
  if (Array.isArray(schema)) return schema.map(moonshotNode)
  if (!isRecord(schema)) return schema
  if (typeof schema.$ref === "string") return { $ref: schema.$ref }
-  return Object.fromEntries(
-    Object.entries(schema).flatMap(([key, value]) => {
+  const type = schema.type === undefined && Array.isArray(schema.enum) ? enumType(schema.enum) : undefined
+  return Object.fromEntries([
+    ...(type === undefined ? [] : [["type", type]]),
+    ...Object.entries(schema).flatMap(([key, value]) => {
      if (key === "items" && Array.isArray(value)) return [[key, tupleItemsSchema(value)]]
      if (key === "prefixItems") {
        if ("items" in schema) return []
@@ -23,7 +35,7 @@ const moonshotNode = (schema: unknown): unknown => {
      if (key === "unevaluatedItems") return []
      return [[key, moonshotNode(value)]]
    }),
-  )
+  ])
 }

 const moonshot = (schema: JsonSchema): JsonSchema => {
@@ -34,25 +46,59 @@ const moonshot = (schema: JsonSchema): JsonSchema => {
 const openAI = (schema: JsonSchema): JsonSchema => schema
 const responses = openAI

-const gemini = (schema: JsonSchema): JsonSchema => GeminiToolSchema.convert(schema) ?? {}
+const gemini = GeminiJsonSchema.normalize

-const modelCompatibility = (
-  schema: JsonSchema,
-  compatibility: LanguageModelToolSchemaCompatibility | undefined,
-): JsonSchema => {
-  if (compatibility === undefined) return schema
-  switch (compatibility) {
+const MODEL_NAMES = [
+  [/gemini/i, "gemini"],
+  [/kimi/i, "moonshot"],
+] as const
+
+// Tool arguments are always a JSON object, and most providers reject a tool schema whose root does not
+// declare `type: "object"`, such as `{}` or a bare `properties` map. Effect encodes an empty struct as
+// `anyOf` object or array; every object matches its bare object branch, so that `anyOf` is dropped.
+const objectRoot = (schema: JsonSchema): JsonSchema => {
+  if (schema.type !== undefined) return schema
+  if (
+    Array.isArray(schema.anyOf) &&
+    schema.anyOf.some((branch) => isRecord(branch) && branch.type === "object" && Object.keys(branch).length === 1)
+  )
+    return { type: "object", ...Object.fromEntries(Object.entries(schema).filter(([key]) => key !== "anyOf")) }
+  return { type: "object", ...schema }
+}
+
+// Every tool schema gets an object root. Then an explicit `sanitizer` wins, and `none` opts out.
+// Otherwise the protocol's own default applies (the Gemini API always uses Gemini's rules), then the
+// model name selects the family's rules so models reached through gateways and OpenAI-compatible
+// endpoints get the same handling.
+const modelCompatibility = (schema: JsonSchema, model: LanguageModel): JsonSchema => {
+  const root = objectRoot(schema)
+  switch (
+    model.compatibility?.sanitizer ??
+    model.route.sanitizer ??
+    MODEL_NAMES.find(([name]) => name.test(model.id))?.[1]
+  ) {
    case "gemini":
-      return gemini(schema)
+      return gemini(root)
    case "moonshot":
-      return moonshot(schema)
+      return moonshot(root)
+    case "none":
+    case undefined:
+      return root
  }
 }

+// Applied once to every request before any protocol builds its body, including tools in namespaces.
+const tools = (entries: LLMRequest["tools"], model: LanguageModel): LLMRequest["tools"] =>
+  entries.map((tool) =>
+    tool.type === "tool"
+      ? new ToolDefinition({ ...tool, inputSchema: modelCompatibility(tool.inputSchema, model) })
+      : { ...tool, tools: tools(tool.tools, model) },
+  )
+
 export const ToolSchemaProjection = {
  gemini,
-  modelCompatibility,
  moonshot,
  openAI,
  responses,
+  tools,
 } as const
@@ -4,13 +4,11 @@ import { ImageModel, ImageResponse, type ImageRequestFor } from "../image.js"
 import { Media } from "../media.js"
 import { MediaProtocol } from "../route/media-protocol.js"
 import { MediaRoute } from "../route/media.js"
-import { ProviderID, mergeJsonRecords, type AIError } from "../schema/index.js"
+import { mergeJsonRecords, type OpenString } from "../schema/index.js"
 import { ProviderShared, optionalNull } from "./shared.js"
 import { MediaInput } from "./utils/media-input.js"

-const ADAPTER = "xai-images"
-const NAME = "xAI Images"
-const PROVIDER = ProviderID.make("xai")
+const route = MediaProtocol.identity({ id: "xai-images", name: "xAI Images", provider: "xai" })
 export const DEFAULT_BASE_URL = "https://api.x.ai/v1"
 export const PATH = "/images/generations"
 export const EDIT_PATH = "/images/edits"
@@ -19,13 +17,11 @@ export const EDIT_PATH = "/images/edits"
 // 1. Public model input
 // ---------------------------------------------------------------------------

-export type XAIImageString<Known extends string> = Known | (string & {})
-
 /** Provider-native options. Common fields (`n`, `aspectRatio`, `images`) live on the request. */
 export type XAIImageOptions = {
-  readonly resolution?: XAIImageString<"1k" | "2k">
-  readonly responseFormat?: XAIImageString<"url" | "b64_json">
-  readonly response_format?: XAIImageString<"url" | "b64_json">
+  readonly resolution?: OpenString<"1k" | "2k">
+  readonly responseFormat?: OpenString<"url" | "b64_json">
+  readonly response_format?: OpenString<"url" | "b64_json">
 } & Record<string, unknown>

 export type Request = ImageRequestFor<XAIImageOptions>
@@ -58,15 +54,12 @@ const nativeOptions = (options: XAIImageOptions | undefined) => {

 const isEdit = (request: Request) => (request.images?.length ?? 0) > 0

-const reference = (asset: Media.Asset): Effect.Effect<Record<string, unknown>, AIError> => {
-  const inline = asset.inline()
-  if (inline) return Effect.succeed({ url: inline.dataUrl, type: "image_url" as const })
-  const url = ProviderShared.mediaUrl(asset)
-  if (url) return Effect.succeed({ url, type: "image_url" as const })
-  const id = MediaInput.refID(asset, PROVIDER)
-  if (id) return Effect.succeed({ file_id: id })
-  return Effect.fail(ProviderShared.invalidRequest(`${NAME} accepts image URLs, data URLs, bytes, and xAI file IDs`))
-}
+const reference = (asset: Media.Asset) =>
+  ProviderShared.mediaReference(asset, route.provider, route.name).pipe(
+    Effect.map((item) =>
+      item.type === "ref" ? { file_id: item.value } : { url: item.value, type: "image_url" as const },
+    ),
+  )

 const fromRequest = Effect.fn("XAIImages.fromRequest")(function* (request: Request) {
  const references = yield* Effect.forEach(request.images ?? [], reference)
@@ -91,33 +84,25 @@ const fromRequest = Effect.fn("XAIImages.fromRequest")(function* (request: Reque
 // 6. Response decoding
 // ---------------------------------------------------------------------------

+const decodeDocument = route.decodeJson(XAIImageResponse)
+
 const decodeResponse = Effect.fn("XAIImages.decodeResponse")(function* (
  response: HttpClientResponse.HttpClientResponse,
 ) {
-  const output = yield* MediaProtocol.decodeJson(ADAPTER, NAME, XAIImageResponse)(response)
+  const output = yield* decodeDocument(response)
  const decoded = output.value
-  const images = yield* Effect.forEach(decoded.data, (item, index) => {
-    const providerMetadata =
-      item.revised_prompt === undefined || item.revised_prompt === null
-        ? undefined
-        : { xai: { revisedPrompt: item.revised_prompt } }
-    if (item.b64_json)
-      return MediaInput.decodedAsset(
-        output.invalid,
-        `${NAME} result ${index}`,
-        item.b64_json,
-        item.mime_type ?? undefined,
-        {
-          providerMetadata,
-        },
-      )
-    if (item.url)
-      return Effect.succeed(Media.url(item.url, { mediaType: item.mime_type ?? undefined, providerMetadata }))
-    return Effect.fail(output.invalid(`${NAME} result ${index} has neither image data nor a URL`))
-  })
-  if (images.length === 0) return yield* output.invalid(`${NAME} returned no images`)
+  const images = yield* Effect.forEach(decoded.data, (item, index) =>
+    MediaInput.imageOutput(output.invalid, `${route.name} result ${index}`, item, item.mime_type ?? undefined, {
+      providerMetadata:
+        item.revised_prompt === undefined || item.revised_prompt === null
+          ? undefined
+          : { xai: { revisedPrompt: item.revised_prompt } },
+    }),
+  )
+  if (images.length === 0) return yield* output.invalid(`${route.name} returned no images`)
  const usage = ProviderShared.isRecord(decoded.usage) ? decoded.usage : undefined
-  // xAI reports image counts rather than tokens, seconds, or credits; the raw record stays in provider metadata.
+  // xAI reports a USD cost (`cost_in_usd_ticks`) rather than tokens, seconds, or credits; the raw record stays in
+  // provider metadata.
  return new ImageResponse({
    images,
    providerMetadata: usage === undefined ? undefined : { xai: { usage } },
@@ -128,9 +113,7 @@ const decodeResponse = Effect.fn("XAIImages.decodeResponse")(function* (
 // 7. Protocol and route
 // ---------------------------------------------------------------------------

-export const protocol = MediaProtocol.inline<Request, ImageResponse>({
-  id: ADAPTER,
-  name: NAME,
+export const protocol = MediaProtocol.inline<Request, ImageResponse>(route, {
  unsupported: ["mask", "size", "seed", "format"],
  body: { from: fromRequest },
  response: { decode: decodeResponse },
@@ -138,13 +121,7 @@ export const protocol = MediaProtocol.inline<Request, ImageResponse>({

 export const model = (input: MediaRoute.ModelInput) =>
  ImageModel.fromRoute<XAIImageOptions>(
-    {
-      id: ADAPTER,
-      provider: PROVIDER,
-      protocol,
-      baseURL: DEFAULT_BASE_URL,
-      path: ({ request }) => (isEdit(request) ? EDIT_PATH : PATH),
-    },
+    { protocol, baseURL: DEFAULT_BASE_URL, path: ({ request }) => (isEdit(request) ? EDIT_PATH : PATH) },
    input,
  )

@@ -50,7 +50,8 @@ const fromRequest = Effect.fn("XAIResponses.fromRequest")(function* (request: LL
      operation: "in-band-compaction",
      provider: request.model.provider,
      route: request.model.route.id,
-      message: "xAI requires explicit compaction through LLMClient.compact; automatic context management is not supported",
+      message:
+        "xAI requires explicit compaction through LLMClient.compact; automatic context management is not supported",
    })
  return yield* decodeBody(yield* OpenResponses.fromRequestWithAdapter(request, adapter))
 })
@@ -93,6 +94,8 @@ export const protocol = Protocol.make({
  },
 })

-export const compact = ResponsesCompaction.make(adapter)
+export const compact = ResponsesCompaction.make(adapter, (request) =>
+  OpenResponses.lowerTools(ProviderShared.flattenTools(request.tools), adapter),
+)

 export * as XAIResponses from "./xai-responses.js"
@@ -0,0 +1,205 @@
+import { Effect, Schema } from "effect"
+import type { HttpClientResponse } from "effect/unstable/http"
+import type { Status } from "../generation.js"
+import { Media } from "../media.js"
+import { MediaProtocol } from "../route/media-protocol.js"
+import { MediaRoute } from "../route/media.js"
+import { mergeJsonRecords } from "../schema/index.js"
+import { VideoModel, VideoResponse, type VideoRequestFor } from "../video.js"
+import { ProviderShared, optionalNull } from "./shared.js"
+
+const route = MediaProtocol.identity({ id: "xai-video", name: "xAI Video", provider: "xai" })
+export const DEFAULT_BASE_URL = "https://api.x.ai/v1"
+export const PATH = "/videos/generations"
+export const EDIT_PATH = "/videos/edits"
+export const EXTEND_PATH = "/videos/extensions"
+export const STATUS_PATH = "/videos"
+
+// ---------------------------------------------------------------------------
+// 1. Public model input
+// ---------------------------------------------------------------------------
+
+/**
+ * Provider-native options. Common fields (`frames`, `references`, `video`, `durationSeconds`, `aspectRatio`,
+ * `resolution`, `audio`) live on the request. `mode` selects the endpoint a `video` input is sent to.
+ */
+export type XAIVideoOptions = {
+  readonly mode?: "edit" | "extend"
+  readonly reference_audios?: ReadonlyArray<{ readonly voice_id: string }>
+} & Record<string, unknown>
+
+export type Request = VideoRequestFor<XAIVideoOptions>
+
+// ---------------------------------------------------------------------------
+// 2. Token and response schemas
+// ---------------------------------------------------------------------------
+
+export const Token = Schema.Struct({ requestID: Schema.String })
+export type Token = Schema.Schema.Type<typeof Token>
+
+const StartResponse = Schema.Struct({ request_id: Schema.String })
+
+const VideoStatus = Schema.Struct({
+  status: Schema.String,
+  progress: optionalNull(Schema.Number),
+  video: optionalNull(
+    Schema.Struct({
+      url: optionalNull(Schema.String),
+      duration: optionalNull(Schema.Number),
+      respect_moderation: optionalNull(Schema.Boolean),
+    }),
+  ),
+  error: optionalNull(
+    Schema.Struct({
+      code: optionalNull(Schema.String),
+      message: optionalNull(Schema.String),
+    }),
+  ),
+  model: optionalNull(Schema.String),
+})
+
+const STATUS = {
+  pending: "running",
+  done: "completed",
+  failed: "failed",
+  expired: "expired",
+} as const satisfies Record<string, Status>
+
+// ---------------------------------------------------------------------------
+// 5. Request body construction
+// ---------------------------------------------------------------------------
+
+const mediaInput = (asset: Media.Asset) =>
+  ProviderShared.mediaReference(asset, route.provider, route.name).pipe(
+    Effect.map((reference) => (reference.type === "ref" ? { file_id: reference.value } : { url: reference.value })),
+  )
+
+const nativeOptions = (options: XAIVideoOptions | undefined) => {
+  if (!options) return undefined
+  const { mode: _mode, ...native } = options
+  return native
+}
+
+const fromRequest = Effect.fn("XAIVideo.fromRequest")(function* (request: Request) {
+  const image = request.frames?.first === undefined ? undefined : yield* mediaInput(request.frames.first)
+  const lastFrame = request.frames?.last === undefined ? undefined : yield* mediaInput(request.frames.last)
+  const video = request.video === undefined ? undefined : yield* mediaInput(request.video)
+  const references = yield* Effect.forEach(request.references ?? [], mediaInput)
+  return MediaProtocol.json(
+    mergeJsonRecords(
+      {
+        model: request.model.id,
+        prompt: request.prompt,
+        image,
+        last_frame: lastFrame,
+        reference_images: references.length === 0 ? undefined : references,
+        video,
+        duration: request.durationSeconds,
+        aspect_ratio: request.aspectRatio,
+        resolution: request.resolution,
+        generate_audio: request.audio,
+      },
+      nativeOptions(request.providerOptions),
+      request.http?.body,
+    ) ?? {},
+  )
+})
+
+// ---------------------------------------------------------------------------
+// 6. Response decoding
+// ---------------------------------------------------------------------------
+
+const decodeStart = route.decodeStarted(StartResponse, (value) => ({
+  token: { requestID: value.request_id },
+  snapshot: { id: value.request_id, status: "running" },
+}))
+
+// `progress` is undocumented but observed live as a 0..100 percentage (recorded cassette: 1 → 10 → 37 → 100).
+const fraction = (progress: number | null | undefined) =>
+  progress !== undefined && progress !== null && progress >= 0 && progress <= 100 ? progress / 100 : undefined
+
+const decodeVideoStatus = route.decodeJson(VideoStatus)
+
+const decodeStatus = Effect.fn("XAIVideo.decodeStatus")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.PollContext<Token>,
+) {
+  const output = yield* decodeVideoStatus(response)
+  const status = yield* MediaProtocol.status(STATUS, output.value.status, output)
+  return { id: context.token.requestID, status, progress: fraction(output.value.progress) }
+})
+
+const decodeResult = Effect.fn("XAIVideo.decodeResult")(function* (
+  response: HttpClientResponse.HttpClientResponse,
+  context: MediaProtocol.PollContext<Token>,
+) {
+  const output = yield* decodeVideoStatus(response)
+  const decoded = output.value
+  const status = yield* MediaProtocol.status(STATUS, decoded.status, output)
+  if (status === "running") return yield* output.pending(context.token.requestID)
+  if (status === "failed") {
+    const code = decoded.error?.code ?? undefined
+    const message = decoded.error?.message ?? undefined
+    return yield* output.ended(
+      "failed",
+      `${route.name} generation failed${code === undefined ? "" : ` (${code})`}${message === undefined ? "" : `: ${message}`}`,
+    )
+  }
+  if (status !== "completed")
+    return yield* output.ended("expired", `${route.name} request ${context.token.requestID} expired`)
+  // `respect_moderation: false` marks a filtered result; a URL may still be present, so report it as a notice.
+  const notices =
+    decoded.video?.respect_moderation === false
+      ? [{ type: "moderated" as const, message: `${route.name} flagged the generated video for moderation` }]
+      : undefined
+  const url = decoded.video?.url ?? undefined
+  if (url === undefined && notices !== undefined)
+    return yield* output.contentPolicy(`${route.name} withheld the video for moderation`)
+  if (url === undefined) return yield* output.invalid(`${route.name} completed without a video URL`)
+  const duration = decoded.video?.duration ?? undefined
+  return new VideoResponse({
+    videos: [
+      Media.url(url, {
+        mediaType: "video/mp4",
+        info: duration === undefined ? undefined : { durationSeconds: duration },
+      }),
+    ],
+    notices,
+    providerMetadata: { xai: { requestId: context.token.requestID, model: decoded.model ?? undefined } },
+  })
+})
+
+// ---------------------------------------------------------------------------
+// 7. Protocol and route
+// ---------------------------------------------------------------------------
+
+const statusPath = (token: Token) => `${STATUS_PATH}/${token.requestID}`
+
+export const protocol = MediaProtocol.queued<Request, VideoResponse, Token>(route, {
+  token: Token,
+  unsupported: ["n", "seed", "negativePrompt"],
+  start: { body: { from: fromRequest }, decode: decodeStart },
+  status: { path: statusPath, decode: decodeStatus },
+  result: { path: statusPath, decode: decodeResult },
+})
+
+// A source video goes to `/videos/edits` unless `providerOptions.mode` asks for an extension.
+const startPath = (request: Request) => {
+  if (request.video === undefined) return PATH
+  return request.providerOptions?.mode === "extend" ? EXTEND_PATH : EDIT_PATH
+}
+
+export const model = (input: MediaRoute.ModelInput) =>
+  VideoModel.fromRoute<XAIVideoOptions, Token>(
+    {
+      protocol,
+      baseURL: DEFAULT_BASE_URL,
+      path: ({ request }) => startPath(request),
+    },
+    input,
+  )
+
+export const XAIVideo = {
+  protocol,
+  model,
+} as const
@@ -1,26 +1,22 @@
-import { Effect, Schema } from "effect"
+import { Duration, Effect, Schema } from "effect"
 import type { HttpClientResponse } from "effect/unstable/http"
 import { ImageModel, ImageResponse, type ImageRequestFor } from "../image.js"
-import { Media } from "../media.js"
 import { MediaProtocol } from "../route/media-protocol.js"
 import { MediaRoute } from "../route/media.js"
-import { ProviderID, mergeJsonRecords } from "../schema/index.js"
+import { mergeJsonRecords, type OpenString } from "../schema/index.js"

-const ADAPTER = "zai-images"
-const NAME = "Z.ai Images"
-const PROVIDER = ProviderID.make("zai")
+const route = MediaProtocol.identity({ id: "zai-images", name: "Z.ai Images", provider: "zai" })
 export const DEFAULT_BASE_URL = "https://api.z.ai/api/paas/v4"
 export const PATH = "/images/generations"
+const OUTPUT_RETENTION = Duration.days(30)

 // ---------------------------------------------------------------------------
 // 1. Public model input
 // ---------------------------------------------------------------------------

-export type ZAIImageString<Known extends string> = Known | (string & {})
-
 /** Provider-native options. The common `size` field lives on the request. */
 export type ZAIImageOptions = {
-  readonly quality?: ZAIImageString<"hd" | "standard">
+  readonly quality?: OpenString<"hd" | "standard">
  readonly userID?: string
 } & Record<string, unknown>

@@ -69,23 +65,25 @@ const fromRequest = Effect.fn("ZAIImages.fromRequest")(function* (request: Reque
 // 6. Response decoding
 // ---------------------------------------------------------------------------

+const decodeDocument = route.decodeJson(ZAIImageResponse)
+
 const decodeResponse = Effect.fn("ZAIImages.decodeResponse")(function* (
  response: HttpClientResponse.HttpClientResponse,
 ) {
-  const output = yield* MediaProtocol.decodeJson(ADAPTER, NAME, ZAIImageResponse)(response)
+  const output = yield* decodeDocument(response)
  const decoded = output.value
-  if (decoded.data.length === 0) return yield* output.invalid(`${NAME} returned no images`)
+  if (decoded.data.length === 0) return yield* output.invalid(`${route.name} returned no images`)
  const filters = decoded.content_filter ?? []
  return new ImageResponse({
    // Z.ai returns only URLs and no content type; the media type resolves when the asset is materialized.
-    images: decoded.data.map((item) => Media.url(item.url)),
+    images: yield* Effect.forEach(decoded.data, (item) => MediaProtocol.expiringUrl(item.url, OUTPUT_RETENTION)),
    // Z.ai reports applied content filters alongside a successful result; surface them instead of dropping them.
    notices:
      filters.length === 0
        ? undefined
        : filters.map((filter) => ({
            type: "moderated" as const,
-            message: `${NAME} applied a content filter${filter.role === undefined ? "" : ` for ${filter.role}`}${
+            message: `${route.name} applied a content filter${filter.role === undefined ? "" : ` for ${filter.role}`}${
              filter.level === undefined ? "" : ` at level ${filter.level}`
            }`,
            providerMetadata: { zai: filter },
@@ -105,19 +103,14 @@ const decodeResponse = Effect.fn("ZAIImages.decodeResponse")(function* (
 // 7. Protocol and route
 // ---------------------------------------------------------------------------

-export const protocol = MediaProtocol.inline<Request, ImageResponse>({
-  id: ADAPTER,
-  name: NAME,
+export const protocol = MediaProtocol.inline<Request, ImageResponse>(route, {
  unsupported: ["images", "mask", "n", "aspectRatio", "seed", "format"],
  body: { from: fromRequest },
  response: { decode: decodeResponse },
 })

 export const model = (input: MediaRoute.ModelInput) =>
-  ImageModel.fromRoute<ZAIImageOptions>(
-    { id: ADAPTER, provider: PROVIDER, protocol, baseURL: DEFAULT_BASE_URL, path: PATH },
-    input,
-  )
+  ImageModel.fromRoute<ZAIImageOptions>({ protocol, baseURL: DEFAULT_BASE_URL, path: PATH }, input)

 export const ZAIImages = {
  protocol,
@@ -80,12 +80,19 @@ const SERVER_CODES = new Set([
  "slow_down",
  "serviceunavailableexception",
 ])
-const INVALID_REQUEST_CODES = new Set(["invalid_prompt", "invalid_request_error", "validationexception"])
+// `invalid_request` is the Vercel AI Gateway's code for an upstream request rejection.
+const INVALID_REQUEST_CODES = new Set([
+  "invalid_prompt",
+  "invalid_request",
+  "invalid_request_error",
+  "validationexception",
+])
 // Azure OpenAI reports `content_filter` with `innererror.code` ResponsibleAIPolicyViolation.
 // OpenRouter tags provider failures with a typed `error_type`; its Responses skin also
 // emits `image_content_policy_violation` as the native code.
 const CONTENT_POLICY_CODES = new Set([
  "content_filter",
+  "content_moderation",
  "responsibleaipolicyviolation",
  "content_policy_violation",
  "image_content_policy_violation",
@@ -204,6 +211,8 @@ function providerCodes(value: unknown) {
  const exception = isRecord(decoded.exception) ? decoded.exception : undefined
  return [
    decoded.code,
+    // Stability's `{ id, name, errors }` bodies carry the code in `name`.
+    Array.isArray(decoded.errors) ? decoded.name : undefined,
    decoded.error_type,
    error?.code,
    error?.type,
@@ -31,6 +31,7 @@ export interface Settings extends ProviderPackage.Settings {
  readonly profile?: string
  readonly region?: string
  readonly topP?: number
+  readonly thinking?: BedrockConverse.OptionsInput["thinking"]
 }
 export const routes = [BedrockConverse.route]

@@ -71,6 +72,7 @@ export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, se
    generation: settings.topP === undefined ? undefined : { topP: settings.topP },
    headers: settings.headers === undefined ? undefined : { ...settings.headers },
    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    providerOptions: settings.thinking === undefined ? undefined : { thinking: settings.thinking },
    profile: settings.profile,
    region: settings.region,
  }).model(modelID)
@@ -0,0 +1,37 @@
+import { Auth } from "../route/auth.js"
+import type { ProviderAuthOption } from "../route/auth-options.js"
+import { MediaRoute } from "../route/media.js"
+import { type HttpOptions, ProviderID, type ModelID } from "../schema/index.js"
+import { AssemblyAITranscription } from "../protocols/assemblyai-transcription.js"
+
+export type { AssemblyAITranscriptionOptions } from "../protocols/assemblyai-transcription.js"
+
+export const id = ProviderID.make("assemblyai")
+
+export type Config = ProviderAuthOption<"optional"> & {
+  /** `https://api.eu.assemblyai.com` for the EU region. */
+  readonly baseURL?: string
+  readonly headers?: Record<string, string>
+  readonly http?: HttpOptions.Input
+}
+
+// The key is the whole `authorization` value, without a scheme.
+const auth = (options: ProviderAuthOption<"optional">) => {
+  if ("auth" in options && options.auth) return options.auth
+  return Auth.optional("apiKey" in options ? options.apiKey : undefined, "apiKey")
+    .orElse(Auth.config("ASSEMBLYAI_API_KEY"))
+    .pipe(Auth.header("authorization"))
+}
+
+export const configure = (input: Config = {}) => {
+  const media = MediaRoute.deployment(input, auth(input))
+  const transcription = (modelID: string | ModelID) => AssemblyAITranscription.model({ ...media, id: modelID })
+  return {
+    id,
+    transcription,
+    configure,
+  }
+}
+
+export const provider = configure()
+export const transcription = provider.transcription
@@ -1,11 +1,11 @@
 import { Auth } from "../route/auth.js"
 import { type AtLeastOne, type ProviderAuthOption } from "../route/auth-options.js"
 import type { Route, RouteDefaultsInput, CompactionOperations } from "../route/client.js"
+import { Endpoint } from "../route/endpoint.js"
 import type { ProviderPackage } from "../provider-package.js"
 import { ProviderConfigurationError, ProviderID, type ModelID } from "../schema/index.js"
 import * as OpenAIChat from "../protocols/openai-chat.js"
 import * as OpenAIResponses from "../protocols/openai-responses.js"
-import { ProviderShared } from "../protocols/shared.js"
 import { withOpenAIOptions, type OpenAIProviderOptionsInput } from "./openai-options.js"

 export const id = ProviderID.make("azure")
@@ -108,7 +108,7 @@ const configuredRoute = <Body, Prepared, Compact extends CompactionOperations |
  })

 function endpoint(input: Config, modelID: string | ModelID) {
-  const baseURL = ProviderShared.trimBaseUrl(input.baseURL ?? resourceBaseURL(input.resourceName!))
+  const baseURL = Endpoint.trimBaseUrl(input.baseURL ?? resourceBaseURL(input.resourceName!))
  const query = { "api-version": input.apiVersion ?? "v1", ...input.queryParams }

  if (input.useDeploymentBasedUrls) return { baseURL: `${baseURL}/deployments/${modelID}`, query }
@@ -0,0 +1,36 @@
+import { Auth } from "../route/auth.js"
+import type { ProviderAuthOption } from "../route/auth-options.js"
+import { MediaRoute } from "../route/media.js"
+import { type HttpOptions, ProviderID, type ModelID } from "../schema/index.js"
+import { BlackForestLabsImages } from "../protocols/bfl-images.js"
+
+export type { BlackForestLabsImageOptions } from "../protocols/bfl-images.js"
+
+export const id = ProviderID.make("black-forest-labs")
+
+export type Config = ProviderAuthOption<"optional"> & {
+  /** `https://api.eu.bfl.ai` or `https://api.us.bfl.ai` pin inference to one region. */
+  readonly baseURL?: string
+  readonly headers?: Record<string, string>
+  readonly http?: HttpOptions.Input
+}
+
+const auth = (options: ProviderAuthOption<"optional">) => {
+  if ("auth" in options && options.auth) return options.auth
+  return Auth.optional("apiKey" in options ? options.apiKey : undefined, "apiKey")
+    .orElse(Auth.config("BFL_API_KEY"))
+    .pipe(Auth.header("x-key"))
+}
+
+export const configure = (input: Config = {}) => {
+  const media = MediaRoute.deployment(input, auth(input))
+  const image = (modelID: string | ModelID) => BlackForestLabsImages.model({ ...media, id: modelID })
+  return {
+    id,
+    image,
+    configure,
+  }
+}
+
+export const provider = configure()
+export const image = provider.image
@@ -0,0 +1,29 @@
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import { MediaRoute } from "../route/media.js"
+import { type HttpOptions, ProviderID, type ModelID } from "../schema/index.js"
+import { CartesiaSpeech } from "../protocols/cartesia-speech.js"
+
+export type { CartesiaEncoding, CartesiaSpeechOptions } from "../protocols/cartesia-speech.js"
+
+export const id = ProviderID.make("cartesia")
+
+export type Config = ProviderAuthOption<"optional"> & {
+  readonly baseURL?: string
+  readonly headers?: Record<string, string>
+  readonly http?: HttpOptions.Input
+}
+
+const auth = (options: ProviderAuthOption<"optional">) => AuthOptions.bearer(options, "CARTESIA_API_KEY")
+
+export const configure = (input: Config = {}) => {
+  const media = MediaRoute.deployment(input, auth(input))
+  const speech = (modelID: string | ModelID) => CartesiaSpeech.model({ ...media, id: modelID })
+  return {
+    id,
+    speech,
+    configure,
+  }
+}
+
+export const provider = configure()
+export const speech = provider.speech
@@ -0,0 +1,38 @@
+import { Auth } from "../route/auth.js"
+import type { ProviderAuthOption } from "../route/auth-options.js"
+import { MediaRoute } from "../route/media.js"
+import { type HttpOptions, ProviderID, type ModelID } from "../schema/index.js"
+import { DeepgramSpeech } from "../protocols/deepgram-speech.js"
+import { DeepgramTranscription } from "../protocols/deepgram-transcription.js"
+
+export type { DeepgramEncoding, DeepgramSpeechOptions } from "../protocols/deepgram-speech.js"
+export type { DeepgramTranscriptionOptions } from "../protocols/deepgram-transcription.js"
+
+export const id = ProviderID.make("deepgram")
+
+export type Config = ProviderAuthOption<"optional"> & {
+  readonly baseURL?: string
+  readonly headers?: Record<string, string>
+  readonly http?: HttpOptions.Input
+}
+
+const auth = (options: ProviderAuthOption<"optional">) => {
+  if ("auth" in options && options.auth) return options.auth
+  return Auth.optional("apiKey" in options ? options.apiKey : undefined, "apiKey")
+    .orElse(Auth.config("DEEPGRAM_API_KEY"))
+    .pipe(Auth.scheme("Token"))
+}
+
+export const configure = (input: Config = {}) => {
+  const media = MediaRoute.deployment(input, auth(input))
+  return {
+    id,
+    speech: (modelID: string | ModelID) => DeepgramSpeech.model({ ...media, id: modelID }),
+    transcription: (modelID: string | ModelID) => DeepgramTranscription.model({ ...media, id: modelID }),
+    configure,
+  }
+}
+
+export const provider = configure()
+export const speech = provider.speech
+export const transcription = provider.transcription
@@ -0,0 +1,35 @@
+import { Auth } from "../route/auth.js"
+import type { ProviderAuthOption } from "../route/auth-options.js"
+import { MediaRoute } from "../route/media.js"
+import { type HttpOptions, ProviderID, type ModelID } from "../schema/index.js"
+import { ElevenLabsSpeech } from "../protocols/elevenlabs-speech.js"
+
+export type { ElevenLabsOutputFormat, ElevenLabsSpeechOptions } from "../protocols/elevenlabs-speech.js"
+
+export const id = ProviderID.make("elevenlabs")
+
+export type Config = ProviderAuthOption<"optional"> & {
+  readonly baseURL?: string
+  readonly headers?: Record<string, string>
+  readonly http?: HttpOptions.Input
+}
+
+const auth = (options: ProviderAuthOption<"optional">) => {
+  if ("auth" in options && options.auth) return options.auth
+  return Auth.optional("apiKey" in options ? options.apiKey : undefined, "apiKey")
+    .orElse(Auth.config("ELEVENLABS_API_KEY"))
+    .pipe(Auth.header("xi-api-key"))
+}
+
+export const configure = (input: Config = {}) => {
+  const media = MediaRoute.deployment(input, auth(input))
+  const speech = (modelID: string | ModelID) => ElevenLabsSpeech.model({ ...media, id: modelID })
+  return {
+    id,
+    speech,
+    configure,
+  }
+}
+
+export const provider = configure()
+export const speech = provider.speech
@@ -0,0 +1,39 @@
+import { Auth } from "../route/auth.js"
+import type { ProviderAuthOption } from "../route/auth-options.js"
+import { MediaRoute } from "../route/media.js"
+import { type HttpOptions, ProviderID, type ModelID } from "../schema/index.js"
+import { FalImages } from "../protocols/fal-images.js"
+import { FalVideo } from "../protocols/fal-video.js"
+
+export type { FalImageOptions } from "../protocols/fal-images.js"
+export type { FalVideoOptions } from "../protocols/fal-video.js"
+
+export const id = ProviderID.make("fal")
+
+export type Config = ProviderAuthOption<"optional"> & {
+  readonly baseURL?: string
+  readonly headers?: Record<string, string>
+  readonly http?: HttpOptions.Input
+}
+
+// fal authenticates with `Authorization: Key <FAL_KEY>` rather than a bearer token.
+const auth = (options: ProviderAuthOption<"optional">) => {
+  if ("auth" in options && options.auth) return options.auth
+  return Auth.optional("apiKey" in options ? options.apiKey : undefined, "apiKey")
+    .orElse(Auth.config("FAL_KEY"))
+    .pipe(Auth.scheme("Key"))
+}
+
+export const configure = (input: Config = {}) => {
+  const media = MediaRoute.deployment(input, auth(input))
+  return {
+    id,
+    image: (modelID: string | ModelID) => FalImages.model({ ...media, id: modelID }),
+    video: (modelID: string | ModelID) => FalVideo.model({ ...media, id: modelID }),
+    configure,
+  }
+}
+
+export const provider = configure()
+export const image = provider.image
+export const video = provider.video
@@ -1,12 +1,19 @@
 import type { RouteDefaultsInput } from "../route/client.js"
 import { Auth } from "../route/auth.js"
 import type { ProviderAuthOption } from "../route/auth-options.js"
+import { MediaRoute } from "../route/media.js"
 import type { ProviderPackage } from "../provider-package.js"
-import { HttpOptions, ProviderID, mergeHttpOptions, type ModelID } from "../schema/index.js"
+import { ProviderID, type ModelID } from "../schema/index.js"
 import { Gemini } from "../protocols/gemini.js"
 import { GoogleImages } from "../protocols/google-images.js"
+import { GoogleSpeech } from "../protocols/google-speech.js"
+import { GoogleTranscription } from "../protocols/google-transcription.js"
+import { GoogleVideo } from "../protocols/google-video.js"

 export type { GoogleImageOptions } from "../protocols/google-images.js"
+export type { GoogleSpeechOptions } from "../protocols/google-speech.js"
+export type { GoogleTranscriptionOptions } from "../protocols/google-transcription.js"
+export type { GoogleVideoOptions } from "../protocols/google-video.js"
 export type GeminiOptionsInput = Gemini.OptionsInput
 export type GeminiProviderOptionsInput = Gemini.ProviderOptionsInput

@@ -40,18 +47,14 @@ const configuredRoute = (input: Config) => {

 export const configure = (input: Config = {}) => {
  const route = configuredRoute(input)
-  const image = (modelID: string | ModelID) =>
-    GoogleImages.model({
-      id: modelID,
-      auth: auth(input),
-      baseURL: input.baseURL,
-      headers: input.headers,
-      http: mergeHttpOptions(input.http === undefined ? undefined : HttpOptions.make(input.http)),
-    })
+  const media = MediaRoute.deployment(input, auth(input))
  return {
    id,
    model: (modelID: string | ModelID) => route.model<Gemini.ProviderOptionsInput>({ id: modelID }),
-    image,
+    image: (modelID: string | ModelID) => GoogleImages.model({ ...media, id: modelID }),
+    video: (modelID: string | ModelID) => GoogleVideo.model({ ...media, id: modelID }),
+    speech: (modelID: string | ModelID) => GoogleSpeech.model({ ...media, id: modelID }),
+    transcription: (modelID: string | ModelID) => GoogleTranscription.model({ ...media, id: modelID }),
    configure,
  }
 }
@@ -70,3 +73,6 @@ export const model: ProviderPackage.Definition<Settings, Gemini.ProviderOptionsI
  }).model(modelID)

 export const image = provider.image
+export const video = provider.video
+export const speech = provider.speech
+export const transcription = provider.transcription
@@ -3,13 +3,19 @@ export * as Anthropic from "./anthropic.js"
 export * as AnthropicCompatible from "./anthropic-compatible.js"
 export * as AmazonBedrock from "./amazon-bedrock.js"
 export * as AmazonBedrockMantle from "./amazon-bedrock-mantle.js"
+export * as AssemblyAI from "./assemblyai.js"
 export * as Azure from "./azure.js"
 export * as Baseten from "./baseten.js"
+export * as BlackForestLabs from "./black-forest-labs.js"
+export * as Cartesia from "./cartesia.js"
 export * as Cerebras from "./cerebras.js"
 export * as CloudflareAIGateway from "./cloudflare-ai-gateway.js"
 export * as CloudflareWorkersAI from "./cloudflare-workers-ai.js"
 export * as DeepInfra from "./deepinfra.js"
+export * as Deepgram from "./deepgram.js"
 export * as DeepSeek from "./deepseek.js"
+export * as ElevenLabs from "./elevenlabs.js"
+export * as Fal from "./fal.js"
 export * as Fireworks from "./fireworks.js"
 export * as Google from "./google.js"
 export * as GoogleVertex from "./google-vertex.js"
@@ -26,6 +32,9 @@ export * as OpenAICompatible from "./openai-compatible.js"
 export * as OpenAICompatibleResponses from "./openai-compatible-responses.js"
 export * as OpenCodeZen from "./opencode-zen.js"
 export * as OpenRouter from "./openrouter.js"
+export * as Replicate from "./replicate.js"
+export * as Runway from "./runway.js"
+export * as Stability from "./stability.js"
 export * as TogetherAI from "./togetherai.js"
 export * as TypeSafeAI from "./typesafe-ai.js"
 export * as VercelAIGateway from "./vercel-ai-gateway.js"
@@ -7,7 +7,8 @@ import { MetaImages } from "../protocols/meta-images.js"
 import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
 import { Route, type RouteDefaultsInput } from "../route/client.js"
 import { Endpoint } from "../route/endpoint.js"
-import { HttpOptions, ProviderID, ToolDefinition, type ModelID } from "../schema/index.js"
+import { MediaRoute } from "../route/media.js"
+import { ProviderID, ToolDefinition, type ModelID, type OpenString } from "../schema/index.js"
 import type { OpenResponsesProviderOptionsInput } from "./open-responses-options.js"

 export const id = ProviderID.make("meta")
@@ -48,8 +49,8 @@ export const webSearch = (options: WebSearchOptions = {}) =>

 export interface ImageGenerationOptions {
  readonly size?: string
-  readonly outputFormat?: "webp" | "png" | "jpeg" | (string & {})
-  readonly reasoningStrength?: "low" | "high" | (string & {})
+  readonly outputFormat?: OpenString<"webp" | "png" | "jpeg">
+  readonly reasoningStrength?: OpenString<"low" | "high">
  readonly enableImageSearch?: boolean
  readonly enableWebSearch?: boolean
  readonly enableShell?: boolean
@@ -139,14 +140,8 @@ export const configure = (input: LanguageModelOptions = {}) => {
      id: modelID,
      compatibility: { requireSignature: false },
    })
-  const image = (modelID: string | ModelID) =>
-    MetaImages.model({
-      id: modelID,
-      baseURL: endpoint ?? baseURL,
-      auth: options.auth,
-      headers: input.headers,
-      http: input.http === undefined ? undefined : HttpOptions.make(input.http),
-    })
+  const media = MediaRoute.deployment(input, options.auth)
+  const image = (modelID: string | ModelID) => MetaImages.model({ ...media, id: modelID })
  return { id, model: responses, responses, chat, messages, image, configure }
 }

@@ -113,19 +113,19 @@ export const configure = (input: Config = {}) => {
        supportsStore: false,
        supportsStrictMode: false,
        supportsPromptCacheKey: true,
-        toolSchema: "moonshot",
+        sanitizer: "moonshot",
        reasoningField: "reasoning_content",
      },
    })
  const messages = (modelID: string | ModelID) =>
    messagesRoute.with(defaults).model<MessagesOptionsInput>({
      id: modelID,
-      compatibility: { requireSignature: false, toolSchema: "moonshot" },
+      compatibility: { requireSignature: false, sanitizer: "moonshot" },
    })
  const responses = (modelID: string | ModelID) =>
    responsesRoute
      .with(defaults)
-      .model<ResponsesOptionsInput>({ id: modelID, compatibility: { toolSchema: "moonshot" } })
+      .model<ResponsesOptionsInput>({ id: modelID, compatibility: { sanitizer: "moonshot" } })
  return { id, model: chat, chat, messages, responses, configure }
 }

@@ -32,10 +32,7 @@ const openAIProviderOptions = (options: OpenAIOptionsInput | undefined): Provide
  return result
 }

-export const gpt5DefaultOptions = (
-  modelID: string,
-  options: { readonly textVerbosity?: boolean } = {},
-): ProviderOptions | undefined => {
+export const gpt5DefaultOptions = (modelID: string): ProviderOptions | undefined => {
  const id = modelID.toLowerCase()
  if (!id.includes("gpt-5") || id.includes("gpt-5-chat") || id.includes("gpt-5-pro")) return undefined
  return openAIProviderOptions({
@@ -47,27 +44,19 @@ export const gpt5DefaultOptions = (
    // this, callers using the default model facade get reasoning summaries
    // they cannot replay statelessly.
    include: ["reasoning.encrypted_content"],
-    textVerbosity:
-      options.textVerbosity === true && id.includes("gpt-5.") && !id.includes("codex") && !id.includes("-chat")
-        ? "low"
-        : undefined,
  })
 }

-export const openAIDefaultOptions = (
-  modelID: string,
-  options: { readonly textVerbosity?: boolean } = {},
-): ProviderOptions | undefined =>
-  mergeProviderOptions(openAIProviderOptions({ store: false }), gpt5DefaultOptions(modelID, options))
+export const openAIDefaultOptions = (modelID: string): ProviderOptions | undefined =>
+  mergeProviderOptions(openAIProviderOptions({ store: false }), gpt5DefaultOptions(modelID))

 export const withOpenAIOptions = <Options extends { readonly providerOptions?: OpenAIProviderOptionsInput }>(
  modelID: string,
  options: Options,
-  defaults: { readonly textVerbosity?: boolean } = {},
 ): Omit<Options, "providerOptions"> & { readonly providerOptions?: ProviderOptions } => {
  return {
    ...options,
-    providerOptions: mergeProviderOptions(openAIDefaultOptions(modelID, defaults), options.providerOptions),
+    providerOptions: mergeProviderOptions(openAIDefaultOptions(modelID), options.providerOptions),
  }
 }

@@ -1,14 +1,26 @@
 import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
 import type { Route, RouteDefaultsInput, CompactionOperations } from "../route/client.js"
+import { MediaRoute } from "../route/media.js"
 import type { ProviderPackage } from "../provider-package.js"
-import { HttpOptions, ProviderID, ToolDefinition, mergeHttpOptions, type ModelID } from "../schema/index.js"
+import {
+  HttpOptions,
+  ProviderID,
+  ToolDefinition,
+  mergeHttpOptions,
+  type ModelID,
+  type OpenString,
+} from "../schema/index.js"
 import * as OpenAIChat from "../protocols/openai-chat.js"
 import * as OpenAIResponses from "../protocols/openai-responses.js"
 import { withOpenAIOptions, type OpenAIProviderOptionsInput } from "./openai-options.js"
-import { OpenAIImages, type OpenAIImageString } from "../protocols/openai-images.js"
+import { OpenAIImages } from "../protocols/openai-images.js"
+import { OpenAISpeech } from "../protocols/openai-speech.js"
+import { OpenAITranscription } from "../protocols/openai-transcription.js"

 export type { OpenAIOptionsInput, OpenAIResponseIncludable } from "./openai-options.js"
 export type { OpenAIImageOptions } from "../protocols/openai-images.js"
+export type { OpenAISpeechOptions } from "../protocols/openai-speech.js"
+export type { OpenAITranscriptionOptions } from "../protocols/openai-transcription.js"

 export const id = ProviderID.make("openai")

@@ -25,14 +37,14 @@ export type Config = RouteDefaultsInput &
  }

 export interface ImageGenerationOptions {
-  readonly action?: OpenAIImageString<"auto" | "generate" | "edit">
-  readonly background?: OpenAIImageString<"auto" | "opaque" | "transparent">
-  readonly inputFidelity?: OpenAIImageString<"low" | "high">
+  readonly action?: OpenString<"auto" | "generate" | "edit">
+  readonly background?: OpenString<"auto" | "opaque" | "transparent">
+  readonly inputFidelity?: OpenString<"low" | "high">
  readonly outputCompression?: number
-  readonly outputFormat?: OpenAIImageString<"png" | "jpeg" | "webp">
+  readonly outputFormat?: OpenString<"png" | "jpeg" | "webp">
  readonly partialImages?: number
-  readonly quality?: OpenAIImageString<"auto" | "low" | "medium" | "high" | "standard" | "hd">
-  readonly size?: OpenAIImageString<
+  readonly quality?: OpenString<"auto" | "low" | "medium" | "high" | "standard" | "hd">
+  readonly size?: OpenString<
    "auto" | "256x256" | "512x512" | "1024x1024" | "1536x1024" | "1024x1536" | "1792x1024" | "1024x1792"
  >
 }
@@ -88,24 +100,24 @@ export const configure = (input: Config = {}) => {
  const modelDefaults = defaults(input)
  const responses = (id: string | ModelID) =>
    responsesRoute
-      .with(withOpenAIOptions(id, modelDefaults, { textVerbosity: true }))
+      .with(withOpenAIOptions(id, modelDefaults))
      .model<OpenAIProviderOptionsInput>({ id })
  const chat = (id: string | ModelID) =>
    chatRoute.with(withOpenAIOptions(id, modelDefaults)).model<OpenAIProviderOptionsInput>({
      id,
      compatibility: { supportsPromptCacheKey: true },
    })
-  const image = (modelID: string | ModelID) =>
-    OpenAIImages.model({
-      id: modelID,
-      auth: auth(input),
-      baseURL: input.baseURL,
-      headers: input.headers,
-      http: mergeHttpOptions(
-        input.http === undefined ? undefined : HttpOptions.make(input.http),
-        input.queryParams === undefined ? undefined : new HttpOptions({ query: input.queryParams }),
-      ),
-    })
+  const deployment = MediaRoute.deployment(input, auth(input))
+  const media = {
+    ...deployment,
+    http: mergeHttpOptions(
+      deployment.http,
+      input.queryParams === undefined ? undefined : new HttpOptions({ query: input.queryParams }),
+    ),
+  }
+  const image = (modelID: string | ModelID) => OpenAIImages.model({ ...media, id: modelID })
+  const speech = (modelID: string | ModelID) => OpenAISpeech.model({ ...media, id: modelID })
+  const transcription = (modelID: string | ModelID) => OpenAITranscription.model({ ...media, id: modelID })

  return {
    id,
@@ -113,6 +125,8 @@ export const configure = (input: Config = {}) => {
    responses,
    chat,
    image,
+    speech,
+    transcription,
    configure,
  }
 }
@@ -159,3 +173,5 @@ export const chatModel: ProviderPackage.Definition<Settings, OpenAIProviderOptio
 export const responses = provider.responses
 export const chat = provider.chat
 export const image = provider.image
+export const speech = provider.speech
+export const transcription = provider.transcription
@@ -20,7 +20,7 @@ export const configure = (input: Options = {}) => {
      auth: AuthOptions.bearer(input, "OPENCODE_API_KEY"),
      baseURL: input.baseURL ?? baseURL,
      headers: input.headers,
-      http: input.http === undefined ? undefined : HttpOptions.make(input.http),
+      http: HttpOptions.make(input.http),
    })
  return { id, experimental: { evaluation }, configure }
 }
@@ -3,29 +3,27 @@ import { Route, type RouteDefaultsInput } from "../route/client.js"
 import { Endpoint } from "../route/endpoint.js"
 import { Protocol } from "../route/protocol.js"
 import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
-import { HttpOptions, ProviderID, type CacheHint, type ModelID } from "../schema/index.js"
+import { HttpOptions, ProviderID, type CacheHint, type ModelID, type OpenString } from "../schema/index.js"
 import type { ProviderPackage } from "../provider-package.js"
 import { SystemOne } from "../experimental/system-one.js"
 import { OpenAIChat } from "../protocols/openai-chat.js"
 import { newBreakpoints, ttlBucket } from "../protocols/utils/cache.js"
-import { isRecord } from "../protocols/shared.js"
+import { isRecord, ProviderShared } from "../protocols/shared.js"

 export const id = ProviderID.make("openrouter")
 const baseURL = "https://openrouter.ai/api/v1"
 const ADAPTER = "openrouter"

-type OpenRouterString<Known extends string> = Known | (string & {})
-
 export interface OpenRouterProviderRouting {
  readonly [key: string]: unknown
  readonly order?: ReadonlyArray<string>
  readonly allow_fallbacks?: boolean
  readonly require_parameters?: boolean
-  readonly data_collection?: OpenRouterString<"allow" | "deny">
+  readonly data_collection?: OpenString<"allow" | "deny">
  readonly only?: ReadonlyArray<string>
  readonly ignore?: ReadonlyArray<string>
  readonly quantizations?: ReadonlyArray<string>
-  readonly sort?: OpenRouterString<"price" | "throughput" | "latency">
+  readonly sort?: OpenString<"price" | "throughput" | "latency">
  readonly max_price?: Readonly<{
    prompt?: number | string
    completion?: number | string
@@ -41,7 +39,7 @@ export type OpenRouterPlugin =
      id: "web"
      max_results?: number
      search_prompt?: string
-      engine?: OpenRouterString<"native" | "exa">
+      engine?: OpenString<"native" | "exa">
    }>
  | Readonly<{ id: "file-parser"; max_files?: number; pdf?: { engine?: string } }>
  | Readonly<{ id: "moderation" }>
@@ -58,7 +56,7 @@ export interface OpenRouterOptions {
  readonly reasoning?: Readonly<{
    enabled?: boolean
    exclude?: boolean
-    effort?: OpenRouterString<"none" | "minimal" | "low" | "medium" | "high" | "xhigh" | "max">
+    effort?: OpenString<"none" | "minimal" | "low" | "medium" | "high" | "xhigh" | "max">
    max_tokens?: number
  }>
  readonly usage?: boolean | Readonly<{ include: boolean }>
@@ -66,7 +64,7 @@ export interface OpenRouterOptions {
  readonly web_search_options?: Readonly<{
    max_results?: number
    search_prompt?: string
-    engine?: OpenRouterString<"native" | "exa">
+    engine?: OpenString<"native" | "exa">
  }>
 }

@@ -125,7 +123,7 @@ export const protocol = Protocol.make({
          return {
            ...body,
            messages,
-            ...bodyOptions(request.providerOptions),
+            ...bodyOptions(request.providerOptions, request.generation?.maxTokens),
          } as OpenRouterBody
        }),
      ),
@@ -145,7 +143,14 @@ const cacheControl = () => {
  }
 }

-const bodyOptions = (input: unknown) => {
+// OpenRouter forwards `reasoning.max_tokens` as the upstream thinking budget. Upstreams such as Anthropic and Alibaba
+// reject one that is not below the output limit; 1,024 is Anthropic's minimum budget.
+const fitReasoning = (reasoning: Record<string, unknown>, maxTokens: number | undefined) =>
+  typeof reasoning.max_tokens === "number"
+    ? { ...reasoning, max_tokens: ProviderShared.fitThinkingBudget(reasoning.max_tokens, maxTokens, 1_024) }
+    : reasoning
+
+const bodyOptions = (input: unknown, maxTokens: number | undefined) => {
  const openrouter = isRecord(input) ? input : {}
  const { usage, models, provider, plugins, web_search_options, debug, user, reasoning, promptCacheKey, ...options } =
    openrouter
@@ -164,7 +169,7 @@ const bodyOptions = (input: unknown) => {
    ...(isRecord(web_search_options) ? { web_search_options } : {}),
    ...(isRecord(debug) ? { debug } : {}),
    ...(typeof user === "string" ? { user } : {}),
-    ...(isRecord(reasoning) ? { reasoning } : {}),
+    ...(isRecord(reasoning) ? { reasoning: fitReasoning(reasoning, maxTokens) } : {}),
  }
 }

@@ -198,7 +203,7 @@ export const configure = (input: LanguageModelOptions = {}) => {
      auth: AuthOptions.bearer(input, "OPENROUTER_API_KEY"),
      baseURL: input.baseURL ?? baseURL,
      headers: input.headers,
-      http: input.http === undefined ? undefined : HttpOptions.make(input.http),
+      http: HttpOptions.make(input.http),
    })
  return {
    id,
@@ -0,0 +1,30 @@
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import { MediaRoute } from "../route/media.js"
+import { type HttpOptions, ProviderID, type ModelID } from "../schema/index.js"
+import { ReplicateImages } from "../protocols/replicate-images.js"
+
+export type { ReplicateImageOptions } from "../protocols/replicate-images.js"
+
+export const id = ProviderID.make("replicate")
+
+export type Config = ProviderAuthOption<"optional"> & {
+  readonly baseURL?: string
+  /** `{ Prefer: "wait=60" }` holds the submission open until the prediction finishes (up to 60 seconds). */
+  readonly headers?: Record<string, string>
+  readonly http?: HttpOptions.Input
+}
+
+const auth = (options: ProviderAuthOption<"optional">) => AuthOptions.bearer(options, "REPLICATE_API_TOKEN")
+
+export const configure = (input: Config = {}) => {
+  const media = MediaRoute.deployment(input, auth(input))
+  const image = (modelID: string | ModelID) => ReplicateImages.model({ ...media, id: modelID })
+  return {
+    id,
+    image,
+    configure,
+  }
+}
+
+export const provider = configure()
+export const image = provider.image
@@ -0,0 +1,29 @@
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import { MediaRoute } from "../route/media.js"
+import { type HttpOptions, ProviderID, type ModelID } from "../schema/index.js"
+import { RunwayVideo } from "../protocols/runway-video.js"
+
+export type { RunwayVideoOptions } from "../protocols/runway-video.js"
+
+export const id = ProviderID.make("runway")
+
+export type Config = ProviderAuthOption<"optional"> & {
+  readonly baseURL?: string
+  readonly headers?: Record<string, string>
+  readonly http?: HttpOptions.Input
+}
+
+const auth = (options: ProviderAuthOption<"optional">) => AuthOptions.bearer(options, "RUNWAYML_API_SECRET")
+
+export const configure = (input: Config = {}) => {
+  const media = MediaRoute.deployment(input, auth(input))
+  const video = (modelID: string | ModelID) => RunwayVideo.model({ ...media, id: modelID })
+  return {
+    id,
+    video,
+    configure,
+  }
+}
+
+export const provider = configure()
+export const video = provider.video
@@ -0,0 +1,30 @@
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import { MediaRoute } from "../route/media.js"
+import { type HttpOptions, ProviderID, type ModelID } from "../schema/index.js"
+import { StabilityImages } from "../protocols/stability-images.js"
+
+export type { StabilityImageOptions, StabilityUpscaleOptions } from "../protocols/stability-images.js"
+
+export const id = ProviderID.make("stability")
+
+export type Config = ProviderAuthOption<"optional"> & {
+  readonly baseURL?: string
+  readonly headers?: Record<string, string>
+  readonly http?: HttpOptions.Input
+}
+
+const auth = (options: ProviderAuthOption<"optional">) => AuthOptions.bearer(options, "STABILITY_API_KEY")
+
+export const configure = (input: Config = {}) => {
+  const media = MediaRoute.deployment(input, auth(input))
+  return {
+    id,
+    image: (modelID: string | ModelID) => StabilityImages.model({ ...media, id: modelID }),
+    upscale: () => StabilityImages.upscaleModel(media),
+    configure,
+  }
+}
+
+export const provider = configure()
+export const image = provider.image
+export const upscale = provider.upscale
@@ -20,7 +20,7 @@ export const configure = (input: Options = {}) => {
      auth: AuthOptions.bearer(input, "TYPESAFE_API_KEY"),
      baseURL: input.baseURL ?? baseURL,
      headers: input.headers,
-      http: input.http === undefined ? undefined : HttpOptions.make(input.http),
+      http: HttpOptions.make(input.http),
    })
  return { id, experimental: { evaluation }, configure }
 }
@@ -67,7 +67,7 @@ export const configure = (input: Options = {}) => {
    EvaluationModel.make<EvaluationOptions>({
      id: modelID,
      provider: id,
-      http: input.http === undefined ? undefined : HttpOptions.make(input.http),
+      http: HttpOptions.make(input.http),
      route: {
        id: "vercel-evaluation",
        evaluate: (req, send) =>
@@ -1,11 +1,13 @@
 import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
 import { Route, type RouteDefaultsInput } from "../route/client.js"
 import { Endpoint } from "../route/endpoint.js"
-import { HttpOptions, ProviderID, type ModelID } from "../schema/index.js"
+import { MediaRoute } from "../route/media.js"
+import { ProviderID, type ModelID } from "../schema/index.js"
 import { OpenAIChat } from "../protocols/openai-chat.js"
 import { OpenResponsesChannel } from "../protocols/open-responses-channel.js"
 import { XAIResponses } from "../protocols/xai-responses.js"
 import { XAIImages } from "../protocols/xai-images.js"
+import { XAIVideo } from "../protocols/xai-video.js"
 import type { OpenAIOptionsInput } from "./openai-options.js"
 import type { ProviderPackage } from "../provider-package.js"

@@ -27,6 +29,7 @@ export type Settings = ProviderPackage.Settings &
  }

 export type { XAIImageOptions } from "../protocols/xai-images.js"
+export type { XAIVideoOptions } from "../protocols/xai-video.js"

 const RESPONSES_WEBSOCKET_ROTATE_AFTER_MS = 24 * 60 * 1000

@@ -87,20 +90,14 @@ export const configure = (input: LanguageModelOptions = {}) => {
  const chatRoute = configuredChatRoute(input)
  const responses = (modelID: string | ModelID) => responsesRoute.model<XAIProviderOptionsInput>({ id: modelID })
  const chat = (modelID: string | ModelID) => chatRoute.model<XAIProviderOptionsInput>({ id: modelID })
-  const image = (modelID: string | ModelID) =>
-    XAIImages.model({
-      id: modelID,
-      auth: auth(input),
-      baseURL: input.baseURL ?? baseURL,
-      headers: input.headers,
-      http: input.http === undefined ? undefined : HttpOptions.make(input.http),
-    })
+  const media = MediaRoute.deployment(input, auth(input))
  return {
    id,
    model: responses,
    responses,
    chat,
-    image,
+    image: (modelID: string | ModelID) => XAIImages.model({ ...media, id: modelID }),
+    video: (modelID: string | ModelID) => XAIVideo.model({ ...media, id: modelID }),
    configure,
  }
 }
@@ -121,3 +118,4 @@ export const model: ProviderPackage.Definition<
 export const responses = provider.responses
 export const chat = provider.chat
 export const image = provider.image
+export const video = provider.video
@@ -5,7 +5,8 @@ import { OpenAIChat } from "../protocols/openai-chat.js"
 import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
 import { Route, type RouteDefaultsInput } from "../route/client.js"
 import { Endpoint } from "../route/endpoint.js"
-import { HttpOptions, ProviderID, type ModelID } from "../schema/index.js"
+import { MediaRoute } from "../route/media.js"
+import { ProviderID, type ModelID } from "../schema/index.js"

 export const id = ProviderID.make("zai")

@@ -48,14 +49,8 @@ export const configure = (input: Config = {}) => {
        auth: auth(input),
      })
      .model<ChatOptionsInput>({ id: modelID, compatibility: ZAIChat.compatibility })
-  const image = (modelID: string | ModelID) =>
-    ZAIImages.model({
-      id: modelID,
-      auth: auth(input),
-      baseURL: input.baseURL,
-      headers: input.headers,
-      http: input.http === undefined ? undefined : HttpOptions.make(input.http),
-    })
+  const media = MediaRoute.deployment(input, auth(input))
+  const image = (modelID: string | ModelID) => ZAIImages.model({ ...media, id: modelID })

  return {
    id,
@@ -1,12 +1,15 @@
-import { Config, Effect, Redacted } from "effect"
+import { Config, Effect, Option, Redacted } from "effect"
 import { Headers } from "effect/unstable/http"
-import { AuthenticationError, InvalidRequestError, AIError, type HttpOptions } from "../schema/index.js"
+import { AuthenticationError, AIError, type HttpOptions } from "../schema/index.js"

 export class MissingCredentialError extends Error {
  readonly _tag = "MissingCredentialError"

-  constructor(readonly source: string) {
-    super(`Missing auth credential: ${source}`)
+  constructor(
+    readonly source: string,
+    message = `Missing auth credential: ${source}`,
+  ) {
+    super(message)
  }
 }

@@ -16,7 +19,7 @@ type Secret = string | Redacted.Redacted | Config.Config<string | Redacted.Redac

 export interface AuthInput {
  readonly request: { readonly http?: HttpOptions }
-  readonly method: "POST" | "GET"
+  readonly method: "POST" | "GET" | "PUT" | "DELETE"
  readonly url: string
  readonly body: string
  readonly headers: Headers.Headers
@@ -89,7 +92,14 @@ export const optional = (secret: Secret | undefined, source = "optional value")
    ? credential(Effect.fail(new MissingCredentialError(source)))
    : credentialFromSecret(secret, source)

-export const config = (name: string) => credentialFromSecret(Config.redacted(name), name)
+export const config = (name: string) =>
+  credential(
+    Effect.gen(function* () {
+      const secret = yield* Config.option(Config.redacted(name))
+      if (Option.isSome(secret) && Redacted.value(secret.value) !== "") return secret.value
+      return yield* Effect.fail(new MissingCredentialError(name, `${name} is not set`))
+    }),
+  )

 export const effect = (load: Effect.Effect<Redacted.Redacted, CredentialError>) => credential(load)

@@ -134,16 +144,21 @@ export function bearerHeader(name: string, source?: Secret | Credential) {
  return render(source)
 }

+/** `Authorization: <scheme> <secret>` for providers whose scheme is not `Bearer`, such as fal's `Key`. */
+export function scheme(name: string): (source: Secret | Credential) => Definition
+export function scheme(name: string, source: Secret | Credential): Definition
+export function scheme(name: string, source?: Secret | Credential) {
+  const render = (input: Secret | Credential) =>
+    fromCredential(credentialInput(input), (secret) => ({ authorization: `${name} ${secret}` }))
+  if (source === undefined) return render
+  return render(source)
+}
+
 const toAIError = (error: AuthError): AIError => {
-  if (error instanceof MissingCredentialError || error instanceof Config.ConfigError) {
-    return new AIError({
-      reason:
-        error instanceof MissingCredentialError
-          ? new AuthenticationError({ message: error.message, cause: error })
-          : new InvalidRequestError({ message: `Failed to resolve auth config: ${error.message}`, cause: error }),
-    })
-  }
-  return error
+  if (error instanceof AIError) return error
+  const message =
+    error instanceof MissingCredentialError ? error.message : `Failed to resolve auth config: ${error.message}`
+  return new AIError({ reason: new AuthenticationError({ message, cause: error }) })
 }

 export const toEffect =
@@ -11,7 +11,8 @@ import { applyEffortUpdates } from "../effort-updates.js"
 import { normalizeToolHistory } from "../tool-history.js"
 import { sanitizeSurrogates } from "../utils/sanitize.js"
 import * as ProviderShared from "../protocols/shared.js"
-import type { ProtocolID, ProviderOptions } from "../schema/index.js"
+import { ToolSchemaProjection } from "../protocols/utils/tool-schema.js"
+import type { LanguageModelSanitizerCompatibility, ProtocolID, ProviderOptions } from "../schema/index.js"
 import {
  AIError,
  CompactionResponse,
@@ -57,6 +58,7 @@ export interface Route<
  readonly defaults: RouteDefaults
  readonly body: RouteBody<Body>
  readonly supportsEffortUpdates?: (request: LLMRequest) => boolean
+  readonly sanitizer?: LanguageModelSanitizerCompatibility
  readonly with: {
    <Next extends CompactionOperations | undefined>(
      patch: RoutePatch<Body, Prepared> & { readonly compact: Next },
@@ -152,7 +154,7 @@ const mergeRouteDefaults = (base: RouteDefaults | undefined, patch: RouteDefault
    providerOptions: mergeProviderOptions(base?.providerOptions, patch.providerOptions),
    http: mergeHttpOptions(
      base?.http,
-      httpOptions(patch.http),
+      HttpOptions.make(patch.http),
      headers === undefined ? undefined : new HttpOptions({ headers }),
    ),
  }
@@ -172,11 +174,6 @@ const mergeHeaders = (...items: ReadonlyArray<Record<string, string> | undefined
 export const generationOptions = (input: GenerationOptions.Input | undefined) =>
  input === undefined ? undefined : GenerationOptions.make(input)

-export const httpOptions = (input: HttpOptionsInput | undefined) => {
-  if (input === undefined) return input
-  return HttpOptions.make(input)
-}
-
 export interface Interface {
  readonly compact: CompactMethod
  readonly stream: StreamMethod
@@ -261,7 +258,9 @@ const unsupportedCompaction = (request: LLMRequest, mechanism: string | undefine
  })
 }

-export class Service extends Context.Service<Service, Interface>()("@opencode/LLMClient") {}
+export class LLMClientService extends Context.Service<LLMClientService, Interface>()("@opencode/LLMClient") {}
+export const Service = LLMClientService
+export type Service = LLMClientService

 const resolveRequestOptions = (request: LLMRequest) => {
  const messages = normalizeToolHistory(request.messages)
@@ -391,6 +390,7 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
      defaults: routeInput.defaults ?? {},
      body: protocol.body,
      supportsEffortUpdates: protocol.supportsEffortUpdates,
+      sanitizer: protocol.sanitizer,
      with: (patch: RoutePatch<Body, Prepared>) => {
        const { compact, id, provider, providerMetadataKey, auth, transport, endpoint, ...defaults } = patch
        return build({
@@ -562,7 +562,9 @@ const prepareRequest = (request: LLMRequest) => {
      tool.type === "tool" ? tool : { ...tool, tools: dedupe(tool.tools) },
    )
  const resolved = applyCachePolicy(
-    applyEffortUpdates(LLMRequest.update(sanitized, { tools: dedupe(sanitized.tools) })),
+    applyEffortUpdates(
+      LLMRequest.update(sanitized, { tools: ToolSchemaProjection.tools(dedupe(sanitized.tools), sanitized.model) }),
+    ),
  )
  const headers = resolved.model.route.headers?.({ request: resolved })
  return headers === undefined
@@ -1,5 +1,4 @@
 import type { LLMRequest } from "../schema/index.js"
-import * as ProviderShared from "../protocols/shared.js"

 export interface EndpointInput<Body, Request = LLMRequest> {
  readonly request: Request
@@ -47,6 +46,8 @@ export const merge = <Body, Request = LLMRequest>(
  query: patch.query === undefined ? base.query : { ...base.query, ...patch.query },
 })

+export const trimBaseUrl = (value: string) => value.replace(/\/+$/, "")
+
 const renderPart = <Body, Request>(part: EndpointPart<Body, Request>, input: EndpointInput<Body, Request>) =>
  typeof part === "function" ? part(input) : part

@@ -54,7 +55,7 @@ export const render = <Body, Request = LLMRequest>(
  endpoint: Definition<Body, Request>,
  input: EndpointInput<Body, Request>,
 ) => {
-  const url = new URL(`${ProviderShared.trimBaseUrl(endpoint.baseURL ?? "")}${renderPart(endpoint.path, input)}`)
+  const url = new URL(`${trimBaseUrl(endpoint.baseURL ?? "")}${renderPart(endpoint.path, input)}`)
  for (const [key, value] of Object.entries(endpoint.query ?? {})) url.searchParams.set(key, value)
  return url
 }
@@ -19,4 +19,8 @@ export type HttpMiddleware = (
  handler: HttpHandler,
 ) => Effect.Effect<HttpClientResponse.HttpClientResponse, Error>

-export class Service extends Context.Service<Service, Interface>()("@opencode/AI/RequestExecutor") {}
+export class RequestExecutorService extends Context.Service<RequestExecutorService, Interface>()(
+  "@opencode/AI/RequestExecutor",
+) {}
+export const Service = RequestExecutorService
+export type Service = RequestExecutorService
@@ -255,4 +255,20 @@ export const layer: Layer.Layer<Service, never, HttpClient.HttpClient> = Layer.e

 export const fetchLayer = layer.pipe(Layer.provide(FetchHttpClient.layer))

+/** Run `fn` on every request: it sees the raw response before status classification, inside middleware already on `executor`, and outside per-call middleware. */
+export const middleware = (fn: HttpMiddleware, executor: Layer.Layer<Service> = fetchLayer): Layer.Layer<Service> =>
+  Layer.effect(
+    Service,
+    Effect.gen(function* () {
+      const inner = yield* Service
+      return Service.of({
+        execute: (request, next) =>
+          inner.execute(
+            request,
+            next === undefined ? fn : (input, handler) => fn(input, (forwarded) => next(forwarded, handler)),
+          ),
+      })
+    }),
+  ).pipe(Layer.provide(executor))
+
 export * as RequestExecutor from "./executor.js"
@@ -1,6 +1,6 @@
-import type { Stream } from "effect"
-import * as ProviderShared from "../protocols/shared.js"
-import type { AIError } from "../schema/index.js"
+import { Effect, Stream } from "effect"
+import { makeParser, type Event } from "effect/unstable/encoding/Sse"
+import { AIError, InvalidProviderOutputError } from "../schema/index.js"

 /**
 * Decode a streaming HTTP response body into provider-protocol frames.
@@ -12,6 +12,8 @@ import type { AIError } from "../schema/index.js"
 *   `[DONE]`; protocols that use it as a terminal select `sseWithDone`.
 * - AWS event stream — length-prefixed binary frames with CRC checksums.
 *   Each emitted frame is one parsed binary event record.
+ * - Media streams — newline-delimited JSON (`lines`) or the whole body as one
+ *   frame (`document`); chunked binary bodies need no framing.
 *
 * The frame type is opaque to this layer; the protocol's event schema decodes
 * each frame before its state machine handles it.
@@ -23,19 +25,90 @@ export interface Definition<Frame> {
  readonly body?: (frame: Frame) => string | undefined
 }

+/**
+ * `framing` step for Server-Sent Events. Decodes UTF-8, runs the SSE channel
+ * decoder, optionally filters named events, and drops empty events and known
+ * keepalives that proxies send as data. `[DONE]` is dropped by default or
+ * retained for protocols that use it as their stream boundary. Retry control events are ignored without
+ * interrupting the stream. Decoder failures become provider output errors so
+ * the public error channel stays `AIError`.
+ */
+export const sseFraming = (
+  bytes: Stream.Stream<Uint8Array, AIError>,
+  events?: ReadonlySet<string>,
+  includeDone = false,
+): Stream.Stream<string, AIError> =>
+  bytes.pipe(
+    Stream.decodeText(),
+    Stream.mapAccumEffect(
+      () => {
+        const output: Event[] = []
+        return {
+          output,
+          parser: makeParser((event) => {
+            if (event._tag === "Event") output.push(event)
+          }),
+        }
+      },
+      (state, chunk) =>
+        Effect.gen(function* () {
+          const error = state.parser.feed(chunk)
+          if (error)
+            return yield* new AIError({
+              reason: new InvalidProviderOutputError({
+                route: "sse",
+                message: error.message,
+                body: chunk,
+                cause: error,
+              }),
+            })
+          return [state, state.output.splice(0)] as const
+        }),
+    ),
+    Stream.filter(
+      (event) =>
+        (events === undefined || events.has(event.event)) &&
+        event.data.length > 0 &&
+        // Some OpenAI-compatible proxies serialize an empty flush as a bare
+        // `data: null`, between events or after `[DONE]`. No protocol has a
+        // null event, so it carries nothing and must not abort the stream.
+        event.data !== "null" &&
+        // Vertex AI partner models (e.g. `xai/grok-4.6`) send their SSE
+        // keepalive comment as `data: : keepalive` while reasoning.
+        event.data !== ": keepalive" &&
+        (event.data !== "[DONE]" || includeDone || (events !== undefined && event.event !== "message")),
+    ),
+    Stream.map((event) => event.data),
+  )
+
 /** Server-Sent Events framing. Used by every JSON-streaming HTTP provider. */
-export const sse: Definition<string> = { id: "sse", frame: ProviderShared.sseFraming }
+export const sse: Definition<string> = { id: "sse", frame: sseFraming }

 /** Server-Sent Events framing that retains the conventional `[DONE]` sentinel. */
 export const sseWithDone: Definition<string> = {
  id: "sse",
-  frame: (bytes) => ProviderShared.sseFraming(bytes, undefined, true),
+  frame: (bytes) => sseFraming(bytes, undefined, true),
 }

 /** SSE framing restricted to protocol-recognized event names. */
 export const sseEvents = (events: ReadonlySet<string>): Definition<string> => ({
  id: "sse",
-  frame: (bytes) => ProviderShared.sseFraming(bytes, events),
+  frame: (bytes) => sseFraming(bytes, events),
 })

+export const lines: Definition<string> = {
+  id: "lines",
+  frame: (bytes) =>
+    bytes.pipe(
+      Stream.decodeText(),
+      Stream.splitLines,
+      Stream.filter((line) => line.trim().length > 0),
+    ),
+}
+
+export const document: Definition<string> = {
+  id: "document",
+  frame: (bytes) => Stream.fromEffect(Stream.mkString(bytes.pipe(Stream.decodeText()))),
+}
+
 export * as Framing from "./framing.js"
@@ -7,7 +7,7 @@ export type {
  RouteDefaultsInput,
  AnyRoute,
  Interface as LLMClientShape,
-  Service as LLMClientService,
+  LLMClientService,
  StreamOptions,
  CompactMethod,
  CompactionOperations,
@@ -1,17 +1,54 @@
-import { Effect, Schema } from "effect"
+import { Clock, Duration, Effect, Schema, type Stream } from "effect"
 import { HttpClientResponse } from "effect/unstable/http"
-import { AIError, HttpContext, InvalidProviderOutputError } from "../schema/index.js"
+import type { Snapshot, Status } from "../generation.js"
+import { Media } from "../media.js"
+import type { AuthInput } from "./auth.js"
+import {
+  AIError,
+  ContentPolicyError,
+  HttpContext,
+  InvalidProviderOutputError,
+  InvalidRequestError,
+  ProviderID,
+  ProviderInternalError,
+  UnsupportedOperationError,
+} from "../schema/index.js"

 // ---------------------------------------------------------------------------
 // Bodies
 // ---------------------------------------------------------------------------

-export type Body =
-  | { readonly type: "json"; readonly value: Record<string, unknown> }
-  | { readonly type: "multipart"; readonly value: FormData }
+/** Array values become repeated parameters (`keyterm=a&keyterm=b`). */
+export type Query = Readonly<Record<string, string | ReadonlyArray<string>>>

-export const json = (value: Record<string, unknown>): Body => ({ type: "json", value })
+/** `query` is appended to the endpoint URL before the route and caller `http.query` overlays. */
+export type Body =
+  | { readonly type: "json"; readonly value: Record<string, unknown>; readonly query?: Query }
+  | { readonly type: "multipart"; readonly value: FormData }
+  | {
+      readonly type: "binary"
+      readonly value: Uint8Array
+      readonly contentType: string
+      readonly query?: Query
+    }
+
+export const json = (value: Record<string, unknown>, query?: Query): Body => ({
+  type: "json",
+  value,
+  query,
+})
 export const multipart = (value: FormData): Body => ({ type: "multipart", value })
+export const binary = (value: Uint8Array, contentType: string, query?: Query): Body => ({
+  type: "binary",
+  value,
+  contentType,
+  query,
+})
+
+export type Send = (path: string, body: Body) => Effect.Effect<HttpClientResponse.HttpClientResponse, AIError>
+
+/** Runs after unsupported-field rejection and before `body.from`, for providers that need an upload first. */
+export type Prepare<Request> = (request: Request, send: Send) => Effect.Effect<Request, AIError>

 // ---------------------------------------------------------------------------
 // Protocol kinds
@@ -26,7 +63,7 @@ export interface DecodeContext<Request> {
 export interface Inline<Request, Response> {
  readonly kind: "inline"
  readonly id: string
-  readonly name: string
+  readonly provider: ProviderID
  /** Common request fields this protocol cannot lower; the route rejects them before `body.from` runs. */
  readonly unsupported?: ReadonlyArray<keyof Request & string>
  readonly body: { readonly from: (request: Request) => Effect.Effect<Body, AIError> }
@@ -39,11 +76,113 @@ export interface Inline<Request, Response> {
 }

 export const inline = <Request, Response>(
-  input: Omit<Inline<Request, Response>, "kind">,
-): Inline<Request, Response> => ({
-  kind: "inline",
-  ...input,
-})
+  route: Identity,
+  input: Omit<Inline<Request, Response>, "kind" | "id" | "provider">,
+): Inline<Request, Response> => ({ kind: "inline", id: route.id, provider: route.provider, ...input })
+
+/** What `start` learned from the submission response: the route-owned handle plus the first observation. */
+export interface Started<Token> {
+  readonly token: Token
+  readonly snapshot: Snapshot
+}
+
+/**
+ * A follow-up call's inputs: the decoded token and the auth headers the route sent, so a protocol can attach them
+ * to output URLs that require the same credentials to download (Veo). `materialize` downloads an output through the
+ * route's executor, for URLs that expire too soon to hand back (BFL).
+ */
+export interface PollContext<Token> {
+  readonly token: Token
+  readonly auth: Record<string, string>
+  readonly materialize: (asset: Media.Asset) => Effect.Effect<Media.Asset, AIError>
+}
+
+/**
+ * Submit, then poll. `start` posts the body to the route endpoint; `status`, `result`, and `cancel` are follow-up
+ * calls addressed by the token. Paths are relative to the route base URL unless the provider hands back absolute
+ * URLs (fal `status_url`), in which case they are used verbatim. `result` is always its own GET: providers that
+ * return the output inside the status body (Veo, xAI, Runway) point `result.path` at the status path and decode the
+ * same document, so `Generation.await` and `Video.resume(...).await()` behave identically everywhere.
+ */
+export interface Queued<Request, Response, Token> {
+  readonly kind: "queued"
+  readonly id: string
+  readonly provider: ProviderID
+  /** Common request fields this protocol cannot lower; the route rejects them before `start.body.from` runs. */
+  readonly unsupported?: ReadonlyArray<keyof Request & string>
+  /** Serializable handle. `Generation.token` carries the encoded form so it can be persisted and resumed elsewhere. */
+  readonly token: Schema.Codec<Token, unknown>
+  readonly start: {
+    readonly prepare?: Prepare<Request>
+    readonly body: { readonly from: (request: Request) => Effect.Effect<Body, AIError> }
+    readonly decode: (
+      response: HttpClientResponse.HttpClientResponse,
+      context: DecodeContext<Request>,
+    ) => Effect.Effect<Started<Token>, AIError>
+  }
+  readonly status: {
+    readonly path: (token: Token) => string
+    readonly decode: (
+      response: HttpClientResponse.HttpClientResponse,
+      context: PollContext<Token>,
+    ) => Effect.Effect<Snapshot, AIError>
+  }
+  readonly result: {
+    readonly path: (token: Token) => string
+    readonly decode: (
+      response: HttpClientResponse.HttpClientResponse,
+      context: PollContext<Token>,
+    ) => Effect.Effect<Response, AIError>
+  }
+  readonly cancel?: {
+    readonly method: AuthInput["method"]
+    readonly path: (token: Token) => string
+    /**
+     * Fetch a fresh status first and skip the call for terminal generations, for providers whose cancel endpoint
+     * destroys finished work (Runway's `DELETE /v1/tasks/{id}` deletes completed tasks and their outputs).
+     */
+    readonly activeOnly?: boolean
+  }
+}
+
+export const queued = <Request, Response, Token>(
+  route: Identity,
+  input: Omit<Queued<Request, Response, Token>, "kind" | "id" | "provider">,
+): Queued<Request, Response, Token> => ({ kind: "queued", id: route.id, provider: route.provider, ...input })
+
+export type Mode = "generate" | "stream"
+
+export type Addressed<Request> = Request & { readonly mode: Mode }
+
+export interface ResponseContext<Request> extends DecodeContext<Addressed<Request>> {
+  readonly http: HttpContext
+}
+
+/**
+ * One request whose body is parsed incrementally, like LLM protocols: `frames` → `step`* → `finish`. `generate` and
+ * `stream` share this state machine; `request.mode` lets a protocol pick a different body, path, or framing.
+ */
+export interface Streamed<Request, Event, Frame, State> {
+  readonly kind: "stream"
+  readonly id: string
+  readonly provider: ProviderID
+  /** Common request fields this protocol cannot lower; the route rejects them before `body.from` runs. */
+  readonly unsupported?: ReadonlyArray<keyof Request & string>
+  readonly body: { readonly from: (request: Addressed<Request>) => Effect.Effect<Body, AIError> }
+  readonly frames: (
+    bytes: Stream.Stream<Uint8Array, AIError>,
+    context: DecodeContext<Addressed<Request>>,
+  ) => Stream.Stream<Frame, AIError>
+  readonly initial: () => State
+  readonly step: (state: State, frame: Frame) => Effect.Effect<readonly [State, ReadonlyArray<Event>], AIError>
+  /** Emit exactly one terminal event, or fail when the provider stopped before completing. */
+  readonly finish: (state: State, context: ResponseContext<Request>) => Effect.Effect<ReadonlyArray<Event>, AIError>
+}
+
+export const stream = <Request, Event, Frame, State>(
+  route: Identity,
+  input: Omit<Streamed<Request, Event, Frame, State>, "kind" | "id" | "provider">,
+): Streamed<Request, Event, Frame, State> => ({ kind: "stream", id: route.id, provider: route.provider, ...input })

 // ---------------------------------------------------------------------------
 // Response helpers
@@ -52,44 +191,122 @@ export const inline = <Request, Response>(
 const context = (response: HttpClientResponse.HttpClientResponse) =>
  new HttpContext({ url: response.request.url, status: response.status, headers: response.headers })

-/** Read a text body while retaining the original payload and HTTP context on every downstream error. */
-export const text = Effect.fn("MediaProtocol.text")(function* (
-  route: string,
-  name: string,
-  response: HttpClientResponse.HttpClientResponse,
-) {
-  const http = context(response)
-  const body = yield* response.text.pipe(
-    Effect.mapError(
-      (cause) =>
+/** One protocol's route id, display name, and provider, with the decoders and errors that carry them. */
+export const identity = (input: { readonly id: string; readonly name: string; readonly provider: string }) => {
+  const provider = ProviderID.make(input.provider)
+  const frameError = (message: string, body?: string, cause?: unknown) =>
+    new AIError({ reason: new InvalidProviderOutputError({ route: input.id, message, body, cause }) })
+
+  /**
+   * Read a text body while retaining the original payload and HTTP context on every downstream error. `invalid` is a
+   * malformed provider document; `ended` is a generation that reached a terminal status without output (`failed` is
+   * provider-side, `cancelled`/`expired` mean the result will never exist); `pending` is a `result()` read before the
+   * generation finished, which is caller misuse; `contentPolicy` is a moderated result.
+   */
+  const text = Effect.fn("MediaProtocol.text")(function* (response: HttpClientResponse.HttpClientResponse) {
+    const http = context(response)
+    const body = yield* response.text.pipe(
+      Effect.mapError(
+        (cause) =>
+          new AIError({
+            reason: new InvalidProviderOutputError({
+              route: input.id,
+              message: `Failed to read the ${input.name} response`,
+              http,
+              cause,
+            }),
+          }),
+      ),
+    )
+    return {
+      body,
+      http,
+      invalid: (message: string, cause?: unknown) =>
+        new AIError({ reason: new InvalidProviderOutputError({ route: input.id, message, body, http, cause }) }),
+      ended: (status: Exclude<Status, "queued" | "running" | "completed">, message: string) =>
        new AIError({
-          reason: new InvalidProviderOutputError({
-            route,
-            message: `Failed to read the ${name} response`,
+          reason:
+            status === "failed"
+              ? new ProviderInternalError({ message, body, http })
+              : new InvalidRequestError({ message, body, http }),
+        }),
+      pending: (id: string) =>
+        new AIError({
+          reason: new InvalidRequestError({
+            message: `${input.name} generation ${id} has not finished; await it before reading the result`,
+            body,
            http,
-            cause,
          }),
        }),
-    ),
-  )
-  return {
-    body,
-    http,
-    invalid: (message: string, cause?: unknown) =>
-      new AIError({ reason: new InvalidProviderOutputError({ route, message, body, http, cause }) }),
-  }
-})
-
-/** Read and Schema-decode a JSON body. Decode failures keep the raw body as `reason.body`. */
-export const decodeJson = <A>(route: string, name: string, schema: Schema.Codec<A, unknown>) => {
-  const decode = Schema.decodeUnknownEffect(Schema.fromJsonString(schema))
-  return Effect.fn("MediaProtocol.decodeJson")(function* (response: HttpClientResponse.HttpClientResponse) {
-    const output = yield* text(route, name, response)
-    const value = yield* decode(output.body).pipe(
-      Effect.mapError((cause) => output.invalid(`${name} returned an invalid response`, cause)),
-    )
-    return { ...output, value }
+      contentPolicy: (message: string) => new AIError({ reason: new ContentPolicyError({ message, body, http }) }),
+    }
  })
+
+  /** Read and Schema-decode a JSON body. Decode failures keep the raw body as `reason.body`. */
+  const decodeJson = <A>(schema: Schema.Codec<A, unknown>) => {
+    const decode = Schema.decodeUnknownEffect(Schema.fromJsonString(schema))
+    return Effect.fn("MediaProtocol.decodeJson")(function* (response: HttpClientResponse.HttpClientResponse) {
+      const output = yield* text(response)
+      const value = yield* decode(output.body).pipe(
+        Effect.mapError((cause) => output.invalid(`${input.name} returned an invalid response`, cause)),
+      )
+      return { ...output, value }
+    })
+  }
+
+  return {
+    id: input.id,
+    name: input.name,
+    provider,
+    text,
+    decodeJson,
+    /** Decode a submission response into the token and first snapshot. */
+    decodeStarted: <A, Token>(schema: Schema.Codec<A, unknown>, started: (value: A) => Started<Token>) => {
+      const decode = decodeJson(schema)
+      return (response: HttpClientResponse.HttpClientResponse) =>
+        decode(response).pipe(Effect.map((output) => started(output.value)))
+    },
+    /** Schema-decode one JSON stream frame. Decode failures keep the frame as `reason.body`. */
+    decodeFrame: <A>(schema: Schema.Codec<A, unknown>) => {
+      const decode = Schema.decodeUnknownEffect(Schema.fromJsonString(schema))
+      return (frame: string) =>
+        decode(frame).pipe(
+          Effect.mapError((cause) => frameError(`${input.name} sent an invalid stream event`, frame, cause)),
+        )
+    },
+    /** A stream-time failure; the frame stays on `reason.body`. */
+    frameError,
+    incomplete: () =>
+      new AIError({
+        reason: new InvalidProviderOutputError({
+          route: input.id,
+          message: "The provider response ended unexpectedly.",
+          classification: "incomplete-stream",
+        }),
+      }),
+    unsupported: (operation: string, message: string) =>
+      new AIError({ reason: new UnsupportedOperationError({ operation, provider, route: input.id, message }) }),
+  }
 }

+export type Identity = ReturnType<typeof identity>
+
+export type Output = Effect.Success<ReturnType<Identity["text"]>>
+
+/** Map a provider status string through the protocol's table; unknown values are an invalid provider document. */
+export const status = <Table extends Record<string, Status>>(
+  table: Table,
+  raw: string,
+  output: Output,
+): Effect.Effect<Status, AIError> => {
+  if (!Object.hasOwn(table, raw)) return Effect.fail(output.invalid(`Unknown generation status "${raw}"`))
+  return Effect.succeed(table[raw])
+}
+
+/** A `url` asset whose provider-declared retention window starts now. */
+export const expiringUrl = (url: string, retention: Duration.Duration, options?: Parameters<typeof Media.url>[1]) =>
+  Clock.currentTimeMillis.pipe(
+    Effect.map((now) => Media.url(url, { ...options, expiresAt: now + Duration.toMillis(retention) })),
+  )
+
 export * as MediaProtocol from "./media-protocol.js"
@@ -1,11 +1,22 @@
-import { Effect } from "effect"
-import { Headers, HttpClientRequest } from "effect/unstable/http"
-import { Auth } from "./auth.js"
+import { Effect, Schema, Stream } from "effect"
+import { Headers, HttpClientRequest, type HttpClientResponse } from "effect/unstable/http"
+import { Auth, type AuthInput } from "./auth.js"
 import { Endpoint } from "./endpoint.js"
-import type { Interface } from "./executor-service.js"
+import { RequestExecutorService, type Interface } from "./executor-service.js"
+import { RequestExecutor } from "./executor.js"
 import { MediaProtocol } from "./media-protocol.js"
-import { ProviderShared } from "../protocols/shared.js"
-import { AIError, HttpOptions, ProviderID, mergeHttpOptions } from "../schema/index.js"
+import { Generation, isTerminal } from "../generation.js"
+import type { Media } from "../media.js"
+import {
+  AIError,
+  AIErrorReason,
+  HttpOptions,
+  InvalidRequestError,
+  ProviderID,
+  UnsupportedOperationError,
+  mergeHttpOptions,
+} from "../schema/index.js"
+import { encodeJson } from "../utils/json.js"
 import { sanitizeSurrogates } from "../utils/sanitize.js"

 export type Execute = Interface["execute"]
@@ -25,73 +36,360 @@ export interface ModelInput {
  readonly http?: HttpOptions
 }

+/** A provider facade's `configure(...)` input as the `ModelInput` every media selector shares, minus the model id. */
+export const deployment = (
+  input: { readonly baseURL?: string; readonly headers?: Record<string, string>; readonly http?: HttpOptions.Input },
+  auth: Auth.Definition,
+): Omit<ModelInput, "id"> => ({
+  auth,
+  baseURL: input.baseURL,
+  headers: input.headers,
+  http: HttpOptions.make(input.http),
+})
+
 // ---------------------------------------------------------------------------
-// Route
+// Routes
 // ---------------------------------------------------------------------------

-export interface Route<Request extends MediaRequest, Response> {
+/** One request, one response. */
+export interface InlineRoute<Request extends MediaRequest, Response> {
+  readonly kind: "inline"
  readonly id: string
  readonly provider: ProviderID
  readonly protocol: string
  readonly generate: (request: Request, execute: Execute) => Effect.Effect<Response, AIError>
 }

-export interface MakeInput<Request extends MediaRequest, Response> {
+/** Submit, then poll through the returned `Generation`. */
+export interface QueuedRoute<Request extends MediaRequest, Response> {
+  readonly kind: "queued"
  readonly id: string
-  readonly provider: string | ProviderID
-  readonly protocol: MediaProtocol.Inline<Request, Response>
+  readonly provider: ProviderID
+  readonly protocol: string
+  readonly start: (request: Request, execute: Execute) => Effect.Effect<Generation<Response>, AIError>
+  /** Rebuild a handle from a persisted `Generation.token`; fails typed when the token is not this route's. */
+  readonly resume: (
+    model: MediaRequest["model"],
+    token: unknown,
+    execute: Execute,
+  ) => Effect.Effect<Generation<Response>, AIError>
+}
+
+/** One request whose response parses into events; `generate` runs the same stream and collects it. */
+export interface StreamRoute<Request extends MediaRequest, Event, Response> {
+  readonly kind: "stream"
+  readonly id: string
+  readonly provider: ProviderID
+  readonly protocol: string
+  readonly stream: (request: Request, execute: Execute) => Stream.Stream<Event, AIError>
+  readonly generate: (request: Request, execute: Execute) => Effect.Effect<Response, AIError>
+}
+
+export type AnyRoute<Request extends MediaRequest, Event, Response> =
+  | InlineRoute<Request, Response>
+  | StreamRoute<Request, Event, Response>
+  | QueuedRoute<Request, Response>
+
+export interface Composition<Request extends MediaRequest> {
  readonly endpoint: Endpoint.Definition<MediaProtocol.Body, Request>
  readonly auth: Auth.Definition
  /** Deployment headers applied before transport authentication. */
  readonly headers?: Record<string, string>
 }

+export interface InlineInput<Request extends MediaRequest, Response> extends Composition<Request> {
+  readonly protocol: MediaProtocol.Inline<Request, Response>
+}
+
+export interface QueuedInput<Request extends MediaRequest, Response, Token> extends Composition<Request> {
+  readonly protocol: MediaProtocol.Queued<Request, Response, Token>
+}
+
+export interface StreamInput<Request extends MediaRequest, Event, Response, Frame, State>
+  extends Composition<MediaProtocol.Addressed<Request>> {
+  readonly protocol: MediaProtocol.Streamed<Request, Event, Frame, State>
+  readonly collect: (events: ReadonlyArray<Event>) => Effect.Effect<Response, AIError>
+}
+
 /**
 * Compose an inline media protocol with an endpoint and auth into a runnable route. The route owns the transport
- * plumbing every media protocol would otherwise duplicate: option merging, surrogate sanitizing, unsupported-field rejection, URL and query
- * rendering, auth headers, JSON vs multipart encoding, and handing the response back to the protocol for decoding.
+ * plumbing every media protocol would otherwise duplicate: option merging, surrogate sanitizing, unsupported-field
+ * rejection, URL and query rendering, auth headers, JSON, multipart, or binary encoding, and handing responses back
+ * to the protocol.
 */
-export const make = <Request extends MediaRequest, Response>(
-  input: MakeInput<Request, Response>,
-): Route<Request, Response> => {
-  const provider = ProviderID.make(input.provider)
-  const routeHttp = input.headers === undefined ? undefined : new HttpOptions({ headers: input.headers })
-  const authorize = Auth.toEffect(input.auth)
+export const inline = <Request extends MediaRequest, Response>(
+  input: InlineInput<Request, Response>,
+): InlineRoute<Request, Response> => {
+  const transport = makeTransport(input)
  return {
-    id: input.id,
-    provider,
+    kind: "inline",
+    id: input.protocol.id,
+    provider: input.protocol.provider,
    protocol: input.protocol.id,
    generate: Effect.fn(`MediaRoute.generate`)(function* (request: Request, execute: Execute) {
-      yield* rejectUnsupported(input.id, provider, request, input.protocol.unsupported)
-      const http = mergeHttpOptions(routeHttp, request.model.http, request.http)
-      // Sanitize after merging so model-level overlays are covered; the model value is restored, not sanitized.
-      const resolved: Request = { ...sanitizeSurrogates({ ...request, http }), model: request.model }
-      const body = yield* input.protocol.body.from(resolved)
-      const url = Endpoint.render(input.endpoint, { request: resolved, body })
-      for (const [key, value] of Object.entries(http?.query ?? {})) url.searchParams.set(key, value)
-      const encoded = body.type === "json" ? ProviderShared.encodeJson(body.value) : "[multipart/form-data]"
-      const baseHeaders = Headers.fromInput(http?.headers)
-      const headers = yield* authorize({
-        request: resolved,
-        method: "POST",
-        url: url.toString(),
-        body: encoded,
-        // The HTTP client sets the multipart boundary; a caller-supplied content-type would corrupt it.
-        headers: body.type === "multipart" ? Headers.remove(baseHeaders, "content-type") : baseHeaders,
-      })
-      const transport = HttpClientRequest.post(url.toString()).pipe(
-        HttpClientRequest.setHeaders(headers),
-        body.type === "json"
-          ? HttpClientRequest.bodyText(encoded, "application/json")
-          : HttpClientRequest.bodyFormData(body.value),
+      const submitted = yield* transport.submit(
+        request,
+        { unsupported: input.protocol.unsupported, from: input.protocol.body.from },
+        execute,
      )
-      const response = yield* execute(transport)
-      return yield* input.protocol.response.decode(response, { request: resolved, body })
+      return yield* input.protocol.response.decode(submitted.response, submitted.context)
    }),
  }
 }

-/** Common fields are never silently dropped: a present field the protocol declared unsupported fails typed. */
+/**
+ * Compose a queued media protocol the same way, adding `start`/`resume` handles whose polls reuse the route's auth,
+ * deployment headers, and (for `start`) the request's `http` overlay. The token is decoded once at the boundary and
+ * closed over by the resulting `Generation.Route`.
+ */
+export const queued = <Request extends MediaRequest, Response, Token>(
+  input: QueuedInput<Request, Response, Token>,
+): QueuedRoute<Request, Response> => {
+  const transport = makeTransport(input)
+  const protocol = input.protocol
+  const decodeToken = Schema.decodeUnknownEffect(protocol.token)
+  // A protocol producing a token its own codec rejects is a programmer defect, not a provider error.
+  const encodeToken = Schema.encodeSync(protocol.token)
+
+  const generationRoute = (token: Token, http: HttpOptions | undefined, execute: Execute) => {
+    const materialize = (asset: Media.Asset) =>
+      asset.materialize().pipe(Effect.provideService(RequestExecutorService, { execute }))
+    const poll = <A>(operation: {
+      readonly path: (token: Token) => string
+      readonly decode: (
+        response: HttpClientResponse.HttpClientResponse,
+        context: MediaProtocol.PollContext<Token>,
+      ) => Effect.Effect<A, AIError>
+    }) =>
+      transport
+        .call("GET", operation.path(token), http, execute)
+        .pipe(Effect.flatMap((sent) => operation.decode(sent.response, { token, auth: sent.auth, materialize })))
+    const status = poll(protocol.status)
+    const cancel = protocol.cancel
+    const send =
+      cancel === undefined
+        ? undefined
+        : transport.call(cancel.method, cancel.path(token), http, execute).pipe(Effect.asVoid)
+    return {
+      status,
+      result: poll(protocol.result),
+      cancel:
+        send !== undefined && cancel?.activeOnly
+          ? status.pipe(Effect.flatMap((snapshot) => (isTerminal(snapshot.status) ? Effect.void : send)))
+          : send,
+    }
+  }
+
+  const start = Effect.fn("MediaRoute.start")(function* (request: Request, execute: Execute) {
+    const submitted = yield* transport.submit(
+      request,
+      { unsupported: protocol.unsupported, prepare: protocol.start.prepare, from: protocol.start.body.from },
+      execute,
+    )
+    const started = yield* protocol.start.decode(submitted.response, submitted.context)
+    const route = generationRoute(started.token, submitted.context.request.http, execute)
+    return new Generation(route, encodeToken(started.token), started.snapshot)
+  })
+
+  const resume = Effect.fn("MediaRoute.resume")(function* (
+    model: MediaRequest["model"],
+    raw: unknown,
+    execute: Execute,
+  ) {
+    const token = yield* decodeToken(raw).pipe(
+      Effect.mapError(
+        (cause) =>
+          new AIError({
+            reason: new InvalidRequestError({
+              message: `${protocol.id} cannot resume a generation from this token`,
+              cause,
+            }),
+          }),
+      ),
+    )
+    const route = generationRoute(token, transport.http(model), execute)
+    return new Generation(route, encodeToken(token), yield* route.status)
+  })
+
+  return { kind: "queued", id: protocol.id, provider: protocol.provider, protocol: protocol.id, start, resume }
+}
+
+/** Compose a streaming media protocol; `generate` runs the same stream in `generate` mode and folds it with `collect`. */
+export const stream = <Request extends MediaRequest, Event, Response, Frame, State>(
+  input: StreamInput<Request, Event, Response, Frame, State>,
+): StreamRoute<Request, Event, Response> => {
+  const transport = makeTransport(input)
+  const protocol = input.protocol
+  const events = (request: Request, execute: Execute, mode: MediaProtocol.Mode) =>
+    Stream.unwrap(
+      Effect.gen(function* () {
+        const submitted = yield* transport.submit(
+          { ...request, mode },
+          { unsupported: protocol.unsupported, from: protocol.body.from },
+          execute,
+        )
+        const http = RequestExecutor.responseHttp(submitted.response)
+        return Stream.suspend(() => {
+          // Parser state is local to one response, exactly like `Route.make`'s LLM stream loop.
+          let state = protocol.initial()
+          return protocol.frames(RequestExecutor.responseStream(submitted.response), submitted.context).pipe(
+            Stream.mapEffect((frame) =>
+              protocol.step(state, frame).pipe(
+                Effect.map(([next, output]) => {
+                  state = next
+                  return output
+                }),
+              ),
+            ),
+            Stream.flattenIterable,
+            Stream.concat(
+              Stream.suspend(() => Stream.fromIterableEffect(protocol.finish(state, { ...submitted.context, http }))),
+            ),
+            Stream.mapError((error) =>
+              error.reason.http !== undefined
+                ? error
+                : new AIError({
+                    reason: AIErrorReason.make({
+                      ...error.reason,
+                      message: error.reason.message,
+                      cause: error.reason.cause,
+                      http,
+                    }),
+                  }),
+            ),
+          )
+        })
+      }),
+    )
+  return {
+    kind: "stream",
+    id: protocol.id,
+    provider: protocol.provider,
+    protocol: protocol.id,
+    stream: (request, execute) => events(request, execute, "stream"),
+    generate: (request, execute) =>
+      events(request, execute, "generate").pipe(Stream.runCollect, Effect.flatMap(input.collect)),
+  }
+}
+
+// ---------------------------------------------------------------------------
+// Transport plumbing shared by every kind
+// ---------------------------------------------------------------------------
+
+const makeTransport = <Request extends MediaRequest>(
+  input: Composition<Request> & { readonly protocol: { readonly id: string; readonly provider: ProviderID } },
+) => {
+  const routeHttp = input.headers === undefined ? undefined : new HttpOptions({ headers: input.headers })
+  const authorize = Auth.toEffect(input.auth)
+  const baseURL = (path: string) => new URL(`${Endpoint.trimBaseUrl(input.endpoint.baseURL ?? "")}${path}`)
+  /** `auth` is only what `Auth` added or changed, never untouched deployment headers. */
+  const send = Effect.fn("MediaRoute.send")(function* (
+    call: {
+      readonly method: AuthInput["method"]
+      readonly url: URL
+      readonly headers: Headers.Headers
+      readonly request: AuthInput["request"]
+      readonly body?: MediaProtocol.Body
+    },
+    execute: Execute,
+  ) {
+    const encoded = encode(call.body, call.headers)
+    const url = call.url.toString()
+    const headers = yield* authorize({
+      request: call.request,
+      method: call.method,
+      url,
+      body: encoded.text,
+      headers: encoded.headers,
+    })
+    const response = yield* execute(
+      encoded.apply(HttpClientRequest.make(call.method)(url).pipe(HttpClientRequest.setHeaders(headers))),
+    )
+    return {
+      response,
+      auth: Object.fromEntries(Object.entries(headers).filter(([key, value]) => encoded.headers[key] !== value)),
+    }
+  })
+  return {
+    /** Route and model overlays; `start` additionally merges the request's own `http`. */
+    http: (model: MediaRequest["model"]) => mergeHttpOptions(routeHttp, model.http),
+    /** POST the protocol body to the route endpoint. */
+    submit: Effect.fn("MediaRoute.submit")(function* (
+      request: Request,
+      protocol: {
+        readonly unsupported?: ReadonlyArray<keyof Request & string>
+        readonly prepare?: MediaProtocol.Prepare<Request>
+        readonly from: (request: Request) => Effect.Effect<MediaProtocol.Body, AIError>
+      },
+      execute: Execute,
+    ) {
+      yield* rejectUnsupported(input.protocol.id, input.protocol.provider, request, protocol.unsupported)
+      const http = mergeHttpOptions(routeHttp, request.model.http, request.http)
+      const headers = Headers.fromInput(http?.headers)
+      const prepared =
+        protocol.prepare === undefined
+          ? request
+          : yield* protocol.prepare(request, (path, body) =>
+              send({ method: "POST", url: baseURL(path), headers, request, body }, execute).pipe(
+                Effect.map((sent) => sent.response),
+              ),
+            )
+      // Sanitize after merging so model-level overlays are covered; the model value is restored, not sanitized.
+      const resolved: Request = { ...sanitizeSurrogates({ ...prepared, http }), model: request.model }
+      const body = yield* protocol.from(resolved)
+      const url = withQuery(
+        withQuery(
+          Endpoint.render(input.endpoint, { request: resolved, body }),
+          body.type === "multipart" ? undefined : body.query,
+        ),
+        http?.query,
+      )
+      const sent = yield* send({ method: "POST", url, headers, request: resolved, body }, execute)
+      return { response: sent.response, context: { request: resolved, body } }
+    }),
+    /** Bodiless follow-up call (status, result, cancel) with the same auth and headers as `submit`. */
+    call: (method: AuthInput["method"], path: string, http: HttpOptions | undefined, execute: Execute) => {
+      // Provider-issued absolute URLs (fal `status_url`) are used as-is; everything else resolves against the base.
+      const url = withQuery(/^https?:\/\//.test(path) ? new URL(path) : baseURL(path), http?.query)
+      for (const [key, value] of Object.entries(input.endpoint.query ?? {})) url.searchParams.set(key, value)
+      return send({ method, url, headers: Headers.fromInput(http?.headers), request: { http } }, execute)
+    },
+  }
+}
+
+const withQuery = (url: URL, query: MediaProtocol.Query | undefined) => {
+  for (const [key, value] of Object.entries(query ?? {})) {
+    url.searchParams.delete(key)
+    for (const item of typeof value === "string" ? [value] : value) url.searchParams.append(key, item)
+  }
+  return url
+}
+
+const encode = (body: MediaProtocol.Body | undefined, headers: Headers.Headers) => {
+  if (body === undefined) return { text: "", headers, apply: (request: HttpClientRequest.HttpClientRequest) => request }
+  if (body.type === "json") {
+    const text = encodeJson(body.value)
+    return { text, headers, apply: HttpClientRequest.bodyText(text, "application/json") }
+  }
+  if (body.type === "binary")
+    return {
+      text: `[${body.contentType}]`,
+      headers,
+      apply: HttpClientRequest.bodyUint8Array(body.value, body.contentType),
+    }
+  return {
+    text: "[multipart/form-data]",
+    // The HTTP client sets the multipart boundary; a caller-supplied content-type would corrupt it.
+    headers: Headers.remove(headers, "content-type"),
+    apply: HttpClientRequest.bodyFormData(body.value),
+  }
+}
+
+/**
+ * Common fields are never silently dropped: a present field the protocol declared unsupported fails typed. `false`
+ * counts as present because some booleans mean something when false (video `audio`); protocols reject opt-in
+ * booleans such as speech `timestamps` with `=== true` in `body.from` instead of listing them.
+ */
 const rejectUnsupported = <Request extends object>(
  route: string,
  provider: ProviderID,
@@ -104,11 +402,13 @@ const rejectUnsupported = <Request extends object>(
  })
  if (present.length === 0) return Effect.void
  return Effect.fail(
-    ProviderShared.unsupportedOperation({
-      operation: `media.${present[0]}`,
-      provider,
-      route,
-      message: `${provider}/${route} does not support ${present.join(", ")}`,
+    new AIError({
+      reason: new UnsupportedOperationError({
+        operation: `media.${present[0]}`,
+        provider,
+        route,
+        message: `${provider}/${route} does not support ${present.join(", ")}`,
+      }),
    }),
  )
 }
@@ -1,5 +1,5 @@
 import { Schema, type Effect } from "effect"
-import type { AIError, LLMEvent, LLMRequest, ProtocolID } from "../schema/index.js"
+import type { AIError, LanguageModelSanitizerCompatibility, LLMEvent, LLMRequest, ProtocolID } from "../schema/index.js"

 /**
 * The semantic API contract of one model server family.
@@ -43,6 +43,8 @@ export interface Protocol<Body, Frame, Event, State> {
  readonly stream: ProtocolStream<Frame, Event, State>
  /** Whether `body.from` lowers `Message.effort(...)` markers; wrappers around another `body.from` must forward it. */
  readonly supportsEffortUpdates?: (request: LLMRequest) => boolean
+  /** Tool schema sanitizer for every model on this protocol unless the model's compatibility sets one; wrappers around another `body.from` must forward it. */
+  readonly sanitizer?: LanguageModelSanitizerCompatibility
 }

 export interface ProtocolBody<Body> {
--- a/Show More
+++ b/Show More