refactor(ai): simplify response fragment guards

fix(ai): enforce exclusive checkpoint representations
fix(ai): preserve compacted image detail on replay
2026-08-31 14:06:21 +00:00 · 2026-08-31 19:30:04 +05:30 · 2026-08-31 19:29:23 +05:30 · 2026-08-31 19:27:34 +05:30 · 2026-08-31 19:26:21 +05:30 · 2026-08-31 19:25:28 +05:30
576 changed files with 28587 additions and 8216 deletions
@@ -2,3 +2,4 @@ packages/core/migration/**/snapshot.json linguist-generated
 packages/core/src/database/migration.gen.ts linguist-generated
 packages/core/src/models-dev/snapshot.txt linguist-generated
 packages/core/src/**/*.txt text eol=lf
+packages/httpapi-codegen/test/generated/*.ts text eol=lf
@@ -5,6 +5,7 @@ on:
    branches:
      - dev
      - production
+      - beta
  workflow_dispatch:

 concurrency: ${{ github.workflow }}-${{ github.ref }}
@@ -15,7 +16,7 @@ permissions:

 jobs:
  deploy:
-    if: github.repository == 'anomalyco/opencode' && (github.ref_name == 'dev' || github.ref_name == 'production')
+    if: github.repository == 'anomalyco/opencode' && (github.ref_name == 'dev' || github.ref_name == 'production' || github.ref_name == 'beta')
    runs-on: ubuntu-latest
    environment: ${{ github.ref_name }}
    steps:
@@ -28,6 +29,7 @@ jobs:
          node-version: "24"

      - uses: aws-actions/configure-aws-credentials@7474bc4690e29a8392af63c5b98e7449536d5c3a # v4.3.1
+        if: github.ref_name != 'beta'
        with:
          role-to-assume: ${{ vars.AWS_DEPLOY_ROLE_ARN }}
          role-session-name: opencode-${{ github.run_id }}
@@ -1,6 +1,7 @@
 - After changing the public Protocol or Server `HttpApi`, run `bun run generate` from `packages/client`. Do not edit generated client files directly.
 - Keep runtime dependencies directed from Schema to Core and Protocol, then from Core and Protocol to Server. Client runtime code may depend on Schema and Protocol but never Core or Server; `sdk` composes Client, Core, and Server.
 - Current implementation changes belong in `packages/core`, `packages/cli`, `packages/server`, `packages/protocol`, `packages/schema`, and related generated client surfaces when required.
+- This repository does not use Changesets. Do not add `.changeset` files; follow the existing release workflow instead.
 - The default branch in this repo is `v2`.
 - Base all new branches and worktrees on `v2`, or `origin/v2` when the local `v2` ref is unavailable. Do not base them on `dev`.
 - Local `main` ref may not exist; use `v2` or `origin/v2` for diffs.
@@ -45,6 +46,7 @@ Examples: `fix(tui): simplify thinking toggle styling`, `docs: update contributi
 ### General Principles

 - Keep things in one function unless composable or reusable
+- Validate unknown values once at the boundary that owns them. Pass typed values inward instead of repeating `typeof value === "object"` and property-existence checks. Do not defensively revalidate values already guaranteed by a schema, constructor, or internal type.
 - Do not extract single-use helpers preemptively. Inline the logic at the call site unless the helper is reused, hides a genuinely complex boundary, or has a clear independent name that improves the caller.
 - Before adding complexity for a speculative or vanishingly unlikely race or security edge case, explain the concrete failure mode, likelihood, and complexity cost to the user and get their buy-in. Do not silently expand scope for theoretical robustness.
 - Avoid `try`/`catch` where possible
@@ -54,6 +54,7 @@
      "name": "@opencode-ai/app",
      "version": "1.18.15",
      "dependencies": {
+        "@corvu/drawer": "catalog:",
        "@dnd-kit/abstract": "0.5.0",
        "@dnd-kit/dom": "0.5.0",
        "@dnd-kit/helpers": "0.5.0",
@@ -183,6 +184,7 @@
        "@typescript/native-preview": "catalog:",
        "effect": "catalog:",
        "solid-js": "catalog:",
+        "zod": "catalog:",
      },
      "peerDependencies": {
        "effect": "4.0.0-rc.112",
@@ -430,6 +432,7 @@
        "@types/bun": "catalog:",
        "@types/node": "catalog:",
        "@typescript/native-preview": "catalog:",
+        "app-builder-lib": "26.15.7",
        "drizzle-orm": "catalog:",
        "effect": "catalog:",
        "electron": "42.10.1",
@@ -1640,6 +1643,10 @@

    "@cloudflare/workers-types": ["@cloudflare/workers-types@4.20251008.0", "", {}, "sha512-dZLkO4PbCL0qcCSKzuW7KE4GYe49lI12LCfQ5y9XeSwgYBoAUbwH4gmJ6A0qUIURiTJTkGkRkhVPqpq2XNgYRA=="],

+    "@corvu/dialog": ["@corvu/dialog@0.2.4", "", { "dependencies": { "@corvu/utils": "~0.4.2", "solid-dismissible": "~0.1.1", "solid-focus-trap": "~0.1.8", "solid-presence": "~0.2.0", "solid-prevent-scroll": "~0.1.10" }, "peerDependencies": { "solid-js": "^1.8" } }, "sha512-n54vJq+fOy8GVrnYBdJpD6JXNuyx7LOeMrRxwzAvZnYGpW8+AA12tnb/P/2emJj/HjOO5otheGKb0breshdFlA=="],
+
+    "@corvu/drawer": ["@corvu/drawer@0.2.4", "", { "dependencies": { "@corvu/dialog": "~0.2.4", "@corvu/utils": "~0.4.2", "@solid-primitives/memo": "^1.4.1", "solid-transition-size": "~0.1.4" }, "peerDependencies": { "solid-js": "^1.8" } }, "sha512-7jQoGZ8ROB9CmXam2nMY2wEskU3IoFwZQywkF/7vrBc/edGsPv7mOVQ1GN6G+4nd7nrMZ3UtqHAhmRh9V0azlw=="],
+
    "@corvu/utils": ["@corvu/utils@0.4.2", "", { "dependencies": { "@floating-ui/dom": "^1.6.11" }, "peerDependencies": { "solid-js": "^1.8" } }, "sha512-Ox2kYyxy7NoXdKWdHeDEjZxClwzO4SKM8plAaVwmAJPxHMqA0rLOoAsa+hBDwRLpctf+ZRnAd/ykguuJidnaTA=="],

    "@cspotcode/source-map-support": ["@cspotcode/source-map-support@0.8.1", "", { "dependencies": { "@jridgewell/trace-mapping": "0.3.9" } }, "sha512-IchNf6dN4tHoMFIn/7OE8LWZ19Y6q/67Bmf6vnGREv8RSbBVb9LPJxEcnwrcwX6ixSvaiGoomAUvu4YSxXrVgw=="],
@@ -1678,7 +1685,7 @@

    "@electron/fuses": ["@electron/fuses@1.8.0", "", { "dependencies": { "chalk": "^4.1.1", "fs-extra": "^9.0.1", "minimist": "^1.2.5" }, "bin": { "electron-fuses": "dist/bin.js" } }, "sha512-zx0EIq78WlY/lBb1uXlziZmDZI4ubcCXIMJ4uGjXzZW0nS19TjSPeXPAjzzTmKQlJUZm0SbmZhPKP7tuQ1SsEw=="],

-    "@electron/get": ["@electron/get@5.1.0", "", { "dependencies": { "debug": "^4.1.1", "env-paths": "^3.0.0", "graceful-fs": "^4.2.11", "progress": "^2.0.3", "semver": "^7.6.3", "sumchecker": "^3.0.1" }, "optionalDependencies": { "undici": "^7.24.4" } }, "sha512-3kSBtG8ObcTVfXanm5vVJ6UnBLEVmVsRk1M+vGqCuMBV+XLCbJYuWQful+yIy0GQDsSlK0kHEriEHn7SPk4EnA=="],
+    "@electron/get": ["@electron/get@3.1.0", "", { "dependencies": { "debug": "^4.1.1", "env-paths": "^2.2.0", "fs-extra": "^8.1.0", "got": "^11.8.5", "progress": "^2.0.3", "semver": "^6.2.0", "sumchecker": "^3.0.1" }, "optionalDependencies": { "global-agent": "^3.0.0" } }, "sha512-F+nKc0xW+kVbBRhFzaMgPy3KwmuNTYX1fx6+FxxoSnNgwYX6LD7AKBTWkU0MQ6IBoe7dz069CNkR673sPAgkCQ=="],

    "@electron/notarize": ["@electron/notarize@2.5.0", "", { "dependencies": { "debug": "^4.1.1", "fs-extra": "^9.0.1", "promise-retry": "^2.0.1" } }, "sha512-jNT8nwH1f9X5GEITXaQ8IF/KdskvIkOFfB2CvwumsveVidzpSc+mvhhTMdAGSYF3O+Nq49lJ7y+ssODRXu06+A=="],

@@ -2774,7 +2781,7 @@

    "@silvia-odwyer/photon-node": ["@silvia-odwyer/photon-node@0.3.4", "", {}, "sha512-bnly4BKB3KDTFxrUIcgCLbaeVVS8lrAkri1pEzskpmxu9MdfGQTy8b8EgcD83ywD3RPMsIulY8xJH5Awa+t9fA=="],

-    "@sindresorhus/is": ["@sindresorhus/is@7.2.0", "", {}, "sha512-P1Cz1dWaFfR4IR+U13mqqiGsLFf1KbayybWwdd2vfctdV6hDpUkgCY0nKOLLTMSoRd/jJNjtbqzf13K8DCCXQw=="],
+    "@sindresorhus/is": ["@sindresorhus/is@4.6.0", "", {}, "sha512-t09vSN3MdfsyCHoFcTRCH/iUtG7OJ0CsjzB8cjAmKc/va/kIgeDI/TxsigdncE/4be734m0cvIYwNaV4i2XqAw=="],

    "@smithy/config-resolver": ["@smithy/config-resolver@4.7.2", "", { "dependencies": { "@smithy/core": "^3.33.2", "tslib": "^2.6.2" } }, "sha512-Y1XfSefHIOub9762qm3ShafdlEE/Va8h3kLUeMq765fNeWeNLcOP2YUPr86H1SlyGwZTOqQ67RlBZPZ3k9Djgg=="],

@@ -2874,6 +2881,8 @@

    "@solid-primitives/media": ["@solid-primitives/media@2.3.6", "", { "dependencies": { "@solid-primitives/event-listener": "^2.4.6", "@solid-primitives/rootless": "^1.5.4", "@solid-primitives/static-store": "^0.1.4", "@solid-primitives/utils": "^6.4.1" }, "peerDependencies": { "solid-js": "^1.6.12" } }, "sha512-pk49gPOq/UMRUJ+pTSrOfBiR8xJjRYHXIf1iR/jSnyQ/KroU+ZXhkZzavC7hvfp2vJeOTW5k2/HN0r3Q1VJ7Pw=="],

+    "@solid-primitives/memo": ["@solid-primitives/memo@1.5.1", "", { "dependencies": { "@solid-primitives/scheduled": "^1.5.3", "@solid-primitives/utils": "^6.4.1" }, "peerDependencies": { "solid-js": "^1.6.12" } }, "sha512-VDPrkl9epp0tbby9MvsqphGFCYCtDRC5J8FKzTqHbQiG5hhR8n6xv4MfjhTW231IaBxxPHLxS43EE8c5Q23mSQ=="],
+
    "@solid-primitives/props": ["@solid-primitives/props@3.2.4", "", { "dependencies": { "@solid-primitives/utils": "^6.4.1" }, "peerDependencies": { "solid-js": "^1.6.12" } }, "sha512-MXXdvi2TSB6d+0N6ueA/HP1j/Kh9SEc4WdF1ZDMLwPagxW6pIHHSS9xbRO0RjqSVpNP4j0nxCWT1hx3CkJNhNA=="],

    "@solid-primitives/refs": ["@solid-primitives/refs@1.1.4", "", { "dependencies": { "@solid-primitives/utils": "^6.4.1" }, "peerDependencies": { "solid-js": "^1.6.12" } }, "sha512-bLjwIs6ZPu8NQnuw04sU3Zc8qKSpbc0umUU/O4SHf6oWOdO4+dHY8vb1T7C4b/Tg103+9WGr6sEE0+NFlbaB/A=="],
@@ -3532,7 +3541,7 @@

    "chromium-pickle-js": ["chromium-pickle-js@0.2.0", "", {}, "sha512-1R5Fho+jBq0DDydt+/vHWj5KJNJCKdARKOCwZUen84I5BreWoLqRLANH1U87eJy1tiASPtMnGqJJq0ZsLoRPOw=="],

-    "ci-info": ["ci-info@4.4.0", "", {}, "sha512-77PSwercCZU2Fc4sX94eF8k8Pxte6JAwL4/ICZLFjJLqegs7kCuAsqqj/70NQF6TvDpgFjkubQB2FW2ZZddvQg=="],
+    "ci-info": ["ci-info@4.3.1", "", {}, "sha512-Wdy2Igu8OcBpI2pZePZ5oWjPC38tmDVx5WKUXKwlLYkA0ozo85sLsLvkBbBn/sZaSCMFOGZJ14fvW9t5/d7kdA=="],

    "classnames": ["classnames@2.3.2", "", {}, "sha512-CSbhY4cFEJRe6/GQzIk5qXZ4Jeg5pcsP7b5peFSDpffpe1cqjASH/n9UTjBwOp6XpMSTwQ8Za2K5V02ueA7Tmw=="],

@@ -3812,7 +3821,7 @@

    "dot-prop": ["dot-prop@8.0.2", "", { "dependencies": { "type-fest": "^3.8.0" } }, "sha512-xaBe6ZT4DHPkg0k4Ytbvn5xoxgpG0jOS1dYxSOwAHPuNLjP3/OzN0gH55SrLqpx8cBfSaVt91lXYkApjb+nYdQ=="],

-    "dotenv": ["dotenv@17.4.2", "", {}, "sha512-nI4U3TottKAcAD9LLud4Cb7b2QztQMUEfHbvhTH09bqXTxnSie8WnjPALV/WMCrJZ6UV/qHJ6L03OqO3LcdYZw=="],
+    "dotenv": ["dotenv@16.6.1", "", {}, "sha512-uBq4egWHTcTt33a72vpSG0z3HnPuIl6NqYcTrKEg2azoEyl2hpW0zqlxysq2pK9HlDIHyHyakeYaYnSAwd8bow=="],

    "dotenv-expand": ["dotenv-expand@11.0.7", "", { "dependencies": { "dotenv": "^16.4.5" } }, "sha512-zIHwmZPRshsCdpMDyVsqGmgyP0yT8GAgXUnkdAoJisxvf33k7yO6OuoKmcTGuXPWSsm8Oh88nZicRLA9Y0rUeA=="],

@@ -3884,7 +3893,7 @@

    "entities": ["entities@7.0.1", "", {}, "sha512-TWrgLOFUQTH994YUyl1yT4uyavY5nNB5muff+RtWaqNVCAK408b5ZnnbNAUEWLTCpum9w6arT70i1XdQ4UeOPA=="],

-    "env-paths": ["env-paths@3.0.0", "", {}, "sha512-dtJUTepzMW3Lm/NPxRf3wP4642UWhjL2sQxc+ym2YMj1m/H2zDNQOlezafzkHwn6sMstjHTwG6iQQsctDW/b1A=="],
+    "env-paths": ["env-paths@2.2.1", "", {}, "sha512-+h1lkLKhZMTYjog1VEpJNG7NZJWcuc2DDk/qsqSTRRCOXiLjeQ1d1/udrUGhqMxUgAlwKNZ0cf2uqan5GLuS2A=="],

    "err-code": ["err-code@2.0.3", "", {}, "sha512-2bmlRpNKBxT/CRmPOlyISQpNj+qSeYvcym/uT0Jx2bMOlKLtSy1ZmLuVxSEKKyor/N5yhvp/ZiG1oE3DEYMSFA=="],

@@ -4206,7 +4215,7 @@

    "hono-openapi": ["hono-openapi@1.1.2", "", { "peerDependencies": { "@hono/standard-validator": "^0.2.0", "@standard-community/standard-json": "^0.3.5", "@standard-community/standard-openapi": "^0.2.9", "@types/json-schema": "^7.0.15", "hono": "^4.8.3", "openapi-types": "^12.1.3" }, "optionalPeers": ["@hono/standard-validator", "hono"] }, "sha512-toUcO60MftRBxqcVyxsHNYs2m4vf4xkQaiARAucQx3TiBPDtMNNkoh+C4I1vAretQZiGyaLOZNWn1YxfSyUA5g=="],

-    "hosted-git-info": ["hosted-git-info@9.0.3", "", { "dependencies": { "lru-cache": "^11.1.0" } }, "sha512-Hc+ghLoSt6QaYZUv0WBiIvmMDZuZZ7oaDvdH8MbfOO4lOsxdXLEvuC6ePoGs9H1X9oCLyq6+NVN0MKqD+ydxyg=="],
+    "hosted-git-info": ["hosted-git-info@4.1.0", "", { "dependencies": { "lru-cache": "^6.0.0" } }, "sha512-kyCuEOWjJqZuDbRHzL8V93NzQhwIB71oFWSyzVo+KPZI+pnQPPxucdkrOZvkLRnrf5URsQM+IJ09Dw29cRALIA=="],

    "html-entities": ["html-entities@2.3.3", "", {}, "sha512-DV5Ln36z34NNTDgnz0EWGBLZENelNAtkiFA4kyNOG2tDI6Mz1uSWiq1wAKdyjnJwyDiDO7Fa2SO1CTxPXL8VxA=="],

@@ -5300,6 +5309,10 @@

    "socks-proxy-agent": ["socks-proxy-agent@8.0.5", "", { "dependencies": { "agent-base": "^7.1.2", "debug": "^4.3.4", "socks": "^2.8.3" } }, "sha512-HehCEsotFqbPW9sJ8WVYB6UbmIMv7kUUORIF2Nncq4VQvBfNBLibW9YZR5dlYCSUhwcD628pRllm7n+E+YTzJw=="],

+    "solid-dismissible": ["solid-dismissible@0.1.1", "", { "dependencies": { "@corvu/utils": "~0.4.1" }, "peerDependencies": { "solid-js": "^1.8" } }, "sha512-9kcKBJIMdS+586cA1g63HYWxKh3h89leeNHbPZ1csYjuni+NvPBtNr11l0iEX2AKKEt6FHk6qNhc/gjoYAW1pA=="],
+
+    "solid-focus-trap": ["solid-focus-trap@0.1.9", "", { "dependencies": { "@corvu/utils": "~0.4.2" }, "peerDependencies": { "solid-js": "^1.8" } }, "sha512-LTyNki6GUJPRLXV5uMWPkYClB07SUMubbr2EkAddiR0CJCF/I283txilMU9RURSr/P8EewMfXWu2o3aWrK7A5A=="],
+
    "solid-js": ["solid-js@1.9.15", "", { "dependencies": { "csstype": "^3.1.0", "seroval": "~1.5.4", "seroval-plugins": "~1.5.4" } }, "sha512-EeiY2xfpZJqPLjXspVEKjAII4yv8NyG//NxZ3IpOFHdUNnnTyL0uJOeS9LWGvA7cFCz5y94cjFwYlmw5Luncsg=="],

    "solid-list": ["solid-list@0.3.0", "", { "dependencies": { "@corvu/utils": "~0.4.0" }, "peerDependencies": { "solid-js": "^1.8" } }, "sha512-t4hx/F/l8Vmq+ib9HtZYl7Z9F1eKxq3eKJTXlvcm7P7yI4Z8O7QSOOEVHb/K6DD7M0RxzVRobK/BS5aSfLRwKg=="],
@@ -5314,6 +5327,8 @@

    "solid-stripe": ["solid-stripe@0.8.1", "", { "peerDependencies": { "@stripe/stripe-js": ">=1.44.1 <8.0.0", "solid-js": "^1.6.0" } }, "sha512-l2SkWoe51rsvk9u1ILBRWyCHODZebChSGMR6zHYJTivTRC0XWrRnNNKs5x1PYXsaIU71KYI6ov5CZB5cOtGLWw=="],

+    "solid-transition-size": ["solid-transition-size@0.1.4", "", { "dependencies": { "@corvu/utils": "~0.3.2" }, "peerDependencies": { "solid-js": "^1.8" } }, "sha512-ocHVnbfy23CgfaH4cEUR/AFg0Y3CEL8Oh3n9Qv8OHFJgPh+zkmERKZQfi/xH5XvxDCizg8VjPrVUhiHB1Gza8g=="],
+
    "solid-use": ["solid-use@0.9.1", "", { "peerDependencies": { "solid-js": "^1.7" } }, "sha512-UwvXDVPlrrbj/9ewG9ys5uL2IO4jSiwys2KPzK4zsnAcmEl7iDafZWW1Mo4BSEWOmQCGK6IvpmGHo1aou8iOFw=="],

    "sort-keys": ["sort-keys@1.1.2", "", { "dependencies": { "is-plain-obj": "^1.0.0" } }, "sha512-vzn8aSqKgytVik0iwdBEi+zevbTYZogewTUM6dtpmGwEcdzbub/TX4bCzRhebDCRC3QzXgJsLRKB2V/Oof7HXg=="],
@@ -5812,7 +5827,7 @@

    "y18n": ["y18n@5.0.8", "", {}, "sha512-0pfFzegeDWJHJIAmTLRP2DwHjdF5s7jo9tuztdQxAhINCdvS+3nGINqPd00AphqJR/0LhANUS6/+7SCb98YOfA=="],

-    "yallist": ["yallist@4.0.0", "", {}, "sha512-3wdGidZyq5PB084XLES5TpOSRA3wjXAlIWMhum2kRcv/41Sn2emQ0dycQW4uZXLejwKvg6EsvbdlVL+FYEct7A=="],
+    "yallist": ["yallist@5.0.0", "", {}, "sha512-YgvUTfwqyc7UXVMrB+SImsVYSmTS8X/tSrtdNZMImM+n7+QTriRXyXim0mBrTXNeqzVF0KWGgHPeiyViFFrNDw=="],

    "yaml": ["yaml@2.9.0", "", { "bin": { "yaml": "bin.mjs" } }, "sha512-2AvhNX3mb8zd6Zy7INTtSpl1F15HW6Wnqj0srWlkKLcpYl/gMIMJiyuGq2KeI2YFxUPjdlB+3Lc10seMLtL4cA=="],

@@ -5960,6 +5975,8 @@

    "@astrojs/starlight/js-yaml": ["js-yaml@4.3.1", "", { "dependencies": { "argparse": "^2.0.1" }, "bin": { "js-yaml": "bin/js-yaml.js" } }, "sha512-CY6crGq313MX8GkwvB7tzgp99vjQxY1++5y10/BKN/GUfHqWaOGQMNZkBvqSzsZKWk/ijwHlWzzkLulsGHhjWQ=="],

+    "@astrojs/telemetry/ci-info": ["ci-info@4.4.0", "", {}, "sha512-77PSwercCZU2Fc4sX94eF8k8Pxte6JAwL4/ICZLFjJLqegs7kCuAsqqj/70NQF6TvDpgFjkubQB2FW2ZZddvQg=="],
+
    "@astrojs/telemetry/is-docker": ["is-docker@4.0.0", "", { "bin": { "is-docker": "cli.js" } }, "sha512-LHE+wROyG/Y/0ZnbktRCoTix2c1RhgWaZraMZ8o1Q7zCh0VSrICJQO5oqIIISrcSBtrXv0o233w1IYwsWCjTzA=="],

    "@aws-crypto/crc32/@aws-sdk/types": ["@aws-sdk/types@3.974.4", "", { "dependencies": { "@smithy/types": "^4.16.1", "tslib": "^2.6.2" } }, "sha512-dSFDNG00MEz0/xl5gxL62giLd1iYyJsTxZ1I1DOj6lC+bbgLB4TRsYClJg3b62dhXT1uATzsTNXPnC+33EJV3A=="],
@@ -6114,7 +6131,9 @@

    "@electron/fuses/fs-extra": ["fs-extra@9.1.0", "", { "dependencies": { "at-least-node": "^1.0.0", "graceful-fs": "^4.2.0", "jsonfile": "^6.0.1", "universalify": "^2.0.0" } }, "sha512-hcg3ZmepS30/7BSFqRvoo3DOMQu7IjqxO5nCDt+zM9XWjb33Wg7ziNT+Qvqbuc3+gWpzO02JubVyk2G4Zvo1OQ=="],

-    "@electron/get/undici": ["undici@7.29.0", "", {}, "sha512-IDxfleLmmbSskfWSUATiN1nfn2rDuvnMOqb5CWR92iIfojA0Ud+ulOAAEQ57LPr9rWmsreUyf5lwyao+7GNNVw=="],
+    "@electron/get/fs-extra": ["fs-extra@8.1.0", "", { "dependencies": { "graceful-fs": "^4.2.0", "jsonfile": "^4.0.0", "universalify": "^0.1.0" } }, "sha512-yhlQgA6mnOJUKOsRUFsgJdQCvkKhcz8tlZG5HBQfReYZy46OwLcY+Zia0mtdHsOo9y/hP+CxMN0TU9QxoOtG4g=="],
+
+    "@electron/get/semver": ["semver@6.3.1", "", { "bin": { "semver": "bin/semver.js" } }, "sha512-BR7VvDCVHO+q2xBEWskxS6DJE1qRnb7DxzUrogb71CWoSficBxYsiAGd+Kl0mmq/MprG9yArRkyrQxTO6XjMzA=="],

    "@electron/notarize/fs-extra": ["fs-extra@9.1.0", "", { "dependencies": { "at-least-node": "^1.0.0", "graceful-fs": "^4.2.0", "jsonfile": "^6.0.1", "universalify": "^2.0.0" } }, "sha512-hcg3ZmepS30/7BSFqRvoo3DOMQu7IjqxO5nCDt+zM9XWjb33Wg7ziNT+Qvqbuc3+gWpzO02JubVyk2G4Zvo1OQ=="],

@@ -6150,6 +6169,12 @@

    "@modelcontextprotocol/sdk/jose": ["jose@6.2.9", "", {}, "sha512-XrchZOFZUl/T3vTwRe8XK+cJrGtMF4th1ARnDfwbBXFKThGhlsxEE4Zu03AD/bjJSt/9jT/mxrOCkJWOg77aPA=="],

+    "@npmcli/arborist/hosted-git-info": ["hosted-git-info@9.0.3", "", { "dependencies": { "lru-cache": "^11.1.0" } }, "sha512-Hc+ghLoSt6QaYZUv0WBiIvmMDZuZZ7oaDvdH8MbfOO4lOsxdXLEvuC6ePoGs9H1X9oCLyq6+NVN0MKqD+ydxyg=="],
+
+    "@npmcli/config/ci-info": ["ci-info@4.4.0", "", {}, "sha512-77PSwercCZU2Fc4sX94eF8k8Pxte6JAwL4/ICZLFjJLqegs7kCuAsqqj/70NQF6TvDpgFjkubQB2FW2ZZddvQg=="],
+
+    "@npmcli/package-json/hosted-git-info": ["hosted-git-info@9.0.3", "", { "dependencies": { "lru-cache": "^11.1.0" } }, "sha512-Hc+ghLoSt6QaYZUv0WBiIvmMDZuZZ7oaDvdH8MbfOO4lOsxdXLEvuC6ePoGs9H1X9oCLyq6+NVN0MKqD+ydxyg=="],
+
    "@npmcli/query/postcss-selector-parser": ["postcss-selector-parser@7.1.5", "", { "dependencies": { "cssesc": "^3.0.0", "util-deprecate": "^1.0.2" } }, "sha512-KvvtD7SrlBP7dlgkBghEE3r84CABm5SmV2aNcG4oCA+qDnJ/tvKonFVvwWAyyWUEwxuNawdfEAZKP9zM3oZ2Uw=="],

    "@octokit/auth-app/@octokit/request": ["@octokit/request@10.0.14", "", { "dependencies": { "@octokit/endpoint": "^11.0.3", "@octokit/request-error": "^7.1.1", "@octokit/types": "^17.0.0", "content-type": "^2.0.0", "json-with-bigint": "^3.5.12", "universal-user-agent": "^7.0.2" } }, "sha512-bgWgiSfFS689/AxQDarU+b3Qu1FYewfVr5vI/jhV84s7GQ3C2OsdRxukGGYKB37MCgwcS50UrfBLlHzhiG13sw=="],
@@ -6290,12 +6315,16 @@

    "@pierre/trees/react-dom": ["react-dom@19.2.8", "", { "dependencies": { "scheduler": "^0.27.0" }, "peerDependencies": { "react": "^19.2.8" } }, "sha512-rVprimfGBG3DR+Tq0IQG2DT5PxKth1WIGDmj5yPmlzr4YBe7uyE+Du4oVqTDXZSHGGGXRtTJEGSSePyQCMBglQ=="],

+    "@poppinss/dumper/@sindresorhus/is": ["@sindresorhus/is@7.2.0", "", {}, "sha512-P1Cz1dWaFfR4IR+U13mqqiGsLFf1KbayybWwdd2vfctdV6hDpUkgCY0nKOLLTMSoRd/jJNjtbqzf13K8DCCXQw=="],
+
    "@poppinss/dumper/supports-color": ["supports-color@10.2.2", "", {}, "sha512-SS+jx45GF1QjgEXQx4NJZV9ImqmO2NPz5FNsIHrsDjh2YsHnawpan7SNQ1o8NuhrbHZy9AZhIoCUiCeaW/C80g=="],

    "@protobuf-ts/plugin/typescript": ["typescript@3.9.10", "", { "bin": { "tsc": "bin/tsc", "tsserver": "bin/tsserver" } }, "sha512-w6fIxVE/H1PkLKcCPsFqKE7Kv7QUwhU8qQY2MueZXWx5cPZdwFupLgKK3vntcK98BtNHZtAF4LA/yl2a7k8R6Q=="],

    "@rollup/pluginutils/estree-walker": ["estree-walker@2.0.2", "", {}, "sha512-Rfkk/Mp/DL7JVje3u18FxFujQlTNR2q6QfMSMB7AvCBx91NGj/ba3kCfza0f6dVDbw7YlRf/nDrn7pQrCCyQ/w=="],

+    "@sentry/bundler-plugins/dotenv": ["dotenv@17.4.2", "", {}, "sha512-nI4U3TottKAcAD9LLud4Cb7b2QztQMUEfHbvhTH09bqXTxnSie8WnjPALV/WMCrJZ6UV/qHJ6L03OqO3LcdYZw=="],
+
    "@sentry/bundler-plugins/glob": ["glob@13.0.6", "", { "dependencies": { "minimatch": "^10.2.2", "minipass": "^7.1.3", "path-scurry": "^2.0.2" } }, "sha512-Wjlyrolmm8uDpm/ogGyXZXb1Z+Ca2B8NbJwqBVg0axK9GbBeoS7yGV6vjXnYdGm6X53iehEuxxbyiKp8QmN4Vw=="],

    "@sentry/cli/https-proxy-agent": ["https-proxy-agent@5.0.1", "", { "dependencies": { "agent-base": "6", "debug": "4" } }, "sha512-dFcAjpTQFgoLMzC2VwU+C/CbS7uRL0lWmxDITmqm7C+7F0Odmj6s9l6alZc6AELXhrnggM2CeWSXHGOdX2YtwA=="],
@@ -6368,14 +6397,6 @@

    "anymatch/picomatch": ["picomatch@2.3.2", "", {}, "sha512-V7+vQEJ06Z+c5tSye8S+nHUfI51xoXIXjHQ99cQtKUkQqqO1kO/KCJUfZXuB47h/YBlDhah2H3hdUGXn8ie0oA=="],

-    "app-builder-lib/@electron/get": ["@electron/get@3.1.0", "", { "dependencies": { "debug": "^4.1.1", "env-paths": "^2.2.0", "fs-extra": "^8.1.0", "got": "^11.8.5", "progress": "^2.0.3", "semver": "^6.2.0", "sumchecker": "^3.0.1" }, "optionalDependencies": { "global-agent": "^3.0.0" } }, "sha512-F+nKc0xW+kVbBRhFzaMgPy3KwmuNTYX1fx6+FxxoSnNgwYX6LD7AKBTWkU0MQ6IBoe7dz069CNkR673sPAgkCQ=="],
-
-    "app-builder-lib/ci-info": ["ci-info@4.3.1", "", {}, "sha512-Wdy2Igu8OcBpI2pZePZ5oWjPC38tmDVx5WKUXKwlLYkA0ozo85sLsLvkBbBn/sZaSCMFOGZJ14fvW9t5/d7kdA=="],
-
-    "app-builder-lib/dotenv": ["dotenv@16.6.1", "", {}, "sha512-uBq4egWHTcTt33a72vpSG0z3HnPuIl6NqYcTrKEg2azoEyl2hpW0zqlxysq2pK9HlDIHyHyakeYaYnSAwd8bow=="],
-
-    "app-builder-lib/hosted-git-info": ["hosted-git-info@4.1.0", "", { "dependencies": { "lru-cache": "^6.0.0" } }, "sha512-kyCuEOWjJqZuDbRHzL8V93NzQhwIB71oFWSyzVo+KPZI+pnQPPxucdkrOZvkLRnrf5URsQM+IJ09Dw29cRALIA=="],
-
    "app-builder-lib/js-yaml": ["js-yaml@4.3.1", "", { "dependencies": { "argparse": "^2.0.1" }, "bin": { "js-yaml": "bin/js-yaml.js" } }, "sha512-CY6crGq313MX8GkwvB7tzgp99vjQxY1++5y10/BKN/GUfHqWaOGQMNZkBvqSzsZKWk/ijwHlWzzkLulsGHhjWQ=="],

    "app-builder-lib/semver": ["semver@7.7.4", "", { "bin": { "semver": "bin/semver.js" } }, "sha512-vFKC2IEtQnVhpT78h1Yp8wzwrf8CM+MzKMHGJZfBtzhZNycRFnXsHk6E5TxIkkMsgNS7mdX3AGB7x2QM2di4lA=="],
@@ -6386,6 +6407,8 @@

    "astro/@clack/prompts": ["@clack/prompts@1.7.0", "", { "dependencies": { "@clack/core": "1.4.3", "fast-string-width": "^3.0.2", "fast-wrap-ansi": "^0.2.0", "sisteransi": "^1.0.5" } }, "sha512-y7/yvZ2TPAnR9+jnc00klvNNLkJiXFFrQA/hlLCcxA9a2A4zQIOimyFQ9XfwYKiGD1fb5GY8vbKIIgO8d5Tb2A=="],

+    "astro/ci-info": ["ci-info@4.4.0", "", {}, "sha512-77PSwercCZU2Fc4sX94eF8k8Pxte6JAwL4/ICZLFjJLqegs7kCuAsqqj/70NQF6TvDpgFjkubQB2FW2ZZddvQg=="],
+
    "astro/diff": ["diff@8.0.3", "", {}, "sha512-qejHi7bcSD4hQAZE0tNAawRK1ZtafHDmMTMkrrIGgSLl7hTnQHmKCeB45xAcbfTqK2zowkM3j3bHt/4b/ARbYQ=="],

    "astro/es-module-lexer": ["es-module-lexer@2.3.2", "", {}, "sha512-poHGpORABojJJucnV9KbOavETW8lBVnphkW77ER5/BQ5Fz7oXSoCNek7IH3vR5nRjdsEz926ibFYX8KtLQmdyw=="],
@@ -6430,6 +6453,8 @@

    "conf/dot-prop": ["dot-prop@10.2.0", "", { "dependencies": { "type-fest": "^5.0.0" } }, "sha512-BTJ9aZYL3vCfZlZOBLy9v8TUqWGQ0pzFnygKwFZt5udj6viBoFIBviKPUoZLDCPn1FoXffv6McQFDenrm5Krfw=="],

+    "conf/env-paths": ["env-paths@3.0.0", "", {}, "sha512-dtJUTepzMW3Lm/NPxRf3wP4642UWhjL2sQxc+ym2YMj1m/H2zDNQOlezafzkHwn6sMstjHTwG6iQQsctDW/b1A=="],
+
    "config-chain/ini": ["ini@1.3.8", "", {}, "sha512-JV/yugV2uzW5iMRSiZAyDtQd+nxtUnjeLt0acNdw98kKLrvuRVyB80tsREOE7yvGVgalhZ6RNXCmEHkUKBKxew=="],

    "cross-spawn/which": ["which@2.0.2", "", { "dependencies": { "isexe": "^2.0.0" }, "bin": { "node-which": "./bin/node-which" } }, "sha512-BLI3Tl1TW3Pvl70l3yq3Y64i+awpwXqsGBYWkkqMtnbXgrMD+yj7rhW0kuEDxzJaYXGjEW5ogapKNMEKNMjibA=="],
@@ -6456,14 +6481,16 @@

    "dot-prop/type-fest": ["type-fest@3.13.1", "", {}, "sha512-tLq3bSNx+xSpwvAJnzrK0Ep5CLNWjvFTOp71URMaAEWBfRb9nnJiBoUe0tF8bI4ZFO3omgBR6NvnbzVUT3Ly4g=="],

-    "dotenv-expand/dotenv": ["dotenv@16.6.1", "", {}, "sha512-uBq4egWHTcTt33a72vpSG0z3HnPuIl6NqYcTrKEg2azoEyl2hpW0zqlxysq2pK9HlDIHyHyakeYaYnSAwd8bow=="],
-
    "duplexer2/readable-stream": ["readable-stream@2.3.8", "", { "dependencies": { "core-util-is": "~1.0.0", "inherits": "~2.0.3", "isarray": "~1.0.0", "process-nextick-args": "~2.0.0", "safe-buffer": "~5.1.1", "string_decoder": "~1.1.1", "util-deprecate": "~1.0.1" } }, "sha512-8p0AUk4XODgIewSi0l8Epjs+EVnWiK7NoDIEGU0HhE7+ZyY8D1IMY7odu5lRrFXGg71L15KG8QrPmum45RTtdA=="],

    "editorconfig/commander": ["commander@10.0.1", "", {}, "sha512-y4Mg2tXshplEbSGzx7amzPwKKOCGuoSRP/CjEdwwk0FOGlUbq6lKuoyDZTNZkmxHdJtp54hdfY/JUrdL7Xfdug=="],

    "editorconfig/minimatch": ["minimatch@9.0.9", "", { "dependencies": { "brace-expansion": "^2.0.2" } }, "sha512-OBwBN9AL4dqmETlpS2zasx+vTeWclWzkblfZk7KTA5j3jeOONz/tRCnZomUyvNg83wL5Zv9Ss6HMJXAgL8R2Yg=="],

+    "electron/@electron/get": ["@electron/get@5.1.0", "", { "dependencies": { "debug": "^4.1.1", "env-paths": "^3.0.0", "graceful-fs": "^4.2.11", "progress": "^2.0.3", "semver": "^7.6.3", "sumchecker": "^3.0.1" }, "optionalDependencies": { "undici": "^7.24.4" } }, "sha512-3kSBtG8ObcTVfXanm5vVJ6UnBLEVmVsRk1M+vGqCuMBV+XLCbJYuWQful+yIy0GQDsSlK0kHEriEHn7SPk4EnA=="],
+
+    "electron-builder/ci-info": ["ci-info@4.4.0", "", {}, "sha512-77PSwercCZU2Fc4sX94eF8k8Pxte6JAwL4/ICZLFjJLqegs7kCuAsqqj/70NQF6TvDpgFjkubQB2FW2ZZddvQg=="],
+
    "electron-publish/mime": ["mime@2.6.0", "", { "bin": { "mime": "cli.js" } }, "sha512-USPkMeET31rOMiarsBNIHZKLGgvKc/LrjofAnBlOttf5ajRvqiRA8QsenbcooctK6d6Ts6aqZXBA+XbkKthiQg=="],

    "electron-store/type-fest": ["type-fest@5.8.0", "", { "dependencies": { "tagged-tag": "^1.0.0" } }, "sha512-YGYEVz3Fm5iy/AybuA0oyNFq7H4CgQNfRp/qfe8nurE1kuCeNm3/vfm9X4Mtl+qLyaKJUh5xrFZwogr41SMjYA=="],
@@ -6510,7 +6537,7 @@

    "globby/ignore": ["ignore@5.3.2", "", {}, "sha512-hsBTNUqQTDwkWtcdYI2i06Y/nUBEsNEDJKjWdigLvegy8kDuJAS8uRlpkkcQpyEXL0Z/pjDy5HBmMjRCJ2gq+g=="],

-    "got/@sindresorhus/is": ["@sindresorhus/is@4.6.0", "", {}, "sha512-t09vSN3MdfsyCHoFcTRCH/iUtG7OJ0CsjzB8cjAmKc/va/kIgeDI/TxsigdncE/4be734m0cvIYwNaV4i2XqAw=="],
+    "hosted-git-info/lru-cache": ["lru-cache@6.0.0", "", { "dependencies": { "yallist": "^4.0.0" } }, "sha512-Jo6dJ04CmSjuznwJSS3pUeWmd/H0ffTlkXXgwZi+eq1UCmqQwCh+eLsYOYCwY991i2Fah4h1BEMCx4qThGbsiA=="],

    "html-minifier-terser/commander": ["commander@10.0.1", "", {}, "sha512-y4Mg2tXshplEbSGzx7amzPwKKOCGuoSRP/CjEdwwk0FOGlUbq6lKuoyDZTNZkmxHdJtp54hdfY/JUrdL7Xfdug=="],

@@ -6556,10 +6583,10 @@

    "nitro/undici": ["undici@7.29.0", "", {}, "sha512-IDxfleLmmbSskfWSUATiN1nfn2rDuvnMOqb5CWR92iIfojA0Ud+ulOAAEQ57LPr9rWmsreUyf5lwyao+7GNNVw=="],

-    "node-gyp/env-paths": ["env-paths@2.2.1", "", {}, "sha512-+h1lkLKhZMTYjog1VEpJNG7NZJWcuc2DDk/qsqSTRRCOXiLjeQ1d1/udrUGhqMxUgAlwKNZ0cf2uqan5GLuS2A=="],
-
    "node-gyp/undici": ["undici@6.28.0", "", {}, "sha512-LIY910g9TI13YS95lrMFrs8Rm/u/irgHeTWoKCoteeJ04CUJ92eEfj0rVn+7VKMPBpUPiUoBKfhNyLI23EE/KA=="],

+    "npm-package-arg/hosted-git-info": ["hosted-git-info@9.0.3", "", { "dependencies": { "lru-cache": "^11.1.0" } }, "sha512-Hc+ghLoSt6QaYZUv0WBiIvmMDZuZZ7oaDvdH8MbfOO4lOsxdXLEvuC6ePoGs9H1X9oCLyq6+NVN0MKqD+ydxyg=="],
+
    "openid-client/jose": ["jose@4.15.9", "", {}, "sha512-1vUQX+IdDMVPj4k8kOxgUqlcK518yluMuGZwqlr44FS1ppZB/5GWh4rZG89erpOBOJjU/OBsnCVFfapsRz6nEA=="],

    "openid-client/lru-cache": ["lru-cache@6.0.0", "", { "dependencies": { "yallist": "^4.0.0" } }, "sha512-Jo6dJ04CmSjuznwJSS3pUeWmd/H0ffTlkXXgwZi+eq1UCmqQwCh+eLsYOYCwY991i2Fah4h1BEMCx4qThGbsiA=="],
@@ -6616,6 +6643,8 @@

    "slice-ansi/is-fullwidth-code-point": ["is-fullwidth-code-point@4.0.0", "", {}, "sha512-O4L094N2/dZ7xqVdrXhh9r1KODPJpFms8B5sGdJLPy664AgvXsreZUyCQQNItZRDlYug4xStLjNp/sz3HvBowQ=="],

+    "solid-transition-size/@corvu/utils": ["@corvu/utils@0.3.2", "", { "dependencies": { "@floating-ui/dom": "^1.6.7" }, "peerDependencies": { "solid-js": "^1.8" } }, "sha512-ZWlyWEE8qV9+CB9OAyo2bTrZGXQN9ZeM+JfYv89zoR+lRACKTDuoOZEdiyL8Uc7U5dUSH1uTqKhTTnaHWb+wZA=="],
+
    "sort-keys/is-plain-obj": ["is-plain-obj@1.1.0", "", {}, "sha512-yvkRyxmFKEOQ4pNXCmJG5AEQNlXJS5LaONXo5/cLdTZdWvsZ1ioJEonLGAosKlMWE8lwUy/bJzMjcw8az73+Fg=="],

    "source-map-support/source-map": ["source-map@0.6.1", "", {}, "sha512-UjgapumWlbMhkBgzT7Ykc5YXUT46F0iKu8SGXq0bcwP5dz/h0Plj6enJqjz1Zbq2l5WaqYnrVbwWOWMyF3F47g=="],
@@ -6642,8 +6671,6 @@

    "svgo/commander": ["commander@11.1.0", "", {}, "sha512-yPVavfyCcRhmorC7rWlkHn15b4wDVgVmBA7kV4QVBsF7kv/9TKJAbAXVTxvTnwP8HHKjRCJDClKbciiYS7p0DQ=="],

-    "tar/yallist": ["yallist@5.0.0", "", {}, "sha512-YgvUTfwqyc7UXVMrB+SImsVYSmTS8X/tSrtdNZMImM+n7+QTriRXyXim0mBrTXNeqzVF0KWGgHPeiyViFFrNDw=="],
-
    "temp/rimraf": ["rimraf@2.6.3", "", { "dependencies": { "glob": "^7.1.3" }, "bin": { "rimraf": "./bin.js" } }, "sha512-mwqeW5XsA2qAejG46gYdENaxXjx9onRNCfn7L0duuP4hCuTIi/QO7PDK07KJfp1d+izWPrzEJDcSqBa0OZQriA=="],

    "tempy/type-fest": ["type-fest@0.16.0", "", {}, "sha512-eaBzG6MxNzEn9kiwvtre90cXaNLkmadMWa1zQMs3XORCXNbsH/OewwbxC5ia9dCxIxnTAsSxXJaa/p5y8DlvJg=="],
@@ -6786,6 +6813,8 @@

    "@astrojs/starlight/astro/acorn": ["acorn@8.18.0", "", { "bin": { "acorn": "bin/acorn" } }, "sha512-lGq+9yr1/GuAWaVYIHRjvvySG5/4VfKIvC8EWxStPdcDh/Ka7FG3twP6v4d5BkravUilhIAsG4Qj83t02LWUPQ=="],

+    "@astrojs/starlight/astro/ci-info": ["ci-info@4.4.0", "", {}, "sha512-77PSwercCZU2Fc4sX94eF8k8Pxte6JAwL4/ICZLFjJLqegs7kCuAsqqj/70NQF6TvDpgFjkubQB2FW2ZZddvQg=="],
+
    "@astrojs/starlight/astro/common-ancestor-path": ["common-ancestor-path@1.0.1", "", {}, "sha512-L3sHRo1pXXEqX8VU28kfgUY+YGsk09hPqZiZmLacNib6XNTCM8ubYeT7ryXQw8asB1sKgcU5lkB7ONug08aB8w=="],

    "@astrojs/starlight/astro/cookie": ["cookie@1.1.1", "", {}, "sha512-ei8Aos7ja0weRpFzJnEA9UHJ/7XQmqglbRwnf2ATjcB9Wq874VKH9kfjjirM6UhU2/E5fFYadylyhFldcqSidQ=="],
@@ -6908,6 +6937,8 @@

    "@electron/fuses/fs-extra/jsonfile": ["jsonfile@6.2.1", "", { "dependencies": { "universalify": "^2.0.0" }, "optionalDependencies": { "graceful-fs": "^4.1.6" } }, "sha512-zwOTdL3rFQ/lRdBnntKVOX6k5cKJwEc1HdilT71BWEu7J41gXIB2MRp+vxduPSwZJPWBxEzv4yH1wYLJGUHX4Q=="],

+    "@electron/get/fs-extra/universalify": ["universalify@0.1.2", "", {}, "sha512-rBJeI5CXAlmy1pV+617WB9J63U6XcazHHF2f2dbJix4XzpUF0RS3Zbj0FGIOCAva5P/d/GBOYaACQ1w+0azUkg=="],
+
    "@electron/notarize/fs-extra/jsonfile": ["jsonfile@6.2.1", "", { "dependencies": { "universalify": "^2.0.0" }, "optionalDependencies": { "graceful-fs": "^4.1.6" } }, "sha512-zwOTdL3rFQ/lRdBnntKVOX6k5cKJwEc1HdilT71BWEu7J41gXIB2MRp+vxduPSwZJPWBxEzv4yH1wYLJGUHX4Q=="],

    "@electron/universal/fs-extra/jsonfile": ["jsonfile@6.2.1", "", { "dependencies": { "universalify": "^2.0.0" }, "optionalDependencies": { "graceful-fs": "^4.1.6" } }, "sha512-zwOTdL3rFQ/lRdBnntKVOX6k5cKJwEc1HdilT71BWEu7J41gXIB2MRp+vxduPSwZJPWBxEzv4yH1wYLJGUHX4Q=="],
@@ -7130,6 +7161,8 @@

    "@opencode-ai/web/astro/acorn": ["acorn@8.18.0", "", { "bin": { "acorn": "bin/acorn" } }, "sha512-lGq+9yr1/GuAWaVYIHRjvvySG5/4VfKIvC8EWxStPdcDh/Ka7FG3twP6v4d5BkravUilhIAsG4Qj83t02LWUPQ=="],

+    "@opencode-ai/web/astro/ci-info": ["ci-info@4.4.0", "", {}, "sha512-77PSwercCZU2Fc4sX94eF8k8Pxte6JAwL4/ICZLFjJLqegs7kCuAsqqj/70NQF6TvDpgFjkubQB2FW2ZZddvQg=="],
+
    "@opencode-ai/web/astro/common-ancestor-path": ["common-ancestor-path@1.0.1", "", {}, "sha512-L3sHRo1pXXEqX8VU28kfgUY+YGsk09hPqZiZmLacNib6XNTCM8ubYeT7ryXQw8asB1sKgcU5lkB7ONug08aB8w=="],

    "@opencode-ai/web/astro/cookie": ["cookie@1.1.1", "", {}, "sha512-ei8Aos7ja0weRpFzJnEA9UHJ/7XQmqglbRwnf2ATjcB9Wq874VKH9kfjjirM6UhU2/E5fFYadylyhFldcqSidQ=="],
@@ -7234,14 +7267,6 @@

    "ansi-align/string-width/strip-ansi": ["strip-ansi@6.0.1", "", { "dependencies": { "ansi-regex": "^5.0.1" } }, "sha512-Y38VPSHcqkFrCpFnQ9vuSXmquuv5oXOKpGeT6aGrr3o3Gc9AlVa6JBfUSOCnbxGGZF+/0ooI7KrPuUSztUdU5A=="],

-    "app-builder-lib/@electron/get/env-paths": ["env-paths@2.2.1", "", {}, "sha512-+h1lkLKhZMTYjog1VEpJNG7NZJWcuc2DDk/qsqSTRRCOXiLjeQ1d1/udrUGhqMxUgAlwKNZ0cf2uqan5GLuS2A=="],
-
-    "app-builder-lib/@electron/get/fs-extra": ["fs-extra@8.1.0", "", { "dependencies": { "graceful-fs": "^4.2.0", "jsonfile": "^4.0.0", "universalify": "^0.1.0" } }, "sha512-yhlQgA6mnOJUKOsRUFsgJdQCvkKhcz8tlZG5HBQfReYZy46OwLcY+Zia0mtdHsOo9y/hP+CxMN0TU9QxoOtG4g=="],
-
-    "app-builder-lib/@electron/get/semver": ["semver@6.3.1", "", { "bin": { "semver": "bin/semver.js" } }, "sha512-BR7VvDCVHO+q2xBEWskxS6DJE1qRnb7DxzUrogb71CWoSficBxYsiAGd+Kl0mmq/MprG9yArRkyrQxTO6XjMzA=="],
-
-    "app-builder-lib/hosted-git-info/lru-cache": ["lru-cache@6.0.0", "", { "dependencies": { "yallist": "^4.0.0" } }, "sha512-Jo6dJ04CmSjuznwJSS3pUeWmd/H0ffTlkXXgwZi+eq1UCmqQwCh+eLsYOYCwY991i2Fah4h1BEMCx4qThGbsiA=="],
-
    "app-builder-lib/js-yaml/argparse": ["argparse@2.0.1", "", {}, "sha512-8+9WqebbFzpX9OR+Wa6O29asIogeRMzcGtAINdpMHHyAg10f05aSFVBbcEqGf/PXw1EjAZ+q2/bEBg3DvurK3Q=="],

    "app-builder-lib/which/isexe": ["isexe@3.1.5", "", {}, "sha512-6B3tLtFqtQS4ekarvLVMZ+X+VlvQekbe4taUkf/rhVO3d/h0M2rfARm/pXLcPEsjjMsFgrFgSrhQIxcSVrBz8w=="],
@@ -7260,6 +7285,8 @@

    "astro-expressive-code/astro/acorn": ["acorn@8.18.0", "", { "bin": { "acorn": "bin/acorn" } }, "sha512-lGq+9yr1/GuAWaVYIHRjvvySG5/4VfKIvC8EWxStPdcDh/Ka7FG3twP6v4d5BkravUilhIAsG4Qj83t02LWUPQ=="],

+    "astro-expressive-code/astro/ci-info": ["ci-info@4.4.0", "", {}, "sha512-77PSwercCZU2Fc4sX94eF8k8Pxte6JAwL4/ICZLFjJLqegs7kCuAsqqj/70NQF6TvDpgFjkubQB2FW2ZZddvQg=="],
+
    "astro-expressive-code/astro/common-ancestor-path": ["common-ancestor-path@1.0.1", "", {}, "sha512-L3sHRo1pXXEqX8VU28kfgUY+YGsk09hPqZiZmLacNib6XNTCM8ubYeT7ryXQw8asB1sKgcU5lkB7ONug08aB8w=="],

    "astro-expressive-code/astro/cookie": ["cookie@1.1.1", "", {}, "sha512-ei8Aos7ja0weRpFzJnEA9UHJ/7XQmqglbRwnf2ATjcB9Wq874VKH9kfjjirM6UhU2/E5fFYadylyhFldcqSidQ=="],
@@ -7394,6 +7421,10 @@

    "electron-winstaller/fs-extra/universalify": ["universalify@0.1.2", "", {}, "sha512-rBJeI5CXAlmy1pV+617WB9J63U6XcazHHF2f2dbJix4XzpUF0RS3Zbj0FGIOCAva5P/d/GBOYaACQ1w+0azUkg=="],

+    "electron/@electron/get/env-paths": ["env-paths@3.0.0", "", {}, "sha512-dtJUTepzMW3Lm/NPxRf3wP4642UWhjL2sQxc+ym2YMj1m/H2zDNQOlezafzkHwn6sMstjHTwG6iQQsctDW/b1A=="],
+
+    "electron/@electron/get/undici": ["undici@7.29.0", "", {}, "sha512-IDxfleLmmbSskfWSUATiN1nfn2rDuvnMOqb5CWR92iIfojA0Ud+ulOAAEQ57LPr9rWmsreUyf5lwyao+7GNNVw=="],
+
    "esbuild-plugin-copy/chokidar/readdirp": ["readdirp@3.6.0", "", { "dependencies": { "picomatch": "^2.2.1" } }, "sha512-hOS089on8RduqdbhvQ5Z37A0ESjsqz6qnRcffsMU3495FuTdqSm+7bhJ29JvIOsBDEEnan5DPu9t3To9VRlMzA=="],

    "filelist/minimatch/brace-expansion": ["brace-expansion@2.1.4", "", { "dependencies": { "balanced-match": "^1.0.0" } }, "sha512-hGfVzPxthbf3+2yjg/RBs60cB0FhqBS/zvdV/4wn4/BmN0bNMMHPc4V/BbFieqf1TKAGGAHnY4eSjajCl0f2Xg=="],
@@ -7402,6 +7433,8 @@

    "gcp-metadata/gaxios/node-fetch": ["node-fetch@3.3.2", "", { "dependencies": { "data-uri-to-buffer": "^4.0.0", "fetch-blob": "^3.1.4", "formdata-polyfill": "^4.0.10" } }, "sha512-dRB78srN/l6gqWulah9SrxeYnxeddIG30+GOqK/9OlLVyLg3HPnr6SqOWTWOXKRwC2eGYCkZ59NNuSgvSrpgOA=="],

+    "hosted-git-info/lru-cache/yallist": ["yallist@4.0.0", "", {}, "sha512-3wdGidZyq5PB084XLES5TpOSRA3wjXAlIWMhum2kRcv/41Sn2emQ0dycQW4uZXLejwKvg6EsvbdlVL+FYEct7A=="],
+
    "js-beautify/glob/jackspeak": ["jackspeak@3.4.3", "", { "dependencies": { "@isaacs/cliui": "^8.0.2" }, "optionalDependencies": { "@pkgjs/parseargs": "^0.11.0" } }, "sha512-OGlZQpz2yfahA/Rd1Y8Cd9SIEsqvXkLVoSw/cgwhnhFMDbsQFeZYoJJ7bIZBS9BcamUW96asq/npPWugM+RQBw=="],

    "js-beautify/glob/minimatch": ["minimatch@9.0.9", "", { "dependencies": { "brace-expansion": "^2.0.2" } }, "sha512-OBwBN9AL4dqmETlpS2zasx+vTeWclWzkblfZk7KTA5j3jeOONz/tRCnZomUyvNg83wL5Zv9Ss6HMJXAgL8R2Yg=="],
@@ -7450,6 +7483,12 @@

    "miniflare/sharp/@img/sharp-win32-x64": ["@img/sharp-win32-x64@0.33.5", "", { "os": "win32", "cpu": "x64" }, "sha512-MpY/o8/8kj+EcnxwvrP4aTJSWw/aZ7JIGR4aBeZkZw5B7/Jn+tY9/VNwtcoGmdT7GfggGIU4kygOMSbYnOrAbg=="],

+    "minipass-flush/minipass/yallist": ["yallist@4.0.0", "", {}, "sha512-3wdGidZyq5PB084XLES5TpOSRA3wjXAlIWMhum2kRcv/41Sn2emQ0dycQW4uZXLejwKvg6EsvbdlVL+FYEct7A=="],
+
+    "minipass-pipeline/minipass/yallist": ["yallist@4.0.0", "", {}, "sha512-3wdGidZyq5PB084XLES5TpOSRA3wjXAlIWMhum2kRcv/41Sn2emQ0dycQW4uZXLejwKvg6EsvbdlVL+FYEct7A=="],
+
+    "openid-client/lru-cache/yallist": ["yallist@4.0.0", "", {}, "sha512-3wdGidZyq5PB084XLES5TpOSRA3wjXAlIWMhum2kRcv/41Sn2emQ0dycQW4uZXLejwKvg6EsvbdlVL+FYEct7A=="],
+
    "p-locate/p-limit/yocto-queue": ["yocto-queue@0.1.0", "", {}, "sha512-rVksvsnNCdJ/ohGc6xgPwyN8eheCxsiLM8mxuE/t/mOVqJewPuO1miLpTHQiRgTKCLexL4MeAFVagts7HmNZ2Q=="],

    "pkg-dir/find-up/locate-path": ["locate-path@5.0.0", "", { "dependencies": { "p-locate": "^4.1.0" } }, "sha512-t7hw9pI+WvuwNJXwk5zVHpyhIqzg2qTlklJOf0mVxGSbe3Fp2VieZcduNYjaLDoy6p9uGpQEGWG87WpMKlNq8g=="],
@@ -7476,6 +7515,8 @@

    "toolbeam-docs-theme/astro/acorn": ["acorn@8.18.0", "", { "bin": { "acorn": "bin/acorn" } }, "sha512-lGq+9yr1/GuAWaVYIHRjvvySG5/4VfKIvC8EWxStPdcDh/Ka7FG3twP6v4d5BkravUilhIAsG4Qj83t02LWUPQ=="],

+    "toolbeam-docs-theme/astro/ci-info": ["ci-info@4.4.0", "", {}, "sha512-77PSwercCZU2Fc4sX94eF8k8Pxte6JAwL4/ICZLFjJLqegs7kCuAsqqj/70NQF6TvDpgFjkubQB2FW2ZZddvQg=="],
+
    "toolbeam-docs-theme/astro/common-ancestor-path": ["common-ancestor-path@1.0.1", "", {}, "sha512-L3sHRo1pXXEqX8VU28kfgUY+YGsk09hPqZiZmLacNib6XNTCM8ubYeT7ryXQw8asB1sKgcU5lkB7ONug08aB8w=="],

    "toolbeam-docs-theme/astro/cookie": ["cookie@1.1.1", "", {}, "sha512-ei8Aos7ja0weRpFzJnEA9UHJ/7XQmqglbRwnf2ATjcB9Wq874VKH9kfjjirM6UhU2/E5fFYadylyhFldcqSidQ=="],
@@ -8154,8 +8195,6 @@

    "ansi-align/string-width/strip-ansi/ansi-regex": ["ansi-regex@5.0.1", "", {}, "sha512-quJQXlTSUGL2LH9SUXo8VwsY4soanhgo6LNSm84E1LBcE8s3O0wpdiRzyR9z/ZZJMlMWv37qOOb9pdJlMUEKFQ=="],

-    "app-builder-lib/@electron/get/fs-extra/universalify": ["universalify@0.1.2", "", {}, "sha512-rBJeI5CXAlmy1pV+617WB9J63U6XcazHHF2f2dbJix4XzpUF0RS3Zbj0FGIOCAva5P/d/GBOYaACQ1w+0azUkg=="],
-
    "archiver-utils/glob/jackspeak/@isaacs/cliui": ["@isaacs/cliui@8.0.2", "", { "dependencies": { "string-width": "^5.1.2", "string-width-cjs": "npm:string-width@^4.2.0", "strip-ansi": "^7.0.1", "strip-ansi-cjs": "npm:strip-ansi@^6.0.1", "wrap-ansi": "^8.1.0", "wrap-ansi-cjs": "npm:wrap-ansi@^7.0.0" } }, "sha512-O8jcjabXaleOG9DQ0+ARXWZBTfnP4WNAqzuiJK7ll44AmxGKv/J2M4TPjxjY3znBCfvBXFzucm1twdyFybFqEA=="],

    "archiver-utils/glob/minimatch/brace-expansion": ["brace-expansion@2.1.4", "", { "dependencies": { "balanced-match": "^1.0.0" } }, "sha512-hGfVzPxthbf3+2yjg/RBs60cB0FhqBS/zvdV/4wn4/BmN0bNMMHPc4V/BbFieqf1TKAGGAHnY4eSjajCl0f2Xg=="],
@@ -1,4 +1,5 @@
 import { domain } from "./stage"
+import { createWebApp } from "./webapp"

 const GITHUB_APP_ID = new sst.Secret("GITHUB_APP_ID")
 const GITHUB_APP_PRIVATE_KEY = new sst.Secret("GITHUB_APP_PRIVATE_KEY")
@@ -59,12 +60,4 @@ new sst.cloudflare.x.Astro("Web", {
  },
 })

-new sst.cloudflare.StaticSite("WebApp", {
-  domain: "app." + domain,
-  path: "packages/app",
-  build: {
-    // Preserve Sentry credentials and run source-map uploads on every deployment.
-    command: "bun run build",
-    output: "./dist",
-  },
-})
+createWebApp("app." + domain)
@@ -0,0 +1,18 @@
+export function createWebApp(domain: string) {
+  return new sst.cloudflare.StaticSite("WebApp", {
+    domain,
+    path: "packages/app",
+    environment:
+      $app.stage === "beta"
+        ? {
+            OPENCODE_CHANNEL: "beta",
+            VITE_SENTRY_ENVIRONMENT: "beta",
+          }
+        : undefined,
+    build: {
+      // Preserve Sentry credentials and run source-map uploads on every deployment.
+      command: "bun run build",
+      output: "./dist",
+    },
+  })
+}
@@ -1,8 +1,8 @@
 {
  "nodeModules": {
-    "x86_64-linux": "sha256-EtUp4pHl9TyPtRrLGvk/X7kd2LuIxNxCpUwF5aLtzN4=",
-    "aarch64-linux": "sha256-m0j/pMZCguclR3/T9JmzCfi11YzmIvBFyR2bVhIO37Y=",
-    "aarch64-darwin": "sha256-nqefk68ZTUfNU15q1WkXaGsFzPNwOjCtMHpp6WrpNqM=",
-    "x86_64-darwin": "sha256-syD7hX62E4yCDV/wux1QKw4q/zZr24f99Y2mmzMJo6o="
+    "x86_64-linux": "sha256-fG6VYtNC0pce4VM9po7vVucPuJul42yuuijTjNSr7rk=",
+    "aarch64-linux": "sha256-3TznrmNqdt25cOxia6vcdi/5qKaeyLPIsNXGYBSJNrs=",
+    "aarch64-darwin": "sha256-8Kmagb5tfECSWZNsIJgrRP1d3X5tuEoWLEWkV3UENZo=",
+    "x86_64-darwin": "sha256-mIV+mDwIGD02BNYZVi37sY4ls1T01N6z76eBtH0sKiA="
  }
 }
@@ -249,6 +249,107 @@ The published legacy `Service`, `layer`, `clientLayer`, and module-level control
 over the same implementation, including the legacy live `requests` array. New tests should use `Test` and
 `testLayer`.

+## Provider compaction
+
+Compaction is opt-in. The package supports automatic compaction in OpenAI/Azure Responses and Anthropic Messages (including Claude on Vertex and Bedrock Messages), and explicit compaction calls in OpenAI/Azure/xAI Responses. Model and deployment support still depends on the provider.
+
+This is different from prompt caching, server-side history storage, or truncation. Compaction returns provider-owned context that must be replayed to continue the conversation.
+
+### Automatic compaction
+
+Inside an `Effect.gen`, enable OpenAI compaction with typed provider options:
+
+```ts
+import { LLM, LLMClient, LLMRequest, Message } from "@opencode-ai/ai"
+import { OpenAI } from "@opencode-ai/ai/providers"
+
+const request = LLM.request({
+  model: OpenAI.configure({ apiKey }).responses("gpt-5.3-codex"),
+  messages,
+  providerOptions: {
+    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")],
+})
+```
+
+`store: false` remains the default. Keep the entire `response.message`, not just `response.text`. Compaction events become ordered `CompactionPart`s alongside text and reasoning. The conversation contains everything needed to continue; there is no separate replay object or hidden provider transcript.
+
+A compaction part has `provider` and exactly one representation: `encrypted` for Responses, or `text` for Anthropic. Responses also preserves the optional checkpoint `id`. These fields survive message serialization without becoming visible assistant text. Sending a checkpoint to another provider or an incompatible API fails rather than silently losing context.
+
+```ts
+import { CompactionPart, ProviderID } from "@opencode-ai/ai"
+
+CompactionPart.make({ provider: ProviderID.make("openai"), id: "cmp_123", encrypted: "..." })
+CompactionPart.make({ provider: ProviderID.make("anthropic"), text: "Summary of the conversation..." })
+```
+
+For Anthropic, use:
+
+```ts
+providerOptions: {
+  contextManagement: {
+    edits: [{
+      type: "compact_20260112",
+      trigger: { type: "input_tokens", value: 150_000 },
+      pauseAfterCompaction: true,
+      instructions: "Summarize the task and decisions. Do not call tools while summarizing.",
+    }],
+  },
+}
+```
+
+- The trigger is optional (provider default: 150,000 tokens), with a minimum of 50,000.
+- Custom instructions replace Anthropic's default summarization instructions.
+- The route adds `compact-2026-01-12` to existing beta headers, including when replaying a checkpoint without enabling new compactions.
+- A pause is exposed as `response.finishReason.raw === "compaction"`. The caller explicitly issues the next request; the package never automatically resumes.
+- Anthropic can return a compaction block with `content: null` when summarization fails. This becomes a compaction part with `text: null`, which is **not** a successful replacement for prior history. The package never prunes history automatically.
+- `Usage` totals include all reported Anthropic `usage.iterations`, including compaction. `contextTokens` separately reports the final message iteration's inclusive input size, when available. A compaction-only pause does not report a post-compaction context size. Raw iteration usage remains in `providerMetadata`.
+
+Bedrock's Converse API does not support this feature. Select the native Claude Messages route explicitly; the default `.model(...)` remains Converse:
+
+```ts
+import { AmazonBedrock } from "@opencode-ai/ai/providers"
+
+const model = AmazonBedrock.configure({ region: "us-east-1", credentials }).messages("us.anthropic.claude-opus-4-6-v1")
+```
+
+The corresponding package entrypoint is `@opencode-ai/ai/providers/amazon-bedrock/messages`. It uses InvokeModelWithResponseStream, AWS event-stream framing, bearer or SigV4 auth, and `anthropic_beta` in the request body.
+
+### Explicit compaction
+
+`LLMClient.compact(request)` performs exactly one HTTP call to `/responses/compact`, using the selected route's endpoint, credentials, query, and HTTP middleware. It returns a `CompactionResponse` containing replacement `messages` and usage, not a normal generation response.
+
+```ts
+const compacted = yield * LLMClient.compact(request)
+const next = LLMRequest.update(request, {
+  messages: [...compacted.messages, Message.user("Continue")],
+})
+const response = yield * LLMClient.generate(next)
+```
+
+Replace the prior window with `compacted.messages`. 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. Generation-only body overlays such as `stream` and `store` are not sent to the compact endpoint.
+
+The input must still fit the model's context window. Explicit compaction is not an overflow-recovery operation. xAI supports this explicit path, not the automatic OpenAI option. Unsupported routes, including Bedrock Mantle, do not inherit an explicit compact endpoint simply because they use a Responses protocol.
+
+### Ownership and verification
+
+The AI package transports options and typed conversation parts. It does not schedule compaction, persist Session checkpoints, select history, switch providers, or replace Core's existing local compaction policy. Native compaction is not enabled for OpenCode Sessions by this feature; Session integration must persist these parts before enabling it. The AI SDK bridge rejects native compaction parts rather than dropping them. Provider-executed tool APIs and persistence changes are a separate follow-up.
+
+Tests cover serialized round trips, real local HTTP plus a tool loop, AWS binary frames and signing, provider errors, malformed blocks, and usage accounting. Live provider tests are gated by `RECORD=true` and the relevant API keys:
+
+```sh
+# Run from packages/ai. Only records the selected new cassette group.
+RECORD=true RECORDED_PREFIX=openai-compaction bun test test/provider/compaction.recorded.test.ts
+RECORD=true RECORDED_PREFIX=xai-compaction bun test test/provider/compaction.recorded.test.ts
+RECORD=true RECORDED_PREFIX=anthropic-compaction bun test test/provider/compaction.recorded.test.ts
+```
+
+Provider references: [OpenAI](https://developers.openai.com/api/docs/guides/compaction), [Azure](https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/responses#server-side-compaction), [Anthropic](https://platform.claude.com/docs/en/build-with-claude/compaction), [Bedrock](https://docs.aws.amazon.com/bedrock/latest/userguide/claude-messages-compaction.html), [xAI](https://docs.x.ai/developers/advanced-api-usage/context-compaction).
+
 ## Caching

 Prompt caching is **on by default**. Every `LLMRequest` resolves to `cache: "auto"` unless the caller opts out with `cache: "none"`. Each protocol translates `CacheHint`s to its wire format (`cache_control` on Anthropic, `cachePoint` on Bedrock; OpenAI and Gemini do implicit caching server-side and don't need inline markers — auto is a no-op there).
@@ -40,6 +40,7 @@ const RESPECTS_INLINE_HINTS = new Set([
  "anthropic-messages",
  "google-vertex-messages",
  "bedrock-converse",
+  "bedrock-messages",
  "openrouter",
 ])

@@ -6,8 +6,12 @@ import { Auth } from "../route/auth.js"
 import { Endpoint } from "../route/endpoint.js"
 import { Framing } from "../route/framing.js"
 import { Protocol } from "../route/protocol.js"
+import { Headers } from "effect/unstable/http"
+import { HttpTransport } from "../route/transport/index.js"
 import {
  AIError,
+  HttpOptions,
+  LLMRequest,
  LLMEvent,
  mergeJsonRecords,
  Usage,
@@ -15,7 +19,6 @@ import {
  type FinishReasonDetails,
  type FinishReason,
  type JsonSchema,
-  type LLMRequest,
  type MediaPart,
  type ProviderMetadata,
  type ToolCallPart,
@@ -61,6 +64,7 @@ export type ThinkingInput =
    ))

 export interface OptionsInput {
+  readonly contextManagement?: ContextManagement
  readonly [key: string]: unknown
  readonly thinking?: ThinkingInput
  readonly effort?: string
@@ -89,6 +93,23 @@ export interface OptionsInput {

 export type ProviderOptionsInput = OptionsInput

+export const ContextManagement = Schema.Struct({
+  edits: Schema.Array(
+    Schema.Struct({
+      type: Schema.Literal("compact_20260112"),
+      trigger: Schema.optional(
+        Schema.Struct({
+          type: Schema.Literal("input_tokens"),
+          value: Schema.Int.check(Schema.isGreaterThanOrEqualTo(50000)),
+        }),
+      ),
+      pauseAfterCompaction: Schema.optional(Schema.Boolean),
+      instructions: Schema.optional(Schema.String),
+    }),
+  ),
+})
+export type ContextManagement = typeof ContextManagement.Type
+
 // =============================================================================
 // Request Body Schema
 // =============================================================================
@@ -236,7 +257,12 @@ const AnthropicUserBlock = Schema.Union([
  AnthropicToolResultBlock,
 ])
 type AnthropicUserBlock = Schema.Schema.Type<typeof AnthropicUserBlock>
+const AnthropicCompactionBlock = Schema.Struct({
+  type: Schema.Literal("compaction"),
+  content: Schema.NullOr(Schema.String),
+})
 const AnthropicAssistantBlock = Schema.Union([
+  AnthropicCompactionBlock,
  AnthropicTextBlock,
  AnthropicThinkingBlock,
  AnthropicRedactedThinkingBlock,
@@ -312,6 +338,18 @@ const AnthropicContainer = Schema.Union([
 ])

 const AnthropicBodyFields = {
+  context_management: Schema.optional(
+    Schema.Struct({
+      edits: Schema.Array(
+        Schema.Struct({
+          type: Schema.Literal("compact_20260112"),
+          trigger: ContextManagement.fields.edits.value.fields.trigger,
+          pause_after_compaction: Schema.optional(Schema.Boolean),
+          instructions: Schema.optional(Schema.String),
+        }),
+      ),
+    }),
+  ),
  model: Schema.String,
  system: optionalArray(AnthropicTextBlock),
  messages: Schema.Array(AnthropicMessage),
@@ -335,7 +373,7 @@ const AnthropicBodyFields = {
 export const AnthropicMessagesBody = Schema.Struct(AnthropicBodyFields)
 export type AnthropicMessagesBody = Schema.Schema.Type<typeof AnthropicMessagesBody>

-const AnthropicUsage = Schema.StructWithRest(
+const AnthropicIterationUsage = Schema.StructWithRest(
  Schema.Struct({
    input_tokens: optionalNull(Schema.Number),
    output_tokens: Schema.optional(Schema.Number),
@@ -354,6 +392,13 @@ const AnthropicUsage = Schema.StructWithRest(
  }),
  [Schema.Record(Schema.String, Schema.Unknown)],
 )
+const AnthropicUsage = Schema.StructWithRest(
+  Schema.Struct({
+    ...AnthropicIterationUsage.schema.fields,
+    iterations: Schema.optional(Schema.Array(AnthropicIterationUsage)),
+  }),
+  [JsonObject],
+)
 type AnthropicUsage = Schema.Schema.Type<typeof AnthropicUsage>

 const AnthropicStreamBlock = Schema.Struct({
@@ -377,6 +422,7 @@ type AnthropicStreamBlock = Schema.Schema.Type<typeof AnthropicStreamBlock>
 const decodeAnthropicStreamBlock = Schema.decodeUnknownOption(AnthropicStreamBlock)

 const AnthropicStreamDelta = Schema.Struct({
+  content: optionalNull(Schema.String),
  type: Schema.optional(Schema.String),
  text: Schema.optional(Schema.String),
  thinking: Schema.optional(Schema.String),
@@ -406,6 +452,8 @@ const AnthropicEvent = Schema.Struct({
 type AnthropicEvent = Schema.Schema.Type<typeof AnthropicEvent>

 interface ParserState {
+  readonly provider: LLMRequest["model"]["provider"]
+  readonly compactions: Readonly<Record<number, string | null>>
  readonly providerMetadataKey: string
  readonly tools: ToolStream.State<number>
  readonly reasoningSignatures: Readonly<Record<number, string>>
@@ -831,6 +879,7 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
      const content: AnthropicUserBlock[] = []
      for (const part of message.content) {
        if (part.type === "text") {
+          if (part.text.trim().length === 0) continue
          content.push({ type: "text", text: part.text, cache_control: cacheControl(breakpoints, part.cache) })
          continue
        }
@@ -840,14 +889,21 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
        }
        return yield* ProviderShared.unsupportedContent("Anthropic Messages", "user", ["text", "media"])
      }
-      messages.push({ role: "user", content })
+      if (content.length > 0) messages.push({ role: "user", content })
      continue
    }

    if (message.role === "assistant") {
      const content: AnthropicAssistantBlock[] = []
      for (const part of message.content) {
+        if (part.type === "compaction") {
+          if (part.provider !== request.model.provider || part.text === undefined)
+            return yield* invalid("Compaction state must be replayed to its originating provider and API")
+          content.push({ type: "compaction", content: part.text })
+          continue
+        }
        if (part.type === "text") {
+          if (part.text.trim().length === 0) continue
          content.push({ type: "text", text: part.text, cache_control: cacheControl(breakpoints, part.cache) })
          continue
        }
@@ -891,7 +947,7 @@ const lowerMessages = Effect.fn("AnthropicMessages.lowerMessages")(function* (
          `Anthropic Messages assistant messages only support text, reasoning, and tool-call content for now`,
        )
      }
-      messages.push({ role: "assistant", content })
+      if (content.length > 0) messages.push({ role: "assistant", content })
      continue
    }

@@ -1001,6 +1057,9 @@ const resolveThinking = Effect.fn("AnthropicMessages.resolveThinking")(function*
 })

 const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (request: LLMRequest) {
+  const management = yield* ProviderShared.validateWith(
+    Schema.decodeUnknownEffect(Schema.UndefinedOr(ContextManagement)),
+  )(request.providerOptions?.contextManagement)
  const generation = request.generation
  const toolSchemaCompatibility = request.model.compatibility?.toolSchema
  // Allocate the 4-breakpoint budget in invalidation order: tools → system →
@@ -1019,10 +1078,11 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
        )
  // 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)
  const system =
-    request.system.length === 0
+    systemParts.length === 0
      ? undefined
-      : request.system.map((part) => ({
+      : systemParts.map((part) => ({
          type: "text" as const,
          text: part.text,
          cache_control: cacheControl(breakpoints, part.cache),
@@ -1034,7 +1094,7 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
    )
  }
  const options = yield* resolveOptions(request)
-  return {
+  const body = {
    model: request.model.id,
    system,
    messages,
@@ -1055,6 +1115,18 @@ const fromRequest = Effect.fn("AnthropicMessages.fromRequest")(function* (reques
    metadata: options.metadata,
    service_tier: options.service_tier,
  }
+  if (!management) return body
+  return {
+    ...body,
+    context_management: {
+      edits: management.edits.map((edit) => ({
+        type: edit.type,
+        trigger: edit.trigger,
+        pause_after_compaction: edit.pauseAfterCompaction,
+        instructions: edit.instructions,
+      })),
+    },
+  }
 })

 // =============================================================================
@@ -1076,18 +1148,31 @@ const mapFinishReason = (reason: string | null | undefined): FinishReason => {
 // expose that subset through `output_tokens_details.thinking_tokens`.
 const mapUsage = (usage: AnthropicUsage | undefined, providerMetadataKey: string): Usage | undefined => {
  if (!usage) return undefined
-  const nonCached = usage.input_tokens ?? undefined
-  const cacheRead = usage.cache_read_input_tokens ?? undefined
-  const cacheWrite = usage.cache_creation_input_tokens ?? undefined
+  const iterations = usage.iterations?.length ? usage.iterations : [usage]
+  const last = usage.iterations?.at(-1)
+  const nonCached = ProviderShared.sumTokens(...iterations.map((item) => item.input_tokens ?? undefined))
+  const cacheRead = ProviderShared.sumTokens(...iterations.map((item) => item.cache_read_input_tokens ?? undefined))
+  const cacheWrite = ProviderShared.sumTokens(
+    ...iterations.map((item) => item.cache_creation_input_tokens ?? undefined),
+  )
  const inputTokens = ProviderShared.sumTokens(nonCached, cacheRead, cacheWrite)
+  const outputTokens = ProviderShared.sumTokens(...iterations.map((item) => item.output_tokens))
  return new Usage({
    inputTokens,
-    outputTokens: usage.output_tokens,
+    outputTokens,
+    contextTokens:
+      last?.type === "message"
+        ? ProviderShared.sumTokens(
+            last.input_tokens ?? undefined,
+            last.cache_read_input_tokens ?? undefined,
+            last.cache_creation_input_tokens ?? undefined,
+          )
+        : undefined,
    nonCachedInputTokens: nonCached,
    cacheReadInputTokens: cacheRead,
    cacheWriteInputTokens: cacheWrite,
-    reasoningTokens: usage.output_tokens_details?.thinking_tokens,
-    totalTokens: ProviderShared.totalTokens(inputTokens, usage.output_tokens, undefined),
+    reasoningTokens: ProviderShared.sumTokens(...iterations.map((item) => item.output_tokens_details?.thinking_tokens)),
+    totalTokens: ProviderShared.totalTokens(inputTokens, outputTokens, undefined),
    providerMetadata: { [providerMetadataKey]: usage },
  })
 }
@@ -1109,6 +1194,7 @@ const mergeUsage = (left: Usage | undefined, right: Usage | undefined, providerM
  return new Usage({
    inputTokens,
    outputTokens,
+    contextTokens: right.contextTokens ?? left.contextTokens,
    nonCachedInputTokens,
    cacheReadInputTokens,
    cacheWriteInputTokens,
@@ -1167,7 +1253,6 @@ const onContentBlockStart = (
  event: AnthropicEvent & { readonly content_block: AnthropicStreamBlock },
 ): StepResult => {
  const block = event.content_block
-  if (!block) return [state, NO_EVENTS]

  if (block.type === "tool_use" || block.type === "server_tool_use") {
    if (event.index === undefined || !block.id) return [state, NO_EVENTS]
@@ -1262,7 +1347,16 @@ const onContentBlockDelta = Effect.fn("AnthropicMessages.onContentBlockDelta")(f
 ) {
  const delta = event.delta

-  if (delta?.type === "text_delta" && delta.text) {
+  if (delta.type === "compaction_delta") {
+    if (event.index === undefined || !(event.index in state.compactions) || delta.content === undefined)
+      return yield* ProviderShared.eventError(ADAPTER, "Compaction delta is missing its block or content")
+    return [
+      { ...state, compactions: { ...state.compactions, [event.index]: delta.content } },
+      NO_EVENTS,
+    ] satisfies StepResult
+  }
+
+  if (delta.type === "text_delta" && delta.text) {
    if (!state.lifecycle.text.has(`text-${event.index ?? 0}`)) return [state, NO_EVENTS] satisfies StepResult
    const events: LLMEvent[] = []
    return [
@@ -1271,7 +1365,7 @@ const onContentBlockDelta = Effect.fn("AnthropicMessages.onContentBlockDelta")(f
    ] satisfies StepResult
  }

-  if (delta?.type === "thinking_delta" && delta.thinking) {
+  if (delta.type === "thinking_delta" && delta.thinking) {
    if (!state.lifecycle.reasoning.has(`reasoning-${event.index ?? 0}`)) return [state, NO_EVENTS] satisfies StepResult
    const events: LLMEvent[] = []
    return [
@@ -1283,7 +1377,7 @@ const onContentBlockDelta = Effect.fn("AnthropicMessages.onContentBlockDelta")(f
    ] satisfies StepResult
  }

-  if (delta?.type === "signature_delta" && delta.signature) {
+  if (delta.type === "signature_delta" && delta.signature) {
    const index = event.index ?? 0
    if (!state.lifecycle.reasoning.has(`reasoning-${index}`)) return [state, NO_EVENTS] satisfies StepResult
    return [
@@ -1295,7 +1389,7 @@ const onContentBlockDelta = Effect.fn("AnthropicMessages.onContentBlockDelta")(f
    ] satisfies StepResult
  }

-  if (delta?.type === "input_json_delta" && event.index !== undefined) {
+  if (delta.type === "input_json_delta" && event.index !== undefined) {
    if (!delta.partial_json) return [state, NO_EVENTS] satisfies StepResult
    if (!state.tools[event.index]) return [state, NO_EVENTS] satisfies StepResult
    const result = ToolStream.appendExisting(
@@ -1320,6 +1414,18 @@ const onContentBlockStop = Effect.fn("AnthropicMessages.onContentBlockStop")(fun
  event: AnthropicEvent,
 ) {
  if (event.index === undefined) return [state, NO_EVENTS] satisfies StepResult
+  if (event.index in state.compactions) {
+    const { [event.index]: content, ...compactions } = state.compactions
+    const events: LLMEvent[] = []
+    const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
+    events.push(
+      LLMEvent.compaction({
+        provider: state.provider,
+        text: content,
+      }),
+    )
+    return [{ ...state, compactions, lifecycle }, events] satisfies StepResult
+  }
  const result = yield* ToolStream.finish(ADAPTER, state.tools, event.index)
  const events: LLMEvent[] = []
  const resultEvents = result.events ?? []
@@ -1343,31 +1449,51 @@ const onMessageDelta = (
  event: AnthropicEvent & { readonly delta?: AnthropicStreamDelta },
 ): StepResult => {
  const usage = mergeUsage(state.usage, mapUsage(event.usage, state.providerMetadataKey), state.providerMetadataKey)
+  const pendingFinish = (() => {
+    const stopReason = event.delta?.stop_reason
+    if (stopReason === null || stopReason === undefined) return state.pendingFinish
+
+    const stopSequence = event.delta?.stop_sequence
+    const finishMetadata =
+      stopSequence === null || stopSequence === undefined
+        ? state.pendingFinish?.providerMetadata
+        : providerMetadata(state.providerMetadataKey, { stopSequence })
+    return {
+      reason: {
+        normalized: mapFinishReason(stopReason),
+        raw: stopReason,
+      },
+      providerMetadata: finishMetadata,
+    }
+  })()
  return [
    {
      ...state,
      usage,
-      pendingFinish: {
-        reason: {
-          normalized: mapFinishReason(event.delta?.stop_reason),
-          raw: event.delta?.stop_reason ?? undefined,
-        },
-        providerMetadata:
-          event.delta?.stop_sequence === null || event.delta?.stop_sequence === undefined
-            ? undefined
-            : providerMetadata(state.providerMetadataKey, { stopSequence: event.delta.stop_sequence }),
-      },
+      pendingFinish,
    },
    NO_EVENTS,
  ]
 }

 const onMessageStop = Effect.fn("AnthropicMessages.onMessageStop")(function* (state: ParserState) {
+  if (Object.keys(state.compactions).length)
+    return yield* ProviderShared.eventError(ADAPTER, "Response ended with an incomplete compaction block")
  const result = yield* ToolStream.finishAll(ADAPTER, state.tools)
  const events: LLMEvent[] = []
  const lifecycle = result.events.length ? Lifecycle.stepStart(state.lifecycle, events) : state.lifecycle
  events.push(...result.events)
-  const finished = Lifecycle.finish(lifecycle, events, {
+  const closed = Object.entries(state.reasoningSignatures).reduce(
+    (current, [index, signature]) =>
+      Lifecycle.reasoningEnd(
+        current,
+        events,
+        `reasoning-${index}`,
+        providerMetadata(state.providerMetadataKey, { signature }),
+      ),
+    lifecycle,
+  )
+  const finished = Lifecycle.finish(closed, events, {
    reason: state.pendingFinish?.reason ?? {
      normalized: "unknown",
      raw: undefined,
@@ -1397,16 +1523,21 @@ const onError = (event: AnthropicEvent) => {
  )
 }

-const isKnownStreamBlockType = (type: string) =>
-  type === "text" ||
-  type === "thinking" ||
-  type === "redacted_thinking" ||
-  type === "tool_use" ||
-  type === "server_tool_use" ||
-  isServerToolResultType(type)
-
-const isKnownStreamDeltaType = (type: string) =>
-  type === "text_delta" || type === "thinking_delta" || type === "signature_delta" || type === "input_json_delta"
+const STREAM_BLOCK_TYPES = new Set([
+  "compaction",
+  "text",
+  "thinking",
+  "redacted_thinking",
+  "tool_use",
+  "server_tool_use",
+])
+const STREAM_DELTA_TYPES = new Set([
+  "compaction_delta",
+  "text_delta",
+  "thinking_delta",
+  "signature_delta",
+  "input_json_delta",
+])

 const invalidStreamEvent = (event: AnthropicEvent) =>
  Effect.fail(
@@ -1435,7 +1566,16 @@ const step = (state: ParserState, event: AnthropicEvent) => {
  if (event.type === "content_block_start") {
    if (!ProviderShared.isRecord(event.content_block) || typeof event.content_block.type !== "string")
      return invalidStreamEvent(event)
-    if (!isKnownStreamBlockType(event.content_block.type)) return Effect.succeed<StepResult>([state, NO_EVENTS])
+    if (event.content_block.type === "compaction") {
+      const decoded = Schema.decodeUnknownOption(AnthropicCompactionBlock)(event.content_block)
+      if (event.index === undefined || Option.isNone(decoded)) return invalidStreamEvent(event)
+      return Effect.succeed<StepResult>([
+        { ...state, compactions: { ...state.compactions, [event.index]: decoded.value.content } },
+        NO_EVENTS,
+      ])
+    }
+    if (!STREAM_BLOCK_TYPES.has(event.content_block.type) && !isServerToolResultType(event.content_block.type))
+      return Effect.succeed<StepResult>([state, NO_EVENTS])
    const decoded = decodeAnthropicStreamBlock(event.content_block)
    if (Option.isNone(decoded)) return invalidStreamEvent(event)
    const block = decoded.value
@@ -1449,7 +1589,7 @@ const step = (state: ParserState, event: AnthropicEvent) => {
  }
  if (event.type === "content_block_delta") {
    if (!ProviderShared.isRecord(event.delta)) return invalidStreamEvent(event)
-    if (typeof event.delta.type === "string" && !isKnownStreamDeltaType(event.delta.type))
+    if (typeof event.delta.type === "string" && !STREAM_DELTA_TYPES.has(event.delta.type))
      return Effect.succeed<StepResult>([state, NO_EVENTS])
    const decoded = decodeAnthropicStreamDelta(event.delta)
    if (Option.isNone(decoded)) return invalidStreamEvent(event)
@@ -1483,6 +1623,8 @@ export const protocol = Protocol.make({
  stream: {
    event: Protocol.jsonEvent(AnthropicEvent),
    initial: (request) => ({
+      provider: request.model.provider,
+      compactions: {},
      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
      tools: ToolStream.empty<number>(),
      reasoningSignatures: {},
@@ -1492,6 +1634,37 @@ export const protocol = Protocol.make({
  },
 })

+export const transport = <Body extends Pick<AnthropicMessagesBody, "messages" | "context_management">>() => {
+  const http = HttpTransport.httpJson<Body, string>({ framing })
+  return {
+    ...http,
+    prepare: (input: Parameters<typeof http.prepare>[0]) => {
+      if (
+        !input.body.context_management?.edits.length &&
+        !input.body.messages.some((message) => message.content.some((block) => block.type === "compaction"))
+      )
+        return http.prepare(input)
+      const headers = Headers.fromInput(input.request.http?.headers)
+      const betas = new Set(
+        (headers["anthropic-beta"] ?? "")
+          .split(",")
+          .map((item) => item.trim())
+          .filter(Boolean),
+      )
+      betas.add("compact-2026-01-12")
+      return http.prepare({
+        ...input,
+        request: LLMRequest.update(input.request, {
+          http: new HttpOptions({
+            ...input.request.http,
+            headers: { ...headers, "anthropic-beta": [...betas].join(",") },
+          }),
+        }),
+      })
+    },
+  }
+}
+
 export const route = Route.make({
  id: ADAPTER,
  provider: "anthropic",
@@ -1501,7 +1674,7 @@ export const route = Route.make({
    baseURL: DEFAULT_BASE_URL,
  }),
  auth: Auth.none,
-  framing,
+  transport: transport<AnthropicMessagesBody>(),
  headers: () => ({ "anthropic-version": "2023-06-01" }),
 })

@@ -1,4 +1,4 @@
-import { Effect, Schema } from "effect"
+import { Effect, Encoding, Schema } from "effect"
 import { Route } from "../route/client.js"
 import { Endpoint } from "../route/endpoint.js"
 import { Protocol } from "../route/protocol.js"
@@ -262,24 +262,28 @@ const providerMetadata = (key: string, metadata: Record<string, unknown>): Provi

 const reasoningSignature = (part: ReasoningPart, providerMetadataKey: string) => {
  const metadata = part.providerMetadata?.[providerMetadataKey]
-  return (
-    part.encrypted ??
-    (ProviderShared.isRecord(metadata) && typeof metadata.signature === "string" ? metadata.signature : undefined)
-  )
+  if (part.encrypted !== undefined) return part.encrypted
+  if (ProviderShared.isRecord(metadata) && typeof metadata.signature === "string") return metadata.signature
 }

 const reasoningRedactedData = (part: ReasoningPart, providerMetadataKey: string) => {
  const metadata = part.providerMetadata?.[providerMetadataKey]
-  return ProviderShared.isRecord(metadata) && typeof metadata.redactedData === "string"
-    ? metadata.redactedData
-    : undefined
+  if (ProviderShared.isRecord(metadata) && typeof metadata.redactedData === "string") return metadata.redactedData
+}
+
+const removeEmptyToolInputKeys = (input: unknown): unknown => {
+  if (Array.isArray(input)) return input.map(removeEmptyToolInputKeys)
+  if (!ProviderShared.isRecord(input)) return input
+  return Object.fromEntries(
+    Object.entries(input).flatMap(([key, value]) => (key === "" ? [] : [[key, removeEmptyToolInputKeys(value)]])),
+  )
 }

 const lowerToolCall = (part: ToolCallPart): BedrockToolUseBlock => ({
  toolUse: {
    toolUseId: part.id,
    name: part.name,
-    input: part.input,
+    input: removeEmptyToolInputKeys(part.input),
  },
 })

@@ -414,7 +418,12 @@ const lowerMessages = Effect.fn("BedrockConverse.lowerMessages")(function* (
 const lowerSystem = (
  breakpoints: BedrockCache.Breakpoints,
  system: ReadonlyArray<LLMRequest["system"][number]>,
-): BedrockSystemBlock[] => system.flatMap((part) => textWithCache(breakpoints, part.text, part.cache))
+) => {
+  const content = system
+    .filter((part) => part.text.length > 0)
+    .flatMap((part) => textWithCache(breakpoints, part.text, part.cache))
+  return content.length === 0 ? undefined : content
+}

 const fromRequest = Effect.fn("BedrockConverse.fromRequest")(function* (request: LLMRequest) {
  const toolChoice = request.toolChoice ? yield* lowerToolChoice(request.toolChoice) : undefined
@@ -422,38 +431,42 @@ const fromRequest = Effect.fn("BedrockConverse.fromRequest")(function* (request:
  // 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()
-  const toolConfig =
-    request.tools.length > 0
-      ? {
-          tools: lowerTools(request.model.compatibility?.toolSchema, breakpoints, request.tools),
-          // Converse has no native "none". Keep definitions stable for prompt
-          // caching and omit only the unsupported choice.
-          toolChoice,
-        }
-      : undefined
-  const system = request.system.length === 0 ? undefined : lowerSystem(breakpoints, request.system)
+  const toolConfig = (() => {
+    if (request.tools.length === 0) return undefined
+    return {
+      tools: lowerTools(request.model.compatibility?.toolSchema, breakpoints, request.tools),
+      // Converse has no native "none". Keep definitions stable for prompt
+      // caching and omit only the unsupported choice.
+      toolChoice,
+    }
+  })()
+  const system = lowerSystem(breakpoints, request.system)
  const messages = yield* lowerMessages(request, breakpoints)
  if (breakpoints.dropped > 0) {
    yield* Effect.logWarning(
      `Bedrock Converse: dropped ${breakpoints.dropped} cache breakpoint(s); the API allows at most ${BedrockCache.BEDROCK_BREAKPOINT_CAP} per request.`,
    )
  }
-  return {
-    modelId: request.model.id,
-    messages,
-    system,
-    inferenceConfig:
+  const inferenceConfig = (() => {
+    if (
      generation?.maxTokens === undefined &&
      generation?.temperature === undefined &&
      generation?.topP === undefined &&
      (generation?.stop === undefined || generation.stop.length === 0)
-        ? undefined
-        : {
-            maxTokens: generation?.maxTokens,
-            temperature: generation?.temperature,
-            topP: generation?.topP,
-            stopSequences: generation?.stop,
-          },
+    )
+      return undefined
+    return {
+      maxTokens: generation?.maxTokens,
+      temperature: generation?.temperature,
+      topP: generation?.topP,
+      stopSequences: generation?.stop,
+    }
+  })()
+  return {
+    modelId: request.model.id,
+    messages,
+    system,
+    inferenceConfig,
    toolConfig,
    // Converse's base inferenceConfig has no topK; Anthropic/Nova accept it
    // as a model-specific field, so it goes through additionalModelRequestFields.
@@ -469,7 +482,6 @@ const mapFinishReason = (reason: string): FinishReason => {
  if (reason === "max_tokens" || reason === "model_context_window_exceeded") return "length"
  if (reason === "tool_use") return "tool-calls"
  if (reason === "content_filtered" || reason === "guardrail_intervened") return "content-filter"
-  if (reason === "malformed_model_output" || reason === "malformed_tool_use") return "error"
  return "unknown"
 }

@@ -504,6 +516,16 @@ interface ParserState {
  readonly hasToolCalls: boolean
  readonly lifecycle: Lifecycle.State
  readonly reasoningSignatures: Readonly<Record<number, string>>
+  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 step = (state: ParserState, event: BedrockEvent) =>
@@ -551,23 +573,46 @@ const step = (state: ParserState, event: BedrockEvent) =>
      const index = event.contentBlockDelta.contentBlockIndex
      const reasoning = event.contentBlockDelta.delta.reasoningContent
      const events: LLMEvent[] = []
-      const redactedData = reasoning.redactedContent ?? reasoning.data
-      const metadata = reasoning.signature
-        ? providerMetadata(state.providerMetadataKey, { signature: reasoning.signature })
-        : redactedData !== undefined
-          ? providerMetadata(state.providerMetadataKey, { redactedData })
-          : undefined
-      const lifecycle =
-        reasoning.text !== undefined || metadata !== undefined
-          ? Lifecycle.reasoningDelta(state.lifecycle, events, `reasoning-${index}`, reasoning.text ?? "", metadata)
-          : state.lifecycle
+      const redactedChunks = yield* (() => {
+        if (reasoning.redactedContent === undefined) return Effect.succeed(undefined)
+        return Effect.fromResult(Encoding.decodeBase64(reasoning.redactedContent)).pipe(
+          Effect.map((chunk) => [...(state.reasoningRedactedContent[index] ?? []), chunk]),
+          Effect.mapError((cause) =>
+            ProviderShared.eventError(
+              ADAPTER,
+              "Bedrock Converse reasoningContent.redactedContent contains invalid base64 data",
+              undefined,
+              cause,
+            ),
+          ),
+        )
+      })()
+      const redactedData = redactedChunks === undefined ? reasoning.data : encodeRedactedContent(redactedChunks)
+      const metadata = (() => {
+        if (reasoning.signature) return providerMetadata(state.providerMetadataKey, { signature: reasoning.signature })
+        if (redactedData !== undefined) return providerMetadata(state.providerMetadataKey, { redactedData })
+      })()
+      const lifecycle = (() => {
+        if (reasoning.text === undefined && metadata === undefined) return state.lifecycle
+        return Lifecycle.reasoningDelta(state.lifecycle, events, `reasoning-${index}`, reasoning.text ?? "", metadata)
+      })()
+      const reasoningRedactedContent = (() => {
+        if (redactedChunks !== undefined) return { ...state.reasoningRedactedContent, [index]: redactedChunks }
+        if (reasoning.data === undefined) return state.reasoningRedactedContent
+        return Object.fromEntries(
+          Object.entries(state.reasoningRedactedContent).filter(([key]) => key !== String(index)),
+        )
+      })()
+      const reasoningSignatures = (() => {
+        if (!reasoning.signature) return state.reasoningSignatures
+        return { ...state.reasoningSignatures, [index]: reasoning.signature }
+      })()
      return [
        {
          ...state,
          lifecycle,
-          reasoningSignatures: reasoning.signature
-            ? { ...state.reasoningSignatures, [index]: reasoning.signature }
-            : state.reasoningSignatures,
+          reasoningSignatures,
+          reasoningRedactedContent,
        },
        events,
      ] as const
@@ -595,16 +640,24 @@ const step = (state: ParserState, event: BedrockEvent) =>
      const result = yield* ToolStream.finish(ADAPTER, state.tools, index)
      const events: LLMEvent[] = []
      const resultEvents = result.events ?? []
-      const lifecycle = resultEvents.length
-        ? Lifecycle.stepStart(state.lifecycle, events)
-        : Lifecycle.reasoningEnd(
-            Lifecycle.textEnd(state.lifecycle, events, `text-${index}`),
-            events,
-            `reasoning-${index}`,
-            state.reasoningSignatures[index]
-              ? providerMetadata(state.providerMetadataKey, { signature: state.reasoningSignatures[index] })
-              : undefined,
-          )
+      const lifecycle = (() => {
+        if (resultEvents.length) return Lifecycle.stepStart(state.lifecycle, events)
+        const metadata = (() => {
+          const signature = state.reasoningSignatures[index]
+          if (signature) return providerMetadata(state.providerMetadataKey, { signature })
+          const redactedContent = state.reasoningRedactedContent[index]
+          if (redactedContent)
+            return providerMetadata(state.providerMetadataKey, {
+              redactedData: encodeRedactedContent(redactedContent),
+            })
+        })()
+        return Lifecycle.reasoningEnd(
+          Lifecycle.textEnd(state.lifecycle, events, `text-${index}`),
+          events,
+          `reasoning-${index}`,
+          metadata,
+        )
+      })()
      events.push(...resultEvents)
      return [
        {
@@ -618,12 +671,24 @@ const step = (state: ParserState, event: BedrockEvent) =>
          reasoningSignatures: Object.fromEntries(
            Object.entries(state.reasoningSignatures).filter(([key]) => key !== String(index)),
          ),
+          reasoningRedactedContent: Object.fromEntries(
+            Object.entries(state.reasoningRedactedContent).filter(([key]) => key !== String(index)),
+          ),
        },
        events,
      ] as const
    }

    if (event.messageStop) {
+      if (
+        event.messageStop.stopReason === "malformed_model_output" ||
+        event.messageStop.stopReason === "malformed_tool_use"
+      )
+        return yield* ProviderShared.eventError(
+          ADAPTER,
+          `Bedrock Converse stopped with ${event.messageStop.stopReason}`,
+          ProviderShared.encodeJson(event),
+        )
      return [
        {
          ...state,
@@ -670,23 +735,22 @@ const step = (state: ParserState, event: BedrockEvent) =>

 const framing = BedrockEventStream.framing(ADAPTER)

-const onHalt = (state: ParserState): ReadonlyArray<LLMEvent> =>
-  state.pendingFinish
-    ? (() => {
-        const events: LLMEvent[] = []
-        Lifecycle.finish(state.lifecycle, events, {
-          reason: {
-            ...state.pendingFinish.reason,
-            normalized:
-              state.pendingFinish.reason.normalized === "stop" && state.hasToolCalls
-                ? "tool-calls"
-                : state.pendingFinish.reason.normalized,
-          },
-          usage: state.pendingFinish.usage,
-        })
-        return events
-      })()
-    : []
+const onHalt = (state: ParserState): ReadonlyArray<LLMEvent> => {
+  if (!state.pendingFinish) return []
+  const normalized = (() => {
+    if (state.pendingFinish.reason.normalized === "stop" && state.hasToolCalls) return "tool-calls"
+    return state.pendingFinish.reason.normalized
+  })()
+  const events: LLMEvent[] = []
+  Lifecycle.finish(state.lifecycle, events, {
+    reason: {
+      ...state.pendingFinish.reason,
+      normalized,
+    },
+    usage: state.pendingFinish.usage,
+  })
+  return events
+}

 // =============================================================================
 // Protocol And Bedrock Route
@@ -711,6 +775,7 @@ export const protocol = Protocol.make({
      hasToolCalls: false,
      lifecycle: Lifecycle.initial(),
      reasoningSignatures: {},
+      reasoningRedactedContent: {},
    }),
    step,
    onHalt: (state) => Effect.succeed(onHalt(state)),
@@ -1,7 +1,7 @@
 import { EventStreamCodec } from "@smithy/eventstream-codec"
 import { fromUtf8, toUtf8 } from "@smithy/util-utf8"
 import { Effect, Encoding, Stream } from "effect"
-import { AIError, AIErrorReason } from "../schema/index.js"
+import { AIError, AIErrorReason, InvalidProviderOutputError } from "../schema/index.js"
 import { Framing } from "../route/framing.js"
 import { ProviderShared } from "./shared.js"

@@ -22,6 +22,10 @@ interface FrameBufferState {

 const initialFrameBuffer: FrameBufferState = { buffer: new Uint8Array(0), offset: 0 }

+type FrameInput = { readonly _tag: "Chunk"; readonly bytes: Uint8Array } | { readonly _tag: "End" }
+
+const endOfStream: FrameInput = { _tag: "End" }
+
 const appendChunk = (state: FrameBufferState, chunk: Uint8Array): FrameBufferState => {
  const remaining = state.buffer.length - state.offset
  // Compact: drop the consumed prefix and append the new chunk in one alloc.
@@ -33,9 +37,23 @@ const appendChunk = (state: FrameBufferState, chunk: Uint8Array): FrameBufferSta
  return { buffer: next, offset: 0 }
 }

-const consumeFrames = (route: string) => (state: FrameBufferState, chunk: Uint8Array) =>
+const consumeFrames = (route: string) => (state: FrameBufferState, input: FrameInput) =>
  Effect.gen(function* () {
-    let cursor = appendChunk(state, chunk)
+    if (input._tag === "End") {
+      const remaining = state.buffer.subarray(state.offset)
+      if (remaining.length > 0)
+        return yield* new AIError({
+          reason: new InvalidProviderOutputError({
+            route,
+            classification: "incomplete-stream",
+            message: `Incomplete Bedrock Converse event-stream frame: ${remaining.length} buffered bytes remain at end of stream`,
+            body: Encoding.encodeBase64(remaining),
+          }),
+        })
+      return [state, []] as const
+    }
+
+    let cursor = appendChunk(state, input.bytes)
    const out: object[] = []
    while (cursor.buffer.length - cursor.offset >= 4) {
      const view = cursor.buffer.subarray(cursor.offset)
@@ -113,7 +131,12 @@ const consumeFrames = (route: string) => (state: FrameBufferState, chunk: Uint8A
 export const framing = (route: string): Framing.Definition<object> => ({
  id: "aws-event-stream",
  body: (frame) => ("rawBody" in frame && typeof frame.rawBody === "string" ? frame.rawBody : undefined),
-  frame: (bytes) => bytes.pipe(Stream.mapAccumEffect(() => initialFrameBuffer, consumeFrames(route))),
+  frame: (bytes) =>
+    bytes.pipe(
+      Stream.map((bytes): FrameInput => ({ _tag: "Chunk", bytes })),
+      Stream.concat(Stream.succeed(endOfStream)),
+      Stream.mapAccumEffect(() => initialFrameBuffer, consumeFrames(route)),
+    ),
 })

 export * as BedrockEventStream from "./bedrock-event-stream.js"
@@ -0,0 +1,96 @@
+import { Effect, Encoding, Schema, Struct } from "effect"
+import { Headers } from "effect/unstable/http"
+import { AIError } from "../schema/index.js"
+import { Route } from "../route/client.js"
+import { Endpoint } from "../route/endpoint.js"
+import { Protocol } from "../route/protocol.js"
+import { classifyProviderFailure } from "../provider-error.js"
+import { AnthropicMessages } from "./anthropic-messages.js"
+import { BedrockEventStream } from "./bedrock-event-stream.js"
+import { BedrockAuth } from "./utils/bedrock-auth.js"
+import { JsonObject, ProviderShared } from "./shared.js"
+
+const ID = "bedrock-messages"
+const VERSION = "bedrock-2023-05-31"
+const Body = Schema.Struct({
+  ...Struct.omit(AnthropicMessages.AnthropicMessagesBody.fields, ["model", "stream"]),
+  anthropic_version: Schema.Literal(VERSION),
+  anthropic_beta: Schema.optional(Schema.Array(Schema.String)),
+})
+const Event = Schema.Struct({
+  chunk: Schema.optional(Schema.Struct({ bytes: Schema.String })),
+  exception: Schema.optional(
+    Schema.Struct({
+      type: Schema.String,
+      details: Schema.StructWithRest(
+        Schema.Struct({ message: Schema.optional(Schema.String), originalMessage: Schema.optional(Schema.String) }),
+        [JsonObject],
+      ),
+    }),
+  ),
+})
+
+export const protocol = Protocol.make({
+  id: ID,
+  body: {
+    schema: Body,
+    from: Effect.fn("BedrockMessages.fromRequest")(function* (request) {
+      const body = yield* AnthropicMessages.protocol.body.from(request)
+      const headers = Headers.fromInput(request.http?.headers)
+      const betas = new Set(
+        (headers["anthropic-beta"] ?? "")
+          .split(",")
+          .map((value) => value.trim())
+          .filter(Boolean),
+      )
+      if (
+        body.context_management?.edits.length ||
+        body.messages.some((message) => message.content.some((block) => block.type === "compaction"))
+      )
+        betas.add("compact-2026-01-12")
+      return {
+        ...Struct.omit(body, ["model", "stream"]),
+        anthropic_version: VERSION,
+        anthropic_beta: betas.size ? [...betas] : undefined,
+      } satisfies typeof Body.Type
+    }),
+  },
+  stream: {
+    event: Event,
+    initial: AnthropicMessages.protocol.stream.initial,
+    step: Effect.fn("BedrockMessages.step")(function* (state, event) {
+      if (event.exception)
+        return yield* new AIError({
+          reason: classifyProviderFailure({
+            message: event.exception.details.message ?? event.exception.details.originalMessage ?? event.exception.type,
+            rawBody: ProviderShared.encodeJson(event),
+          }),
+        })
+      if (!event.chunk) return yield* ProviderShared.eventError(ID, "Bedrock Messages event is missing its chunk")
+      const text = yield* Effect.fromResult(Encoding.decodeBase64String(event.chunk.bytes)).pipe(
+        Effect.mapError((cause) =>
+          ProviderShared.eventError(ID, "Invalid Bedrock Messages chunk encoding", undefined, cause),
+        ),
+      )
+      const decoded = yield* Schema.decodeUnknownEffect(AnthropicMessages.protocol.stream.event)(text).pipe(
+        Effect.mapError((cause) => ProviderShared.eventError(ID, "Invalid Bedrock Messages event", undefined, cause)),
+      )
+      return yield* AnthropicMessages.protocol.stream.step(state, decoded)
+    }),
+  },
+})
+
+export const route = Route.make({
+  id: ID,
+  provider: "amazon-bedrock",
+  providerMetadataKey: "anthropic",
+  protocol,
+  endpoint: Endpoint.path(
+    ({ request }) => `/model/${encodeURIComponent(request.model.id)}/invoke-with-response-stream`,
+    { baseURL: "https://bedrock-runtime.us-east-1.amazonaws.com" },
+  ),
+  auth: BedrockAuth.auth,
+  framing: BedrockEventStream.framing(ID),
+})
+
+export * as BedrockMessages from "./bedrock-messages.js"
@@ -240,6 +240,10 @@ interface ParserState {
  readonly lifecycle: Lifecycle.State
  readonly reasoningSignature?: string
  readonly textSignature?: string
+  readonly reasoningId?: string
+  readonly textId?: string
+  readonly nextReasoningId: number
+  readonly nextTextId: number
  readonly seenCallIds?: ReadonlySet<string>
 }

@@ -571,19 +575,23 @@ const finish = (state: ParserState): ReadonlyArray<LLMEvent> => {

  const events: LLMEvent[] = []
  let lifecycle = state.lifecycle
-  if (state.reasoningSignature !== undefined)
+  if (state.reasoningId !== undefined)
    lifecycle = Lifecycle.reasoningEnd(
      lifecycle,
      events,
-      "reasoning-0",
-      providerMetadata(state.providerMetadataKey, { thoughtSignature: state.reasoningSignature }),
+      state.reasoningId,
+      state.reasoningSignature === undefined
+        ? undefined
+        : providerMetadata(state.providerMetadataKey, { thoughtSignature: state.reasoningSignature }),
    )
-  if (state.textSignature !== undefined)
+  if (state.textId !== undefined)
    lifecycle = Lifecycle.textEnd(
      lifecycle,
      events,
-      "text-0",
-      providerMetadata(state.providerMetadataKey, { thoughtSignature: state.textSignature }),
+      state.textId,
+      state.textSignature === undefined
+        ? undefined
+        : providerMetadata(state.providerMetadataKey, { thoughtSignature: state.textSignature }),
    )
  Lifecycle.finish(lifecycle, events, {
    reason: {
@@ -632,6 +640,10 @@ const step = (state: ParserState, event: GeminiEvent) => {
  let lifecycle = nextState.lifecycle
  let reasoningSignature = nextState.reasoningSignature
  let textSignature = nextState.textSignature
+  let reasoningId = nextState.reasoningId
+  let textId = nextState.textId
+  let nextReasoningId = nextState.nextReasoningId
+  let nextTextId = nextState.nextTextId
  // Supplier ids must be tracked across chunks of the same response, not just within one event's parts.
  const seenCallIds = new Set(nextState.seenCallIds)

@@ -657,27 +669,51 @@ const step = (state: ParserState, event: GeminiEvent) => {
    else if (signature !== undefined && "text" in part) textSignature = signature
    if ("text" in part && part.text.length > 0) {
      if (part.thought) {
+        if (textId !== undefined) {
+          lifecycle = Lifecycle.textEnd(
+            lifecycle,
+            events,
+            textId,
+            textSignature
+              ? providerMetadata(state.providerMetadataKey, { thoughtSignature: textSignature })
+              : undefined,
+          )
+          textId = undefined
+          textSignature = undefined
+        }
+        if (reasoningId === undefined) {
+          reasoningId = `reasoning-${nextReasoningId}`
+          nextReasoningId += 1
+        }
        lifecycle = Lifecycle.reasoningDelta(
          lifecycle,
          events,
-          "reasoning-0",
+          reasoningId,
          part.text,
          signature ? providerMetadata(state.providerMetadataKey, { thoughtSignature: signature }) : undefined,
        )
        continue
      }
-      lifecycle = Lifecycle.reasoningEnd(
-        lifecycle,
-        events,
-        "reasoning-0",
-        reasoningSignature
-          ? providerMetadata(state.providerMetadataKey, { thoughtSignature: reasoningSignature })
-          : undefined,
-      )
+      if (reasoningId !== undefined) {
+        lifecycle = Lifecycle.reasoningEnd(
+          lifecycle,
+          events,
+          reasoningId,
+          reasoningSignature
+            ? providerMetadata(state.providerMetadataKey, { thoughtSignature: reasoningSignature })
+            : undefined,
+        )
+        reasoningId = undefined
+        reasoningSignature = undefined
+      }
+      if (textId === undefined) {
+        textId = `text-${nextTextId}`
+        nextTextId += 1
+      }
      lifecycle = Lifecycle.textDelta(
        lifecycle,
        events,
-        "text-0",
+        textId,
        part.text,
        textSignature ? providerMetadata(state.providerMetadataKey, { thoughtSignature: textSignature }) : undefined,
      )
@@ -695,14 +731,28 @@ const step = (state: ParserState, event: GeminiEvent) => {
      const duplicate = supplied !== undefined && seenCallIds.has(supplied)
      if (supplied !== undefined) seenCallIds.add(supplied)
      const id = supplied !== undefined && !duplicate ? supplied : `tool_${crypto.randomUUID().replaceAll("-", "")}`
-      lifecycle = Lifecycle.reasoningEnd(
-        lifecycle,
-        events,
-        "reasoning-0",
-        reasoningSignature
-          ? providerMetadata(state.providerMetadataKey, { thoughtSignature: reasoningSignature })
-          : undefined,
-      )
+      if (reasoningId !== undefined) {
+        lifecycle = Lifecycle.reasoningEnd(
+          lifecycle,
+          events,
+          reasoningId,
+          reasoningSignature
+            ? providerMetadata(state.providerMetadataKey, { thoughtSignature: reasoningSignature })
+            : undefined,
+        )
+        reasoningId = undefined
+        reasoningSignature = undefined
+      }
+      if (textId !== undefined) {
+        lifecycle = Lifecycle.textEnd(
+          lifecycle,
+          events,
+          textId,
+          textSignature ? providerMetadata(state.providerMetadataKey, { thoughtSignature: textSignature }) : undefined,
+        )
+        textId = undefined
+        textSignature = undefined
+      }
      lifecycle = Lifecycle.stepStart(lifecycle, events)
      events.push(
        LLMEvent.toolCall({
@@ -725,6 +775,10 @@ const step = (state: ParserState, event: GeminiEvent) => {
      lifecycle,
      reasoningSignature,
      textSignature,
+      reasoningId,
+      textId,
+      nextReasoningId,
+      nextTextId,
      seenCallIds,
      finishReason: candidate.finishReason ?? nextState.finishReason,
    },
@@ -752,6 +806,8 @@ export const protocol = Protocol.make({
      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
      hasToolCalls: false,
      lifecycle: Lifecycle.initial(),
+      nextReasoningId: 0,
+      nextTextId: 0,
    }),
    step,
    onHalt: (state) => Effect.succeed(finish(state)),
@@ -1,6 +1,8 @@
 export * as AnthropicMessages from "./anthropic-messages.js"
 export * as BedrockConverse from "./bedrock-converse.js"
+export { BedrockMessages } from "./bedrock-messages.js"
 export * as Gemini from "./gemini.js"
+export * as MistralChat from "./mistral-chat.js"
 export * as OpenAIChat from "./openai-chat.js"
 export * as OpenAIImages from "./openai-images.js"
 export * as OpenAICompatibleChat from "./openai-compatible-chat.js"
@@ -0,0 +1,780 @@
+import { Effect, Schema } from "effect"
+import { Auth } from "../route/auth.js"
+import { Route } from "../route/client.js"
+import { Endpoint } from "../route/endpoint.js"
+import { Framing } from "../route/framing.js"
+import { Protocol } from "../route/protocol.js"
+import { HttpTransport } from "../route/transport/index.js"
+import {
+  AIError,
+  InvalidProviderOutputError,
+  LLMEvent,
+  Usage,
+  type FinishReasonDetails,
+  type LLMRequest,
+  type MediaPart,
+  type ToolCallPart,
+  type ToolDefinition,
+} from "../schema/index.js"
+import { classifyProviderFailure } from "../provider-error.js"
+import { JsonObject, optionalArray, optionalNull, ProviderShared } from "./shared.js"
+import { Lifecycle } from "./utils/lifecycle.js"
+import { ToolStream } from "./utils/tool-stream.js"
+
+const ADAPTER = "mistral-chat"
+const DONE = "[DONE]" as const
+const TOOL_ID = /^[A-Za-z0-9]{9}$/
+export const DEFAULT_BASE_URL = "https://api.mistral.ai/v1"
+export const PATH = "/chat/completions"
+
+const MistralTextContent = Schema.Struct({
+  type: Schema.Literal("text"),
+  text: Schema.String,
+})
+
+const MistralThinkingUnit = Schema.StructWithRest(
+  Schema.Struct({
+    type: Schema.optional(Schema.String),
+    text: Schema.optional(Schema.String),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+type MistralThinkingUnit = Schema.Schema.Type<typeof MistralThinkingUnit>
+
+const MistralThinkingContent = Schema.StructWithRest(
+  Schema.Struct({
+    type: Schema.Literal("thinking"),
+    thinking: Schema.Array(MistralThinkingUnit),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+type MistralThinkingContent = Schema.Schema.Type<typeof MistralThinkingContent>
+const isMistralThinkingContent = Schema.is(MistralThinkingContent)
+
+const MistralUserContent = Schema.Union([
+  MistralTextContent,
+  Schema.Struct({ type: Schema.Literal("image_url"), image_url: Schema.String }),
+  Schema.Struct({ type: Schema.Literal("document_url"), document_url: Schema.String }),
+])
+type MistralUserContent = Schema.Schema.Type<typeof MistralUserContent>
+
+const MistralAssistantToolCall = Schema.Struct({
+  id: Schema.String,
+  type: Schema.Literal("function"),
+  function: Schema.Struct({ name: Schema.String, arguments: Schema.String }),
+})
+type MistralAssistantToolCall = Schema.Schema.Type<typeof MistralAssistantToolCall>
+
+const MistralMessage = Schema.Union([
+  Schema.Struct({ role: Schema.Literal("system"), content: Schema.String }),
+  Schema.Struct({
+    role: Schema.Literal("user"),
+    content: Schema.Union([Schema.String, Schema.Array(MistralUserContent)]),
+  }),
+  Schema.Struct({
+    role: Schema.Literal("assistant"),
+    content: Schema.Union([Schema.String, Schema.Array(Schema.Union([MistralTextContent, MistralThinkingContent]))]),
+    tool_calls: optionalArray(MistralAssistantToolCall),
+    prefix: Schema.optional(Schema.Literal(true)),
+  }),
+  Schema.Struct({
+    role: Schema.Literal("tool"),
+    tool_call_id: Schema.String,
+    name: Schema.String,
+    content: Schema.Union([Schema.String, Schema.Array(MistralUserContent)]),
+  }),
+]).pipe(Schema.toTaggedUnion("role"))
+type MistralMessage = Schema.Schema.Type<typeof MistralMessage>
+
+const MistralTool = Schema.Struct({
+  type: Schema.Literal("function"),
+  function: Schema.Struct({
+    name: Schema.String,
+    description: Schema.String,
+    parameters: JsonObject,
+    strict: Schema.Literal(false),
+  }),
+})
+type MistralTool = Schema.Schema.Type<typeof MistralTool>
+
+const MistralOptions = Schema.Struct({
+  safePrompt: Schema.optional(Schema.Boolean),
+  documentImageLimit: Schema.optional(Schema.Number),
+  documentPageLimit: Schema.optional(Schema.Number),
+  parallelToolCalls: Schema.optional(Schema.Boolean),
+  reasoningEffort: Schema.optional(Schema.String),
+  promptMode: Schema.optional(Schema.Literal("reasoning")),
+  promptCacheKey: Schema.optional(Schema.String),
+})
+
+export type ReasoningEffort = "none" | "minimal" | "low" | "medium" | "high" | "xhigh" | (string & {})
+
+export type ProviderOptionsInput = {
+  readonly safePrompt?: boolean
+  readonly documentImageLimit?: number
+  readonly documentPageLimit?: number
+  readonly parallelToolCalls?: boolean
+  readonly reasoningEffort?: ReasoningEffort
+  readonly promptMode?: "reasoning"
+  readonly promptCacheKey?: string
+  readonly [key: string]: unknown
+}
+
+const MistralBody = Schema.Struct({
+  model: Schema.String,
+  messages: Schema.Array(MistralMessage),
+  tools: optionalArray(MistralTool),
+  tool_choice: Schema.optional(
+    Schema.Union([
+      Schema.Literals(["auto", "none", "any"]),
+      Schema.Struct({ type: Schema.Literal("function"), function: Schema.Struct({ name: Schema.String }) }),
+    ]),
+  ),
+  stream: Schema.Literal(true),
+  max_tokens: Schema.optional(Schema.Number),
+  random_seed: Schema.optional(Schema.Number),
+  temperature: Schema.optional(Schema.Number),
+  top_p: Schema.optional(Schema.Number),
+  frequency_penalty: Schema.optional(Schema.Number),
+  presence_penalty: Schema.optional(Schema.Number),
+  stop: optionalArray(Schema.String),
+  prompt_cache_key: Schema.optional(Schema.String),
+  safe_prompt: Schema.optional(Schema.Boolean),
+  document_image_limit: Schema.optional(Schema.Number),
+  document_page_limit: Schema.optional(Schema.Number),
+  parallel_tool_calls: Schema.optional(Schema.Boolean),
+  reasoning_effort: Schema.optional(Schema.String),
+  prompt_mode: Schema.optional(Schema.Literal("reasoning")),
+})
+export type MistralBody = Schema.Schema.Type<typeof MistralBody>
+
+const MistralUsageDetails = Schema.StructWithRest(Schema.Struct({ cached_tokens: optionalNull(Schema.Number) }), [
+  Schema.Record(Schema.String, Schema.Unknown),
+])
+
+const MistralUsage = Schema.StructWithRest(
+  Schema.Struct({
+    prompt_tokens: optionalNull(Schema.Number),
+    completion_tokens: optionalNull(Schema.Number),
+    total_tokens: optionalNull(Schema.Number),
+    num_cached_tokens: optionalNull(Schema.Number),
+    prompt_token_details: optionalNull(MistralUsageDetails),
+    prompt_tokens_details: optionalNull(MistralUsageDetails),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+
+const MistralOutputContent = Schema.StructWithRest(
+  Schema.Struct({
+    type: Schema.String,
+    text: optionalNull(Schema.String),
+    thinking: optionalNull(Schema.Unknown),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+type MistralOutputContent = Schema.Schema.Type<typeof MistralOutputContent>
+
+const MistralToolDelta = Schema.Struct({
+  index: optionalNull(Schema.Number),
+  id: optionalNull(Schema.String),
+  function: optionalNull(
+    Schema.Struct({
+      name: optionalNull(Schema.String),
+      arguments: optionalNull(Schema.Union([Schema.String, JsonObject])),
+    }),
+  ),
+})
+type MistralToolDelta = Schema.Schema.Type<typeof MistralToolDelta>
+
+const MistralChoice = Schema.StructWithRest(
+  Schema.Struct({
+    delta: optionalNull(
+      Schema.StructWithRest(
+        Schema.Struct({
+          content: optionalNull(Schema.Union([Schema.String, Schema.Array(MistralOutputContent)])),
+          tool_calls: optionalNull(Schema.Array(MistralToolDelta)),
+        }),
+        [Schema.Record(Schema.String, Schema.Unknown)],
+      ),
+    ),
+    finish_reason: optionalNull(Schema.String),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+
+const MistralError = Schema.StructWithRest(
+  Schema.Struct({
+    message: Schema.String,
+    code: optionalNull(Schema.Union([Schema.String, Schema.Number])),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+
+const MistralEvent = Schema.StructWithRest(
+  Schema.Struct({
+    choices: optionalNull(Schema.Array(MistralChoice)),
+    usage: optionalNull(MistralUsage),
+    error: optionalNull(MistralError),
+  }),
+  [Schema.Record(Schema.String, Schema.Unknown)],
+)
+type MistralEvent = Schema.Schema.Type<typeof MistralEvent>
+const MistralStreamEvent = Schema.Union([Schema.Literal(DONE), Protocol.jsonEvent(MistralEvent)])
+
+const hashID = (value: string) => {
+  const hash = (seed: number) => {
+    let result = seed
+    for (const char of value) result = Math.imul(result ^ char.charCodeAt(0), 16777619)
+    return (result >>> 0).toString(36)
+  }
+  return `${hash(2166136261).padStart(7, "0")}${hash(2246822519).padStart(7, "0")}`.slice(-9)
+}
+
+const toolIDNormalizer = (request: LLMRequest) => {
+  const ids = request.messages.flatMap((message) =>
+    message.content.flatMap((part) => (part.type === "tool-call" || part.type === "tool-result" ? [part.id] : [])),
+  )
+  const used = new Set(ids.filter((id) => TOOL_ID.test(id)))
+  const normalized = new Map<string, string>()
+  return (id: string) => {
+    if (TOOL_ID.test(id)) return id
+    const previous = normalized.get(id)
+    if (previous) return previous
+    let attempt = 0
+    let candidate = hashID(id)
+    while (used.has(candidate)) candidate = hashID(`${id}:${++attempt}`)
+    used.add(candidate)
+    normalized.set(id, candidate)
+    return candidate
+  }
+}
+
+const lowerMedia = Effect.fn("MistralChat.lowerMedia")(function* (part: MediaPart) {
+  const media = ProviderShared.normalizeMedia(part)
+  const url = typeof part.data === "string" && /^(?:https?:|data:)/.test(part.data) ? part.data : media.dataUrl
+  if (media.mime.startsWith("image/")) return { type: "image_url" as const, image_url: url }
+  if (media.mime === "application/pdf") return { type: "document_url" as const, document_url: url }
+  return yield* ProviderShared.invalidRequest(`Mistral Chat does not support media type ${part.mediaType}`)
+})
+
+const lowerUser = Effect.fn("MistralChat.lowerUser")(function* (message: LLMRequest["messages"][number]) {
+  const content: MistralUserContent[] = []
+  for (const part of message.content) {
+    if (part.type === "text") {
+      content.push({ type: "text", text: part.text })
+      continue
+    }
+    if (part.type === "media") {
+      content.push(yield* lowerMedia(part))
+      continue
+    }
+    return yield* ProviderShared.unsupportedContent("Mistral Chat", "user", ["text", "media"])
+  }
+  if (content.every((part) => part.type === "text"))
+    return { role: "user" as const, content: content.map((part) => part.text).join("") }
+  return { role: "user" as const, content }
+})
+
+const lowerToolCall = (part: ToolCallPart, normalizeID: (id: string) => string): MistralAssistantToolCall => ({
+  id: normalizeID(part.id),
+  type: "function",
+  function: { name: part.name, arguments: ProviderShared.encodeJson(part.input) },
+})
+
+const lowerAssistant = Effect.fn("MistralChat.lowerAssistant")(function* (
+  message: LLMRequest["messages"][number],
+  normalizeID: (id: string) => string,
+  prefix: boolean,
+) {
+  const structured = message.content.some(
+    (part) => part.type === "reasoning" && isMistralThinkingContent(part.providerMetadata?.mistral?.thinking),
+  )
+  const content: Array<Schema.Schema.Type<typeof MistralTextContent> | MistralThinkingContent> = []
+  const text: string[] = []
+  const toolCalls: MistralAssistantToolCall[] = []
+  for (const part of message.content) {
+    if (part.type === "text") {
+      if (structured) content.push({ type: "text", text: part.text })
+      else text.push(part.text)
+      continue
+    }
+    if (part.type === "reasoning") {
+      const native = part.providerMetadata?.mistral?.thinking
+      if (structured && isMistralThinkingContent(native)) content.push(native)
+      else if (structured) content.push({ type: "text", text: part.text })
+      else text.push(part.text)
+      continue
+    }
+    if (part.type === "tool-call") {
+      toolCalls.push(lowerToolCall(part, normalizeID))
+      continue
+    }
+    return yield* ProviderShared.unsupportedContent("Mistral Chat", "assistant", ["text", "reasoning", "tool-call"])
+  }
+  return {
+    role: "assistant" as const,
+    content: structured ? content : text.join(""),
+    ...(toolCalls.length > 0 ? { tool_calls: toolCalls } : {}),
+    ...(prefix ? { prefix: true as const } : {}),
+  }
+})
+
+const lowerToolResults = Effect.fn("MistralChat.lowerToolResults")(function* (
+  message: LLMRequest["messages"][number],
+  normalizeID: (id: string) => string,
+) {
+  const output: MistralMessage[] = []
+  for (const part of message.content) {
+    if (part.type !== "tool-result")
+      return yield* ProviderShared.unsupportedContent("Mistral Chat", "tool", ["tool-result"])
+    if (part.result.type !== "content") {
+      output.push({
+        role: "tool",
+        tool_call_id: normalizeID(part.id),
+        name: part.name,
+        content: ProviderShared.toolResultText(part),
+      })
+      continue
+    }
+    const content: MistralUserContent[] = []
+    for (const item of part.result.value) {
+      if (item.type === "text") {
+        content.push({ type: "text", text: item.text })
+        continue
+      }
+      content.push(yield* lowerMedia({ type: "media", mediaType: item.mime, data: item.uri, filename: item.name }))
+    }
+    output.push({
+      role: "tool",
+      tool_call_id: normalizeID(part.id),
+      name: part.name,
+      content: content.some((item) => item.type !== "text")
+        ? content
+        : content.map((item) => (item.type === "text" ? item.text : "")).join(""),
+    })
+  }
+  return output
+})
+
+const lowerMessages = Effect.fn("MistralChat.lowerMessages")(function* (request: LLMRequest) {
+  const normalizeID = toolIDNormalizer(request)
+  const messages: MistralMessage[] =
+    request.system.length === 0 ? [] : [{ role: "system", content: ProviderShared.joinText(request.system) }]
+  for (const message of request.messages) {
+    if (message.role === "system") {
+      const update = yield* ProviderShared.wrappedSystemUpdate("Mistral Chat", message)
+      messages.push({
+        role: "user",
+        content: update.text,
+      })
+      continue
+    }
+    if (message.role === "user") {
+      messages.push(yield* lowerUser(message))
+      continue
+    }
+    if (message.role === "assistant") {
+      const hasToolCalls = message.content.some((part) => part.type === "tool-call")
+      const hasNativeThinking = message.content.some(
+        (part) => part.type === "reasoning" && isMistralThinkingContent(part.providerMetadata?.mistral?.thinking),
+      )
+      const text = message.content
+        .flatMap((part) => (part.type === "text" || part.type === "reasoning" ? [part.text] : []))
+        .join("")
+      if (!hasToolCalls && !hasNativeThinking && text.trim() === "") continue
+      messages.push(yield* lowerAssistant(message, normalizeID, !hasToolCalls && message === request.messages.at(-1)))
+      continue
+    }
+    messages.push(...(yield* lowerToolResults(message, normalizeID)))
+  }
+  return messages
+})
+
+const lowerTool = (tool: ToolDefinition): MistralTool => ({
+  type: "function",
+  function: { name: tool.name, description: tool.description, parameters: tool.inputSchema, strict: false },
+})
+
+export const fromRequest = Effect.fn("MistralChat.fromRequest")(function* (request: LLMRequest) {
+  const options = yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(MistralOptions))(
+    request.providerOptions ?? {},
+  )
+  const selected = request.toolChoice?.type === "tool" ? request.toolChoice.name : undefined
+  if (request.toolChoice?.type === "tool" && !selected)
+    return yield* ProviderShared.invalidRequest("Mistral Chat tool choice requires a tool name")
+  if (options.reasoningEffort !== undefined && options.promptMode !== undefined)
+    return yield* ProviderShared.invalidRequest(
+      "Mistral Chat reasoningEffort and promptMode provider options are mutually exclusive",
+    )
+  const toolChoice = request.toolChoice
+    ? yield* ProviderShared.matchToolChoice("Mistral Chat", request.toolChoice, {
+        auto: () => "auto" as const,
+        none: () => "none" as const,
+        required: () => "any" as const,
+        tool: (name) => ({ type: "function" as const, function: { name } }),
+      })
+    : undefined
+  return {
+    model: request.model.id,
+    messages: yield* lowerMessages(request),
+    tools: request.tools.length > 0 ? request.tools.map(lowerTool) : undefined,
+    tool_choice: toolChoice,
+    stream: true as const,
+    max_tokens: request.generation?.maxTokens,
+    random_seed: request.generation?.seed,
+    temperature: request.generation?.temperature,
+    top_p: request.generation?.topP,
+    frequency_penalty: request.generation?.frequencyPenalty,
+    presence_penalty: request.generation?.presencePenalty,
+    stop: request.generation?.stop,
+    prompt_cache_key: request.cache === "none" ? undefined : (options.promptCacheKey ?? request.promptCacheKey),
+    safe_prompt: options.safePrompt,
+    document_image_limit: options.documentImageLimit,
+    document_page_limit: options.documentPageLimit,
+    parallel_tool_calls:
+      options.parallelToolCalls ?? (request.toolChoice?.disableParallelToolUse === true ? false : undefined),
+    reasoning_effort: options.reasoningEffort,
+    prompt_mode: options.promptMode,
+  }
+})
+
+type ToolKey = string | number
+interface PendingTool {
+  readonly id: string
+  readonly name?: string
+  readonly input: string
+}
+
+interface ActiveContent {
+  readonly type: "text" | "reasoning"
+  readonly id: string
+  readonly thinking?: MistralThinkingContent
+}
+
+export interface ParserState {
+  readonly tools: ToolStream.State<ToolKey>
+  readonly pendingTools: Partial<Record<ToolKey, PendingTool>>
+  readonly toolIDs: ReadonlyMap<string, string>
+  readonly usedToolIDs: ReadonlySet<string>
+  readonly completedTools: ReadonlyArray<LLMEvent>
+  readonly latestToolKey?: ToolKey
+  readonly generatedTools: number
+  readonly lifecycle: Lifecycle.State
+  readonly active?: ActiveContent
+  readonly nextContent: number
+  readonly usage?: Usage
+  readonly finishReason?: FinishReasonDetails
+}
+
+const mapUsage = (usage: MistralEvent["usage"]): Usage | undefined => {
+  if (!usage) return undefined
+  const input = usage.prompt_tokens ?? undefined
+  const reported =
+    usage.num_cached_tokens ??
+    usage.prompt_tokens_details?.cached_tokens ??
+    usage.prompt_token_details?.cached_tokens ??
+    undefined
+  const cached = input === undefined || reported === undefined ? undefined : Math.max(0, Math.min(input, reported))
+  const output = usage.completion_tokens ?? undefined
+  return new Usage({
+    inputTokens: input,
+    outputTokens: output,
+    nonCachedInputTokens: ProviderShared.subtractTokens(input, cached),
+    cacheReadInputTokens: cached,
+    totalTokens: ProviderShared.totalTokens(input, output, usage.total_tokens ?? undefined),
+    providerMetadata: { mistral: usage },
+  })
+}
+
+const mapFinishReason = (reason: string) => {
+  switch (reason) {
+    case "stop":
+      return "stop" as const
+    case "length":
+    case "model_length":
+      return "length" as const
+    case "tool_calls":
+      return "tool-calls" as const
+    case "content_filter":
+      return "content-filter" as const
+    case "error":
+    case "network_error":
+      return "error" as const
+    default:
+      return "unknown" as const
+  }
+}
+
+const thinkingUnits = (value: unknown): ReadonlyArray<MistralThinkingUnit> => {
+  if (typeof value === "string") return [{ type: "text", text: value }]
+  if (!Array.isArray(value)) return []
+  return value.filter(Schema.is(MistralThinkingUnit))
+}
+
+const thinkingText = (thinking: ReadonlyArray<MistralThinkingUnit>) =>
+  thinking.flatMap((unit) => (typeof unit.text === "string" ? [unit.text] : [])).join("")
+
+const thinkingMetadata = (thinking: MistralThinkingContent) => ({ mistral: { thinking } })
+
+const closeActive = (state: ParserState, events: LLMEvent[]) => {
+  if (!state.active) return state
+  const lifecycle =
+    state.active.type === "text"
+      ? Lifecycle.textEnd(state.lifecycle, events, state.active.id)
+      : Lifecycle.reasoningEnd(
+          state.lifecycle,
+          events,
+          state.active.id,
+          thinkingMetadata(state.active.thinking ?? { type: "thinking", thinking: [] }),
+          thinkingText(state.active.thinking?.thinking ?? []),
+        )
+  return { ...state, lifecycle, active: undefined }
+}
+
+const appendText = (state: ParserState, events: LLMEvent[], text: string) => {
+  if (text.length === 0) return state
+  const current = state.active?.type === "text" ? state : closeActive(state, events)
+  const active = current.active ?? { type: "text" as const, id: `text-${current.nextContent}` }
+  return {
+    ...current,
+    lifecycle: Lifecycle.textDelta(current.lifecycle, events, active.id, text),
+    active,
+    nextContent: current.active ? current.nextContent : current.nextContent + 1,
+  }
+}
+
+const appendThinking = (state: ParserState, events: LLMEvent[], part: MistralOutputContent) => {
+  const current = state.active?.type === "reasoning" ? state : closeActive(state, events)
+  const units = thinkingUnits(part.thinking)
+  const active = current.active ?? { type: "reasoning" as const, id: `reasoning-${current.nextContent}` }
+  const thinking = {
+    ...active.thinking,
+    ...part,
+    type: "thinking" as const,
+    thinking: [...(active.thinking?.thinking ?? []), ...units],
+  }
+  const text = thinkingText(units)
+  return {
+    ...current,
+    lifecycle:
+      text.length > 0
+        ? Lifecycle.reasoningDelta(current.lifecycle, events, active.id, text, thinkingMetadata(thinking))
+        : Lifecycle.reasoningStart(current.lifecycle, events, active.id, thinkingMetadata(thinking)),
+    active: { ...active, thinking },
+    nextContent: current.active ? current.nextContent : current.nextContent + 1,
+  }
+}
+
+const appendContent = (
+  state: ParserState,
+  events: LLMEvent[],
+  content: string | ReadonlyArray<MistralOutputContent>,
+) => {
+  if (typeof content === "string") return appendText(state, events, content)
+  return content.reduce((current, part) => {
+    if (part.type === "text") return appendText(current, events, part.text ?? "")
+    if (part.type === "thinking") return appendThinking(current, events, part)
+    return closeActive(current, events)
+  }, state)
+}
+
+const normalizeStreamToolID = (state: ParserState, source: string) => {
+  if (TOOL_ID.test(source))
+    return { id: source, state: { ...state, usedToolIDs: new Set([...state.usedToolIDs, source]) } }
+  const previous = state.toolIDs.get(source)
+  if (previous) return { id: previous, state }
+  let attempt = 0
+  let id = hashID(source)
+  while (state.usedToolIDs.has(id)) id = hashID(`${source}:${++attempt}`)
+  return {
+    id,
+    state: {
+      ...state,
+      toolIDs: new Map([...state.toolIDs, [source, id]]),
+      usedToolIDs: new Set([...state.usedToolIDs, id]),
+    },
+  }
+}
+
+const toolText = (tool: MistralToolDelta) => {
+  const value = tool.function?.arguments
+  if (typeof value === "string") return value
+  return value === null || value === undefined ? "" : ProviderShared.encodeJson(value)
+}
+
+const appendTools = Effect.fn("MistralChat.appendTools")(function* (
+  initial: ParserState,
+  events: LLMEvent[],
+  deltas: ReadonlyArray<MistralToolDelta>,
+) {
+  if (deltas.length === 0) return initial
+  let state = closeActive(initial, events)
+  for (const [position, delta] of deltas.entries()) {
+    const wireID = delta.id?.trim() || undefined
+    const providedID = wireID === "null" ? undefined : wireID
+    const key =
+      delta.index ??
+      (providedID
+        ? `id:${providedID}`
+        : deltas.length > 1
+          ? `position:${position}`
+          : (state.latestToolKey ?? `missing:${state.generatedTools}`))
+    const existing = state.tools[key]
+    const pending = state.pendingTools[key]
+    const source = providedID ?? `generated:${String(key)}`
+    const normalized =
+      existing || pending ? { id: existing?.id ?? pending?.id ?? "", state } : normalizeStreamToolID(state, source)
+    state = normalized.state
+    const name = existing?.name ?? pending?.name ?? (delta.function?.name?.trim() || undefined)
+    const text = `${pending?.input ?? ""}${toolText(delta)}`
+    if (!name) {
+      state = {
+        ...state,
+        pendingTools: { ...state.pendingTools, [key]: { id: normalized.id, input: text } },
+        latestToolKey: key,
+        generatedTools: state.generatedTools + (!providedID && !pending ? 1 : 0),
+      }
+      continue
+    }
+    const result = ToolStream.appendOrStart(
+      ADAPTER,
+      state.tools,
+      key,
+      { id: normalized.id, name, text },
+      "Mistral Chat tool call delta is missing a name",
+    )
+    if (ToolStream.isError(result)) return yield* result
+    if (result.events.length > 0) state = { ...state, lifecycle: Lifecycle.stepStart(state.lifecycle, events) }
+    events.push(...result.events)
+    const pendingTools = { ...state.pendingTools }
+    delete pendingTools[key]
+    state = {
+      ...state,
+      tools: result.tools,
+      pendingTools,
+      latestToolKey: key,
+      generatedTools: state.generatedTools + (!providedID && !existing && !pending ? 1 : 0),
+    }
+  }
+  return state
+})
+
+const hasLateContent = (event: MistralEvent) => {
+  const delta = event.choices?.[0]?.delta
+  if (typeof delta?.content === "string" && delta.content.length > 0) return true
+  if (Array.isArray(delta?.content) && delta.content.length > 0) return true
+  return (delta?.tool_calls ?? []).some(
+    (tool) => Boolean(tool.id) || Boolean(tool.function?.name) || tool.function?.arguments !== undefined,
+  )
+}
+
+const step = Effect.fn("MistralChat.step")(function* (state: ParserState, event: MistralEvent) {
+  if (event.error) {
+    const body = ProviderShared.encodeJson(event)
+    return yield* new AIError({
+      reason: classifyProviderFailure({
+        message: event.error.message,
+        status: typeof event.error.code === "number" ? event.error.code : undefined,
+        rawBody: body,
+      }),
+    })
+  }
+  const events: LLMEvent[] = []
+  const usage = mapUsage(event.usage) ?? state.usage
+  if (state.finishReason) {
+    if (hasLateContent(event))
+      return yield* ProviderShared.eventError(
+        ADAPTER,
+        "Mistral Chat received content after the finish reason",
+        ProviderShared.encodeJson(event),
+      )
+    return [{ ...state, usage }, events] as const
+  }
+  const choice = event.choices?.[0]
+  const withContent = choice?.delta?.content == null ? state : appendContent(state, events, choice.delta.content)
+  const withTools = yield* appendTools(withContent, events, choice?.delta?.tool_calls ?? [])
+  if (!choice?.finish_reason) return [{ ...withTools, usage }, events] as const
+
+  const finishReason = {
+    normalized: mapFinishReason(choice.finish_reason),
+    raw: choice.finish_reason,
+  }
+  const incomplete = finishReason.normalized === "length" || finishReason.normalized === "content-filter"
+  if (!incomplete && Object.keys(withTools.pendingTools).length > 0)
+    return yield* ProviderShared.eventError(
+      ADAPTER,
+      "Mistral Chat tool call delta is missing a name",
+      ProviderShared.encodeJson(event),
+    )
+  const finished =
+    !incomplete && Object.keys(withTools.tools).length > 0
+      ? yield* ToolStream.finishAll(ADAPTER, withTools.tools)
+      : undefined
+  return [
+    {
+      ...withTools,
+      tools: finished?.tools ?? withTools.tools,
+      completedTools: finished?.events ?? withTools.completedTools,
+      usage,
+      finishReason,
+    },
+    events,
+  ] as const
+})
+
+const finishEvents = Effect.fn("MistralChat.finishEvents")(function* (state: ParserState) {
+  if (!state.finishReason)
+    return yield* new AIError({
+      reason: new InvalidProviderOutputError({
+        message: "Mistral Chat stream ended without finish_reason",
+        classification: "incomplete-stream",
+        route: ADAPTER,
+      }),
+    })
+  const events: LLMEvent[] = []
+  const closed = closeActive(state, events)
+  const lifecycle = closed.completedTools.length > 0 ? Lifecycle.stepStart(closed.lifecycle, events) : closed.lifecycle
+  events.push(...closed.completedTools)
+  const reason =
+    state.finishReason.normalized === "stop" && closed.completedTools.some(LLMEvent.is.toolCall)
+      ? { ...state.finishReason, normalized: "tool-calls" as const }
+      : state.finishReason
+  Lifecycle.finish(lifecycle, events, { reason, usage: closed.usage })
+  return events
+})
+
+export const protocol = Protocol.make({
+  id: ADAPTER,
+  body: { schema: MistralBody, from: fromRequest },
+  stream: {
+    event: MistralStreamEvent,
+    initial: (): ParserState => ({
+      tools: ToolStream.empty<ToolKey>(),
+      pendingTools: {},
+      toolIDs: new Map(),
+      usedToolIDs: new Set(),
+      completedTools: [],
+      generatedTools: 0,
+      lifecycle: Lifecycle.initial(),
+      nextContent: 0,
+    }),
+    step: (state: ParserState, event) => (event === DONE ? Effect.succeed([state, []] as const) : step(state, event)),
+    terminal: (event) => event === DONE,
+    onHalt: finishEvents,
+  },
+})
+
+export const framing = Framing.sseWithDone
+export const httpTransport = HttpTransport.sseJson.with<MistralBody>().with({ framing })
+
+export const route = Route.make({
+  id: ADAPTER,
+  provider: "mistral",
+  providerMetadataKey: "mistral",
+  protocol,
+  endpoint: Endpoint.path(PATH, { baseURL: DEFAULT_BASE_URL }),
+  auth: Auth.none,
+  transport: httpTransport,
+})
+
+export * as MistralChat from "./mistral-chat.js"
@@ -1,4 +1,4 @@
-import { Effect, Schema } from "effect"
+import { Effect, Option, Schema } from "effect"
 import type { Content } from "@opencode-ai/schema/tool"
 import { HttpTransport } from "../route/transport/index.js"
 import { Protocol } from "../route/protocol.js"
@@ -32,13 +32,14 @@ export const PATH = "/responses"
 // =============================================================================
 // Request Body Schema
 // =============================================================================
-const OpenResponsesInputText = Schema.Struct({
+export const OpenResponsesInputText = Schema.Struct({
  type: Schema.tag("input_text"),
  text: Schema.String,
 })
-const OpenResponsesInputImage = Schema.Struct({
+export const OpenResponsesInputImage = Schema.Struct({
  type: Schema.tag("input_image"),
  image_url: Schema.String,
+  detail: Schema.optional(Schema.String),
 })
 const OpenResponsesInputFile = Schema.Struct({
  type: Schema.tag("input_file"),
@@ -54,7 +55,7 @@ const MediaInput = Schema.Union([OpenResponsesInputImage, OpenResponsesInputFile
 export type MediaInput = Schema.Schema.Type<typeof MediaInput>
 const OpenResponsesInputContent = Schema.Union([OpenResponsesInputText, MediaInput])

-const OpenResponsesOutputText = Schema.Struct({
+export const OpenResponsesOutputText = Schema.Struct({
  type: Schema.tag("output_text"),
  text: Schema.String,
 })
@@ -62,6 +63,13 @@ const OpenResponsesOutputText = Schema.Struct({
 export const MessagePhase = Schema.NullOr(Schema.Literals(["commentary", "final_answer"]))
 type MessagePhase = Schema.Schema.Type<typeof MessagePhase>

+export const MessageMetadata = Schema.Struct({
+  itemId: Schema.optional(Schema.String),
+  type: Schema.optional(Schema.Literal("message")),
+  status: Schema.optional(Schema.String),
+  phase: Schema.optional(MessagePhase),
+})
+
 const messagePhase = (value: unknown): MessagePhase | undefined => {
  if (value === null || value === "commentary" || value === "final_answer") return value
  return undefined
@@ -72,7 +80,7 @@ const OpenResponsesReasoningSummaryText = Schema.Struct({
  text: Schema.String,
 })

-const OpenResponsesReasoningItem = Schema.Struct({
+export const OpenResponsesReasoningItem = Schema.Struct({
  type: Schema.tag("reasoning"),
  id: Schema.optionalKey(Schema.String),
  summary: Schema.Array(OpenResponsesReasoningSummaryText),
@@ -149,16 +157,30 @@ const OpenResponsesFunctionCallOutput = Schema.Union([
  Schema.Array(OpenResponsesFunctionCallOutputContent),
 ])

+export const CompactionItem = Schema.Struct({
+  type: Schema.Literal("compaction"),
+  id: optionalNull(Schema.String),
+  encrypted_content: Schema.String,
+})
+
 export const InputItem = Schema.Union([
+  CompactionItem,
  Schema.Struct({ role: Schema.tag("system"), content: Schema.String }),
  Schema.Struct({ role: Schema.tag("developer"), content: Schema.String }),
-  Schema.Struct({ role: Schema.tag("user"), content: Schema.Array(OpenResponsesInputContent) }),
+  Schema.Struct({
+    role: Schema.tag("user"),
+    content: Schema.Array(OpenResponsesInputContent),
+    type: Schema.optional(Schema.Literal("message")),
+    id: Schema.optional(Schema.String),
+    status: Schema.optional(Schema.String),
+  }),
  Schema.Struct({
    type: Schema.tag("message"),
    id: Schema.optionalKey(Schema.String),
    role: Schema.tag("assistant"),
    content: Schema.Array(OpenResponsesOutputText),
    phase: Schema.optionalKey(MessagePhase),
+    status: Schema.optional(Schema.String),
  }),
  OpenResponsesReasoningItem,
  Schema.Struct({
@@ -267,7 +289,7 @@ const OpenResponsesBody = Schema.Struct({
 })
 export type OpenResponsesBody = Schema.Schema.Type<typeof OpenResponsesBody>

-const OpenResponsesUsage = Schema.Struct({
+export const OpenResponsesUsage = Schema.Struct({
  input_tokens: Schema.optional(Schema.Number),
  input_tokens_details: optionalNull(
    Schema.Struct({
@@ -387,6 +409,8 @@ export interface Extension {
 const BASE: Extension = { id: ADAPTER, name: NAME }

 export interface ParserState {
+  readonly provider: LLMRequest["model"]["provider"]
+  readonly completedCompactions: ReadonlySet<string>
  readonly id: string
  readonly name: string
  readonly providerMetadataKey: string
@@ -501,7 +525,13 @@ const lowerMedia = Effect.fn("OpenResponses.lowerMedia")(function* (
      ...(url ? { file_url: url } : { file_data: media.dataUrl }),
    }
  }
-  return { type: "input_image" as const, image_url: url ?? media.dataUrl }
+  return {
+    type: "input_image" as const,
+    image_url: url ?? media.dataUrl,
+    detail: yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(OpenResponsesInputImage.fields.detail))(
+      part.providerMetadata?.[request.model.route.providerMetadataKey ?? "openresponses"]?.detail,
+    ),
+  }
 })

 const lowerUserContent = Effect.fnUntraced(function* (
@@ -568,6 +598,9 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
  const providerMetadataKey = request.model.route.providerMetadataKey ?? "openresponses"

  for (const message of request.messages) {
+    const metadata = yield* ProviderShared.validateWith(
+      Schema.decodeUnknownEffect(Schema.UndefinedOr(MessageMetadata)),
+    )(message.providerMetadata?.[providerMetadataKey])
    if (message.role === "system") {
      input.push({
        role: "developer",
@@ -577,10 +610,9 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
    }

    if (message.role === "user") {
-      input.push({
-        role: "user",
-        content: yield* Effect.forEach(message.content, (part) => lowerUserContent(part, request, extension)),
-      })
+      const content = yield* Effect.forEach(message.content, (part) => lowerUserContent(part, request, extension))
+      if (content.length > 0)
+        input.push({ role: "user", content, type: metadata?.type, id: metadata?.itemId, status: metadata?.status })
      continue
    }

@@ -593,9 +625,10 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
        const groups = content.reduce<
          Array<{ id: string | undefined; phase: MessagePhase | null | undefined; parts: TextPart[] }>
        >((groups, part) => {
-          const metadata = part.providerMetadata?.[providerMetadataKey]
-          const id = itemID(part.providerMetadata, providerMetadataKey)
-          const phase = ProviderShared.isRecord(metadata) ? messagePhase(metadata.phase) : undefined
+          const partMetadata = part.providerMetadata?.[providerMetadataKey]
+          const id = itemID(part.providerMetadata, providerMetadataKey) ?? metadata?.itemId
+          const partPhase = messagePhase(partMetadata?.phase)
+          const phase = partPhase === undefined ? metadata?.phase : partPhase
          const group = groups.at(-1)
          if (group && group.id === id && group.phase === phase) group.parts.push(part)
          else groups.push({ id, phase, parts: [part] })
@@ -606,6 +639,7 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
            type: "message" as const,
            ...(group.id === undefined ? {} : { id: group.id }),
            role: "assistant" as const,
+            status: metadata?.status,
            content: group.parts.map((part) => ({ type: "output_text" as const, text: part.text })),
            ...(group.phase === undefined ? {} : { phase: group.phase }),
          })),
@@ -613,6 +647,15 @@ const lowerMessages = Effect.fn("OpenResponses.lowerMessages")(function* (reques
        content.splice(0, content.length)
      }
      for (const part of message.content) {
+        if (part.type === "compaction") {
+          flushText()
+          if (part.provider !== request.model.provider || part.encrypted === undefined)
+            return yield* ProviderShared.invalidRequest(
+              "Compaction state must be replayed to its originating provider and API",
+            )
+          input.push({ type: "compaction", id: part.id, encrypted_content: part.encrypted })
+          continue
+        }
        if (part.type === "text") {
          content.push(part)
          continue
@@ -780,7 +823,7 @@ export const fromRequest = Effect.fn("OpenResponses.fromRequest")(function* (req
 // cached-read and cache-write subsets, and `output_tokens` (inclusive total)
 // with a `reasoning_tokens` subset. Pass the totals through and derive the
 // non-cached breakdown.
-const mapUsage = (usage: OpenResponsesUsage | null | undefined, providerMetadataKey: string) => {
+export const mapUsage = (usage: OpenResponsesUsage | null | undefined, providerMetadataKey: string) => {
  if (!usage) return undefined
  const cached = usage.input_tokens_details?.cached_tokens
  const cacheWrite = usage.input_tokens_details?.cache_write_tokens
@@ -845,6 +888,21 @@ const onOutputTextDone = (state: ParserState, event: Event, id: string): StepRes
  return [{ ...state, lifecycle: Lifecycle.textEnd(state.lifecycle, events, id) }, events]
 }

+const decodeMessagePart = Schema.decodeUnknownOption(
+  Schema.Union([OpenResponsesOutputText, Schema.Struct({ type: Schema.tag("refusal"), refusal: Schema.String })]),
+)
+
+const decodeSummaryPart = Schema.decodeUnknownOption(OpenResponsesReasoningSummaryText)
+
+const decodeReasoningPart = Schema.decodeUnknownOption(
+  Schema.Struct({ type: Schema.tag("reasoning_text"), text: Schema.String }),
+)
+
+const joinReasoningText = (parts: ReadonlyArray<string | undefined>) => {
+  if (!parts.some((part) => part !== undefined && part.length > 0)) return undefined
+  return parts.filter((part) => part !== undefined).join("\n\n")
+}
+
 export const outputItemID = (state: ParserState, event: Event) =>
  event.output_index === undefined ? event.item_id : (state.outputItems[event.output_index] ?? event.item_id)

@@ -1065,24 +1123,52 @@ const onFunctionCallArgumentsDelta = Effect.fn("OpenResponses.onFunctionCallArgu
  return [{ ...state, lifecycle, tools: result.tools }, events] satisfies StepResult
 })

-const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (state: ParserState, event: Event) {
-  const item = event.item
+const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
+  state: ParserState,
+  item: Event["item"],
+) {
  if (!item) return [state, NO_EVENTS] satisfies StepResult

-  if (item.type === "message" && item.id !== undefined) {
-    const itemPhase = messagePhase(item.phase)
-    const phase = itemPhase === undefined && state.message?.id === item.id ? state.message.phase : itemPhase
+  if (item.type === "compaction") {
+    if (!item.id || typeof item.encrypted_content !== "string")
+      return yield* ProviderShared.eventError(state.id, "Compaction output is missing its id or encrypted content")
+    if (state.completedCompactions.has(item.id)) return [state, NO_EVENTS] satisfies StepResult
    const events: LLMEvent[] = []
+    const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
+    events.push(
+      LLMEvent.compaction({
+        provider: state.provider,
+        id: item.id,
+        encrypted: item.encrypted_content,
+      }),
+    )
+    return [
+      { ...state, lifecycle, completedCompactions: new Set([...state.completedCompactions, item.id]) },
+      events,
+    ] satisfies StepResult
+  }
+
+  if (item.type === "message" && item.id !== undefined) {
+    const message = state.message?.id === item.id ? state.message : undefined
+    const itemPhase = messagePhase(item.phase)
+    const phase = itemPhase === undefined ? message?.phase : itemPhase
+    const parts: ReadonlyArray<unknown> = Array.isArray(item.content) ? item.content : []
+    const content: string[] = []
+    for (const part of parts) {
+      const decoded = Option.getOrUndefined(decodeMessagePart(part))
+      if (!decoded) continue
+      content.push(decoded.type === "output_text" ? decoded.text : decoded.refusal)
+    }
+    const text = content.length > 0 ? content.join("") : undefined
+    const metadata = providerMetadata(state, { itemId: item.id, ...(phase === undefined ? {} : { phase }) })
+    const events: LLMEvent[] = []
+    const lifecycle =
+      message && text ? Lifecycle.textStart(state.lifecycle, events, item.id, metadata) : state.lifecycle
    return [
      {
        ...state,
-        lifecycle: Lifecycle.textEnd(
-          state.lifecycle,
-          events,
-          item.id,
-          providerMetadata(state, { itemId: item.id, ...(phase === undefined ? {} : { phase }) }),
-        ),
-        message: state.message?.id === item.id ? undefined : state.message,
+        lifecycle: Lifecycle.textEnd(lifecycle, events, item.id, metadata, text),
+        message: message ? undefined : state.message,
      },
      events,
    ] satisfies StepResult
@@ -1137,17 +1223,33 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
  }

  if (isReasoningItem(item)) {
-    const events: LLMEvent[] = []
+    if (state.reasoningItems[item.id]?.open === false) return [state, NO_EVENTS] satisfies StepResult
    const metadata = reasoningMetadata(state, item)
+    const summaryParts: ReadonlyArray<unknown> = Array.isArray(item.summary) ? item.summary : []
+    const summary: Array<string | undefined> = []
+    for (const part of summaryParts) {
+      const decoded = Option.getOrUndefined(decodeSummaryPart(part))
+      // Keep missing entries so the array still matches the provider's summary indexes.
+      summary.push(decoded?.text)
+    }
+    const reasoningParts: ReadonlyArray<unknown> = Array.isArray(item.content) ? item.content : []
+    const content: string[] = []
+    for (const part of reasoningParts) {
+      const decoded = Option.getOrUndefined(decodeReasoningPart(part))
+      if (decoded) content.push(decoded.text)
+    }
+    const itemText = joinReasoningText(summary) ?? joinReasoningText(content)
+    const events: LLMEvent[] = []
    const reasoningItem = state.reasoningItems[item.id]
    if (reasoningItem) {
-      if (!reasoningItem.open) return [state, NO_EVENTS] satisfies StepResult
-      const lifecycle = Object.entries(reasoningItem.summaryParts)
-        .filter((entry) => entry[1] === "active" || entry[1] === "can-conclude")
-        .reduce(
-          (lifecycle, entry) => Lifecycle.reasoningEnd(lifecycle, events, `${item.id}:${entry[0]}`, metadata),
-          state.lifecycle,
-        )
+      const fragments = Object.entries(reasoningItem.summaryParts)
+      let lifecycle = state.lifecycle
+      for (const [index, status] of fragments) {
+        if (status === "concluded") continue
+        // Do not repeat earlier summaries that were already emitted as separate fragments.
+        const finalText = fragments.length === 1 ? itemText : summary[Number(index)]
+        lifecycle = Lifecycle.reasoningEnd(lifecycle, events, `${item.id}:${index}`, metadata, finalText || undefined)
+      }
      return [
        {
          ...state,
@@ -1167,7 +1269,13 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
    if (!state.lifecycle.reasoning.has(item.id)) {
      const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
      events.push(LLMEvent.reasoningStart({ id: item.id, providerMetadata: metadata }))
-      events.push(LLMEvent.reasoningEnd({ id: item.id, providerMetadata: metadata }))
+      events.push(
+        LLMEvent.reasoningEnd({
+          id: item.id,
+          providerMetadata: metadata,
+          text: itemText,
+        }),
+      )
      return [
        {
          ...state,
@@ -1195,38 +1303,38 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
 })

 const onResponseFinish = Effect.fn("OpenResponses.onResponseFinish")(function* (state: ParserState, event: Event) {
-  const reconciled =
-    event.type === "response.completed"
-      ? yield* Effect.reduce(
-          event.response?.output ?? [],
-          () => [state, NO_EVENTS] satisfies StepResult,
-          ([current, events], item) => {
-            const id = item.id ?? (item.type === "function_call" ? item.call_id : undefined)
-            if (
-              id === undefined ||
-              ((item.type !== "function_call" || !current.tools[id]) &&
-                (item.type !== "reasoning" || !current.reasoningItems[id]?.open))
-            )
-              return Effect.succeed([current, events] satisfies StepResult)
-            return onOutputItemDone(current, { type: "response.output_item.done", item }).pipe(
-              Effect.map(([next, emitted]) => [next, [...events, ...emitted]] satisfies StepResult),
-            )
-          },
-        )
-      : ([state, NO_EVENTS] satisfies StepResult)
-  const current = reconciled[0]
-  // Some compatible providers omit output_item.done even after completing the response.
-  const pending =
-    event.type === "response.completed"
-      ? yield* ToolStream.finishAll(current.id, current.tools)
-      : { tools: current.tools, events: NO_EVENTS }
-  const events: LLMEvent[] = [...reconciled[1], ...pending.events]
-  const hasFunctionCall =
-    pending.events.some((event) => LLMEvent.is.toolCall(event) || LLMEvent.is.toolInputError(event)) ||
-    current.hasFunctionCall
+  let current = state
+  const events: LLMEvent[] = []
+  if (event.type === "response.completed") {
+    for (const item of event.response?.output ?? []) {
+      if (item.type !== "compaction" && item.type !== "function_call") continue
+      if (item.type === "compaction") {
+        // Terminal recovery cannot insert a checkpoint before already-emitted content.
+        if (state.lifecycle.stepStarted && !state.completedCompactions.has(item.id ?? ""))
+          return yield* ProviderShared.eventError(
+            state.id,
+            "Cannot recover a compaction checkpoint after output has been emitted",
+          )
+      }
+      if (item.type === "function_call" && !current.tools[item.id ?? item.call_id ?? ""]) continue
+      const [next, emitted] = yield* onOutputItemDone(current, item)
+      current = next
+      events.push(...emitted)
+    }
+    // Some compatible providers omit output_item.done even after completing the response.
+    const pending = yield* ToolStream.finishAll(current.id, current.tools)
+    current = {
+      ...current,
+      tools: pending.tools,
+      hasFunctionCall:
+        current.hasFunctionCall ||
+        pending.events.some((event) => LLMEvent.is.toolCall(event) || LLMEvent.is.toolInputError(event)),
+    }
+    events.push(...pending.events)
+  }
  const lifecycle = Lifecycle.finish(current.lifecycle, events, {
    reason: {
-      normalized: mapFinishReason(event, hasFunctionCall),
+      normalized: mapFinishReason(event, current.hasFunctionCall),
      raw: event.response?.incomplete_details?.reason,
    },
    usage: mapUsage(event.response?.usage, current.providerMetadataKey),
@@ -1238,7 +1346,7 @@ const onResponseFinish = Effect.fn("OpenResponses.onResponseFinish")(function* (
          })
        : undefined,
  })
-  return [{ ...current, lifecycle, hasFunctionCall, tools: pending.tools }, events] satisfies StepResult
+  return [{ ...current, lifecycle }, events] satisfies StepResult
 })

 // Build the prettiest summary available from whatever the provider supplied.
@@ -1346,7 +1454,7 @@ export const step = (state: ParserState, input: Event) => {
  if (event.type === "response.output_item.done") {
    if (event.item?.type === "message" && event.item.id === undefined)
      return ProviderShared.eventError(state.id, `${event.type} message is missing id`)
-    return onOutputItemDone(state, event)
+    return onOutputItemDone(state, event.item)
  }
  if (event.type === "response.completed" || event.type === "response.incomplete") return onResponseFinish(state, event)
  if (event.type === "response.failed") return providerFailure(event, `${state.name} response failed`)
@@ -1373,6 +1481,8 @@ export const step = (state: ParserState, input: Event) => {
 * implementations compose this baseline with their own tools and event variants.
 */
 export const initial = (request: LLMRequest, extension: Extension = BASE): ParserState => ({
+  provider: request.model.provider,
+  completedCompactions: new Set<string>(),
  id: extension.id,
  name: extension.name,
  providerMetadataKey: request.model.route.providerMetadataKey ?? "openresponses",
@@ -815,7 +815,12 @@ const mapFinishReason = Effect.fn("OpenAIChat.mapFinishReason")(function* (event
    case "tool_calls":
      return "tool-calls" as const
    default:
-      return "unknown" as const
+      return yield* new AIError({
+        reason: new UnknownProviderError({
+          message: `Provider finish_reason: ${reason}`,
+          body: ProviderShared.encodeJson(event),
+        }),
+      })
  }
 })

@@ -1054,17 +1059,25 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
      events.push(...result.events)
    }

-    if (finishReason !== undefined && state.finishReason === undefined && Object.keys(pendingTools).length > 0)
+    const incompleteTools = finishReason?.normalized === "content-filter" || finishReason?.normalized === "length"
+    if (
+      finishReason !== undefined &&
+      !incompleteTools &&
+      state.finishReason === undefined &&
+      Object.keys(pendingTools).length
+    )
      return yield* ProviderShared.eventError(
        ADAPTER,
        "OpenAI Chat tool call delta is missing id or name",
        ProviderShared.encodeJson(event),
      )

-    // Finalize accumulated tool inputs eagerly when finish_reason arrives so
-    // valid calls and malformed local calls settle independently.
+    // Filtering or truncation terminates the response without confirming pending tool calls.
    const finished =
-      finishReason !== undefined && state.finishReason === undefined && Object.keys(tools).length > 0
+      finishReason !== undefined &&
+      !incompleteTools &&
+      state.finishReason === undefined &&
+      Object.keys(tools).length > 0
        ? yield* ToolStream.finishAll(ADAPTER, tools)
        : undefined

@@ -12,6 +12,7 @@ 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"

 const ADAPTER = "openai-responses"
 const NAME = "OpenAI Responses"
@@ -20,6 +21,14 @@ const WEBSOCKET_ROTATE_AFTER_MS = 55 * 60 * 1000
 export const DEFAULT_BASE_URL = "https://api.openai.com/v1"
 export const PATH = OpenResponses.PATH

+export const ContextManagement = Schema.Array(
+  Schema.Struct({
+    type: Schema.Literal("compaction"),
+    compactThreshold: Schema.optional(Schema.Int.check(Schema.isGreaterThan(0))),
+  }),
+)
+export type ContextManagement = typeof ContextManagement.Type
+
 const OpenAIResponsesImageGenerationTool = Schema.Struct({
  type: Schema.tag("image_generation"),
  action: Schema.optional(Schema.Literals(["auto", "generate", "edit"])),
@@ -78,6 +87,14 @@ const OpenAIResponsesCoreFields = {
  input: Schema.Array(Schema.Union([OpenResponses.InputItem, OpenAIResponsesHostedToolItem])),
  tools: optionalArray(OpenAIResponsesTools),
  tool_choice: Schema.optional(OpenAIResponsesToolChoice),
+  context_management: Schema.optional(
+    Schema.Array(
+      Schema.Struct({
+        type: Schema.Literal("compaction"),
+        compact_threshold: Schema.optional(Schema.Int.check(Schema.isGreaterThan(0))),
+      }),
+    ),
+  ),
 }

 const OpenAIResponsesBody = Schema.Struct({
@@ -125,6 +142,9 @@ const lowerToolChoice = (toolChoice: NonNullable<LLMRequest["toolChoice"]>, tool
 const decodeBody = ProviderShared.validateWith(Schema.decodeUnknownEffect(OpenAIResponsesBody))

 const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request: LLMRequest) {
+  const management = yield* ProviderShared.validateWith(
+    Schema.decodeUnknownEffect(Schema.UndefinedOr(ContextManagement)),
+  )(request.providerOptions?.contextManagement)
  const body = yield* OpenResponses.fromRequestWithExtension(
    LLMRequest.update(request, { tools: [], toolChoice: undefined }),
    extension,
@@ -133,6 +153,7 @@ const fromRequest = Effect.fn("OpenAIResponses.fromRequest")(function* (request:
  const parallelToolCalls = OpenResponses.resolveParallelToolCalls(request)
  return yield* decodeBody({
    ...body,
+    context_management: management?.map((edit) => ({ type: edit.type, compact_threshold: edit.compactThreshold })),
    ...(parallelToolCalls === undefined ? {} : { parallel_tool_calls: parallelToolCalls }),
    tools:
      request.tools.length === 0
@@ -223,6 +244,7 @@ export const transport = channelTransport({
 })

 export const route = Route.make({
+  compact: ResponsesCompaction.execute,
  id: ADAPTER,
  provider: "openai",
  providerMetadataKey: "openai",
@@ -1,4 +1,4 @@
-import { Effect, Schema } from "effect"
+import { Effect, Encoding, Schema } from "effect"
 import type { MediaPart } from "../../schema/index.js"
 import { ProviderShared } from "../shared.js"

@@ -57,6 +57,16 @@ const documentBlock = (name: string, format: DocumentFormat, bytes: string): Doc
  },
 })

+const mediaBase64 = Effect.fn("BedrockMedia.mediaBase64")(function* (part: MediaPart) {
+  const media = ProviderShared.normalizeMedia(part)
+  const bytes = yield* Effect.fromResult(Encoding.decodeBase64(media.base64)).pipe(
+    Effect.mapError((cause) =>
+      ProviderShared.invalidRequest("Bedrock Converse media data must be valid base64", cause),
+    ),
+  )
+  return Encoding.encodeBase64(bytes)
+})
+
 // Route by MIME. Known image/document formats lower into a typed block; anything
 // else fails with a clear error instead of silently degrading to a malformed
 // document block. Image MIME types not in `IMAGE_FORMATS` (e.g. `image/svg+xml`)
@@ -66,8 +76,7 @@ export const lower = Effect.fn("BedrockMedia.lower")(function* (part: MediaPart)
  const mime = part.mediaType.toLowerCase()
  const imageFormat = IMAGE_FORMATS[mime as keyof typeof IMAGE_FORMATS]
  if (imageFormat) {
-    const media = ProviderShared.normalizeMedia(part)
-    return { image: { format: imageFormat, source: { bytes: media.base64 } } } satisfies ImageBlock
+    return { image: { format: imageFormat, source: { bytes: yield* mediaBase64(part) } } } satisfies ImageBlock
  }
  if (mime.startsWith("image/"))
    return yield* ProviderShared.invalidRequest(`Bedrock Converse does not support image media type ${part.mediaType}`)
@@ -75,8 +84,7 @@ export const lower = Effect.fn("BedrockMedia.lower")(function* (part: MediaPart)
  if (documentFormat) {
    if (!part.filename)
      return yield* ProviderShared.invalidRequest("Bedrock Converse document media requires a filename")
-    const media = ProviderShared.normalizeMedia(part)
-    return documentBlock(part.filename, documentFormat, media.base64)
+    return documentBlock(part.filename, documentFormat, yield* mediaBase64(part))
  }
  return yield* ProviderShared.invalidRequest(`Bedrock Converse does not support media type ${part.mediaType}`)
 })
@@ -0,0 +1,153 @@
+import { Effect, Schema, Stream } from "effect"
+import {
+  AIError,
+  InvalidProviderOutputError,
+  CompactionPart,
+  CompactionResponse,
+  HttpOptions,
+  LLMRequest,
+  Message,
+  type ContentPart,
+  mergeJsonRecords,
+} from "../../schema/index.js"
+import type { CompactOperation } from "../../route/client.js"
+import { Endpoint } from "../../route/endpoint.js"
+import { RequestExecutor } from "../../route/executor.js"
+import { HttpTransport } from "../../route/transport/index.js"
+import { OpenResponses } from "../open-responses.js"
+import { JsonObject, ProviderShared } from "../shared.js"
+
+const Body = Schema.Struct({
+  model: Schema.String,
+  input: Schema.Array(Schema.Unknown),
+  instructions: Schema.optional(Schema.String),
+  previous_response_id: Schema.optional(Schema.String),
+})
+
+const Text = Schema.Union([OpenResponses.OpenResponsesInputText, OpenResponses.OpenResponsesOutputText])
+const File = Schema.Union([
+  Schema.Struct({ type: Schema.Literal("input_file"), filename: Schema.String, file_url: Schema.String }),
+  Schema.Struct({ type: Schema.Literal("input_file"), filename: Schema.String, file_data: Schema.String }),
+])
+const MessageFields = {
+  type: Schema.Literal("message"),
+  id: Schema.optional(Schema.String),
+  status: Schema.optional(Schema.String),
+  phase: Schema.optional(OpenResponses.MessagePhase),
+}
+const Response = Schema.Struct({
+  object: Schema.Literal("response.compaction"),
+  output: Schema.Array(
+    Schema.Union([
+      OpenResponses.CompactionItem,
+      OpenResponses.OpenResponsesReasoningItem,
+      Schema.Struct({
+        ...MessageFields,
+        role: Schema.Literal("user"),
+        content: Schema.Array(Schema.Union([Text, OpenResponses.OpenResponsesInputImage, File])).check(
+          Schema.isMinLength(1),
+        ),
+      }),
+      Schema.Struct({
+        ...MessageFields,
+        role: Schema.Literal("assistant"),
+        content: Schema.Array(Text).check(Schema.isMinLength(1)),
+      }),
+    ]),
+  ),
+  usage: Schema.optional(Schema.StructWithRest(OpenResponses.OpenResponsesUsage, [JsonObject])),
+})
+
+export const execute: CompactOperation = Effect.fn("ResponsesCompaction.execute")(
+  function* (request, executor, options) {
+    const route = request.model.route
+    const native = yield* route.body.from(request)
+    const body = yield* ProviderShared.validateWith(Schema.decodeUnknownEffect(Body))(
+      mergeJsonRecords(native, request.http?.body),
+    )
+    const url = Endpoint.render(route.endpoint, { request, body: native })
+    url.pathname = `${url.pathname.replace(/\/$/, "")}/compact`
+    const parts = yield* HttpTransport.jsonRequestParts({
+      request: LLMRequest.update(request, {
+        http: request.http === undefined ? undefined : new HttpOptions({ ...request.http, body: undefined }),
+      }),
+      body,
+      endpoint: Endpoint.path(url.toString()),
+      auth: route.auth,
+      encodeBody: Schema.encodeSync(Schema.fromJsonString(Body)),
+    })
+    const response = yield* executor.execute(
+      ProviderShared.jsonPost({ url: parts.url, body: parts.bodyText, headers: parts.headers }),
+      options?.http,
+    )
+    const text = yield* RequestExecutor.responseStream(response).pipe(
+      Stream.decodeText(),
+      Stream.runFold(
+        () => "",
+        (text, chunk) => text + chunk,
+      ),
+    )
+    const invalid = (message: string, cause?: unknown) =>
+      new AIError({
+        reason: new InvalidProviderOutputError({
+          route: route.id,
+          message,
+          body: text,
+          cause,
+          http: RequestExecutor.responseHttp(response),
+        }),
+      })
+    const result = yield* Schema.decodeUnknownEffect(Schema.fromJsonString(Response))(text).pipe(
+      Effect.mapError((cause) => invalid("Invalid compaction response", cause)),
+    )
+    if (!result.output.some((item) => item.type === "compaction"))
+      return yield* invalid("Compaction response did not contain a checkpoint")
+    return new CompactionResponse({
+      messages: result.output.map((item) => toMessage(item, request.model)),
+      usage: OpenResponses.mapUsage(result.usage, route.providerMetadataKey ?? String(request.model.provider)),
+    })
+  },
+)
+
+function toMessage(item: (typeof Response.Type.output)[number], model: LLMRequest["model"]): Message {
+  if (item.type === "compaction")
+    return Message.assistant(
+      CompactionPart.make({ provider: model.provider, id: item.id ?? undefined, encrypted: item.encrypted_content }),
+    )
+
+  const key = model.route.providerMetadataKey ?? String(model.provider)
+  if (item.type === "reasoning") {
+    const summary = item.summary.length ? item.summary : [{ text: "" }]
+    return Message.assistant(
+      summary.map((part) => ({
+        type: "reasoning" as const,
+        text: part.text,
+        providerMetadata: { [key]: { itemId: item.id, reasoningEncryptedContent: item.encrypted_content } },
+      })),
+    )
+  }
+
+  return Message.make({
+    role: item.role,
+    providerMetadata: { [key]: { itemId: item.id, type: item.type, status: item.status, phase: item.phase } },
+    content: item.content.map((part): ContentPart => {
+      if (part.type === "input_text" || part.type === "output_text") return { type: "text", text: part.text }
+      if (part.type === "input_image")
+        return {
+          type: "media",
+          data: part.image_url,
+          mediaType: /^data:([^;,]+)/.exec(part.image_url)?.[1] ?? "image/*",
+          providerMetadata: part.detail === undefined ? undefined : { [key]: { detail: part.detail } },
+        }
+      const data = "file_url" in part ? part.file_url : part.file_data
+      return {
+        type: "media",
+        data,
+        filename: part.filename,
+        mediaType: /^data:([^;,]+)/.exec(data)?.[1] ?? "application/octet-stream",
+      }
+    }),
+  })
+}
+
+export * as ResponsesCompaction from "./responses-compaction.js"
@@ -44,6 +44,10 @@ const extension = {

 const decodeBody = ProviderShared.validateWith(Schema.decodeUnknownEffect(XAIResponsesBody))
 const fromRequest = Effect.fn("XAIResponses.fromRequest")(function* (request: LLMRequest) {
+  if (request.providerOptions?.contextManagement !== undefined)
+    return yield* ProviderShared.invalidRequest(
+      "xAI requires explicit compaction through LLMClient.compact; automatic context management is not supported",
+    )
  return yield* decodeBody(yield* OpenResponses.fromRequestWithExtension(request, extension))
 })

@@ -76,6 +76,8 @@ const INVALID_REQUEST_CODES = new Set(["invalid_prompt", "invalid_request_error"
 const RATE_LIMIT_TEXT = /rate increased too quickly|rate[-_\s]?limit|too[_\s]?many[_\s]?requests/i
 const QUOTA_TEXT = /insufficient[-_\s]?quota|quota[-_\s]?exceeded/i
 const CONTENT_POLICY_TEXT = /content[-_\s]?policy|content_filter|safety/i
+const SERVER_ERROR_TEXT =
+  /\b(?:try again|(?:please |you can )?retry (?:the |this |your )?request|try (?:the |this |your )?request again|(?:currently |temporarily )?at capacity|overloaded|temporarily unavailable|service[-_\s]?unavailable|(?:server|internal)[-_\s]?error|server (?:is )?busy|provider returned (?:an )?error|resource[-_\s]?exhausted|upstream (?:connect|connection|request)|request buffer limit while retrying upstream)\b/i

 export interface ProviderFailure {
  readonly message: string
@@ -139,6 +141,9 @@ export function classifyProviderFailure(input: ProviderFailure): AIError["reason
    input.status === 408 ||
    input.status === 409 ||
    (input.status !== undefined && input.status >= 500) ||
+    ((input.status === undefined || input.status < 400) &&
+      !codes.some((code) => INVALID_REQUEST_CODES.has(code)) &&
+      SERVER_ERROR_TEXT.test(text)) ||
    codes.some((code) => SERVER_CODES.has(code) || code.includes("exhausted") || code.includes("unavailable"))
  )
    return new ProviderInternalError({
@@ -1,9 +1,11 @@
-import type { RouteDefaultsInput } from "../route/client.js"
+import type { Route, RouteDefaultsInput } from "../route/client.js"
 import { Auth } from "../route/auth.js"
 import type { ProviderPackage } from "../provider-package.js"
 import { ProviderID, type ModelID } from "../schema/index.js"
 import * as BedrockConverse from "../protocols/bedrock-converse.js"
 import type { BedrockCredentials } from "../protocols/bedrock-converse.js"
+import { BedrockMessages } from "../protocols/bedrock-messages.js"
+import type { AnthropicMessages } from "../protocols/anthropic-messages.js"

 export const id = ProviderID.make("amazon-bedrock")

@@ -25,38 +27,40 @@ export interface Settings extends ProviderPackage.Settings {
  readonly region?: string
  readonly topP?: number
 }
-export const routes = [BedrockConverse.route]
+export const routes = [BedrockConverse.route, BedrockMessages.route]

 const bedrockBaseURL = (region: string) => `https://bedrock-runtime.${region}.amazonaws.com`

-const configuredRoute = (input: Config) => {
+const configuredRoute = <Body, Prepared>(route: Route<Body, Prepared>, input: Config) => {
  const { apiKey, credentials, region, baseURL, ...rest } = input
  const resolvedRegion = region ?? credentials?.region ?? "us-east-1"
-  return BedrockConverse.route.with({
+  return route.with({
    ...rest,
    provider: id,
-    providerMetadataKey: "bedrock",
+    providerMetadataKey: route.providerMetadataKey,
    endpoint: { baseURL: baseURL ?? bedrockBaseURL(resolvedRegion) },
    auth: apiKey === undefined ? BedrockConverse.sigV4Auth(credentials) : Auth.bearer(apiKey),
  })
 }

 export const configure = (input: Config = {}) => {
-  const route = configuredRoute(input)
+  const route = configuredRoute(BedrockConverse.route, input)
+  const messages = configuredRoute(BedrockMessages.route, input)
  return {
    id,
    model: (modelID: string | ModelID) => route.model({ id: modelID }),
+    messages: (modelID: string | ModelID) => messages.model<AnthropicMessages.ProviderOptionsInput>({ id: modelID }),
    configure,
  }
 }

 export const provider = configure()
-export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) => {
+const config = (settings: Settings): Config => {
  if (settings.auth === "bearer" && settings.apiKey === undefined)
    throw new Error("Amazon Bedrock bearer auth requires apiKey")
  if (settings.auth === "sigv4" && settings.apiKey !== undefined)
    throw new Error("Amazon Bedrock SigV4 auth does not accept apiKey")
-  return configure({
+  return {
    apiKey: settings.auth === "sigv4" ? undefined : settings.apiKey,
    baseURL: settings.baseURL,
    credentials: settings.credentials,
@@ -64,5 +68,13 @@ export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, se
    headers: settings.headers === undefined ? undefined : { ...settings.headers },
    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
    region: settings.region,
-  }).model(modelID)
+  }
 }
+
+export const model: ProviderPackage.Definition<Settings>["model"] = (modelID, settings) =>
+  configure(config(settings)).model(modelID)
+export const messagesModel: ProviderPackage.Definition<
+  Settings & { readonly providerOptions?: AnthropicMessages.ProviderOptionsInput },
+  AnthropicMessages.ProviderOptionsInput
+>["model"] = (modelID, settings) =>
+  configure({ ...config(settings), providerOptions: settings.providerOptions }).messages(modelID)
@@ -0,0 +1 @@
+export { messagesModel as model } from "../amazon-bedrock.js"
@@ -57,7 +57,9 @@ const route = Route.make({
  }),
  endpoint: Endpoint.path(({ request }) => `/${request.model.id}:streamRawPredict`),
  auth: Auth.none,
-  framing: AnthropicMessages.framing,
+  transport: AnthropicMessages.transport<
+    Omit<AnthropicMessages.AnthropicMessagesBody, "model"> & { readonly anthropic_version: typeof VERSION }
+  >(),
  headers: () => ({ "anthropic-version": HEADER_VERSION }),
 })

@@ -13,6 +13,7 @@ export * as GoogleVertexChat from "./google-vertex-chat.js"
 export * as GoogleVertexMessages from "./google-vertex-messages.js"
 export * as GoogleVertexResponses from "./google-vertex-responses.js"
 export * as Groq from "./groq.js"
+export * as Mistral from "./mistral.js"
 export * as OpenAI from "./openai.js"
 export * as OpenAICompatible from "./openai-compatible.js"
 export * as OpenAICompatibleResponses from "./openai-compatible-responses.js"
@@ -0,0 +1,51 @@
+import type { ProviderPackage } from "../provider-package.js"
+import { MistralChat } from "../protocols/mistral-chat.js"
+import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
+import type { RouteDefaultsInput } from "../route/client.js"
+import { ProviderID, type ModelID } from "../schema/index.js"
+
+export const id = ProviderID.make("mistral")
+
+export type ProviderOptions = MistralChat.ProviderOptionsInput
+
+export type LanguageModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
+  ProviderAuthOption<"optional"> & {
+    readonly baseURL?: string
+    readonly providerOptions?: ProviderOptions
+  }
+
+export interface Settings extends ProviderPackage.Settings {
+  readonly apiKey?: string
+  readonly baseURL?: string
+  readonly providerOptions?: ProviderOptions
+}
+
+export const route = MistralChat.route
+export const routes = [route]
+
+export const configure = (input: LanguageModelOptions = {}) => {
+  const { apiKey: _apiKey, auth: _auth, baseURL, ...defaults } = input
+  const configured = route.with({
+    ...defaults,
+    endpoint: { baseURL: baseURL ?? MistralChat.DEFAULT_BASE_URL },
+    auth: AuthOptions.bearer(input, "MISTRAL_API_KEY"),
+  })
+  return {
+    id,
+    model: (modelID: string | ModelID) => configured.model<ProviderOptions>({ id: modelID }),
+    configure,
+  }
+}
+
+export const provider = configure()
+
+export const model: ProviderPackage.Definition<Settings, ProviderOptions>["model"] = (modelID, settings) =>
+  configure({
+    apiKey: settings.apiKey,
+    baseURL: settings.baseURL,
+    headers: settings.headers === undefined ? undefined : { ...settings.headers },
+    http: settings.body === undefined ? undefined : { body: { ...settings.body } },
+    providerOptions: settings.providerOptions,
+  }).model(modelID)
+
+export * as Mistral from "./mistral.js"
@@ -1,10 +1,12 @@
 import { mergeProviderOptions, type ProviderOptions } from "../schema/index.js"
 import type { OpenAIServiceTier } from "../protocols/utils/openai-options.js"
 import type { Options } from "../protocols/utils/open-responses-options.js"
+import type { ContextManagement } from "../protocols/openai-responses.js"

 export type { OpenAIResponseIncludable, OpenAIServiceTier } from "../protocols/utils/openai-options.js"

 export type OpenAIOptionsInput = Omit<Options, "serviceTier"> & {
+  readonly contextManagement?: ContextManagement
  readonly serviceTier?: OpenAIServiceTier
  readonly [key: string]: unknown
 }
@@ -7,13 +7,14 @@ import * as OpenAICompatibleChat from "../protocols/openai-compatible-chat.js"
 import * as OpenAIChat from "../protocols/openai-chat.js"
 import { OpenResponsesChannel } from "../protocols/open-responses-channel.js"
 import { XAIResponses } from "../protocols/xai-responses.js"
+import { ResponsesCompaction } from "../protocols/utils/responses-compaction.js"
 import { XAIImages } from "../protocols/xai-images.js"
 import type { OpenAIOptionsInput } from "./openai-options.js"
 import type { ProviderPackage } from "../provider-package.js"

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

-export type XAIProviderOptionsInput = OpenAIOptionsInput
+export type XAIProviderOptionsInput = OpenAIOptionsInput & { readonly contextManagement?: never }

 export type LanguageModelOptions = Omit<RouteDefaultsInput, "providerOptions"> &
  ProviderAuthOption<"optional"> & {
@@ -32,6 +33,7 @@ export type { XAIImageOptions } from "../protocols/xai-images.js"
 const RESPONSES_WEBSOCKET_ROTATE_AFTER_MS = 24 * 60 * 1000

 const responsesRoute = Route.make({
+  compact: ResponsesCompaction.execute,
  id: "openai-responses",
  provider: id,
  providerMetadataKey: "xai",
@@ -7,11 +7,13 @@ import { HttpTransport } from "./transport/index.js"
 import type { HttpMiddleware, Transport, TransportRuntime, WebSocketChannelExecutor } from "./transport/index.js"
 import type { Protocol } from "./protocol.js"
 import { applyCachePolicy } from "../cache-policy.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 {
  AIError,
+  CompactionResponse,
  AIErrorReason,
  GenerationOptions,
  HttpOptions,
@@ -34,6 +36,7 @@ export interface RouteBody<Body> {
 }

 export interface Route<Body, Prepared = unknown> {
+  readonly compact?: CompactOperation
  readonly id: string
  readonly provider?: ProviderID
  /** ProviderMetadata namespace emitted and consumed by this route. */
@@ -41,6 +44,8 @@ export interface Route<Body, Prepared = unknown> {
  readonly protocol: ProtocolID
  readonly endpoint: Endpoint.Definition<Body>
  readonly auth: Auth.Definition
+  /** Deployment headers resolved once for every operation, before transport authentication. */
+  readonly headers?: (input: { readonly request: LLMRequest }) => Record<string, string>
  readonly transport: Transport<Body, Prepared, unknown>
  readonly defaults: RouteDefaults
  readonly body: RouteBody<Body>
@@ -149,6 +154,10 @@ export const httpOptions = (input: HttpOptionsInput | undefined) => {
 }

 export interface Interface {
+  readonly compact: (
+    request: LLMRequest,
+    options?: Pick<StreamOptions, "http">,
+  ) => Effect.Effect<CompactionResponse, AIError>
  readonly stream: StreamMethod
  readonly generate: GenerateMethod
 }
@@ -166,24 +175,33 @@ export interface GenerateMethod {
  (request: LLMRequest, options?: StreamOptions): Effect.Effect<LLMResponse, AIError>
 }

+export type CompactOperation = (
+  request: LLMRequest,
+  executor: RequestExecutor.Interface,
+  options?: Pick<StreamOptions, "http">,
+) => Effect.Effect<CompactionResponse, AIError>
+
 export class Service extends Context.Service<Service, Interface>()("@opencode/LLMClient") {}

 const resolveRequestOptions = (request: LLMRequest) => {
-  const routeDefaults = request.model.route.defaults
-  const modelDefaults = request.model.defaults
-  const generation = mergeGenerationOptions(routeDefaults.generation, modelDefaults?.generation, request.generation)
-  return LLMRequest.update(request, {
+  const messages = normalizeToolHistory(request.messages)
+  const normalized = messages === request.messages ? request : LLMRequest.update(request, { messages })
+  const routeDefaults = normalized.model.route.defaults
+  const modelDefaults = normalized.model.defaults
+  const generation = mergeGenerationOptions(routeDefaults.generation, modelDefaults?.generation, normalized.generation)
+  return LLMRequest.update(normalized, {
    generation: generation ?? new GenerationOptions({}),
    providerOptions: mergeProviderOptions(
      routeDefaults.providerOptions,
      modelDefaults?.providerOptions,
-      request.providerOptions,
+      normalized.providerOptions,
    ),
-    http: mergeHttpOptions(routeDefaults.http, modelDefaults?.http, request.http),
+    http: mergeHttpOptions(routeDefaults.http, modelDefaults?.http, normalized.http),
  })
 }

 export interface MakeInput<Body, Frame, Event, State> {
+  readonly compact?: CompactOperation
  /** Route id used in diagnostics and prepared request metadata. */
  readonly id: string
  /** Provider identity for route-owned model construction. */
@@ -205,6 +223,7 @@ export interface MakeInput<Body, Frame, Event, State> {
 }

 export interface MakeTransportInput<Body, Prepared, Frame, Event, State> {
+  readonly compact?: CompactOperation
  /** Route id used in diagnostics and prepared request metadata. */
  readonly id: string
  /** Provider identity for route-owned model construction. */
@@ -280,12 +299,14 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(

  const build = (routeInput: BuiltRouteInput): Route<Body, Prepared> => {
    const route: Route<Body, Prepared> = {
+      compact: routeInput.compact,
      id: routeInput.id,
      provider: routeInput.provider === undefined ? undefined : ProviderID.make(routeInput.provider),
      providerMetadataKey: routeInput.providerMetadataKey,
      protocol: protocol.id,
      endpoint: routeInput.endpoint,
      auth: routeInput.auth ?? Auth.none,
+      headers: routeInput.headers,
      transport: routeInput.transport,
      defaults: routeInput.defaults ?? {},
      body: protocol.body,
@@ -315,7 +336,6 @@ function makeFromTransport<Body, Prepared, Frame, Event, State>(
          endpoint: routeInput.endpoint,
          auth: routeInput.auth ?? Auth.none,
          encodeBody,
-          headers: routeInput.headers,
          middleware: options?.http,
          webSocket: options?.webSocket,
        }),
@@ -432,6 +452,7 @@ export function make<Body, Prepared, Frame, Event, State>(
  if ("transport" in input) return makeFromTransport(input)
  const protocol = input.protocol
  return makeFromTransport({
+    compact: input.compact,
    id: input.id,
    provider: input.provider,
    providerMetadataKey: input.providerMetadataKey,
@@ -444,9 +465,19 @@ export function make<Body, Prepared, Frame, Event, State>(
  })
 }

-const compile = Effect.fn("LLM.compile")(function* (request: LLMRequest, options?: StreamOptions) {
+const prepareRequest = (request: LLMRequest) => {
  const original = applyCachePolicy(resolveRequestOptions(request))
-  const resolved = LLMRequest.update(original, sanitizeSurrogates({ ...LLMRequest.input(original), model: undefined }))
+  const sanitized = LLMRequest.update(original, sanitizeSurrogates({ ...LLMRequest.input(original), model: undefined }))
+  const tools = [...new Map(sanitized.tools.map((tool) => [tool.name, tool])).values()]
+  const resolved = tools.length === sanitized.tools.length ? sanitized : LLMRequest.update(sanitized, { tools })
+  const headers = resolved.model.route.headers?.({ request: resolved })
+  return headers === undefined
+    ? resolved
+    : LLMRequest.update(resolved, { http: mergeHttpOptions(new HttpOptions({ headers }), resolved.http) })
+}
+
+const compile = Effect.fn("LLM.compile")(function* (request: LLMRequest, options?: StreamOptions) {
+  const resolved = prepareRequest(request)
  const route = resolved.model.route

  const body = yield* route.body
@@ -505,6 +536,15 @@ export function generate(request: LLMRequest, options?: StreamOptions): Effect.E
  })
 }

+export const compact = (
+  request: LLMRequest,
+  options?: Pick<StreamOptions, "http">,
+): Effect.Effect<CompactionResponse, AIError, Service> =>
+  Effect.gen(function* () {
+    const client = yield* Service
+    return yield* client.compact(request, options)
+  })
+
 export const streamRequest = (request: LLMRequest, options?: StreamOptions) =>
  Stream.unwrap(
    Effect.gen(function* () {
@@ -515,16 +555,28 @@ export const streamRequest = (request: LLMRequest, options?: StreamOptions) =>
 export const layer: Layer.Layer<Service, never, RequestExecutor.Service> = Layer.effect(
  Service,
  Effect.gen(function* () {
-    const stream = streamRequestWith({
-      http: yield* RequestExecutor.Service,
+    const executor = yield* RequestExecutor.Service
+    const stream = streamRequestWith({ http: executor })
+    return Service.of({
+      stream,
+      generate: generateWith(stream),
+      compact: (request, options) =>
+        Effect.suspend(() => {
+          const operation = request.model.route.compact
+          if (!operation)
+            return ProviderShared.invalidRequest(
+              `${request.model.provider}/${request.model.route.id} does not support explicit compaction`,
+            )
+          return operation(prepareRequest(request), executor, options)
+        }),
    })
-    return Service.of({ stream, generate: generateWith(stream) })
  }),
 )

 export const Route = { make } as const

 export const LLMClient = {
+  compact,
  Service,
  layer,
  stream,
@@ -3,6 +3,7 @@ import { LLM } from "@opencode-ai/schema/llm"
 import { ContentBlockID, ToolCallID } from "./ids.js"
 import {
  Message,
+  CompactionPart,
  ProviderMetadata,
  ToolCallPart,
  ToolOutput,
@@ -62,6 +63,8 @@ export { ProviderMetadata } from "./messages.js"
 * Matches the same escape-hatch field on `LLMEvent`.
 */
 export class Usage extends Schema.Class<Usage>("AI.Usage")({
+  /** Effective input size of the final message iteration, when reported; not billed totals. */
+  contextTokens: Schema.optional(Schema.Number),
  inputTokens: Schema.optional(Schema.Number),
  outputTokens: Schema.optional(Schema.Number),
  nonCachedInputTokens: Schema.optional(Schema.Number),
@@ -72,7 +75,7 @@ export class Usage extends Schema.Class<Usage>("AI.Usage")({
  providerMetadata: Schema.optional(ProviderMetadata),
 }) {
  /**
-   * Visible output tokens — `outputTokens` minus `reasoningTokens`, clamped
+   * Non-reasoning output tokens (including compaction summaries) — `outputTokens` minus `reasoningTokens`, clamped
   * to zero. The one place subtraction happens in this contract; the clamp
   * means a provider reporting `reasoningTokens > outputTokens` produces a
   * harmless zero rather than a negative that crashes downstream schemas.
@@ -88,6 +91,12 @@ export class Usage extends Schema.Class<Usage>("AI.Usage")({

 export type UsageInput = Usage | ConstructorParameters<typeof Usage>[0]

+/** A replacement context window. Replace prior history with these messages. */
+export class CompactionResponse extends Schema.Class<CompactionResponse>("LLM.CompactionResponse")({
+  messages: Schema.Array(Message),
+  usage: Schema.optional(Usage),
+}) {}
+
 export const StepStart = Schema.Struct({
  type: Schema.tag("step-start"),
  index: Schema.Number,
@@ -241,6 +250,7 @@ export const ProviderErrorEvent = Schema.Struct({
 export type ProviderErrorEvent = Schema.Schema.Type<typeof ProviderErrorEvent>

 const llmEventTagged = Schema.Union([
+  CompactionPart,
  StepStart,
  TextStart,
  TextDelta,
@@ -274,6 +284,7 @@ const toolCallID = (value: ToolCallID | string) => ToolCallID.make(value)
 * `events.filter(LLMEvent.guards["tool-call"])`.
 */
 export const LLMEvent = Object.assign(llmEventTagged, {
+  compaction: CompactionPart.make,
  stepStart: StepStart.make,
  textStart: (input: WithID<TextStart, ContentBlockID>) => TextStart.make({ ...input, id: contentBlockID(input.id) }),
  textDelta: (input: WithID<TextDelta, ContentBlockID>) => TextDelta.make({ ...input, id: contentBlockID(input.id) }),
@@ -311,6 +322,7 @@ export const LLMEvent = Object.assign(llmEventTagged, {
    }),
  providerError: ProviderErrorEvent.make,
  is: {
+    compaction: llmEventTagged.guards.compaction,
    stepStart: llmEventTagged.guards["step-start"],
    textStart: llmEventTagged.guards["text-start"],
    textDelta: llmEventTagged.guards["text-delta"],
@@ -333,10 +345,10 @@ export const LLMEvent = Object.assign(llmEventTagged, {
 export type LLMEvent = Schema.Schema.Type<typeof llmEventTagged>

 /** Joins deltas per fragment, letting an authoritative end value replace that fragment's accumulated deltas. */
-const joinFragments = <Delta extends { id: string; text: string }, End extends { id: string; text?: string }>(
+const joinFragments = (
  events: ReadonlyArray<LLMEvent>,
-  isDelta: (event: LLMEvent) => event is Extract<LLMEvent, Delta>,
-  isEnd: (event: LLMEvent) => event is Extract<LLMEvent, End>,
+  isDelta: (event: LLMEvent) => event is LLMEvent & { id: string; text: string },
+  isEnd: (event: LLMEvent) => event is LLMEvent & { id: string; text?: string },
 ) => {
  const order: string[] = []
  const parts = new Map<string, string>()
@@ -563,6 +575,8 @@ const reduceToolCall = (state: ResponseState, event: ToolCall): ResponseState =>
 const reduceResponseState = (state: ResponseState, event: LLMEvent): ResponseState => {
  const next = appendEvent(state, event)
  switch (event.type) {
+    case "compaction":
+      return appendContent(next, event)
    case "text-start":
      return ensureText(next, event.id, event.providerMetadata)
    case "text-delta":
@@ -9,7 +9,7 @@ import {
  LanguageModelSchema,
  ProviderOptions,
 } from "./options.js"
-import { isRecord } from "../utils/record.js"
+import { ProviderID } from "./ids.js"

 export const MessageRole = Schema.Literals(["system", "user", "assistant", "tool"])
 export type MessageRole = Schema.Schema.Type<typeof MessageRole>
@@ -53,14 +53,10 @@ export const MediaPart = Schema.Struct({
  filename: Schema.optional(Schema.String),
  cache: Schema.optional(CacheHint),
  metadata: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
+  providerMetadata: Schema.optional(ProviderMetadata),
 }).annotate({ identifier: "LLM.Content.Media" })
 export type MediaPart = Schema.Schema.Type<typeof MediaPart>

-const isToolResultValue = (value: unknown): value is ToolResultValue =>
-  isRecord(value) &&
-  (value.type === "text" || value.type === "json" || value.type === "error" || value.type === "content") &&
-  "value" in value
-
 const toolResultValueSchema = Schema.Union([
  Schema.Struct({
    type: Schema.Literal("json"),
@@ -80,6 +76,7 @@ const toolResultValueSchema = Schema.Union([
  }),
 ]).annotate({ identifier: "LLM.ToolResult" })
 export type ToolResultValue = Schema.Schema.Type<typeof toolResultValueSchema>
+const isToolResultValue = Schema.is(toolResultValueSchema)

 export const ToolResultValue = Object.assign(toolResultValueSchema, {
  is: isToolResultValue,
@@ -190,9 +187,40 @@ export const ReasoningPart = Schema.Struct({
 }).annotate({ identifier: "LLM.Content.Reasoning" })
 export type ReasoningPart = Schema.Schema.Type<typeof ReasoningPart>

-export const ContentPart = Schema.Union([TextPart, MediaPart, ToolCallPart, ToolResultPart, ReasoningPart]).pipe(
-  Schema.toTaggedUnion("type"),
-)
+/** A provider-generated context checkpoint, distinct from visible assistant text. */
+type CompactionContent =
+  | { readonly encrypted: string; readonly text?: never }
+  | { readonly text: string | null; readonly encrypted?: never }
+
+const compactionPartSchema = Schema.Struct({
+  type: Schema.Literal("compaction"),
+  provider: ProviderID,
+  id: Schema.optional(Schema.String),
+  encrypted: Schema.optional(Schema.String),
+  /** Null means the provider failed to produce a summary; prior history must be retained. */
+  text: Schema.optional(Schema.NullOr(Schema.String)),
+})
+  .pipe(
+    Schema.refine(
+      (part): part is typeof part & CompactionContent => (part.encrypted !== undefined) !== (part.text !== undefined),
+      { message: "Compaction requires either encrypted content or a summary" },
+    ),
+  )
+  .annotate({ identifier: "LLM.Content.Compaction" })
+export type CompactionPart = typeof compactionPartSchema.Type
+export const CompactionPart = Object.assign(compactionPartSchema, {
+  make: (input: Omit<CompactionPart, "type" | "encrypted" | "text"> & CompactionContent): CompactionPart =>
+    Schema.decodeUnknownSync(compactionPartSchema)({ type: "compaction", ...input }),
+})
+
+export const ContentPart = Schema.Union([
+  TextPart,
+  MediaPart,
+  ToolCallPart,
+  ToolResultPart,
+  ReasoningPart,
+  CompactionPart,
+]).pipe(Schema.toTaggedUnion("type"))
 export type ContentPart = Schema.Schema.Type<typeof ContentPart>

 export class Message extends Schema.Class<Message>("LLM.Message")({
@@ -200,6 +228,7 @@ export class Message extends Schema.Class<Message>("LLM.Message")({
  role: MessageRole,
  content: Schema.Array(ContentPart),
  metadata: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
+  providerMetadata: Schema.optional(ProviderMetadata),
  native: Schema.optional(Schema.Record(Schema.String, Schema.Unknown)),
 }) {}

@@ -4,6 +4,7 @@ import { LLMClient } from "./route/client.js"
 import {
  LLMEvent,
  LLMResponse,
+  CompactionResponse,
  type FinishReasonDetails,
  type AIError,
  type LLMRequest,
@@ -133,6 +134,16 @@ const make = (options: LayerOptions) =>
        }
      })
    const test = Test.of({
+      compact: (request) =>
+        stream(request).pipe(
+          Stream.runFold(LLMResponse.empty, LLMResponse.reduce),
+          Effect.flatMap((state) => {
+            const response = LLMResponse.complete(state)
+            if (!response?.message.content.some((part) => part.type === "compaction"))
+              return Effect.die("TestLLM compaction response must contain a checkpoint and terminal finish event")
+            return Effect.succeed(new CompactionResponse({ messages: [response.message], usage: response.usage }))
+          }),
+        ),
      stream,
      generate: (request) =>
        stream(request).pipe(
@@ -0,0 +1,75 @@
+import { Message, ToolResultPart, type ToolCallPart } from "./schema/messages.js"
+
+const EMPTY_TOOL_OUTPUT = "(no tool output)"
+const MISSING_TOOL_RESULT = "Tool result missing"
+
+export function normalizeToolHistory(messages: ReadonlyArray<Message>) {
+  const normalized: Message[] = []
+  const pending = new Map<string, ToolCallPart>()
+  const appendMissingResults = () => {
+    if (pending.size === 0) return
+    normalized.push(missingToolResults(pending.values()))
+    pending.clear()
+  }
+
+  for (const message of messages) {
+    if (message.role === "user" || message.role === "assistant") appendMissingResults()
+
+    if (message.role === "tool") {
+      const tool = normalizeToolMessage(message, pending)
+      if (tool) normalized.push(tool)
+      continue
+    }
+
+    normalized.push(message)
+    if (message.role !== "assistant") continue
+    for (const part of message.content) {
+      if (part.type === "tool-call" && part.providerExecuted !== true) pending.set(part.id, part)
+    }
+  }
+
+  return normalized.length === messages.length && normalized.every((message, index) => message === messages[index])
+    ? messages
+    : normalized
+}
+
+function missingToolResults(calls: Iterable<ToolCallPart>) {
+  return new Message({
+    role: "tool",
+    content: [...calls].map((call) =>
+      ToolResultPart.make({ id: call.id, name: call.name, result: MISSING_TOOL_RESULT, resultType: "error" }),
+    ),
+  })
+}
+
+function normalizeToolMessage(message: Message, pending: Map<string, ToolCallPart>): Message | undefined {
+  const content = message.content.map((part) => {
+    if (part.type !== "tool-result" || part.providerExecuted === true) return part
+    const call = pending.get(part.id)
+    if (call) pending.delete(part.id)
+    return normalizeToolResult(part, call?.name ?? part.name)
+  })
+  if (content.length === 0) return undefined
+  if (content.every((part, index) => part === message.content[index])) return message
+  return new Message({
+    id: message.id,
+    role: message.role,
+    content,
+    metadata: message.metadata,
+    providerMetadata: message.providerMetadata,
+    native: message.native,
+  })
+}
+
+function normalizeToolResult(part: ToolResultPart, name: string): ToolResultPart {
+  const named = part.name === name ? part : { ...part, name }
+  if (named.result.type === "text" && named.result.value === "")
+    return { ...named, result: { type: "text", value: EMPTY_TOOL_OUTPUT } }
+  if (named.result.type === "error" && named.result.value === "")
+    return { ...named, result: { type: "error", value: EMPTY_TOOL_OUTPUT } }
+  if (named.result.type !== "content") return named
+  const value = named.result.value.filter((item) => item.type !== "text" || item.text !== "")
+  if (value.length === 0) return { ...named, result: { type: "text", value: EMPTY_TOOL_OUTPUT } }
+  if (value.length === named.result.value.length) return named
+  return { ...named, result: { type: "content", value } }
+}
@@ -126,6 +126,38 @@ describe("applyCachePolicy", () => {
    }),
  )

+  for (const fixture of [
+    { name: "default", cache: undefined, control: { type: "ephemeral" } },
+    { name: "auto", cache: "auto", control: { type: "ephemeral" } },
+    {
+      name: "explicit one-hour",
+      cache: { tools: true, system: true, messages: { tail: 1 }, ttlSeconds: 3600 },
+      control: { type: "ephemeral", ttl: "1h" },
+    },
+    { name: "disabled", cache: "none", control: undefined },
+  ] as const) {
+    it.effect(`Bedrock Messages respects ${fixture.name} caching`, () =>
+      Effect.gen(function* () {
+        const prepared = yield* compileRequest(
+          LLM.request({
+            model: AmazonBedrock.configure({ apiKey: "test" }).messages("anthropic.claude-opus-4-6-v1"),
+            system: "Stable instructions",
+            tools: [
+              { name: "lookup", description: "Look up a value", inputSchema: { type: "object", properties: {} } },
+            ],
+            prompt: "hello",
+            cache: fixture.cache,
+          }),
+        )
+        expect(prepared.body).toMatchObject({
+          tools: [{ name: "lookup", cache_control: fixture.control }],
+          system: [{ type: "text", text: "Stable instructions", cache_control: fixture.control }],
+          messages: [{ role: "user", content: [{ type: "text", text: "hello", cache_control: fixture.control }] }],
+        })
+      }),
+    )
+  }
+
  it.effect("'auto' is a no-op on Gemini (out-of-band caching protocol)", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
@@ -0,0 +1,97 @@
+import { expect } from "bun:test"
+import { Effect } from "effect"
+import { FetchHttpClient } from "effect/unstable/http"
+import { LLM, LLMRequest, Message } from "../src/index.js"
+import { LLMClient } from "../src/route/client.js"
+import { OpenAI } from "../src/providers.js"
+import { testEffect } from "./lib/effect.js"
+import { runtimeLayer } from "./lib/http.js"
+import { sseEvents } from "./lib/sse.js"
+
+testEffect(runtimeLayer(FetchHttpClient.layer)).live("compaction and a tool loop work end to end over HTTP", () =>
+  Effect.gen(function* () {
+    const checkpoint = { type: "compaction", id: "cmp_local", encrypted_content: "opaque-local-state" }
+    const calls: string[] = []
+    const server = yield* Effect.acquireRelease(
+      Effect.sync(() =>
+        Bun.serve({
+          hostname: "127.0.0.1",
+          port: 0,
+          async fetch(request) {
+            const path = new URL(request.url).pathname
+            calls.push(path)
+            const body = await request.json()
+            expect(request.headers.get("authorization")).toBe("Bearer fixture")
+            if (path === "/v1/responses/compact") {
+              expect(body.stream).toBeUndefined()
+              return Response.json({
+                object: "response.compaction",
+                output: [checkpoint],
+                usage: { input_tokens: 100, output_tokens: 10, total_tokens: 110 },
+              })
+            }
+            expect(body.input[0]).toEqual(checkpoint)
+            expect(body.stream).toBe(true)
+            if (calls.length === 2)
+              return new Response(
+                sseEvents(
+                  {
+                    type: "response.output_item.done",
+                    item: { type: "function_call", id: "fc_1", call_id: "call_1", name: "lookup", arguments: "{}" },
+                  },
+                  { type: "response.completed", response: { id: "resp_1" } },
+                ),
+                { headers: { "content-type": "text/event-stream" } },
+              )
+            expect(body.input.at(-2)).toMatchObject({ type: "function_call", call_id: "call_1" })
+            expect(body.input.at(-1)).toEqual({ type: "function_call_output", call_id: "call_1", output: "42" })
+            const output = sseEvents(
+              { type: "response.output_item.added", item: { type: "message", id: "msg_1" } },
+              { type: "response.output_text.delta", item_id: "msg_1", delta: "The answer is 42." },
+              {
+                type: "response.output_item.done",
+                item: { type: "message", id: "msg_1", content: [{ type: "output_text", text: "The answer is 42." }] },
+              },
+              { type: "response.completed", response: { id: "resp_2" } },
+            )
+            return new Response(
+              new ReadableStream({
+                start(controller) {
+                  controller.enqueue(new TextEncoder().encode(output.slice(0, 37)))
+                  controller.enqueue(new TextEncoder().encode(output.slice(37)))
+                  controller.close()
+                },
+              }),
+              { headers: { "content-type": "text/event-stream" } },
+            )
+          },
+        }),
+      ),
+      (server) => Effect.sync(() => server.stop(true)),
+    )
+    const model = OpenAI.configure({ apiKey: "fixture", baseURL: `http://127.0.0.1:${server.port}/v1` }).responses(
+      "fixture",
+    )
+    const request = LLM.request({
+      model,
+      prompt: "original",
+      tools: [{ name: "lookup", description: "Lookup a number", inputSchema: { type: "object", properties: {} } }],
+    })
+    const compacted = yield* LLMClient.compact(request)
+    const messages = [...compacted.messages, Message.user("Look up the answer")]
+    const first = yield* LLMClient.generate(LLMRequest.update(request, { messages }))
+    expect(first.toolCalls).toHaveLength(1)
+    const call = first.toolCalls[0]!
+    const last = yield* LLMClient.generate(
+      LLMRequest.update(request, {
+        messages: [
+          ...messages,
+          first.message,
+          Message.tool({ id: call.id, name: call.name, result: "42", resultType: "text" }),
+        ],
+      }),
+    )
+    expect(last.text).toBe("The answer is 42.")
+    expect(calls).toEqual(["/v1/responses/compact", "/v1/responses", "/v1/responses"])
+  }),
+)
@@ -0,0 +1,49 @@
+import { expect, test } from "bun:test"
+import { Schema } from "effect"
+import { CompactionPart, LLMEvent, LLMResponse, Message, ProviderID } from "../src/schema/index.js"
+
+test("compaction survives event assembly and message serialization without becoming text", () => {
+  const part = CompactionPart.make({
+    provider: ProviderID.make("openai"),
+    id: "cmp_1",
+    encrypted: "opaque",
+  })
+  const response = LLMResponse.fromEvents([
+    LLMEvent.textStart({ id: "before" }),
+    LLMEvent.textDelta({ id: "before", text: "Before" }),
+    LLMEvent.textEnd({ id: "before" }),
+    part,
+    LLMEvent.textStart({ id: "after" }),
+    LLMEvent.textDelta({ id: "after", text: "After" }),
+    LLMEvent.textEnd({ id: "after" }),
+    LLMEvent.finish({ reason: { normalized: "stop" } }),
+  ])!
+  expect(response.message.content.map((part) => part.type)).toEqual(["text", "compaction", "text"])
+  expect(response.text).toBe("BeforeAfter")
+  expect(response.reasoning).toBe("")
+  expect(response.events.filter(LLMEvent.is.compaction)).toEqual([part])
+  const codec = Schema.fromJsonString(Message)
+  expect(Schema.decodeSync(codec)(Schema.encodeSync(codec)(response.message))).toEqual(response.message)
+})
+
+test("compaction requires exactly one typed representation", () => {
+  const provider = ProviderID.make("anthropic")
+  expect(CompactionPart.make({ provider, text: null })).toEqual({ type: "compaction", provider, text: null })
+  const decode = Schema.decodeUnknownSync(CompactionPart)
+  expect(() => decode({ type: "compaction", provider })).toThrow()
+  expect(() => decode({ type: "compaction", provider, text: "summary", encrypted: "opaque" })).toThrow()
+})
+
+test("tagged content and event guards accept both checkpoint representations", () => {
+  for (const part of [
+    CompactionPart.make({ provider: ProviderID.make("openai"), encrypted: "opaque" }),
+    CompactionPart.make({ provider: ProviderID.make("anthropic"), text: "summary" }),
+    CompactionPart.make({ provider: ProviderID.make("anthropic"), text: null }),
+  ]) {
+    expect(LLMEvent.is.compaction(part)).toBe(true)
+    expect(LLMEvent.guards.compaction(part)).toBe(true)
+    const codec = Schema.fromJsonString(Message)
+    const message = Message.assistant(part)
+    expect(Schema.decodeSync(codec)(Schema.encodeSync(codec)(message))).toEqual(message)
+  }
+})
@@ -1,7 +1,7 @@
 import { describe, expect, test } from "bun:test"
 import { Effect, Ref, Schema } from "effect"
 import { HttpClientRequest, HttpClientResponse } from "effect/unstable/http"
-import { LLM, Message, ToolCallPart, mergeProviderOptions } from "../src/index.js"
+import { LLM, LLMRequest, Message, ToolCallPart, ToolDefinition, mergeProviderOptions } from "../src/index.js"
 import { AnthropicMessages, OpenAIChat } from "../src/protocols.js"
 import { Auth, LLMClient } from "../src/route.js"
 import { compileRequest } from "../src/route/client.js"
@@ -77,6 +77,55 @@ describe("request option precedence", () => {
    }),
  )

+  it.effect("keeps the last tool definition for duplicate names", () =>
+    Effect.gen(function* () {
+      const request = LLM.request({
+        model: OpenAIChat.route.model({ id: "gpt-4o-mini" }),
+        prompt: "Use a tool.",
+      })
+      const prepared = yield* compileRequest(
+        LLMRequest.update(request, {
+          tools: [
+            ToolDefinition.make({ name: "lookup", description: "old", inputSchema: { type: "object" } }),
+            ToolDefinition.make({ name: "search", description: "search", inputSchema: { type: "object" } }),
+            ToolDefinition.make({ name: "lookup", description: "new", inputSchema: { type: "object" } }),
+          ],
+        }),
+      )
+
+      expect(prepared.body.tools).toEqual([
+        {
+          type: "function",
+          function: { name: "lookup", description: "new", parameters: { type: "object" }, strict: false },
+        },
+        {
+          type: "function",
+          function: { name: "search", description: "search", parameters: { type: "object" }, strict: false },
+        },
+      ])
+    }),
+  )
+
+  it.effect("normalizes tool history before protocol lowering", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model: OpenAIChat.route.model({ id: "gpt-4o-mini" }),
+          messages: [
+            Message.assistant(ToolCallPart.make({ id: "call_1", name: "lookup", input: {} })),
+            Message.user("Continue."),
+          ],
+        }),
+      )
+
+      expect(prepared.body.messages).toMatchObject([
+        { role: "assistant", tool_calls: [{ id: "call_1", function: { name: "lookup" } }] },
+        { role: "tool", tool_call_id: "call_1", content: "Tool result missing" },
+        { role: "user", content: "Continue." },
+      ])
+    }),
+  )
+
  it.effect("applies model HTTP defaults before request HTTP overlays", () =>
    LLMClient.generate(
      LLM.request({
@@ -0,0 +1,36 @@
+{
+  "version": 1,
+  "metadata": {
+    "model": "zai-glm-5-2",
+    "tags": [
+      "prefix:mistral-chat-glm",
+      "provider:mistral",
+      "protocol:mistral-chat",
+      "hosted-model",
+      "tool",
+      "tool-call"
+    ],
+    "name": "mistral-chat-glm/streams-an-indexed-tool-call",
+    "recordedAt": "2026-08-30T17:38:02.921Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.mistral.ai/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"zai-glm-5-2\",\"messages\":[{\"role\":\"system\",\"content\":\"Call lookup_weather exactly once with Paris.\"},{\"role\":\"user\",\"content\":\"What is the weather?\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"lookup_weather\",\"description\":\"Look up the current weather for a city\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\",\"enum\":[\"Paris\"]}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"tool_choice\":{\"type\":\"function\",\"function\":{\"name\":\"lookup_weather\"}},\"stream\":true,\"max_tokens\":256,\"temperature\":0}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"id\":\"f139bf0e4b984e51aabf6a83c237674d\",\"object\":\"chat.completion.chunk\",\"created\":1788111482,\"model\":\"zai-glm-5-2\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\"},\"finish_reason\":null}]}\n\ndata: {\"id\":\"f139bf0e4b984e51aabf6a83c237674d\",\"object\":\"chat.completion.chunk\",\"created\":1788111482,\"model\":\"zai-glm-5-2\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"index\":0,\"content\":\"\"},\"finish_reason\":null,\"logprobs\":null}]}\n\ndata: {\"id\":\"f139bf0e4b984e51aabf6a83c237674d\",\"object\":\"chat.completion.chunk\",\"created\":1788111482,\"model\":\"zai-glm-5-2\",\"choices\":[{\"index\":0,\"delta\":{\"tool_calls\":[{\"id\":\"chatcmpl-tool-8cc4d8f9f07b298a\",\"type\":\"function\",\"function\":{\"name\":\"lookup_weather\",\"arguments\":\"{\\\"city\\\": \\\"\"},\"index\":0}],\"index\":0,\"content\":\"\"},\"finish_reason\":null,\"logprobs\":null}]}\n\ndata: {\"id\":\"f139bf0e4b984e51aabf6a83c237674d\",\"object\":\"chat.completion.chunk\",\"created\":1788111482,\"model\":\"zai-glm-5-2\",\"choices\":[{\"index\":0,\"delta\":{\"tool_calls\":[{\"type\":\"function\",\"function\":{\"name\":\"\",\"arguments\":\"Paris\\\"}\"},\"index\":0}],\"index\":0,\"content\":\"\"},\"finish_reason\":null,\"logprobs\":null}]}\n\ndata: {\"id\":\"f139bf0e4b984e51aabf6a83c237674d\",\"object\":\"chat.completion.chunk\",\"created\":1788111482,\"model\":\"zai-glm-5-2\",\"choices\":[{\"index\":0,\"delta\":{\"index\":0,\"content\":\"\"},\"finish_reason\":\"stop\",\"logprobs\":null}],\"usage\":{\"prompt_tokens\":171,\"total_tokens\":182,\"completion_tokens\":11,\"prompt_tokens_details\":{\"cached_tokens\":0}}}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,47 @@
+{
+  "version": 1,
+  "metadata": {
+    "model": "mistral-small-latest",
+    "tags": ["prefix:mistral-chat", "provider:mistral", "protocol:mistral-chat", "tool", "tool-loop", "usage"],
+    "name": "mistral-chat/drives-a-tool-loop",
+    "recordedAt": "2026-08-30T17:18:49.552Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.mistral.ai/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"mistral-small-latest\",\"messages\":[{\"role\":\"system\",\"content\":\"Call lookup_weather exactly once with Paris.\"},{\"role\":\"user\",\"content\":\"What is the weather?\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"lookup_weather\",\"description\":\"Look up the current weather for a city\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\",\"enum\":[\"Paris\"]}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"tool_choice\":{\"type\":\"function\",\"function\":{\"name\":\"lookup_weather\"}},\"stream\":true,\"max_tokens\":160,\"temperature\":0,\"reasoning_effort\":\"none\"}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"id\":\"07491e37a5ed48f9987f1583753a466b\",\"object\":\"chat.completion.chunk\",\"created\":1788110328,\"model\":\"mistral-small-latest\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\"},\"finish_reason\":null}]}\n\ndata: {\"id\":\"07491e37a5ed48f9987f1583753a466b\",\"object\":\"chat.completion.chunk\",\"created\":1788110328,\"model\":\"mistral-small-latest\",\"choices\":[{\"index\":0,\"delta\":{\"tool_calls\":[{\"id\":\"ffJovBNqY\",\"type\":\"function\",\"function\":{\"name\":\"lookup_weather\",\"arguments\":\"{\\\"city\\\": \\\"Paris\\\"}\"},\"index\":0}]},\"finish_reason\":\"tool_calls\"}],\"usage\":{\"prompt_tokens\":110,\"total_tokens\":122,\"completion_tokens\":12,\"prompt_tokens_details\":{\"cached_tokens\":0},\"service_tier\":\"standard\"},\"p\":\"abcdefghijklm\"}\n\ndata: [DONE]\n\n"
+      }
+    },
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.mistral.ai/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"mistral-small-latest\",\"messages\":[{\"role\":\"system\",\"content\":\"Call lookup_weather exactly once with Paris.\"},{\"role\":\"user\",\"content\":\"What is the weather?\"},{\"role\":\"assistant\",\"content\":\"\",\"tool_calls\":[{\"id\":\"ffJovBNqY\",\"type\":\"function\",\"function\":{\"name\":\"lookup_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"}}]},{\"role\":\"tool\",\"tool_call_id\":\"ffJovBNqY\",\"name\":\"lookup_weather\",\"content\":\"{\\\"condition\\\":\\\"sunny\\\",\\\"temperature\\\":\\\"18C\\\"}\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"lookup_weather\",\"description\":\"Look up the current weather for a city\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\",\"enum\":[\"Paris\"]}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"tool_choice\":\"none\",\"stream\":true,\"max_tokens\":160,\"temperature\":0,\"reasoning_effort\":\"none\"}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"id\":\"8fcd293093b849139fc0893a48bbc7ce\",\"object\":\"chat.completion.chunk\",\"created\":1788110328,\"model\":\"mistral-small-latest\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\"},\"finish_reason\":null}]}\n\ndata: {\"id\":\"8fcd293093b849139fc0893a48bbc7ce\",\"object\":\"chat.completion.chunk\",\"created\":1788110328,\"model\":\"mistral-small-latest\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"The\"},\"finish_reason\":null}],\"p\":\"abcdefghijklmn\"}\n\ndata: {\"id\":\"8fcd293093b849139fc0893a48bbc7ce\",\"object\":\"chat.completion.chunk\",\"created\":1788110328,\"model\":\"mistral-small-latest\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\" weather in Paris is\"},\"finish_reason\":null}],\"p\":\"abcdefghijklmn\"}\n\ndata: {\"id\":\"8fcd293093b849139fc0893a48bbc7ce\",\"object\":\"chat.completion.chunk\",\"created\":1788110328,\"model\":\"mistral-small-latest\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\" currently sunny with\"},\"finish_reason\":null}],\"p\":\"abcdefghijklmnopqrstu\"}\n\ndata: {\"id\":\"8fcd293093b849139fc0893a48bbc7ce\",\"object\":\"chat.completion.chunk\",\"created\":1788110328,\"model\":\"mistral-small-latest\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\" a temperature of \"},\"finish_reason\":null}],\"p\":\"abcdef\"}\n\ndata: {\"id\":\"8fcd293093b849139fc0893a48bbc7ce\",\"object\":\"chat.completion.chunk\",\"created\":1788110328,\"model\":\"mistral-small-latest\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"18°C\"},\"finish_reason\":null}],\"p\":\"abcdefghijklmnopqr\"}\n\ndata: {\"id\":\"8fcd293093b849139fc0893a48bbc7ce\",\"object\":\"chat.completion.chunk\",\"created\":1788110328,\"model\":\"mistral-small-latest\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\".\"},\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":57,\"total_tokens\":74,\"completion_tokens\":17,\"prompt_tokens_details\":{\"cached_tokens\":0},\"service_tier\":\"standard\"},\"p\":\"abcdefghijklmnopqrstuvwxyz\"}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -0,0 +1,29 @@
+{
+  "version": 1,
+  "metadata": {
+    "model": "mistral-small-latest",
+    "tags": ["prefix:mistral-chat", "provider:mistral", "protocol:mistral-chat", "text", "usage"],
+    "name": "mistral-chat/streams-text-with-usage",
+    "recordedAt": "2026-08-30T17:18:45.432Z"
+  },
+  "interactions": [
+    {
+      "transport": "http",
+      "request": {
+        "method": "POST",
+        "url": "https://api.mistral.ai/v1/chat/completions",
+        "headers": {
+          "content-type": "application/json"
+        },
+        "body": "{\"model\":\"mistral-small-latest\",\"messages\":[{\"role\":\"user\",\"content\":\"Reply with exactly one word: hello\"}],\"stream\":true,\"max_tokens\":40,\"temperature\":0,\"reasoning_effort\":\"none\"}"
+      },
+      "response": {
+        "status": 200,
+        "headers": {
+          "content-type": "text/event-stream; charset=utf-8"
+        },
+        "body": "data: {\"id\":\"9a4d16bdddb74e5e89c2cf9e9b91e065\",\"object\":\"chat.completion.chunk\",\"created\":1788110325,\"model\":\"mistral-small-latest\",\"choices\":[{\"index\":0,\"delta\":{\"role\":\"assistant\",\"content\":\"\"},\"finish_reason\":null}]}\n\ndata: {\"id\":\"9a4d16bdddb74e5e89c2cf9e9b91e065\",\"object\":\"chat.completion.chunk\",\"created\":1788110325,\"model\":\"mistral-small-latest\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"Hi\"},\"finish_reason\":null}],\"p\":\"abcdefghijklmnopqrs\"}\n\ndata: {\"id\":\"9a4d16bdddb74e5e89c2cf9e9b91e065\",\"object\":\"chat.completion.chunk\",\"created\":1788110325,\"model\":\"mistral-small-latest\",\"choices\":[{\"index\":0,\"delta\":{\"content\":\"\"},\"finish_reason\":\"stop\"}],\"usage\":{\"prompt_tokens\":22,\"total_tokens\":24,\"completion_tokens\":2,\"prompt_tokens_details\":{\"cached_tokens\":0},\"service_tier\":\"standard\"},\"p\":\"abcdefghijklmnopqrstuvwxyz0\"}\n\ndata: [DONE]\n\n"
+      }
+    }
+  ]
+}
@@ -81,6 +81,46 @@ describe("provider error classification", () => {
    ).toEqual(["ProviderInternal", "ProviderInternal", "ProviderInternal"])
  })

+  test("classifies retryable server messages as provider internal", () => {
+    const message =
+      "The model is currently at capacity due to high demand. Please try again in a few minutes, or use a higher service tier for priority processing."
+
+    expect(
+      [
+        message,
+        "Try again",
+        "Please retry your request shortly.",
+        "You can retry the request.",
+        "Try your request again.",
+        "The service is temporarily at capacity.",
+        "The model is overloaded.",
+        "Service unavailable",
+        "Internal server error",
+        "The server is busy.",
+        "Provider returned error",
+        "Provider returned an error",
+        "ResourceExhausted",
+        "Upstream connection failed",
+        "Exceeded request buffer limit while retrying upstream",
+      ].map((message) => classifyProviderFailure({ message })._tag),
+    ).toEqual(Array(15).fill("ProviderInternal"))
+    expect(
+      classifyProviderFailure({ message: "Provider request failed", rawBody: "Please try again later." })._tag,
+    ).toBe("ProviderInternal")
+  })
+
+  test("prioritizes specific failures over retryable server text", () => {
+    expect(
+      [
+        classifyProviderFailure({ message: "Invalid credentials, try again", status: 401 }),
+        classifyProviderFailure({ message: "Quota exceeded, try again", status: 429 }),
+        classifyProviderFailure({ message: "Rate limit exceeded, try again" }),
+        classifyProviderFailure({ message: "Upstream request failed: validation failed", status: 400 }),
+        classifyProviderFailure({ message: "Try again", status: 200 }),
+      ].map((failure) => failure._tag),
+    ).toEqual(["Authentication", "QuotaExceeded", "RateLimit", "InvalidRequest", "ProviderInternal"])
+  })
+
  test("classifies transient client statuses as provider internal", () => {
    expect([408, 409].map((status) => classifyProviderFailure({ message: `HTTP ${status}`, status })._tag)).toEqual([
      "ProviderInternal",
@@ -111,7 +151,6 @@ describe("provider error classification", () => {
    expect(classifyProviderFailure({ message: '{"type":"error","error":{"code":123}}' })._tag).toBe("UnknownProvider")
    expect(classifyProviderFailure({ message: "not-json" })._tag).toBe("UnknownProvider")
    expect(classifyProviderFailure({ message: "network error" })._tag).toBe("UnknownProvider")
-    expect(classifyProviderFailure({ message: "Provider returned error" })._tag).toBe("UnknownProvider")
  })
 })

@@ -0,0 +1,69 @@
+import { Effect } from "effect"
+import { CompactionPart, LLM, LLMClient, LLMEvent, Message, ProviderID } from "../../src/index.js"
+import { OpenAI, Anthropic, AmazonBedrock } from "../../src/providers.js"
+
+const openai = OpenAI.configure({
+  apiKey: "test",
+  providerOptions: { contextManagement: [{ type: "compaction", compactThreshold: 100000 }] },
+}).responses("gpt-5.3-codex")
+LLMClient.compact(LLM.request({ model: openai, prompt: "hello" }))
+
+const checkpoint = CompactionPart.make({ provider: ProviderID.make("openai"), id: "cmp_1", encrypted: "opaque" })
+const provider = ProviderID.make("anthropic")
+CompactionPart.make({ provider, text: "summary" })
+CompactionPart.make({ provider, text: null })
+// @ts-expect-error A checkpoint must have a representation.
+CompactionPart.make({ provider })
+// @ts-expect-error Encrypted and summary representations are mutually exclusive.
+CompactionPart.make({ provider, encrypted: "opaque", text: "summary" })
+// @ts-expect-error A failed summary cannot also carry encrypted content.
+LLMEvent.compaction({ provider, encrypted: "opaque", text: null })
+// @ts-expect-error The canonical message type also enforces the invariant.
+Message.assistant({ type: "compaction", provider })
+if (checkpoint.encrypted !== undefined) {
+  checkpoint.encrypted satisfies string
+  checkpoint.text satisfies undefined
+}
+if (checkpoint.text !== undefined) {
+  checkpoint.text satisfies string | null
+  checkpoint.encrypted satisfies undefined
+}
+checkpoint.encrypted
+// @ts-expect-error Compaction parts do not contain a generic provider payload.
+checkpoint.value
+LLMClient.compact(LLM.request({ model: openai, prompt: "hello" })).pipe(
+  Effect.map((result) => {
+    result.messages
+    // @ts-expect-error Compaction returns replacement history, not a synthetic assistant message.
+    result.message
+  }),
+)
+LLM.request({
+  model: openai,
+  providerOptions: {
+    // @ts-expect-error A token threshold is numeric.
+    contextManagement: [{ type: "compaction", compactThreshold: "100000" }],
+  },
+})
+for (const model of [
+  Anthropic.configure().model("claude-opus-4-6"),
+  AmazonBedrock.configure().messages("anthropic.claude-opus-4-6-v1"),
+]) {
+  LLM.request({
+    model,
+    providerOptions: {
+      contextManagement: {
+        edits: [
+          { type: "compact_20260112", pauseAfterCompaction: true, instructions: "Summarize without using tools" },
+        ],
+      },
+    },
+  })
+  LLM.request({
+    model,
+    providerOptions: {
+      // @ts-expect-error A pause setting is boolean.
+      contextManagement: { edits: [{ type: "compact_20260112", pauseAfterCompaction: "yes" }] },
+    },
+  })
+}
@@ -0,0 +1,24 @@
+import { LLM } from "../../src/index.js"
+import { Mistral } from "../../src/providers.js"
+
+const selected = Mistral.provider.model("mistral-small-latest")
+
+LLM.request({ model: selected, prompt: "Hello", providerOptions: { reasoningEffort: "high" } })
+LLM.request({ model: selected, prompt: "Hello", providerOptions: { reasoningEffort: "future-effort" } })
+LLM.request({ model: selected, prompt: "Hello", providerOptions: { promptMode: "reasoning" } })
+LLM.request({ model: selected, prompt: "Hello", providerOptions: { parallelToolCalls: false } })
+LLM.request({ model: selected, prompt: "Hello", providerOptions: { promptCacheKey: "session-1" } })
+
+LLM.request({
+  model: selected,
+  prompt: "Hello",
+  // @ts-expect-error Mistral reasoning effort must be a string.
+  providerOptions: { reasoningEffort: 1 },
+})
+
+LLM.request({
+  model: selected,
+  prompt: "Hello",
+  // @ts-expect-error Mistral prompt mode only supports reasoning.
+  providerOptions: { promptMode: "standard" },
+})
@@ -0,0 +1,170 @@
+import { expect } from "bun:test"
+import { Effect, Schema } from "effect"
+import { LLM, LLMRequest, Message } from "../../src/index.js"
+import { LLMClient } from "../../src/route/client.js"
+import { Anthropic, GoogleVertexMessages } from "../../src/providers/index.js"
+import { testEffect } from "../lib/effect.js"
+import { dynamicResponse, fixedResponse } from "../lib/http.js"
+import { sseEvents } from "../lib/sse.js"
+
+for (const fixture of [
+  {
+    name: "empty iterations fall back to top-level usage",
+    usage: { input_tokens: 2, output_tokens: 3, cache_read_input_tokens: null, iterations: [] },
+    expected: { inputTokens: 2, outputTokens: 3, totalTokens: 5, contextTokens: undefined },
+  },
+  {
+    name: "compaction-only usage has no post-compaction context size",
+    usage: {
+      input_tokens: 0,
+      output_tokens: 0,
+      iterations: [{ type: "compaction", input_tokens: 7, cache_read_input_tokens: 3, output_tokens: 2 }],
+    },
+    expected: { inputTokens: 10, outputTokens: 2, totalTokens: 12, contextTokens: undefined },
+  },
+  {
+    name: "partially reported iterations preserve known totals",
+    usage: {
+      iterations: [
+        { type: "compaction", input_tokens: 7, cache_creation_input_tokens: 2 },
+        { type: "message", output_tokens: 3 },
+      ],
+    },
+    expected: { inputTokens: 9, outputTokens: 3, totalTokens: 12, contextTokens: undefined },
+  },
+  {
+    name: "missing counters remain unknown rather than zero",
+    usage: { iterations: [{ type: "message" }] },
+    expected: { inputTokens: undefined, outputTokens: undefined, totalTokens: undefined, contextTokens: undefined },
+  },
+]) {
+  testEffect(
+    fixedResponse(
+      sseEvents(
+        { type: "message_start", message: { usage: fixture.usage } },
+        { type: "message_delta", delta: { stop_reason: "end_turn" } },
+        { type: "message_stop" },
+      ),
+    ),
+  ).effect(fixture.name, () =>
+    Effect.gen(function* () {
+      const result = yield* LLMClient.generate(
+        LLM.request({
+          model: Anthropic.configure({ apiKey: "test" }).model("claude-opus-4-6"),
+          prompt: "hello",
+        }),
+      )
+      expect(result.usage).toMatchObject(fixture.expected)
+    }),
+  )
+}
+
+for (const model of [
+  Anthropic.configure({ apiKey: "test" }).model("claude-opus-4-6"),
+  GoogleVertexMessages.configure({ accessToken: "test", project: "test" }).model("claude-opus-4-6"),
+]) {
+  for (const summary of ["Summary of the conversation", null]) {
+    const block = { type: "compaction", content: summary }
+    testEffect(
+      dynamicResponse(({ request, text, respond }) =>
+        Effect.sync(() => {
+          const body = JSON.parse(text)
+          expect(request.headers["anthropic-beta"]).toBe("existing-beta,compact-2026-01-12")
+          if (body.messages.length === 1) {
+            expect(body.context_management.edits).toEqual([
+              {
+                type: "compact_20260112",
+                trigger: { type: "input_tokens", value: 50000 },
+                pause_after_compaction: true,
+              },
+            ])
+          }
+          if (body.messages.length > 1) {
+            expect(body.messages[1].content).toEqual([block])
+            expect(body.context_management).toBeUndefined()
+          }
+          return respond(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 50000, output_tokens: 0 } } },
+              { type: "content_block_start", index: 0, content_block: { type: "compaction", content: null } },
+              { type: "content_block_delta", index: 0, delta: { type: "compaction_delta", content: summary } },
+              { type: "content_block_stop", index: 0 },
+              {
+                type: "message_delta",
+                delta: { stop_reason: "compaction" },
+                usage: {
+                  input_tokens: 1000,
+                  output_tokens: 5,
+                  iterations: [
+                    { type: "compaction", input_tokens: 50000, output_tokens: 1000, cache_read_input_tokens: 10 },
+                    { type: "message", input_tokens: 1000, output_tokens: 5 },
+                  ],
+                },
+              },
+              { type: "message_stop" },
+            ),
+            { headers: { "content-type": "text/event-stream" } },
+          )
+        }),
+      ),
+    ).effect(
+      `${model.provider} replays ${summary === null ? "failed" : "successful"} compaction with billing and context usage`,
+      () =>
+        Effect.gen(function* () {
+          const request = LLM.request({
+            model,
+            prompt: "hello",
+            http: { headers: { "anthropic-beta": "existing-beta" } },
+            providerOptions: {
+              contextManagement: {
+                edits: [
+                  {
+                    type: "compact_20260112",
+                    trigger: { type: "input_tokens", value: 50000 },
+                    pauseAfterCompaction: true,
+                  },
+                ],
+              },
+            },
+          })
+          const first = yield* LLMClient.generate(request)
+          expect(first.finishReason.raw).toBe("compaction")
+          expect(first.message.content).toEqual([{ type: "compaction", provider: model.provider, text: summary }])
+          expect(first.text).toBe("")
+          expect(first.usage?.inputTokens).toBe(51010)
+          expect(first.usage?.outputTokens).toBe(1005)
+          expect(first.usage?.totalTokens).toBe(52015)
+          expect(first.usage?.contextTokens).toBe(1000)
+          const codec = Schema.fromJsonString(Message)
+          const message = Schema.decodeSync(codec)(Schema.encodeSync(codec)(first.message))
+          yield* LLMClient.generate(
+            LLMRequest.update(request, {
+              providerOptions: {},
+              messages: [...request.messages, message, Message.user("continue")],
+            }),
+          )
+        }),
+    )
+  }
+}
+
+for (const events of [
+  [{ type: "content_block_start", index: 0, content_block: { type: "compaction", content: 42 } }],
+  [{ type: "content_block_delta", index: 0, delta: { type: "compaction_delta", content: "no start" } }],
+  [
+    { type: "content_block_start", index: 0, content_block: { type: "compaction", content: null } },
+    { type: "message_stop" },
+  ],
+]) {
+  testEffect(fixedResponse(sseEvents(...events))).effect(
+    `rejects malformed compaction lifecycle: ${JSON.stringify(events)}`,
+    () =>
+      Effect.gen(function* () {
+        const error = yield* LLMClient.generate(
+          LLM.request({ model: Anthropic.configure({ apiKey: "test" }).model("claude-opus-4-6"), prompt: "hello" }),
+        ).pipe(Effect.flip)
+        expect(error.reason._tag).toBe("InvalidProviderOutput")
+        expect(error.reason.http?.status).toBe(200)
+      }),
+  )
+}
@@ -74,6 +74,48 @@ describe("Anthropic Messages route", () => {
    }),
  )

+  it.effect("omits empty system text while preserving whitespace", () =>
+    Effect.gen(function* () {
+      const empty = yield* compileRequest(LLMRequest.update(request, { system: [{ type: "text", text: "" }] }))
+      const whitespace = yield* compileRequest(LLMRequest.update(request, { system: [{ type: "text", text: " " }] }))
+
+      expect(empty.body.system).toBeUndefined()
+      expect(whitespace.body.system).toEqual([{ type: "text", text: " " }])
+    }),
+  )
+
+  it.effect("filters whitespace-only text and removes empty messages", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.user(" \n\t"),
+            Message.user([]),
+            Message.user([
+              { type: "text", text: "" },
+              { type: "text", text: "  Keep this spacing.  " },
+              { type: "text", text: " \n\t" },
+            ]),
+            Message.assistant(" \n\t"),
+            Message.assistant([]),
+            Message.assistant([{ type: "reasoning", text: "" }]),
+            Message.assistant([
+              { type: "text", text: "" },
+              { type: "reasoning", text: "", providerMetadata: { anthropic: { signature: "sig_1" } } },
+            ]),
+          ],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        { role: "user", content: [{ type: "text", text: "  Keep this spacing.  " }] },
+        { role: "assistant", content: [{ type: "thinking", thinking: "", signature: "sig_1" }] },
+      ])
+    }),
+  )
+
  it.effect("lowers adaptive thinking settings with effort", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
@@ -903,6 +945,54 @@ describe("Anthropic Messages route", () => {
    }),
  )

+  it.effect("preserves a reasoning signature when message_stop closes the block", () =>
+    Effect.gen(function* () {
+      const compatible = Route.make({
+        id: "custom-anthropic-messages",
+        provider: "custom-anthropic",
+        protocol: AnthropicMessages.protocol,
+        endpoint: Endpoint.path("/messages", { baseURL: "https://compatible.test/v1" }),
+        auth: Auth.header("x-api-key", "test"),
+        framing: AnthropicMessages.framing,
+      }).model({ id: "custom-model" })
+      const body = sseEvents(
+        { type: "message_start", message: { usage: { input_tokens: 5 } } },
+        {
+          type: "content_block_start",
+          index: 0,
+          content_block: { type: "thinking", thinking: "", signature: "" },
+        },
+        { type: "content_block_delta", index: 0, delta: { type: "thinking_delta", thinking: "Reasoning." } },
+        { type: "content_block_delta", index: 0, delta: { type: "signature_delta", signature: "sig_1" } },
+        { type: "message_delta", delta: { stop_reason: "end_turn" }, usage: { output_tokens: 2 } },
+        { type: "message_stop" },
+      )
+      const response = yield* LLMClient.generate(LLM.request({ model: compatible, prompt: "Think." })).pipe(
+        Effect.provide(fixedResponse(body)),
+      )
+
+      const reasoningEnds = response.events.filter((event) => event.type === "reasoning-end")
+      expect(reasoningEnds).toHaveLength(1)
+      expect(reasoningEnds[0]).toMatchObject({
+        providerMetadata: { "custom-anthropic": { signature: "sig_1" } },
+      })
+      expect(response.message.content).toEqual([
+        {
+          type: "reasoning",
+          text: "Reasoning.",
+          providerMetadata: { "custom-anthropic": { signature: "sig_1" } },
+        },
+      ])
+
+      const prepared = yield* compileRequest(
+        LLM.request({ model: compatible, messages: [response.message], cache: "none" }),
+      )
+      expect(prepared.body.messages).toEqual([
+        { role: "assistant", content: [{ type: "thinking", thinking: "Reasoning.", signature: "sig_1" }] },
+      ])
+    }),
+  )
+
  it.effect("parses text, reasoning, and usage stream fixtures", () =>
    Effect.gen(function* () {
      const body = sseEvents(
@@ -936,6 +1026,7 @@ describe("Anthropic Messages route", () => {
      expect(response.events.find((event) => event.type === "reasoning-end")).toMatchObject({
        providerMetadata: { anthropic: { signature: "sig_1" } },
      })
+      expect(response.events.filter((event) => event.type === "reasoning-end")).toHaveLength(1)
      expect(response.events.find((event) => event.type === "reasoning-delta" && event.text === "")).toBeUndefined()
      expect(response.message.content).toEqual([
        { type: "text", text: "Hello!" },
@@ -949,6 +1040,41 @@ describe("Anthropic Messages route", () => {
    }),
  )

+  it.effect("preserves terminal state across usage-only message deltas", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "message_start", message: { usage: { input_tokens: 5 } } },
+              {
+                type: "message_delta",
+                delta: { stop_reason: "end_turn", stop_sequence: "X" },
+                usage: { output_tokens: 8 },
+              },
+              { type: "message_delta", delta: {}, usage: { output_tokens: 10 } },
+              { type: "message_stop" },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.usage).toMatchObject({ inputTokens: 5, outputTokens: 10, totalTokens: 15 })
+      expect(response.finishReason).toEqual({ normalized: "stop", raw: "end_turn" })
+      expect(response.events.find((event) => event.type === "step-finish")).toMatchObject({
+        reason: { normalized: "stop", raw: "end_turn" },
+        usage: { inputTokens: 5, outputTokens: 10, totalTokens: 15 },
+        providerMetadata: { anthropic: { stopSequence: "X" } },
+      })
+      expect(response.events.at(-1)).toMatchObject({
+        type: "finish",
+        reason: { normalized: "stop", raw: "end_turn" },
+        usage: { inputTokens: 5, outputTokens: 10, totalTokens: 15 },
+        providerMetadata: { anthropic: { stopSequence: "X" } },
+      })
+    }),
+  )
+
  it.effect("requires message_stop before completing a streamed message", () =>
    Effect.gen(function* () {
      const error = yield* LLMClient.generate(request).pipe(
@@ -1,11 +1,12 @@
 import { EventStreamCodec } from "@smithy/eventstream-codec"
 import { fromUtf8, toUtf8 } from "@smithy/util-utf8"
 import { describe, expect } from "bun:test"
-import { Effect } from "effect"
+import { Effect, Encoding, Ref, Stream } from "effect"
 import {
  CacheHint,
  GenerationOptions,
  LLM,
+  LLMEvent,
  LLMRequest,
  Message,
  ToolCallPart,
@@ -83,6 +84,17 @@ const eventStreamBody = (...payloads: ReadonlyArray<readonly [string, object]>)
 const fixedBytes = (bytes: Uint8Array) =>
  fixedResponse(bytes.slice().buffer, { headers: { "content-type": "application/vnd.amazon.eventstream" } })

+const fixedByteChunks = (...chunks: ReadonlyArray<Uint8Array>) =>
+  fixedResponse(
+    new ReadableStream<Uint8Array>({
+      start(controller) {
+        chunks.forEach((chunk) => controller.enqueue(chunk))
+        controller.close()
+      },
+    }),
+    { headers: { "content-type": "application/vnd.amazon.eventstream" } },
+  )
+
 const model = AmazonBedrock.configure({
  baseURL: "https://bedrock-runtime.test",
  apiKey: "test-bearer",
@@ -113,6 +125,50 @@ describe("Bedrock Converse route", () => {
    }),
  )

+  it.effect("omits empty initial system blocks", () =>
+    Effect.gen(function* () {
+      const empty = yield* compileRequest(LLM.request({ model, system: "", prompt: "hello" }))
+      const cachedEmpty = yield* compileRequest(
+        LLM.request({
+          model,
+          system: [{ type: "text", text: "", cache: new CacheHint({ type: "ephemeral" }) }],
+          prompt: "hello",
+          cache: "none",
+        }),
+      )
+
+      expect(empty.body.system).toBeUndefined()
+      expect(cachedEmpty.body.system).toBeUndefined()
+    }),
+  )
+
+  it.effect("omits empty system blocks while preserving order and cache hints", () =>
+    Effect.gen(function* () {
+      const cache = new CacheHint({ type: "ephemeral" })
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          system: [
+            { type: "text", text: "", cache },
+            { type: "text", text: "First." },
+            { type: "text", text: " " },
+            { type: "text", text: "" },
+            { type: "text", text: "Second.", cache },
+          ],
+          prompt: "hello",
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.system).toEqual([
+        { text: "First." },
+        { text: " " },
+        { text: "Second." },
+        { cachePoint: { type: "default" } },
+      ])
+    }),
+  )
+
  it.effect("passes topK through additionalModelRequestFields as top_k", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
@@ -255,6 +311,79 @@ describe("Bedrock Converse route", () => {
    }),
  )

+  it.effect("removes empty keys recursively from outbound tool inputs without mutating history", () =>
+    Effect.gen(function* () {
+      const input = {
+        path: "file.ts",
+        edits: [
+          { oldText: "a", newText: "b", "": "" },
+          null,
+          true,
+          7,
+          "text",
+          ["kept", { "": false, nested: { "": null, value: "ok" } }],
+        ],
+        nested: { "": "drop", empty: {}, onlyEmpty: { "": 1 } },
+        " ": "preserve whitespace key",
+        "": "drop",
+      }
+      const original = structuredClone(input)
+      const call = ToolCallPart.make({ id: "tool_1", name: "edit", input })
+      const prepared = yield* compileRequest(
+        LLM.request({ model, messages: [Message.assistant([call])], cache: "none" }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [
+            {
+              toolUse: {
+                toolUseId: "tool_1",
+                name: "edit",
+                input: {
+                  path: "file.ts",
+                  edits: [{ oldText: "a", newText: "b" }, null, true, 7, "text", ["kept", { nested: { value: "ok" } }]],
+                  nested: { empty: {}, onlyEmpty: {} },
+                  " ": "preserve whitespace key",
+                },
+              },
+            },
+          ],
+        },
+      ])
+      expect(input).toEqual(original)
+      expect(call.input).toBe(input)
+    }),
+  )
+
+  it.effect("keeps empty tool inputs and empties inputs containing only empty keys", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([
+              ToolCallPart.make({ id: "tool_empty_key", name: "first", input: { "": { value: true } } }),
+              ToolCallPart.make({ id: "tool_empty_object", name: "second", input: {} }),
+            ]),
+          ],
+          cache: "none",
+        }),
+      )
+
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [
+            { toolUse: { toolUseId: "tool_empty_key", name: "first", input: {} } },
+            { toolUse: { toolUseId: "tool_empty_object", name: "second", input: {} } },
+          ],
+        },
+      ])
+    }),
+  )
+
  it.effect("merges parallel tool results into one user message", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
@@ -385,19 +514,77 @@ describe("Bedrock Converse route", () => {
    }),
  )

-  it.effect("maps truncation and malformed output stop reasons", () =>
+  it.effect("rejects truncated event-stream frames after message stop", () =>
    Effect.gen(function* () {
-      const reasons = [
-        ["model_context_window_exceeded", "length"],
-        ["malformed_model_output", "error"],
-        ["malformed_tool_use", "error"],
-      ] as const
+      const partialFrames = [
+        eventFrame("metadata", { usage: { inputTokens: 5, outputTokens: 2, totalTokens: 7 } }).subarray(0, 3),
+        exceptionFrame("modelStreamErrorException", { originalMessage: "Upstream model failed" }).subarray(0, -1),
+      ]

-      for (const [raw, normalized] of reasons) {
-        const response = yield* LLMClient.generate(baseRequest).pipe(
-          Effect.provide(fixedBytes(eventStreamBody(["messageStop", { stopReason: raw }]))),
+      for (const partial of partialFrames) {
+        const error = yield* LLMClient.generate(baseRequest).pipe(
+          Effect.provide(fixedBytes(concat([eventFrame("messageStop", { stopReason: "end_turn" }), partial]))),
+          Effect.flip,
        )
-        expect(response.finishReason).toEqual({ normalized, raw })
+
+        expect(error).toMatchObject({
+          reason: { _tag: "InvalidProviderOutput", classification: "incomplete-stream" },
+          message: `Incomplete Bedrock Converse event-stream frame: ${partial.length} buffered bytes remain at end of stream`,
+        })
+        expect(error.reason.body).toBe(Encoding.encodeBase64(partial))
+      }
+    }),
+  )
+
+  it.effect("decodes frames split across transport chunks through exact-boundary EOF", () =>
+    Effect.gen(function* () {
+      const body = eventStreamBody(
+        ["messageStart", { role: "assistant" }],
+        ["contentBlockDelta", { contentBlockIndex: 0, delta: { text: "Hello" } }],
+        ["messageStop", { stopReason: "end_turn" }],
+      )
+      const response = yield* LLMClient.generate(baseRequest).pipe(
+        Effect.provide(fixedByteChunks(body.subarray(0, 2), body.subarray(2, 17), body.subarray(17))),
+      )
+
+      expect(response.text).toBe("Hello")
+      expect(response.finishReason).toEqual({ normalized: "stop", raw: "end_turn" })
+    }),
+  )
+
+  it.effect("maps model context window exhaustion to length", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(baseRequest).pipe(
+        Effect.provide(fixedBytes(eventStreamBody(["messageStop", { stopReason: "model_context_window_exceeded" }]))),
+      )
+
+      expect(response.finishReason).toEqual({
+        normalized: "length",
+        raw: "model_context_window_exceeded",
+      })
+    }),
+  )
+
+  it.effect("fails malformed output stop reasons", () =>
+    Effect.gen(function* () {
+      for (const reason of ["malformed_model_output", "malformed_tool_use"] as const) {
+        const events = yield* Ref.make<ReadonlyArray<LLMEvent>>([])
+        const error = yield* LLMClient.stream(baseRequest).pipe(
+          Stream.tap((event) => Ref.update(events, (current) => [...current, event])),
+          Stream.runDrain,
+          Effect.provide(fixedBytes(eventStreamBody(["messageStop", { stopReason: reason }]))),
+          Effect.flip,
+        )
+
+        expect(error).toMatchObject({
+          reason: { _tag: "InvalidProviderOutput" },
+          message: `Bedrock Converse stopped with ${reason}`,
+        })
+        expect(JSON.parse(error.reason.body ?? "")).toMatchObject({
+          headers: { ":event-type": { value: "messageStop" } },
+          body: JSON.stringify({ stopReason: reason }),
+        })
+        expect((yield* Ref.get(events)).some((event) => event.type === "finish")).toBeFalse()
      }
    }),
  )
@@ -853,7 +1040,7 @@ describe("Bedrock Converse route", () => {
    Effect.gen(function* () {
      // Bedrock represents redactedContent blobs as base64 strings on its JSON
      // wire. The provider owns the payload and requires byte-exact replay.
-      const redactedData = "cmVkYWN0ZWQtdGhpbmtpbmc="
+      const redactedData = "AQID"
      const response = yield* LLMClient.generate(
        LLMRequest.update(baseRequest, {
          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
@@ -863,10 +1050,8 @@ describe("Bedrock Converse route", () => {
          fixedBytes(
            eventStreamBody(
              ["messageStart", { role: "assistant" }],
-              [
-                "contentBlockDelta",
-                { contentBlockIndex: 0, delta: { reasoningContent: { redactedContent: redactedData } } },
-              ],
+              ["contentBlockDelta", { contentBlockIndex: 0, delta: { reasoningContent: { redactedContent: "AQ==" } } }],
+              ["contentBlockDelta", { contentBlockIndex: 0, delta: { reasoningContent: { redactedContent: "AgM=" } } }],
              ["contentBlockStop", { contentBlockIndex: 0 }],
              [
                "contentBlockStart",
@@ -882,12 +1067,17 @@ describe("Bedrock Converse route", () => {
          ),
        ),
      )
-      expect(response.events.find((event) => event.type === "reasoning-delta" && event.text === "")).toEqual({
+      expect(response.events.filter((event) => event.type === "reasoning-delta" && event.text === "").at(-1)).toEqual({
        type: "reasoning-delta",
        id: "reasoning-0",
        text: "",
        providerMetadata: { bedrock: { redactedData } },
      })
+      expect(response.events.find((event) => event.type === "reasoning-end")).toEqual({
+        type: "reasoning-end",
+        id: "reasoning-0",
+        providerMetadata: { bedrock: { redactedData } },
+      })
      const prepared = yield* compileRequest(
        LLM.request({
          model,
@@ -918,6 +1108,73 @@ describe("Bedrock Converse route", () => {
    }),
  )

+  it.effect("keeps redacted reasoning accumulation separate by content block index", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(baseRequest).pipe(
+        Effect.provide(
+          fixedBytes(
+            eventStreamBody(
+              ["messageStart", { role: "assistant" }],
+              ["contentBlockDelta", { contentBlockIndex: 2, delta: { reasoningContent: { redactedContent: "AQ==" } } }],
+              ["contentBlockDelta", { contentBlockIndex: 2, delta: { reasoningContent: { redactedContent: "Ag==" } } }],
+              ["contentBlockStop", { contentBlockIndex: 2 }],
+              ["contentBlockDelta", { contentBlockIndex: 7, delta: { reasoningContent: { redactedContent: "Aw==" } } }],
+              ["contentBlockDelta", { contentBlockIndex: 7, delta: { reasoningContent: { redactedContent: "BA==" } } }],
+              ["contentBlockStop", { contentBlockIndex: 7 }],
+              ["messageStop", { stopReason: "end_turn" }],
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toEqual([
+        { type: "reasoning", text: "", providerMetadata: { bedrock: { redactedData: "AQI=" } } },
+        { type: "reasoning", text: "", providerMetadata: { bedrock: { redactedData: "AwQ=" } } },
+      ])
+    }),
+  )
+
+  it.effect("preserves split redacted reasoning when contentBlockStop is missing", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(baseRequest).pipe(
+        Effect.provide(
+          fixedBytes(
+            eventStreamBody(
+              ["messageStart", { role: "assistant" }],
+              ["contentBlockDelta", { contentBlockIndex: 0, delta: { reasoningContent: { redactedContent: "AQ==" } } }],
+              ["contentBlockDelta", { contentBlockIndex: 0, delta: { reasoningContent: { redactedContent: "AgM=" } } }],
+              ["messageStop", { stopReason: "end_turn" }],
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toEqual([
+        { type: "reasoning", text: "", providerMetadata: { bedrock: { redactedData: "AQID" } } },
+      ])
+    }),
+  )
+
+  it.effect("rejects invalid redacted reasoning base64 with the triggering event", () =>
+    Effect.gen(function* () {
+      const payload = { contentBlockIndex: 0, delta: { reasoningContent: { redactedContent: "%%==" } } }
+      const error = yield* LLMClient.generate(baseRequest).pipe(
+        Effect.provide(fixedBytes(eventStreamBody(["contentBlockDelta", payload]))),
+        Effect.flip,
+      )
+
+      expect(error).toMatchObject({
+        reason: { _tag: "InvalidProviderOutput" },
+        message: "Bedrock Converse reasoningContent.redactedContent contains invalid base64 data",
+      })
+      expect(JSON.parse(error.reason.body ?? "")).toMatchObject({
+        headers: { ":event-type": { value: "contentBlockDelta" } },
+        body: JSON.stringify(payload),
+      })
+      expect(error.reason.cause).toBeInstanceOf(Error)
+    }),
+  )
+
  it.effect("ignores unknown normal stream events", () =>
    Effect.gen(function* () {
      const body = concat([
@@ -1172,6 +1429,20 @@ describe("Bedrock Converse route", () => {
    }),
  )

+  it.effect("rejects image media that is not valid base64", () =>
+    Effect.gen(function* () {
+      const error = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [Message.user({ type: "media", mediaType: "image/png", data: "https://example.test/image.png" })],
+        }),
+      ).pipe(Effect.flip)
+
+      expect(error).toMatchObject({ reason: { _tag: "InvalidRequest" } })
+      expect(error.message).toContain("Bedrock Converse media data must be valid base64")
+    }),
+  )
+
  it.effect("lowers document media into Bedrock document blocks with format and name", () =>
    Effect.gen(function* () {
      const prepared = yield* compileRequest(
@@ -1295,6 +1566,37 @@ describe("Bedrock Converse route", () => {
    }),
  )

+  it.effect("rejects remote media URLs in tool results", () =>
+    Effect.gen(function* () {
+      const error = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant([ToolCallPart.make({ id: "call_1", name: "read", input: {} })]),
+            Message.tool({
+              id: "call_1",
+              name: "read",
+              result: {
+                type: "content",
+                value: [
+                  {
+                    type: "file",
+                    uri: "https://example.test/report.pdf",
+                    mime: "application/pdf",
+                    name: "report.pdf",
+                  },
+                ],
+              },
+            }),
+          ],
+        }),
+      ).pipe(Effect.flip)
+
+      expect(error).toMatchObject({ reason: { _tag: "InvalidRequest" } })
+      expect(error.message).toContain("Bedrock Converse media data must be valid base64")
+    }),
+  )
+
  it.effect("rejects unsupported image media types", () =>
    Effect.gen(function* () {
      const error = yield* compileRequest(
@@ -0,0 +1,89 @@
+import { EventStreamCodec } from "@smithy/eventstream-codec"
+import { fromUtf8, toUtf8 } from "@smithy/util-utf8"
+import { expect } from "bun:test"
+import { Effect } from "effect"
+import { LLM, LLMRequest, Message } from "../../src/index.js"
+import { LLMClient } from "../../src/route/client.js"
+import { AmazonBedrock } from "../../src/providers/index.js"
+import { testEffect } from "../lib/effect.js"
+import { dynamicResponse, fixedResponse } from "../lib/http.js"
+
+const codec = new EventStreamCodec(toUtf8, fromUtf8)
+const frame = (event: object) =>
+  codec.encode({
+    headers: { ":message-type": { type: "string", value: "event" }, ":event-type": { type: "string", value: "chunk" } },
+    body: new TextEncoder().encode(JSON.stringify({ bytes: Buffer.from(JSON.stringify(event)).toString("base64") })),
+  })
+const response = Buffer.concat(
+  [
+    { type: "message_start", message: { usage: { input_tokens: 60000 } } },
+    { type: "content_block_start", index: 0, content_block: { type: "compaction", content: null } },
+    { type: "content_block_delta", index: 0, delta: { type: "compaction_delta", content: "Summary" } },
+    { type: "content_block_stop", index: 0 },
+    { type: "content_block_start", index: 1, content_block: { type: "text", text: "" } },
+    { type: "content_block_delta", index: 1, delta: { type: "text_delta", text: "Hello" } },
+    { type: "content_block_stop", index: 1 },
+    { type: "message_delta", delta: { stop_reason: "end_turn" }, usage: { output_tokens: 10 } },
+    { type: "message_stop" },
+  ].map(frame),
+)
+
+for (const auth of [
+  { apiKey: "test" },
+  { credentials: { accessKeyId: "test", secretAccessKey: "test", region: "us-west-2" } },
+]) {
+  testEffect(
+    dynamicResponse(({ request, text, respond }) =>
+      Effect.sync(() => {
+        expect(request.url).toBe(
+          "https://bedrock-runtime.us-west-2.amazonaws.com/model/anthropic.claude-opus-4-6-v1%3A0/invoke-with-response-stream",
+        )
+        expect(request.headers.authorization).toStartWith(auth.apiKey ? "Bearer test" : "AWS4-HMAC-SHA256")
+        const body = JSON.parse(text)
+        expect(body.model).toBeUndefined()
+        expect(body.stream).toBeUndefined()
+        expect(body.anthropic_version).toBe("bedrock-2023-05-31")
+        expect(body.anthropic_beta).toEqual(["compact-2026-01-12"])
+        expect(body.context_management.edits).toEqual([{ type: "compact_20260112" }])
+        if (body.messages.length > 1)
+          expect(body.messages[1].content[0]).toEqual({ type: "compaction", content: "Summary" })
+        return respond(response, { headers: { "content-type": "application/vnd.amazon.eventstream" } })
+      }),
+    ),
+  ).effect(`Bedrock Messages compaction round trip with ${auth.apiKey ? "bearer" : "SigV4"} authentication`, () =>
+    Effect.gen(function* () {
+      const model = AmazonBedrock.configure({ ...auth, region: "us-west-2" }).messages("anthropic.claude-opus-4-6-v1:0")
+      const request = LLM.request({
+        model,
+        prompt: "hello",
+        providerOptions: { contextManagement: { edits: [{ type: "compact_20260112" }] } },
+      })
+      const first = yield* LLMClient.generate(request)
+      expect(first.text).toBe("Hello")
+      expect(first.message.content.map((part) => part.type)).toEqual(["compaction", "text"])
+      yield* LLMClient.generate(
+        LLMRequest.update(request, { messages: [...request.messages, first.message, Message.user("continue")] }),
+      )
+    }),
+  )
+}
+
+testEffect(
+  fixedResponse(
+    codec.encode({
+      headers: {
+        ":message-type": { type: "string", value: "exception" },
+        ":exception-type": { type: "string", value: "throttlingException" },
+      },
+      body: new TextEncoder().encode(JSON.stringify({ message: "Too many requests", trace: "keep-original" })),
+    }),
+  ),
+).effect("Bedrock Messages retains the original exception frame", () =>
+  Effect.gen(function* () {
+    const error = yield* LLMClient.generate(
+      LLM.request({ model: AmazonBedrock.configure({ apiKey: "test" }).messages("claude"), prompt: "hello" }),
+    ).pipe(Effect.flip)
+    expect(error.reason.body).toContain("keep-original")
+    expect(error.reason.http?.status).toBe(200)
+  }),
+)
@@ -0,0 +1,50 @@
+import { expect } from "bun:test"
+import { Effect, Stream } from "effect"
+import { LLM, LLMRequest, Message } from "../../src/index.js"
+import { LLMClient, WebSocketTransport } from "../../src/route.js"
+import { OpenAI } from "../../src/providers.js"
+import { testEffect } from "../lib/effect.js"
+import { fixedResponse } from "../lib/http.js"
+
+testEffect(fixedResponse("unexpected HTTP fallback")).effect(
+  "WebSocket responses preserve compaction options and replay state",
+  () =>
+    Effect.gen(function* () {
+      const checkpoint = { type: "compaction", id: "cmp_ws", encrypted_content: "opaque" }
+      const sent: unknown[] = []
+      const webSocket = WebSocketTransport.makeDirect({
+        open: () =>
+          Effect.succeed({
+            sendText: (message) =>
+              Effect.sync(() => {
+                const body = JSON.parse(message)
+                expect(body.context_management).toEqual([{ type: "compaction", compact_threshold: 100000 }])
+                expect(body.stream).toBeUndefined()
+                if (sent.length) expect(body.input[1]).toEqual(checkpoint)
+                sent.push(body)
+              }),
+            messages: Stream.fromIterable(
+              [
+                { type: "response.created", response: { id: "resp_ws" } },
+                { type: "response.output_item.done", item: checkpoint },
+                { type: "response.completed", response: { id: "resp_ws", output: [checkpoint] } },
+              ].map((event) => JSON.stringify(event)),
+            ),
+            close: Effect.void,
+          }),
+      })
+      const request = LLM.request({
+        model: OpenAI.configure({ apiKey: "test" }).responses("gpt-5.3-codex"),
+        prompt: "hello",
+        providerOptions: { contextManagement: [{ type: "compaction", compactThreshold: 100000 }] },
+      })
+      const first = yield* LLMClient.generate(request, { webSocket })
+      expect(first.message.content).toHaveLength(1)
+      expect(first.message.content[0]?.type).toBe("compaction")
+      yield* LLMClient.generate(
+        LLMRequest.update(request, { messages: [...request.messages, first.message, Message.user("continue")] }),
+        { webSocket },
+      )
+      expect(sent).toHaveLength(2)
+    }),
+)
@@ -0,0 +1,91 @@
+import { expect } from "bun:test"
+import { Effect } from "effect"
+import { LLM, LLMRequest, Message } from "../../src/index.js"
+import { LLMClient } from "../../src/route/client.js"
+import { OpenAI, XAI, Anthropic } from "../../src/providers.js"
+import { recordedTests } from "../recorded-test.js"
+
+const history = [
+  Message.user("Remember the project codename COPPER-ORBIT-42."),
+  Message.assistant(
+    "The project codename is COPPER-ORBIT-42. " + "We reviewed the implementation and tests. ".repeat(1000),
+  ),
+]
+
+for (const provider of [
+  {
+    id: "openai",
+    key: "OPENAI_API_KEY",
+    model: OpenAI.configure({ apiKey: process.env.OPENAI_API_KEY ?? "fixture" }).responses("gpt-5.3-codex"),
+  },
+  {
+    id: "xai",
+    key: "XAI_API_KEY",
+    model: XAI.configure({ apiKey: process.env.XAI_API_KEY ?? "fixture" }).responses("grok-4.6"),
+  },
+]) {
+  recordedTests({ prefix: `${provider.id}-compaction`, provider: provider.id, requires: [provider.key] }).effect(
+    "compacts and continues with the provider checkpoint",
+    () =>
+      Effect.gen(function* () {
+        const request = LLM.request({ model: provider.model, messages: history, generation: { maxTokens: 1024 } })
+        const compacted = yield* LLMClient.compact(request)
+        const result = yield* LLMClient.generate(
+          LLMRequest.update(request, {
+            messages: [
+              ...compacted.messages,
+              Message.user("What is the project codename? Reply only with the codename."),
+            ],
+          }),
+        )
+        expect(result.text).toContain("COPPER-ORBIT-42")
+      }),
+    120000,
+  )
+}
+
+recordedTests({
+  prefix: "anthropic-compaction",
+  provider: "anthropic",
+  requires: ["ANTHROPIC_API_KEY"],
+  options: { redact: { allowRequestHeaders: ["anthropic-version", "anthropic-beta"] } },
+}).effect(
+  "automatically compacts and continues after a pause",
+  () =>
+    Effect.gen(function* () {
+      const model = Anthropic.configure({ apiKey: process.env.ANTHROPIC_API_KEY ?? "fixture" }).model(
+        "claude-sonnet-4-6",
+      )
+      const request = LLM.request({
+        model,
+        messages: [
+          Message.user(
+            "Remember the project codename COPPER-ORBIT-42. " +
+              "The implementation and tests were reviewed. ".repeat(10000),
+          ),
+        ],
+        generation: { maxTokens: 4096 },
+        providerOptions: {
+          contextManagement: {
+            edits: [
+              { type: "compact_20260112", trigger: { type: "input_tokens", value: 50000 }, pauseAfterCompaction: true },
+            ],
+          },
+        },
+      })
+      const first = yield* LLMClient.generate(request)
+      expect(first.finishReason.raw).toBe("compaction")
+      expect(first.message.content.some((part) => part.type === "compaction")).toBe(true)
+      const result = yield* LLMClient.generate(
+        LLMRequest.update(request, {
+          messages: [
+            ...request.messages,
+            first.message,
+            Message.user("What is the project codename? Reply only with the codename."),
+          ],
+        }),
+      )
+      expect(result.text).toContain("COPPER-ORBIT-42")
+    }),
+  120000,
+)
@@ -0,0 +1,136 @@
+import { expect } from "bun:test"
+import { Effect, Schema } from "effect"
+import { LLM, LLMRequest, Message } from "../../src/index.js"
+import { LLMClient } from "../../src/route/client.js"
+import { OpenAI, Azure, XAI } from "../../src/providers/index.js"
+import { testEffect } from "../lib/effect.js"
+import { dynamicResponse, fixedResponse } from "../lib/http.js"
+import { sseEvents } from "../lib/sse.js"
+
+const checkpoint = { type: "compaction", id: "cmp_1", encrypted_content: "opaque" }
+const response = sseEvents(
+  { type: "response.output_item.done", item: checkpoint },
+  {
+    type: "response.completed",
+    response: { id: "resp_1", output: [checkpoint], usage: { input_tokens: 10, output_tokens: 2, total_tokens: 12 } },
+  },
+)
+
+for (const model of [
+  OpenAI.configure({ apiKey: "test" }).responses("gpt-5.3-codex"),
+  Azure.configure({ apiKey: "test", resourceName: "test" }).responses("deployment"),
+]) {
+  testEffect(
+    dynamicResponse(({ text, respond }) =>
+      Effect.sync(() => {
+        const body = JSON.parse(text)
+        expect(body.context_management).toEqual([{ type: "compaction", compact_threshold: 100000 }])
+        expect(body.store).toBe(false)
+        if (body.input.length > 1) expect(body.input[1]).toEqual(checkpoint)
+        return respond(response, { headers: { "content-type": "text/event-stream" } })
+      }),
+    ),
+  ).effect(`${model.provider} compaction survives generation, serialization, and a second request`, () =>
+    Effect.gen(function* () {
+      const request = LLM.request({
+        model,
+        prompt: "hello",
+        providerOptions: { contextManagement: [{ type: "compaction", compactThreshold: 100000 }] },
+      })
+      const first = yield* LLMClient.generate(request)
+      expect(first.message.content).toHaveLength(1)
+      expect(first.message.content[0]?.type).toBe("compaction")
+      expect(first.text).toBe("")
+      const codec = Schema.fromJsonString(Message)
+      const message = Schema.decodeSync(codec)(Schema.encodeSync(codec)(first.message))
+      yield* LLMClient.generate(
+        LLMRequest.update(request, { messages: [...request.messages, message, Message.user("continue")] }),
+      )
+      const rejected = yield* LLMClient.generate(
+        LLMRequest.update(request, {
+          model: XAI.configure({ apiKey: "test" }).responses("grok-4.6"),
+          providerOptions: {},
+          messages: [message],
+        }),
+      ).pipe(Effect.flip)
+      expect(rejected.reason._tag).toBe("InvalidRequest")
+    }),
+  )
+}
+
+testEffect(fixedResponse(sseEvents({ type: "response.completed", response: { output: [checkpoint] } }))).effect(
+  "recovers compaction from the terminal output when item completion is absent",
+  () =>
+    Effect.gen(function* () {
+      const result = yield* LLMClient.generate(
+        LLM.request({ model: OpenAI.configure({ apiKey: "test" }).responses("gpt-5.3-codex"), prompt: "hello" }),
+      )
+      expect(result.message.content).toHaveLength(1)
+      expect(result.message.content[0]?.type).toBe("compaction")
+    }),
+)
+
+testEffect(
+  fixedResponse(sseEvents({ type: "response.output_item.done", item: { type: "compaction", id: "cmp_bad" } })),
+).effect("rejects incomplete compaction payloads without publishing a checkpoint", () =>
+  Effect.gen(function* () {
+    const error = yield* LLMClient.generate(
+      LLM.request({ model: OpenAI.configure({ apiKey: "test" }).responses("gpt-5.3-codex"), prompt: "hello" }),
+    ).pipe(Effect.flip)
+    expect(error.reason._tag).toBe("InvalidProviderOutput")
+    expect(error.reason.body).toContain("cmp_bad")
+  }),
+)
+
+const textItem = {
+  type: "message",
+  id: "msg_after",
+  role: "assistant",
+  content: [{ type: "output_text", text: "After checkpoint" }],
+}
+
+for (const completed of [false, true]) {
+  testEffect(
+    fixedResponse(
+      sseEvents(
+        { type: "response.output_item.added", output_index: 0, item: { type: "compaction", id: checkpoint.id } },
+        ...(completed ? [{ type: "response.output_item.done", output_index: 0, item: checkpoint }] : []),
+        { type: "response.output_item.added", output_index: 1, item: textItem },
+        { type: "response.output_text.delta", output_index: 1, item_id: textItem.id, delta: "After checkpoint" },
+        { type: "response.output_item.done", output_index: 1, item: textItem },
+        { type: "response.completed", response: { id: "resp_1", output: [checkpoint, textItem] } },
+      ),
+    ),
+  ).effect(completed ? "keeps streamed checkpoints before later text" : "rejects order-unsafe terminal recovery", () =>
+    Effect.gen(function* () {
+      const request = LLM.request({ model: OpenAI.configure({ apiKey: "test" }).responses("fixture"), prompt: "hello" })
+      if (completed) {
+        const response = yield* LLMClient.generate(request)
+        expect(response.message.content.map((part) => part.type)).toEqual(["compaction", "text"])
+        return
+      }
+      const error = yield* LLMClient.generate(request).pipe(Effect.flip)
+      expect(error.reason._tag).toBe("InvalidProviderOutput")
+      expect(error.message).toContain("Cannot recover a compaction checkpoint")
+      expect(error.reason.body).toContain("response.completed")
+      expect(error.reason.http?.status).toBe(200)
+    }),
+  )
+}
+
+testEffect(
+  fixedResponse(
+    sseEvents({
+      type: "response.completed",
+      response: { output: [{ type: "compaction", encrypted_content: "opaque" }] },
+    }),
+  ),
+).effect("rejects terminal checkpoints missing an id", () =>
+  Effect.gen(function* () {
+    const error = yield* LLMClient.generate(
+      LLM.request({ model: OpenAI.configure({ apiKey: "test" }).responses("fixture"), prompt: "hello" }),
+    ).pipe(Effect.flip)
+    expect(error.reason._tag).toBe("InvalidProviderOutput")
+    expect(error.message).toContain("missing its id")
+  }),
+)
@@ -0,0 +1,309 @@
+import { expect } from "bun:test"
+import { Effect, Schema } from "effect"
+import { LLM, LLMRequest, Message } from "../../src/index.js"
+import { LLMClient, Route } from "../../src/route/client.js"
+import { Auth } from "../../src/route/auth.js"
+import { Endpoint } from "../../src/route/endpoint.js"
+import { OpenAIResponses } from "../../src/protocols/openai-responses.js"
+import { OpenAI, Azure, XAI, Anthropic, AmazonBedrockMantle } from "../../src/providers/index.js"
+import { testEffect } from "../lib/effect.js"
+import { dynamicResponse, fixedResponse } from "../lib/http.js"
+import { sseEvents } from "../lib/sse.js"
+
+const checkpoint = { type: "compaction", id: "cmp_1", encrypted_content: "opaque" }
+const retained = {
+  type: "message",
+  role: "user",
+  id: "msg_1",
+  status: "completed",
+  content: [{ type: "input_text", text: "retained" }],
+}
+const output = [retained, checkpoint]
+
+testEffect(
+  dynamicResponse(({ request, text, respond }) =>
+    Effect.sync(() => {
+      expect(request.headers["x-deployment"]).toBe("fixture")
+      expect(request.headers["x-override"]).toBe("request")
+      expect(request.headers["x-default"]).toBe("configured")
+      expect(request.headers.authorization).toBe("Bearer test")
+      expect(new URL(request.url).searchParams.get("api-version")).toBe("fixture")
+      expect(new URL(request.url).searchParams.get("trace")).toBe("request")
+      if (new URL(request.url).pathname.endsWith("/compact")) {
+        expect(JSON.parse(text)).toEqual({
+          model: "fixture",
+          input: [{ role: "user", content: [{ type: "input_text", text: "hello" }] }],
+        })
+        return respond(JSON.stringify({ object: "response.compaction", output }))
+      }
+      return respond(sseEvents({ type: "response.completed", response: { id: "resp_1" } }))
+    }),
+  ),
+).effect("generation and compaction share deployment headers, defaults, auth, query, and middleware", () =>
+  Effect.gen(function* () {
+    const headers: string[] = []
+    const middleware: string[] = []
+    const route = Route.make({
+      id: "compaction-headers",
+      provider: "openai",
+      protocol: OpenAIResponses.protocol,
+      compact: OpenAIResponses.route.compact,
+      transport: OpenAIResponses.httpTransport,
+      endpoint: Endpoint.path("/responses", { baseURL: "https://example.com", query: { "api-version": "fixture" } }),
+      auth: Auth.bearer("test"),
+      headers: ({ request }) => {
+        expect(request.providerOptions?.store).toBe(false)
+        headers.push(String(request.model.id))
+        return { "x-deployment": "fixture", "x-override": "route" }
+      },
+      defaults: {
+        headers: { "x-default": "configured", "x-override": "configured" },
+        providerOptions: { store: false },
+      },
+    })
+    const request = LLM.request({
+      model: route.model({ id: "fixture" }),
+      prompt: "hello",
+      http: { headers: { "x-override": "request" }, query: { trace: "request" }, body: { store: false, stream: true } },
+    })
+    const options: Parameters<typeof LLMClient.compact>[1] = {
+      http: (request, next) => {
+        middleware.push(new URL(request.url).pathname)
+        return next(request)
+      },
+    }
+    yield* LLMClient.generate(request, options)
+    yield* LLMClient.compact(request, options)
+    expect(headers).toEqual(["fixture", "fixture"])
+    expect(middleware).toEqual(["/responses", "/responses/compact"])
+  }),
+)
+
+const retainedItems = [
+  retained,
+  {
+    type: "message",
+    id: "msg_assistant",
+    role: "assistant",
+    status: "completed",
+    phase: "commentary",
+    content: [
+      { type: "output_text", text: "First" },
+      { type: "output_text", text: "Second" },
+    ],
+  },
+  {
+    type: "reasoning",
+    id: "rs_1",
+    summary: [
+      { type: "summary_text", text: "Thinking" },
+      { type: "summary_text", text: "More thinking" },
+    ],
+    encrypted_content: "reasoning-state",
+  },
+  { type: "reasoning", id: "rs_2", summary: [], encrypted_content: "hidden-reasoning" },
+  {
+    type: "message",
+    id: "msg_media",
+    role: "user",
+    content: [
+      { type: "input_image", image_url: "https://example.com/image.png" },
+      { type: "input_file", filename: "report.pdf", file_data: "data:application/pdf;base64,cGRm" },
+      { type: "input_file", filename: "other.pdf", file_url: "https://example.com/report.pdf" },
+    ],
+  },
+  checkpoint,
+]
+
+testEffect(
+  dynamicResponse(({ request, text, respond }) =>
+    Effect.sync(() => {
+      if (new URL(request.url).pathname.endsWith("/compact"))
+        return respond(JSON.stringify({ object: "response.compaction", output: retainedItems }))
+      expect(JSON.parse(text).input).toEqual(retainedItems)
+      return respond(sseEvents({ type: "response.completed", response: { id: "resp_1" } }), {
+        headers: { "content-type": "text/event-stream" },
+      })
+    }),
+  ),
+).effect("retained messages, reasoning, and media are ordinary typed conversation parts", () =>
+  Effect.gen(function* () {
+    const request = LLM.request({
+      model: OpenAI.configure({ apiKey: "test" }).responses("gpt-5.3-codex"),
+      prompt: "hello",
+    })
+    const compacted = yield* LLMClient.compact(request)
+    expect(compacted.messages.map((message) => message.role)).toEqual([
+      "user",
+      "assistant",
+      "assistant",
+      "assistant",
+      "user",
+      "assistant",
+    ])
+    expect(compacted.messages[1]?.content).toEqual([
+      { type: "text", text: "First" },
+      { type: "text", text: "Second" },
+    ])
+    expect(compacted.messages[2]?.content.map((part) => part.type)).toEqual(["reasoning", "reasoning"])
+    expect(compacted.messages[4]?.content.map((part) => part.type)).toEqual(["media", "media", "media"])
+    const codec = Schema.fromJsonString(Schema.Array(Message))
+    const messages = Schema.decodeSync(codec)(Schema.encodeSync(codec)(compacted.messages))
+    yield* LLMClient.generate(LLMRequest.update(request, { messages }))
+  }),
+)
+
+for (const item of [
+  { type: "unknown_provider_item", data: "do not hide in a compaction part" },
+  {
+    type: "message",
+    role: "user",
+    content: [{ type: "input_image", image_url: "https://example.com/image.png", detail: 42 }],
+  },
+  { type: "message", role: "user", content: [] },
+  {
+    type: "message",
+    role: "assistant",
+    content: [{ type: "input_image", image_url: "https://example.com/image.png" }],
+  },
+  { type: "message", role: "user", content: [{ type: "input_file", filename: "missing.pdf" }] },
+]) {
+  testEffect(fixedResponse(JSON.stringify({ object: "response.compaction", output: [item, checkpoint] }))).effect(
+    `rejects unsupported compact output: ${JSON.stringify(item)}`,
+    () =>
+      Effect.gen(function* () {
+        const error = yield* LLMClient.compact(
+          LLM.request({ model: OpenAI.configure({ apiKey: "test" }).responses("gpt-5.3-codex"), prompt: "hello" }),
+        ).pipe(Effect.flip)
+        expect(error.reason._tag).toBe("InvalidProviderOutput")
+        expect(error.reason.body).toContain(JSON.stringify(item))
+        expect(error.reason.http?.status).toBe(200)
+      }),
+  )
+}
+
+for (const model of [
+  OpenAI.configure({ apiKey: "test" }).responses("fixture"),
+  Azure.configure({ apiKey: "test", resourceName: "test" }).responses("fixture"),
+  XAI.configure({ apiKey: "test" }).responses("fixture"),
+]) {
+  const images = [undefined, "low", "high", "auto"].map((detail) => ({
+    type: "input_image",
+    image_url: "https://example.com/image.png",
+    ...(detail === undefined ? {} : { detail }),
+  }))
+  testEffect(
+    dynamicResponse(({ request, text, respond }) =>
+      Effect.sync(() => {
+        if (new URL(request.url).pathname.endsWith("/compact"))
+          return respond(
+            JSON.stringify({
+              object: "response.compaction",
+              output: [{ type: "message", role: "user", content: images }, checkpoint],
+            }),
+          )
+        expect(JSON.parse(text).input[0].content).toEqual(images)
+        return respond(sseEvents({ type: "response.completed", response: { id: "resp_1" } }))
+      }),
+    ),
+  ).effect(`${model.provider} preserves retained image detail through serialization and replay`, () =>
+    Effect.gen(function* () {
+      const request = LLM.request({ model, prompt: "hello" })
+      const compacted = yield* LLMClient.compact(request)
+      const codec = Schema.fromJsonString(Schema.Array(Message))
+      const messages = Schema.decodeSync(codec)(Schema.encodeSync(codec)(compacted.messages))
+      yield* LLMClient.generate(LLMRequest.update(request, { messages }))
+    }),
+  )
+}
+
+testEffect(fixedResponse("must not execute")).effect("xAI rejects automatic compaction options", () =>
+  Effect.gen(function* () {
+    const request = LLMRequest.update(
+      LLM.request({ model: XAI.configure({ apiKey: "test" }).responses("grok-4.6"), prompt: "hello" }),
+      { providerOptions: { contextManagement: [{ type: "compaction" }] } },
+    )
+    const error = yield* LLMClient.generate(request).pipe(Effect.flip)
+    expect(error.reason._tag).toBe("InvalidRequest")
+    expect(error.message).toContain("LLMClient.compact")
+  }),
+)
+
+for (const model of [
+  OpenAI.configure({ apiKey: "test" }).responses("gpt-5.3-codex"),
+  Azure.configure({ apiKey: "test", resourceName: "test" }).responses("deployment"),
+  XAI.configure({ apiKey: "test" }).responses("grok-4.6"),
+]) {
+  testEffect(
+    dynamicResponse(({ request, text, respond }) =>
+      Effect.sync(() => {
+        const body = JSON.parse(text)
+        expect(request.method).toBe("POST")
+        expect(request.headers[model.provider === "azure" ? "api-key" : "authorization"]).toBe(
+          model.provider === "azure" ? "test" : "Bearer test",
+        )
+        if (new URL(request.url).pathname.endsWith("/responses/compact")) {
+          expect(body).toEqual({
+            model: model.id,
+            input: [{ role: "user", content: [{ type: "input_text", text: "original" }] }],
+            instructions: "system",
+          })
+          return respond(
+            JSON.stringify({
+              object: "response.compaction",
+              output,
+              usage: { input_tokens: 1000, output_tokens: 10, total_tokens: 1010 },
+            }),
+            { headers: { "content-type": "application/json" } },
+          )
+        }
+        expect(new URL(request.url).pathname.endsWith("/responses")).toBe(true)
+        expect(body.input).toEqual([...output, { role: "user", content: [{ type: "input_text", text: "continue" }] }])
+        return respond(sseEvents({ type: "response.completed", response: { id: "resp_1", output: [] } }), {
+          headers: { "content-type": "text/event-stream" },
+        })
+      }),
+    ),
+  ).effect(`${model.provider} explicitly compacts and replays the entire canonical window`, () =>
+    Effect.gen(function* () {
+      const request = LLM.request({ model, prompt: "original", system: "system", http: { body: { store: false } } })
+      const compacted = yield* LLMClient.compact(request)
+      expect(compacted.usage?.totalTokens).toBe(1010)
+      expect(compacted.messages.map((message) => message.role)).toEqual(["user", "assistant"])
+      expect(compacted.messages[0]?.content).toEqual([{ type: "text", text: "retained" }])
+      expect(compacted.messages[1]?.content).toEqual([
+        { type: "compaction", provider: model.provider, id: "cmp_1", encrypted: "opaque" },
+      ])
+      const codec = Schema.fromJsonString(Schema.Array(Message))
+      const messages = Schema.decodeSync(codec)(Schema.encodeSync(codec)(compacted.messages))
+      yield* LLMClient.generate(LLMRequest.update(request, { messages: [...messages, Message.user("continue")] }))
+    }),
+  )
+}
+
+for (const model of [
+  Anthropic.configure({ apiKey: "test" }).model("claude-opus-4-6"),
+  AmazonBedrockMantle.configure({ apiKey: "test" }).responses("model"),
+]) {
+  testEffect(fixedResponse("must not execute")).effect(
+    `${model.route.id} does not inherit an unsupported compact endpoint`,
+    () =>
+      Effect.gen(function* () {
+        const error = yield* LLMClient.compact(LLM.request({ model, prompt: "hello" })).pipe(Effect.flip)
+        expect(error.reason._tag).toBe("InvalidRequest")
+      }),
+  )
+}
+
+testEffect(
+  fixedResponse(JSON.stringify({ object: "response.compaction", output: [retained], debug: "original payload" })),
+).effect("invalid explicit compaction preserves the original response and HTTP context", () =>
+  Effect.gen(function* () {
+    const error = yield* LLMClient.compact(
+      LLM.request({ model: OpenAI.configure({ apiKey: "test" }).responses("gpt-5.3-codex"), prompt: "hello" }),
+    ).pipe(Effect.flip)
+    expect(error.reason._tag).toBe("InvalidProviderOutput")
+    expect(error.reason.body).toContain("original payload")
+    expect(error.reason.http?.status).toBe(200)
+  }),
+)
@@ -906,6 +906,94 @@ describe("Gemini route", () => {
    }),
  )

+  it.effect("assigns unique ids to separated reasoning blocks", () =>
+    Effect.gen(function* () {
+      const body = sseEvents(
+        {
+          candidates: [
+            {
+              content: {
+                role: "model",
+                parts: [{ text: "A", thought: true, thoughtSignature: "reasoning_sig_a" }],
+              },
+            },
+          ],
+        },
+        {
+          candidates: [
+            {
+              content: { role: "model", parts: [{ text: "X", thoughtSignature: "text_sig_x" }] },
+            },
+          ],
+        },
+        {
+          candidates: [
+            {
+              content: {
+                role: "model",
+                parts: [{ text: "B", thought: true, thoughtSignature: "reasoning_sig_b" }],
+              },
+            },
+          ],
+        },
+        {
+          candidates: [
+            {
+              content: { role: "model", parts: [{ text: "Y", thoughtSignature: "text_sig_y" }] },
+              finishReason: "STOP",
+            },
+          ],
+        },
+      )
+      const response = yield* LLMClient.generate(request).pipe(Effect.provide(fixedResponse(body)))
+      const starts = response.events.filter((event) => event.type === "reasoning-start")
+      const deltas = response.events.filter((event) => event.type === "reasoning-delta")
+      const ends = response.events.filter((event) => event.type === "reasoning-end")
+
+      expect(starts.map((event) => event.id)).toEqual(["reasoning-0", "reasoning-1"])
+      expect(starts[0]?.id).not.toBe(starts[1]?.id)
+      expect(deltas.map((event) => ({ id: event.id, text: event.text }))).toEqual([
+        { id: "reasoning-0", text: "A" },
+        { id: "reasoning-1", text: "B" },
+      ])
+      expect(ends.map((event) => event.id)).toEqual(["reasoning-0", "reasoning-1"])
+      expect(response.events.filter((event) => event.type === "text-start").map((event) => event.id)).toEqual([
+        "text-0",
+        "text-1",
+      ])
+      expect(response.events.filter((event) => event.type === "text-delta").map((event) => event.id)).toEqual([
+        "text-0",
+        "text-1",
+      ])
+      expect(response.events.filter((event) => event.type === "text-end").map((event) => event.id)).toEqual([
+        "text-0",
+        "text-1",
+      ])
+      expect(response.message.content).toEqual([
+        {
+          type: "reasoning",
+          text: "A",
+          providerMetadata: { google: { thoughtSignature: "reasoning_sig_a" } },
+        },
+        {
+          type: "text",
+          text: "X",
+          providerMetadata: { google: { thoughtSignature: "text_sig_x" } },
+        },
+        {
+          type: "reasoning",
+          text: "B",
+          providerMetadata: { google: { thoughtSignature: "reasoning_sig_b" } },
+        },
+        {
+          type: "text",
+          text: "Y",
+          providerMetadata: { google: { thoughtSignature: "text_sig_y" } },
+        },
+      ])
+    }),
+  )
+
  it.effect("ignores unknown response parts", () =>
    Effect.gen(function* () {
      const response = yield* LLMClient.generate(request).pipe(
@@ -1280,6 +1368,56 @@ describe("Gemini route", () => {
    }),
  )

+  it.effect("separates text blocks around streamed tool calls", () =>
+    Effect.gen(function* () {
+      const body = sseEvents({
+        candidates: [
+          {
+            content: {
+              role: "model",
+              parts: [
+                { text: "before" },
+                { functionCall: { id: "call_1", name: "lookup", args: { query: "weather" } } },
+                { text: "after" },
+              ],
+            },
+            finishReason: "STOP",
+          },
+        ],
+      })
+      const response = yield* LLMClient.generate(
+        LLMRequest.update(request, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
+        }),
+      ).pipe(Effect.provide(fixedResponse(body)))
+
+      expect(response.events.slice(1, 8)).toEqual([
+        { type: "text-start", id: "text-0" },
+        { type: "text-delta", id: "text-0", text: "before" },
+        { type: "text-end", id: "text-0" },
+        {
+          type: "tool-call",
+          id: "call_1",
+          name: "lookup",
+          input: { query: "weather" },
+          providerExecuted: undefined,
+          providerMetadata: undefined,
+        },
+        { type: "text-start", id: "text-1" },
+        { type: "text-delta", id: "text-1", text: "after" },
+        { type: "text-end", id: "text-1" },
+      ])
+      const textStarts = response.events.filter((event) => event.type === "text-start")
+      expect(textStarts[0]?.id).not.toBe(textStarts[1]?.id)
+      expect(response.message.content).toEqual([
+        { type: "text", text: "before" },
+        { type: "tool-call", id: "call_1", name: "lookup", input: { query: "weather" } },
+        { type: "text", text: "after" },
+      ])
+      expect(response.finishReason).toEqual({ normalized: "tool-calls", raw: "STOP" })
+    }),
+  )
+
  it.effect("defaults omitted function call args to an empty object", () =>
    Effect.gen(function* () {
      const response = yield* LLMClient.generate(
@@ -0,0 +1,694 @@
+import { describe, expect, test } from "bun:test"
+import { ConfigProvider, Effect } from "effect"
+import { HttpClientRequest } from "effect/unstable/http"
+import { LLM, LLMEvent, Message, ToolDefinition } from "../../src/index.js"
+import { Mistral } from "../../src/providers/index.js"
+import { MistralChat } from "../../src/protocols/index.js"
+import { LLMClient } from "../../src/route.js"
+import { compileRequest } from "../../src/route/client.js"
+import { it } from "../lib/effect.js"
+import { dynamicResponse, fixedResponse } from "../lib/http.js"
+import { sseEvents } from "../lib/sse.js"
+
+const model = Mistral.configure({ apiKey: "fixture" }).model("mistral-large-latest")
+const request = LLM.request({ model, prompt: "Hello" })
+const chunk = (delta: object, finishReason: string | null = null, usage?: object) => ({
+  choices: [{ delta, finish_reason: finishReason }],
+  usage,
+})
+
+describe("Mistral Chat", () => {
+  test("exposes native provider and protocol identities", async () => {
+    const entrypoint = await import("@opencode-ai/ai/providers/mistral")
+
+    expect(Mistral.id).toBe("mistral")
+    expect(MistralChat.protocol.id).toBe("mistral-chat")
+    expect(Mistral.route).toMatchObject({
+      id: "mistral-chat",
+      provider: "mistral",
+      providerMetadataKey: "mistral",
+      protocol: "mistral-chat",
+    })
+    expect(Mistral.route.endpoint).toMatchObject({
+      baseURL: "https://api.mistral.ai/v1",
+      path: "/chat/completions",
+    })
+    expect(entrypoint.model).toBeFunction()
+  })
+
+  it.effect("lowers native messages, media, tool choice, options, and replay IDs", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          system: "Initial",
+          messages: [
+            Message.system("Updated"),
+            Message.user([
+              { type: "text", text: "Inspect" },
+              { type: "media", mediaType: "image/png", data: "aW1hZ2U=" },
+              { type: "media", mediaType: "application/pdf", data: "cGRm" },
+            ]),
+            Message.assistant([
+              { type: "reasoning", text: "Think" },
+              { type: "text", text: "Calling" },
+              { type: "tool-call", id: "call.same-prefix-1", name: "lookup", input: { city: "Paris" } },
+              { type: "tool-call", id: "call.same-prefix-2", name: "other", input: {} },
+            ]),
+            Message.tool({ id: "call.same-prefix-1", name: "lookup", result: { ok: true } }),
+          ],
+          tools: [
+            ToolDefinition.make({ name: "lookup", description: "Look up a city", inputSchema: { type: "object" } }),
+            ToolDefinition.make({ name: "other", description: "Other operation", inputSchema: { type: "object" } }),
+          ],
+          toolChoice: "lookup",
+          promptCacheKey: "session-1",
+          generation: {
+            maxTokens: 64,
+            seed: 7,
+            temperature: 0.2,
+            topP: 0.8,
+            frequencyPenalty: 0.1,
+            presencePenalty: 0.3,
+            stop: ["done"],
+          },
+          providerOptions: {
+            safePrompt: true,
+            documentImageLimit: 3,
+            documentPageLimit: 8,
+            parallelToolCalls: false,
+            reasoningEffort: "high",
+          },
+        }),
+      )
+
+      expect(prepared.body).toMatchObject({
+        model: "mistral-large-latest",
+        tools: [{ function: { name: "lookup", strict: false } }, { function: { name: "other", strict: false } }],
+        tool_choice: { type: "function", function: { name: "lookup" } },
+        stream: true,
+        max_tokens: 64,
+        random_seed: 7,
+        temperature: 0.2,
+        top_p: 0.8,
+        frequency_penalty: 0.1,
+        presence_penalty: 0.3,
+        stop: ["done"],
+        prompt_cache_key: "session-1",
+        safe_prompt: true,
+        document_image_limit: 3,
+        document_page_limit: 8,
+        parallel_tool_calls: false,
+        reasoning_effort: "high",
+      })
+      expect(prepared.body.messages.slice(0, 4)).toMatchObject([
+        { role: "system", content: "Initial" },
+        { role: "user", content: "<system-update>\nUpdated\n</system-update>" },
+        {
+          role: "user",
+          content: [
+            { type: "text", text: "Inspect" },
+            { type: "image_url", image_url: "data:image/png;base64,aW1hZ2U=" },
+            { type: "document_url", document_url: "data:application/pdf;base64,cGRm" },
+          ],
+        },
+        {
+          role: "assistant",
+          content: "ThinkCalling",
+        },
+      ])
+      const assistant = prepared.body.messages[3]
+      const toolResult = prepared.body.messages[4]
+      expect(assistant?.role).toBe("assistant")
+      expect(toolResult?.role).toBe("tool")
+      if (assistant?.role !== "assistant" || toolResult?.role !== "tool") return
+      const ids = assistant.tool_calls?.map((tool) => tool.id) ?? []
+      expect(ids).toHaveLength(2)
+      expect(ids[0]).toMatch(/^[A-Za-z0-9]{9}$/)
+      expect(ids[1]).toMatch(/^[A-Za-z0-9]{9}$/)
+      expect(ids[0]).not.toBe(ids[1])
+      expect(toolResult.tool_call_id).toBe(ids[0])
+      expect(toolResult.name).toBe("lookup")
+    }),
+  )
+
+  it.effect("preserves valid replay IDs", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant({ type: "tool-call", id: "Ab12Cd34E", name: "lookup", input: {} }),
+            Message.tool({ id: "Ab12Cd34E", name: "lookup", result: "ok" }),
+          ],
+        }),
+      )
+      expect(prepared.body.messages).toMatchObject([
+        { tool_calls: [{ id: "Ab12Cd34E" }] },
+        { tool_call_id: "Ab12Cd34E" },
+      ])
+    }),
+  )
+
+  it.effect("applies trailing prefix, cache, and reasoning options without changing earlier assistants", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          promptCacheKey: "common-key",
+          messages: [Message.assistant("Earlier"), Message.user("Continue"), Message.assistant("Prefix")],
+          providerOptions: { promptCacheKey: "native-key", promptMode: "reasoning" },
+        }),
+      )
+      expect(prepared.body.prompt_cache_key).toBe("native-key")
+      expect(prepared.body.prompt_mode).toBe("reasoning")
+      expect(prepared.body.messages).toEqual([
+        { role: "assistant", content: "Earlier" },
+        { role: "user", content: "Continue" },
+        { role: "assistant", content: "Prefix", prefix: true },
+      ])
+
+      const uncached = yield* compileRequest(
+        LLM.request({
+          model,
+          prompt: "Hello",
+          promptCacheKey: "common-key",
+          cache: "none",
+          providerOptions: { promptCacheKey: "native-key" },
+        }),
+      )
+      expect(uncached.body.prompt_cache_key).toBeUndefined()
+
+      const longKey = "cache-key-".repeat(10)
+      const unbounded = yield* compileRequest(
+        LLM.request({
+          model,
+          prompt: "Hello",
+          promptCacheKey: longKey,
+        }),
+      )
+      expect(unbounded.body.prompt_cache_key).toBe(longKey)
+
+      const conflict = yield* compileRequest(
+        LLM.request({
+          model,
+          prompt: "Hello",
+          providerOptions: { reasoningEffort: "high", promptMode: "reasoning" },
+        }),
+      ).pipe(Effect.flip)
+      expect(conflict.message).toContain("mutually exclusive")
+    }),
+  )
+
+  it.effect("omits empty assistant history unless it carries a tool call", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.assistant(" \n "),
+            Message.assistant({ type: "reasoning", text: "\t" }),
+            Message.assistant({ type: "tool-call", id: "Ab12Cd34E", name: "lookup", input: {} }),
+          ],
+        }),
+      )
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: "",
+          tool_calls: [{ id: "Ab12Cd34E", type: "function", function: { name: "lookup", arguments: "{}" } }],
+        },
+      ])
+    }),
+  )
+
+  it.effect("preserves remote media URLs and structured tool-result media", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.user({
+              type: "media",
+              mediaType: "image/png",
+              data: "https://assets.example.test/input.png",
+            }),
+            Message.tool({
+              id: "Ab12Cd34E",
+              name: "inspect",
+              resultType: "content",
+              result: [
+                { type: "text", text: "Result" },
+                { type: "file", mime: "image/jpeg", uri: "https://assets.example.test/output.jpg" },
+                { type: "file", mime: "application/pdf", uri: "cGRm" },
+              ],
+            }),
+          ],
+        }),
+      )
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "user",
+          content: [{ type: "image_url", image_url: "https://assets.example.test/input.png" }],
+        },
+        {
+          role: "tool",
+          tool_call_id: "Ab12Cd34E",
+          name: "inspect",
+          content: [
+            { type: "text", text: "Result" },
+            { type: "image_url", image_url: "https://assets.example.test/output.jpg" },
+            { type: "document_url", document_url: "data:application/pdf;base64,cGRm" },
+          ],
+        },
+      ])
+    }),
+  )
+
+  it.effect("concatenates text-only user and tool content without separators", () =>
+    Effect.gen(function* () {
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [
+            Message.user([
+              { type: "text", text: "first" },
+              { type: "text", text: "second" },
+            ]),
+            Message.tool({
+              id: "Ab12Cd34E",
+              name: "lookup",
+              resultType: "content",
+              result: [
+                { type: "text", text: "third" },
+                { type: "text", text: "fourth" },
+              ],
+            }),
+          ],
+        }),
+      )
+      expect(prepared.body.messages).toEqual([
+        { role: "user", content: "firstsecond" },
+        { role: "tool", tool_call_id: "Ab12Cd34E", name: "lookup", content: "thirdfourth" },
+      ])
+    }),
+  )
+
+  it.effect("streams ordered thinking and text and replays native thinking metadata", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              chunk({ content: [{ type: "thinking", thinking: [], marker: "empty" }] }),
+              chunk({ content: [{ type: "thinking", thinking: [{ type: "text", text: "Consider" }] }] }),
+              chunk({ content: [{ type: "text", text: "Answer" }] }),
+              chunk({}, "stop"),
+            ),
+          ),
+        ),
+      )
+
+      expect(response.reasoning).toBe("Consider")
+      expect(response.text).toBe("Answer")
+      expect(response.message.content).toEqual([
+        {
+          type: "reasoning",
+          text: "Consider",
+          providerMetadata: {
+            mistral: {
+              thinking: {
+                type: "thinking",
+                thinking: [{ type: "text", text: "Consider" }],
+                marker: "empty",
+              },
+            },
+          },
+        },
+        { type: "text", text: "Answer" },
+      ])
+
+      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
+      expect(replay.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [
+            {
+              type: "thinking",
+              thinking: [{ type: "text", text: "Consider" }],
+              marker: "empty",
+            },
+            { type: "text", text: "Answer" },
+          ],
+          prefix: true,
+        },
+      ])
+    }),
+  )
+
+  it.effect("replays metadata-only native thinking", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(chunk({ content: [{ type: "thinking", thinking: [], marker: "opaque" }] }), chunk({}, "stop")),
+          ),
+        ),
+      )
+      expect(response.message.content).toEqual([
+        {
+          type: "reasoning",
+          text: "",
+          providerMetadata: {
+            mistral: { thinking: { type: "thinking", thinking: [], marker: "opaque" } },
+          },
+        },
+      ])
+
+      const replay = yield* compileRequest(LLM.request({ model, messages: [response.message] }))
+      expect(replay.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: [{ type: "thinking", thinking: [], marker: "opaque" }],
+          prefix: true,
+        },
+      ])
+    }),
+  )
+
+  it.effect("merges indexed argument fragments with missing continuation identity", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              chunk({
+                tool_calls: [{ index: 0, id: "Ab12Cd34E", function: { name: "lookup", arguments: '{"city":' } }],
+              }),
+              chunk({ tool_calls: [{ index: 0, function: { name: "", arguments: '"Paris"}' } }] }),
+              chunk({}, "tool_calls"),
+            ),
+          ),
+        ),
+      )
+
+      expect(response.message.content).toContainEqual({
+        type: "tool-call",
+        id: "Ab12Cd34E",
+        name: "lookup",
+        input: { city: "Paris" },
+      })
+      expect(
+        response.events.filter(
+          (event) =>
+            LLMEvent.is.toolInputStart(event) ||
+            LLMEvent.is.toolInputDelta(event) ||
+            LLMEvent.is.toolInputEnd(event) ||
+            LLMEvent.is.toolCall(event),
+        ),
+      ).toEqual([
+        { type: "tool-input-start", id: "Ab12Cd34E", name: "lookup", providerMetadata: undefined },
+        {
+          type: "tool-input-delta",
+          id: "Ab12Cd34E",
+          name: "lookup",
+          text: '{"city":',
+          input: {},
+        },
+        {
+          type: "tool-input-delta",
+          id: "Ab12Cd34E",
+          name: "lookup",
+          text: '"Paris"}',
+          input: { city: "Paris" },
+        },
+        { type: "tool-input-end", id: "Ab12Cd34E", name: "lookup", providerMetadata: undefined },
+        {
+          type: "tool-call",
+          id: "Ab12Cd34E",
+          name: "lookup",
+          input: { city: "Paris" },
+          providerExecuted: undefined,
+          providerMetadata: undefined,
+        },
+      ])
+      expect(response.events.filter(LLMEvent.is.toolCall)).toHaveLength(1)
+    }),
+  )
+
+  it.effect("normalizes stop to tool calls when a hosted model emits indexed tool fragments", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              chunk({
+                tool_calls: [
+                  {
+                    index: 0,
+                    id: "chatcmpl-tool-8cc4d8f9f07b298a",
+                    function: { name: "lookup", arguments: '{"city":"' },
+                  },
+                ],
+              }),
+              chunk({ tool_calls: [{ index: 0, function: { name: "", arguments: 'Paris"}' } }] }),
+              chunk({}, "stop"),
+            ),
+          ),
+        ),
+      )
+
+      expect(response.finishReason).toEqual({ normalized: "tool-calls", raw: "stop" })
+      expect(response.toolCalls).toMatchObject([{ name: "lookup", input: { city: "Paris" } }])
+      expect(response.events.filter(LLMEvent.is.toolCall)).toHaveLength(1)
+    }),
+  )
+
+  it.effect("generates a stable ID when the first indexed fragment has null identity", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              chunk({
+                tool_calls: [{ index: 0, id: null, function: { name: "lookup", arguments: { city: "Paris" } } }],
+              }),
+              chunk({}, "tool_calls"),
+            ),
+          ),
+        ),
+      )
+      const call = response.message.content.find((part) => part.type === "tool-call")
+      expect(call?.id).toMatch(/^[A-Za-z0-9]{9}$/)
+      expect(call).toMatchObject({ name: "lookup", input: { city: "Paris" } })
+    }),
+  )
+
+  it.effect("generates distinct IDs for parallel null and literal-null identities", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              chunk({
+                tool_calls: [
+                  { index: 0, id: null, function: { name: "first", arguments: {} } },
+                  { index: 1, id: "null", function: { name: "second", arguments: {} } },
+                ],
+              }),
+              chunk({}, "tool_calls"),
+            ),
+          ),
+        ),
+      )
+      const calls = response.message.content.filter((part) => part.type === "tool-call")
+      expect(calls).toHaveLength(2)
+      expect(calls[0]?.id).toMatch(/^[A-Za-z0-9]{9}$/)
+      expect(calls[1]?.id).toMatch(/^[A-Za-z0-9]{9}$/)
+      expect(calls[0]?.id).not.toBe(calls[1]?.id)
+    }),
+  )
+
+  it.effect("keeps parallel indexed calls independent", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              chunk({
+                tool_calls: [
+                  { index: 0, id: "Ab12Cd34E", function: { name: "first", arguments: '{"n":' } },
+                  { index: 1, id: "Fg56Hi78J", function: { name: "second", arguments: '{"n":' } },
+                ],
+              }),
+              chunk({
+                tool_calls: [
+                  { index: 0, function: { arguments: "1}" } },
+                  { index: 1, function: { arguments: "2}" } },
+                ],
+              }),
+              chunk({}, "tool_calls"),
+            ),
+          ),
+        ),
+      )
+      expect(response.message.content.filter((part) => part.type === "tool-call")).toEqual([
+        { type: "tool-call", id: "Ab12Cd34E", name: "first", input: { n: 1 } },
+        { type: "tool-call", id: "Fg56Hi78J", name: "second", input: { n: 2 } },
+      ])
+    }),
+  )
+
+  it.effect("correlates parallel identity-less fragments by batch position", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              chunk({
+                tool_calls: [
+                  { function: { name: "first", arguments: '{"n":' } },
+                  { function: { name: "second", arguments: '{"n":' } },
+                ],
+              }),
+              chunk({
+                tool_calls: [{ function: { arguments: "1}" } }, { function: { arguments: "2}" } }],
+              }),
+              chunk({}, "tool_calls"),
+            ),
+          ),
+        ),
+      )
+      expect(response.message.content.filter((part) => part.type === "tool-call")).toMatchObject([
+        { name: "first", input: { n: 1 } },
+        { name: "second", input: { n: 2 } },
+      ])
+    }),
+  )
+
+  it.effect("maps usage variants and clamps cache reads", () =>
+    Effect.gen(function* () {
+      for (const usage of [
+        { prompt_tokens: 5, completion_tokens: 2, total_tokens: 7, num_cached_tokens: 9 },
+        { prompt_tokens: 5, completion_tokens: 2, prompt_token_details: { cached_tokens: 2 } },
+        { prompt_tokens: 5, completion_tokens: 2, prompt_tokens_details: { cached_tokens: 3 } },
+      ]) {
+        const response = yield* LLMClient.generate(request).pipe(
+          Effect.provide(fixedResponse(sseEvents(chunk({}, "stop", usage)))),
+        )
+        expect(response.usage).toMatchObject({
+          inputTokens: 5,
+          outputTokens: 2,
+          totalTokens: 7,
+        })
+        expect(response.usage?.cacheReadInputTokens).toBe(
+          Math.min(
+            5,
+            usage.num_cached_tokens ??
+              usage.prompt_token_details?.cached_tokens ??
+              usage.prompt_tokens_details?.cached_tokens ??
+              0,
+          ),
+        )
+      }
+    }),
+  )
+
+  it.effect("maps finish reasons and does not finalize truncated tool calls", () =>
+    Effect.gen(function* () {
+      for (const [raw, normalized] of [
+        ["stop", "stop"],
+        ["model_length", "length"],
+        ["tool_calls", "tool-calls"],
+        ["error", "error"],
+        ["future_reason", "unknown"],
+      ] as const) {
+        const response = yield* LLMClient.generate(request).pipe(
+          Effect.provide(fixedResponse(sseEvents(chunk({}, raw)))),
+        )
+        expect(response.finishReason).toEqual({ normalized, raw })
+      }
+
+      const truncated = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              chunk({
+                tool_calls: [{ index: 0, id: "Ab12Cd34E", function: { name: "lookup", arguments: '{"city":' } }],
+              }),
+              chunk({}, "length"),
+            ),
+          ),
+        ),
+      )
+      expect(truncated.finishReason).toEqual({ normalized: "length", raw: "length" })
+      expect(truncated.events.some(LLMEvent.is.toolCall)).toBe(false)
+      expect(truncated.events.some(LLMEvent.is.toolInputEnd)).toBe(false)
+    }),
+  )
+
+  it.effect("ignores non-text output parts and rejects invalid stream endings", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              chunk({ content: null }),
+              chunk({
+                content: [
+                  { type: "reference", reference_ids: [1] },
+                  { type: "image_url", image_url: "https://example.test/image.png" },
+                  { type: "text", text: "Answer" },
+                ],
+              }),
+              chunk({}, "stop"),
+            ),
+          ),
+        ),
+      )
+      expect(response.text).toBe("Answer")
+
+      const missingFinish = yield* LLMClient.generate(request).pipe(
+        Effect.provide(fixedResponse(sseEvents(chunk({ content: "partial" })))),
+        Effect.flip,
+      )
+      expect(missingFinish.message).toContain("without finish_reason")
+
+      const lateContent = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(sseEvents(chunk({}, "stop"), chunk({ content: [{ type: "text", text: "late" }] }))),
+        ),
+        Effect.flip,
+      )
+      expect(lateContent.message).toContain("content after the finish reason")
+    }),
+  )
+
+  it.effect("uses environment bearer auth and custom package settings", () =>
+    LLMClient.generate(
+      LLM.request({
+        model: Mistral.model("fixture-model", {
+          baseURL: "https://mistral.test/v1",
+          headers: { "x-app": "test" },
+          body: { service_tier: "priority" },
+          providerOptions: { safePrompt: true },
+        }),
+        prompt: "Hello",
+      }),
+    ).pipe(
+      Effect.provide(
+        dynamicResponse((input) =>
+          Effect.gen(function* () {
+            const web = yield* HttpClientRequest.toWeb(input.request).pipe(Effect.orDie)
+            expect(web.url).toBe("https://mistral.test/v1/chat/completions")
+            expect(web.headers.get("authorization")).toBe("Bearer secret")
+            expect(web.headers.get("x-app")).toBe("test")
+            expect(input.text).toContain('"service_tier":"priority"')
+            return input.respond(sseEvents(chunk({}, "stop")), { headers: { "content-type": "text/event-stream" } })
+          }),
+        ),
+      ),
+      Effect.provide(ConfigProvider.layer(ConfigProvider.fromEnv({ env: { MISTRAL_API_KEY: "secret" } }))),
+    ),
+  )
+})
@@ -0,0 +1,159 @@
+import { configure } from "@opencode-ai/ai/providers/mistral"
+import { describe, expect } from "bun:test"
+import { Effect } from "effect"
+import { LLM, LLMEvent, LLMRequest, Message, ToolChoice, ToolDefinition } from "../../src/index.js"
+import { LLMClient } from "../../src/route.js"
+import { compileRequest } from "../../src/route/client.js"
+import { recordedTests } from "../recorded-test.js"
+
+const apiKey = process.env.MISTRAL_API_KEY ?? "fixture"
+const recorded = recordedTests({
+  prefix: "mistral-chat",
+  provider: "mistral",
+  protocol: "mistral-chat",
+  requires: ["MISTRAL_API_KEY"],
+})
+const glmRecorded = recordedTests({
+  prefix: "mistral-chat-glm",
+  provider: "mistral",
+  protocol: "mistral-chat",
+  requires: ["MISTRAL_API_KEY"],
+})
+
+const weather = ToolDefinition.make({
+  name: "lookup_weather",
+  description: "Look up the current weather for a city",
+  inputSchema: {
+    type: "object",
+    properties: { city: { type: "string", enum: ["Paris"] } },
+    required: ["city"],
+    additionalProperties: false,
+  },
+})
+
+describe("Mistral recorded", () => {
+  recorded.effect.with(
+    "streams text with usage",
+    { tags: ["text", "usage"], metadata: { model: "mistral-small-latest" } },
+    () =>
+      Effect.gen(function* () {
+        const response = yield* LLMClient.generate(
+          LLM.request({
+            model: configure({ apiKey, providerOptions: { reasoningEffort: "none" } }).model("mistral-small-latest"),
+            prompt: "Reply with exactly one word: hello",
+            generation: { maxTokens: 40, temperature: 0 },
+          }),
+        )
+
+        expect(response.text.trim()).toMatch(/^(?:hello|hi)[!.]?$/i)
+        expect(response.finishReason.normalized).toBe("stop")
+        expect(response.usage?.inputTokens).toBeGreaterThan(0)
+        expect(response.usage?.outputTokens).toBeGreaterThan(0)
+      }),
+    60_000,
+  )
+
+  recorded.effect.with(
+    "replays native reasoning",
+    { tags: ["reasoning", "replay", "usage"], metadata: { model: "mistral-small-latest" } },
+    () =>
+      Effect.gen(function* () {
+        const model = configure({ apiKey, providerOptions: { reasoningEffort: "high" } }).model("mistral-small-latest")
+        const firstRequest = LLM.request({
+          model,
+          prompt: "Calculate 17 multiplied by 23. Think briefly, then reply with only the integer.",
+          generation: { maxTokens: 512, temperature: 0 },
+        })
+        const first = yield* LLMClient.generate(firstRequest)
+
+        expect(first.text.trim()).toBe("391")
+        expect(first.reasoning.length).toBeGreaterThan(0)
+        expect(first.events.some(LLMEvent.is.reasoningDelta)).toBe(true)
+
+        const followUp = LLMRequest.update(firstRequest, {
+          messages: [...firstRequest.messages, first.message, Message.user("Reply with exactly: Done.")],
+          generation: { maxTokens: 256, temperature: 0 },
+        })
+        const replay = yield* compileRequest(followUp)
+        expect(replay.body.messages).toContainEqual(
+          expect.objectContaining({
+            role: "assistant",
+            content: expect.arrayContaining([expect.objectContaining({ type: "thinking" })]),
+          }),
+        )
+
+        const second = yield* LLMClient.generate(followUp)
+        expect(second.text.trim()).toMatch(/Done\.?$/)
+        expect(second.finishReason.normalized).toBe("stop")
+      }),
+    60_000,
+  )
+
+  recorded.effect.with(
+    "drives a tool loop",
+    { tags: ["tool", "tool-loop", "usage"], metadata: { model: "mistral-small-latest" } },
+    () =>
+      Effect.gen(function* () {
+        const model = configure({ apiKey, providerOptions: { reasoningEffort: "none" } }).model("mistral-small-latest")
+        const firstRequest = LLM.request({
+          model,
+          system: "Call lookup_weather exactly once with Paris.",
+          prompt: "What is the weather?",
+          tools: [weather],
+          toolChoice: weather,
+          generation: { maxTokens: 160, temperature: 0 },
+        })
+        const first = yield* LLMClient.generate(firstRequest)
+
+        expect(first.finishReason.normalized).toBe("tool-calls")
+        expect(first.toolCalls).toMatchObject([{ name: "lookup_weather", input: { city: "Paris" } }])
+        expect(first.events.filter(LLMEvent.is.toolCall)).toHaveLength(1)
+
+        const call = first.toolCalls[0]
+        if (!call) throw new Error("Mistral did not return a tool call")
+        const followUp = LLMRequest.update(firstRequest, {
+          toolChoice: ToolChoice.make("none"),
+          messages: [
+            ...firstRequest.messages,
+            first.message,
+            Message.tool({ id: call.id, name: call.name, result: { condition: "sunny", temperature: "18C" } }),
+          ],
+          generation: { maxTokens: 160, temperature: 0 },
+        })
+        const second = yield* LLMClient.generate(followUp)
+
+        expect(second.finishReason.normalized).toBe("stop")
+        expect(second.toolCalls).toHaveLength(0)
+        expect(second.text.toLowerCase()).toContain("sunny")
+      }),
+    60_000,
+  )
+})
+
+describe("Mistral hosted GLM recorded", () => {
+  glmRecorded.effect.with(
+    "streams an indexed tool call",
+    { tags: ["hosted-model", "tool", "tool-call"], metadata: { model: "zai-glm-5-2" } },
+    () =>
+      Effect.gen(function* () {
+        const response = yield* LLMClient.generate(
+          LLM.request({
+            model: configure({ apiKey }).model("zai-glm-5-2"),
+            system: "Call lookup_weather exactly once with Paris.",
+            prompt: "What is the weather?",
+            tools: [weather],
+            toolChoice: weather,
+            generation: { maxTokens: 256, temperature: 0 },
+          }),
+        )
+
+        expect(response.finishReason.normalized).toBe("tool-calls")
+        expect(response.toolCalls).toMatchObject([{ name: "lookup_weather", input: { city: "Paris" } }])
+        expect(response.events.filter(LLMEvent.is.toolInputStart)).toHaveLength(1)
+        expect(response.events.filter(LLMEvent.is.toolInputDelta).length).toBeGreaterThan(0)
+        expect(response.events.filter(LLMEvent.is.toolInputEnd)).toHaveLength(1)
+        expect(response.events.filter(LLMEvent.is.toolCall)).toHaveLength(1)
+      }),
+    60_000,
+  )
+})
@@ -0,0 +1,209 @@
+import { describe, expect } from "bun:test"
+import { Effect } from "effect"
+import { LLM, LLMEvent } from "../../src/index.js"
+import { OpenResponses } from "../../src/protocols/open-responses.js"
+import { configure } from "../../src/providers/openai-compatible-responses.js"
+import { LLMClient } from "../../src/route.js"
+import { it } from "../lib/effect.js"
+import { fixedResponse } from "../lib/http.js"
+import { sseEvents } from "../lib/sse.js"
+
+const request = LLM.request({
+  model: configure({ apiKey: "test-key", baseURL: "https://responses.example.test/v1" }).model("example-model"),
+  prompt: "Respond.",
+})
+const completed = { type: "response.completed", response: { id: "resp_1" } }
+const generate = (...events: OpenResponses.Event[]) =>
+  LLMClient.generate(request).pipe(Effect.provide(fixedResponse(sseEvents(...events))))
+
+describe("Open Responses completed item text", () => {
+  ;["Draft expanded", "D", "Replacement", ""].forEach((text) => {
+    it.effect(`replaces streamed text with completed item text ${JSON.stringify(text)}`, () =>
+      Effect.gen(function* () {
+        const response = yield* generate(
+          { type: "response.output_item.added", item: { type: "message", id: "msg_1", phase: "commentary" } },
+          { type: "response.output_text.delta", item_id: "msg_1", delta: "Draft" },
+          { type: "response.output_text.done", item_id: "msg_1", text: "Part final" },
+          {
+            type: "response.output_item.done",
+            item: { type: "message", id: "msg_1", phase: "final_answer", content: [{ type: "output_text", text }] },
+          },
+          completed,
+        )
+        expect(response.text).toBe(text)
+        expect(response.events.filter(LLMEvent.is.textDelta).map((event) => event.text)).toEqual(["Draft"])
+        expect(response.events.filter(LLMEvent.is.textEnd)).toEqual([
+          {
+            type: "text-end",
+            id: "msg_1",
+            text,
+            providerMetadata: { "openai-compatible": { itemId: "msg_1", phase: "final_answer" } },
+          },
+        ])
+      }),
+    )
+  })
+
+  it.effect("joins completed text and refusal parts without streamed text", () =>
+    Effect.gen(function* () {
+      const response = yield* generate(
+        { type: "response.output_item.added", item: { type: "message", id: "msg_1" } },
+        {
+          type: "response.output_item.done",
+          item: {
+            type: "message",
+            id: "msg_1",
+            content: [
+              { type: "output_text", text: "Answer. " },
+              { type: "refusal", refusal: "Cannot help." },
+            ],
+          },
+        },
+        completed,
+      )
+      expect(response.text).toBe("Answer. Cannot help.")
+      expect(response.events.filter(LLMEvent.is.textStart)).toHaveLength(1)
+      expect(response.events.filter(LLMEvent.is.textEnd)).toHaveLength(1)
+    }),
+  )
+
+  it.effect("does not create an empty text fragment for an empty completed message", () =>
+    Effect.gen(function* () {
+      const response = yield* generate(
+        { type: "response.output_item.added", item: { type: "message", id: "msg_1" } },
+        { type: "response.output_text.done", item_id: "msg_1", text: "" },
+        {
+          type: "response.output_item.done",
+          item: { type: "message", id: "msg_1", content: [{ type: "output_text", text: "" }] },
+        },
+        completed,
+      )
+      expect(response.message.content).toEqual([])
+      expect(response.events.filter(LLMEvent.is.textStart)).toEqual([])
+    }),
+  )
+})
+
+describe("Open Responses completed item reasoning", () => {
+  ;[
+    {
+      name: "summary",
+      summary: [
+        { type: "summary_text", text: "Final" },
+        { type: "summary_text", text: "summary" },
+      ],
+      content: [{ type: "reasoning_text", text: "Raw" }],
+      text: "Final\n\nsummary",
+    },
+    {
+      name: "raw text",
+      summary: [
+        { type: "summary_text", text: "" },
+        { type: "summary_text", text: "" },
+      ],
+      content: [{ type: "reasoning_text", text: "Raw" }],
+      text: "Raw",
+    },
+    {
+      name: "streamed fallback",
+      summary: [
+        { type: "summary_text", text: "" },
+        { type: "summary_text", text: "" },
+      ],
+      content: [
+        { type: "reasoning_text", text: "" },
+        { type: "reasoning_text", text: "" },
+      ],
+      text: "Draft",
+    },
+  ].forEach((fixture) => {
+    it.effect(`uses ${fixture.name} at item completion`, () =>
+      Effect.gen(function* () {
+        const response = yield* generate(
+          { type: "response.output_item.added", item: { type: "reasoning", id: "rs_1" } },
+          { type: "response.reasoning_summary_text.delta", item_id: "rs_1", delta: "Draft" },
+          { type: "response.reasoning_summary_text.done", item_id: "rs_1", text: "Part final" },
+          {
+            type: "response.output_item.done",
+            item: {
+              type: "reasoning",
+              id: "rs_1",
+              summary: fixture.summary,
+              content: fixture.content,
+              encrypted_content: "encrypted",
+            },
+          },
+          completed,
+        )
+        expect(response.reasoning).toBe(fixture.text)
+        expect(response.events.filter(LLMEvent.is.reasoningEnd)).toHaveLength(1)
+        expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
+          "openai-compatible": { itemId: "rs_1", reasoningEncryptedContent: "encrypted" },
+        })
+      }),
+    )
+  })
+
+  it.effect("replaces only the still-open summary without repeating earlier text", () =>
+    Effect.gen(function* () {
+      const response = yield* generate(
+        { type: "response.output_item.added", item: { type: "reasoning", id: "rs_1" } },
+        { type: "response.reasoning_summary_text.delta", item_id: "rs_1", summary_index: 0, delta: "First " },
+        { type: "response.reasoning_summary_part.added", item_id: "rs_1", summary_index: 1 },
+        { type: "response.reasoning_summary_text.delta", item_id: "rs_1", summary_index: 1, delta: "draft" },
+        {
+          type: "response.output_item.done",
+          item: {
+            type: "reasoning",
+            id: "rs_1",
+            summary: [
+              { type: "summary_text", text: "First " },
+              { type: "summary_text", text: "final" },
+            ],
+          },
+        },
+        completed,
+      )
+      expect(response.reasoning).toBe("First final")
+      expect(response.events.filter(LLMEvent.is.reasoningEnd).map((event) => event.text)).toEqual([undefined, "final"])
+    }),
+  )
+})
+;["response.completed", "response.incomplete"].forEach((type) => {
+  it.effect(`keeps streamed text when part finals are followed by ${type} without item completion`, () =>
+    Effect.gen(function* () {
+      const response = yield* generate(
+        { type: "response.output_item.added", item: { type: "message", id: "msg_1" } },
+        { type: "response.output_text.delta", item_id: "msg_1", content_index: 0, delta: "Hel" },
+        { type: "response.output_text.delta", item_id: "msg_1", content_index: 1, delta: "world" },
+        { type: "response.output_text.done", item_id: "msg_1", content_index: 0, text: "Hello " },
+        {
+          type: "response.content_part.done",
+          item_id: "msg_1",
+          content_index: 0,
+          part: { type: "output_text", text: "Hello " },
+        },
+        { type: "response.output_item.added", item: { type: "reasoning", id: "rs_1" } },
+        { type: "response.reasoning_summary_text.delta", item_id: "rs_1", delta: "Draft" },
+        { type: "response.reasoning_summary_text.done", item_id: "rs_1", text: "Part final" },
+        {
+          type: "response.reasoning_summary_part.done",
+          item_id: "rs_1",
+          summary_index: 0,
+          part: { type: "summary_text", text: "Part final" },
+        },
+        {
+          type,
+          response: {
+            id: "resp_1",
+            incomplete_details: type === "response.incomplete" ? { reason: "max_output_tokens" } : undefined,
+          },
+        },
+      )
+      expect(response.text).toBe("Helworld")
+      expect(response.reasoning).toBe("Draft")
+      expect(response.events.filter(LLMEvent.is.textEnd).map((event) => event.text)).toEqual([undefined])
+      expect(response.events.filter(LLMEvent.is.reasoningEnd).map((event) => event.text)).toEqual([undefined])
+    }),
+  )
+})
@@ -129,7 +129,7 @@ describe("Open Responses basic-item lifecycles", () => {
    }),
  )

-  it.effect("preserves done-only encrypted reasoning without replaying its summary or late events", () =>
+  it.effect("preserves done-only reasoning text and encryption without replaying late events", () =>
    Effect.gen(function* () {
      const item = {
        type: "reasoning",
@@ -157,6 +157,7 @@ describe("Open Responses basic-item lifecycles", () => {
        {
          type: "reasoning-end",
          id: "rs_1",
+          text: "Not streamed",
          providerMetadata: { "openai-compatible": { itemId: "rs_1", reasoningEncryptedContent: "encrypted-state" } },
        },
      ])
@@ -301,7 +302,7 @@ describe("Open Responses basic-item lifecycles", () => {
    )
  })

-  it.effect("recovers pending items in completed output order with terminal encrypted metadata", () =>
+  it.effect("recovers pending calls without reconciling terminal reasoning", () =>
    Effect.gen(function* () {
      const events = yield* collect(
        {
@@ -325,11 +326,6 @@ describe("Open Responses basic-item lifecycles", () => {
        },
      )
      expect(events.slice(5, -2)).toEqual([
-        {
-          type: "reasoning-end",
-          id: "rs_1:0",
-          providerMetadata: { "openai-compatible": { itemId: "rs_1", reasoningEncryptedContent: "terminal-state" } },
-        },
        {
          type: "tool-input-end",
          id: "call_1",
@@ -341,8 +337,10 @@ describe("Open Responses basic-item lifecycles", () => {
          id: "call_1",
          name: "lookup",
          input: { query: "final" },
+          providerExecuted: undefined,
          providerMetadata: { "openai-compatible": { itemId: "fc_1" } },
        },
+        { type: "reasoning-end", id: "rs_1:0" },
      ])
    }),
  )
@@ -1460,6 +1460,162 @@ describe("OpenAI Chat route", () => {
    }),
  )

+  it.effect("does not finalize streamed tool calls when content is filtered", () =>
+    Effect.gen(function* () {
+      const body = sseEvents(
+        deltaChunk({
+          tool_calls: [{ index: 0, id: "call_1", function: { name: "lookup", arguments: '{"query":"weather"' } }],
+        }),
+        deltaChunk({}, "content_filter"),
+      )
+      const response = yield* LLMClient.generate(
+        LLMRequest.update(request, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
+        }),
+      ).pipe(Effect.provide(fixedResponse(body)))
+
+      expect(response.events).toEqual([
+        { type: "step-start", index: 0 },
+        {
+          type: "tool-input-start",
+          id: "call_1",
+          name: "lookup",
+          providerExecuted: undefined,
+          providerMetadata: undefined,
+        },
+        {
+          type: "tool-input-delta",
+          id: "call_1",
+          name: "lookup",
+          text: '{"query":"weather"',
+          input: { query: "weather" },
+        },
+        {
+          type: "step-finish",
+          index: 0,
+          reason: { normalized: "content-filter", raw: "content_filter" },
+          usage: undefined,
+          providerMetadata: undefined,
+        },
+        { type: "finish", reason: { normalized: "content-filter", raw: "content_filter" }, usage: undefined },
+      ])
+      expect(response.toolCalls).toEqual([])
+
+      const missingIdentity = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              deltaChunk({ tool_calls: [{ index: 0, id: "call_2", function: { arguments: "{}" } }] }),
+              deltaChunk({}, "content_filter"),
+            ),
+          ),
+        ),
+      )
+      expect(missingIdentity.finishReason).toEqual({ normalized: "content-filter", raw: "content_filter" })
+      expect(missingIdentity.toolCalls).toEqual([])
+    }),
+  )
+
+  it.effect("does not finalize streamed tool calls when output is truncated", () =>
+    Effect.gen(function* () {
+      const body = sseEvents(
+        deltaChunk({
+          tool_calls: [{ index: 0, id: "call_1", function: { name: "lookup", arguments: '{"query":"weather"}' } }],
+        }),
+        deltaChunk({}, "length"),
+      )
+      const response = yield* LLMClient.generate(
+        LLMRequest.update(request, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
+        }),
+      ).pipe(Effect.provide(fixedResponse(body)))
+
+      expect(response.events).toEqual([
+        { type: "step-start", index: 0 },
+        {
+          type: "tool-input-start",
+          id: "call_1",
+          name: "lookup",
+          providerExecuted: undefined,
+          providerMetadata: undefined,
+        },
+        {
+          type: "tool-input-delta",
+          id: "call_1",
+          name: "lookup",
+          text: '{"query":"weather"}',
+          input: { query: "weather" },
+        },
+        {
+          type: "step-finish",
+          index: 0,
+          reason: { normalized: "length", raw: "length" },
+          usage: undefined,
+          providerMetadata: undefined,
+        },
+        { type: "finish", reason: { normalized: "length", raw: "length" }, usage: undefined },
+      ])
+      expect(response.toolCalls).toEqual([])
+
+      const missingIdentity = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              deltaChunk({ tool_calls: [{ index: 0, id: "call_2", function: { arguments: "{}" } }] }),
+              deltaChunk({}, "length"),
+            ),
+          ),
+        ),
+      )
+      expect(missingIdentity.finishReason).toEqual({ normalized: "length", raw: "length" })
+      expect(missingIdentity.toolCalls).toEqual([])
+    }),
+  )
+
+  it.effect("rejects unknown finish reasons without finalizing streamed tool calls", () =>
+    Effect.gen(function* () {
+      const body = sseEvents(
+        deltaChunk({
+          tool_calls: [{ index: 0, id: "call_1", function: { name: "lookup", arguments: '{"query":"weather"' } }],
+        }),
+        deltaChunk({}, "future_reason"),
+      )
+      const events = yield* Ref.make<ReadonlyArray<LLMEvent>>([])
+      const error = yield* LLMClient.stream(
+        LLMRequest.update(request, {
+          tools: [ToolDefinition.make({ name: "lookup", description: "Lookup data", inputSchema: { type: "object" } })],
+        }),
+      ).pipe(
+        Stream.tap((event) => Ref.update(events, (current) => [...current, event])),
+        Stream.runDrain,
+        Effect.provide(fixedResponse(body)),
+        Effect.flip,
+      )
+
+      expect(error).toMatchObject({
+        reason: { _tag: "UnknownProvider" },
+        message: "Provider finish_reason: future_reason",
+      })
+      expect(yield* Ref.get(events)).toEqual([
+        { type: "step-start", index: 0 },
+        {
+          type: "tool-input-start",
+          id: "call_1",
+          name: "lookup",
+          providerExecuted: undefined,
+          providerMetadata: undefined,
+        },
+        {
+          type: "tool-input-delta",
+          id: "call_1",
+          name: "lookup",
+          text: '{"query":"weather"',
+          input: { query: "weather" },
+        },
+      ])
+    }),
+  )
+
  it.effect("ignores empty identity fields on later tool call deltas", () =>
    Effect.gen(function* () {
      const body = sseEvents(
@@ -554,6 +554,19 @@ describe("OpenAI-compatible Chat route", () => {
        reason: { _tag: "UnknownProvider" },
        message: "Provider reported an error (finish_reason: error)",
      })
+
+      const unknown = yield* LLMClient.generate(request).pipe(
+        Effect.provide(fixedResponse(sseEvents(deltaChunk({}, "future_reason")))),
+        Effect.flip,
+      )
+      expect(unknown).toMatchObject({
+        reason: { _tag: "UnknownProvider" },
+        message: "Provider finish_reason: future_reason",
+      })
+      expect(decodeJson(unknown.reason.body ?? "")).toMatchObject({
+        id: "chatcmpl_fixture",
+        choices: [{ finish_reason: "future_reason" }],
+      })
    }),
  )

@@ -581,17 +594,13 @@ describe("OpenAI-compatible Chat route", () => {
    }),
  )

-  it.effect("preserves provider finish outcomes in the common reason algebra", () =>
+  it.effect("preserves content-filter finishes in the common reason algebra", () =>
    Effect.gen(function* () {
      const filtered = yield* LLMClient.generate(request).pipe(
        Effect.provide(fixedResponse(sseEvents(deltaChunk({}, "content_filter")))),
      )
-      const future = yield* LLMClient.generate(request).pipe(
-        Effect.provide(fixedResponse(sseEvents(deltaChunk({}, "future_reason")))),
-      )

      expect(filtered.finishReason).toEqual({ normalized: "content-filter", raw: "content_filter" })
-      expect(future.finishReason).toEqual({ normalized: "unknown", raw: "future_reason" })
    }),
  )

@@ -96,6 +96,27 @@ describe("Open Responses-compatible route", () => {
    }),
  )

+  it.effect("omits user messages with no content", () =>
+    Effect.gen(function* () {
+      const model = configure({
+        apiKey: "test-key",
+        baseURL: "https://responses.example.test/v1",
+        provider: "example",
+      }).model("example-model")
+      const prepared = yield* compileRequest(
+        LLM.request({
+          model,
+          messages: [Message.user("Before."), Message.user([]), Message.user("After.")],
+        }),
+      )
+
+      expect(prepared.body.input).toEqual([
+        { role: "user", content: [{ type: "input_text", text: "Before." }] },
+        { role: "user", content: [{ type: "input_text", text: "After." }] },
+      ])
+    }),
+  )
+
  it.effect("uses data URLs for embedded PDF messages and tool results", () =>
    Effect.gen(function* () {
      const model = configure({
@@ -406,7 +427,7 @@ describe("Open Responses-compatible route", () => {
    })

    routings.forEach((routing) => {
-      it.effect(`preserves reasoning summary boundaries and terminal metadata with ${routing.name}`, () =>
+      it.effect(`preserves reasoning summary boundaries without terminal reconciliation with ${routing.name}`, () =>
        Effect.gen(function* () {
          const address = { item_id: routing.item_id, output_index: routing.output_index }
          const response = yield* LLMClient.generate(request).pipe(
@@ -444,21 +465,18 @@ describe("Open Responses-compatible route", () => {
              type: "reasoning",
              text: "Second.",
              providerMetadata: {
-                "openai-compatible": { itemId: routing.id, reasoningEncryptedContent: "final-state" },
+                "openai-compatible": { itemId: routing.id, reasoningEncryptedContent: null },
              },
            },
          ])
          expect(response.events.filter(LLMEvent.is.reasoningEnd)).toEqual([
-            expect.objectContaining({
+            {
+              type: "reasoning-end",
              id: `${routing.id}:0`,
+              text: undefined,
              providerMetadata: { "openai-compatible": { itemId: routing.id } },
-            }),
-            expect.objectContaining({
-              id: `${routing.id}:1`,
-              providerMetadata: {
-                "openai-compatible": { itemId: routing.id, reasoningEncryptedContent: "final-state" },
-              },
-            }),
+            },
+            { type: "reasoning-end", id: `${routing.id}:1` },
          ])
        }),
      )
@@ -671,7 +689,7 @@ describe("Open Responses-compatible route", () => {
    }),
  )

-  it.effect("preserves terminal reasoning metadata when item completion is missing", () =>
+  it.effect("ignores terminal reasoning output when item completion is missing", () =>
    Effect.gen(function* () {
      const model = configure({
        apiKey: "test-key",
@@ -697,8 +715,9 @@ describe("Open Responses-compatible route", () => {
        ),
      )

-      expect(response.events.find((event) => event.type === "reasoning-end")).toMatchObject({
-        providerMetadata: { "openai-compatible": { itemId: "rs_raw", reasoningEncryptedContent: "raw-state" } },
+      expect(response.events.find((event) => event.type === "reasoning-end")).toEqual({
+        type: "reasoning-end",
+        id: "rs_raw:0",
      })
    }),
  )
@@ -2554,7 +2554,7 @@ describe("OpenAI Responses route", () => {
    }),
  )

-  it.effect("preserves terminal reasoning metadata when output item completion is missing", () =>
+  it.effect("ignores terminal reasoning output when item completion is missing", () =>
    Effect.gen(function* () {
      const response = yield* LLMClient.generate(
        LLMRequest.update(request, { providerOptions: { store: false } }),
@@ -2595,29 +2595,13 @@ describe("OpenAI Responses route", () => {

      expect(response.reasoning).toBe("Checked the diff.")
      expect(response.events.filter((event) => event.type === "reasoning-end")).toEqual([
-        {
-          type: "reasoning-end",
-          id: "rs_1:0",
-          providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: "terminal-state" } },
-        },
+        { type: "reasoning-end", id: "rs_1:0" },
      ])
      expect(response.message.content).toContainEqual({
        type: "reasoning",
        text: "Checked the diff.",
-        providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: "terminal-state" } },
+        providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: null } },
      })
-
-      const prepared = yield* compileRequest(
-        LLM.request({ model, messages: [response.message], providerOptions: { store: false } }),
-      )
-      expect(prepared.body.input).toEqual([
-        {
-          type: "reasoning",
-          id: "rs_1",
-          summary: [{ type: "summary_text", text: "Checked the diff." }],
-          encrypted_content: "terminal-state",
-        },
-      ])
    }),
  )

@@ -2644,7 +2628,7 @@ describe("OpenAI Responses route", () => {
    }),
  )

-  it.effect("reconciles pending reasoning and function calls in completed output order", () =>
+  it.effect("recovers pending function calls without reconciling terminal reasoning", () =>
    Effect.gen(function* () {
      const response = yield* LLMClient.generate(
        LLMRequest.update(request, { providerOptions: { store: false } }),
@@ -2682,14 +2666,15 @@ describe("OpenAI Responses route", () => {
        ),
      )

-      expect(response.events.find((event) => event.type === "reasoning-end")).toMatchObject({
-        providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: "terminal-state" } },
+      expect(response.events.find((event) => event.type === "reasoning-end")).toEqual({
+        type: "reasoning-end",
+        id: "rs_1:0",
      })
      expect(response.events.filter(LLMEvent.is.toolCall)).toEqual([
        expect.objectContaining({ id: "call_1", input: { query: "weather" } }),
      ])
-      expect(response.events.findIndex((event) => event.type === "reasoning-end")).toBeLessThan(
-        response.events.findIndex(LLMEvent.is.toolCall),
+      expect(response.events.findIndex(LLMEvent.is.toolCall)).toBeLessThan(
+        response.events.findIndex((event) => event.type === "reasoning-end"),
      )
      expect(response.finishReason.normalized).toBe("tool-calls")
    }),
@@ -3019,6 +3004,7 @@ describe("OpenAI Responses route", () => {
        {
          type: "reasoning-end",
          id: "rs_1:0",
+          text: "Checked the diff.",
          providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: "encrypted-state" } },
        },
      ])
@@ -21,6 +21,7 @@ import {
  QuotaExceededError,
  RateLimitError,
  RouteID,
+  ToolResultValue,
  TransportError,
  UnknownProviderError,
  Usage,
@@ -83,6 +84,37 @@ describe("llm schema", () => {
  })
 })

+describe("ToolResultValue", () => {
+  test("uses the canonical schema guard", () => {
+    const cases: ReadonlyArray<{ readonly value: unknown; readonly expected: boolean }> = [
+      { value: { type: "json", value: { ok: true } }, expected: true },
+      { value: { type: "text", value: "done" }, expected: true },
+      { value: { type: "error", value: "failed" }, expected: true },
+      { value: { type: "content", value: [{ type: "text", text: "done" }] }, expected: true },
+      { value: { type: "content", value: [{ type: "text" }] }, expected: false },
+      { value: { type: "content", value: "done" }, expected: false },
+      { value: { type: "json" }, expected: false },
+      { value: { type: "unknown", value: "done" }, expected: false },
+    ]
+
+    for (const item of cases) {
+      expect(Schema.is(ToolResultValue)(item.value)).toBe(item.expected)
+      expect(ToolResultValue.is(item.value)).toBe(item.expected)
+    }
+  })
+
+  test("accepts canonical results with extra fields", () => {
+    expect(ToolResultValue.is({ type: "json", value: { ok: true }, metadata: { source: "tool" } })).toBe(true)
+    expect(
+      ToolResultValue.is({
+        type: "content",
+        value: [{ type: "file", uri: "https://example.test/result.txt", mime: "text/plain", checksum: "abc" }],
+        metadata: { source: "tool" },
+      }),
+    ).toBe(true)
+  })
+})
+
 describe("AI.Usage", () => {
  test("subtractTokens clamps non-sensical breakdowns to zero", () => {
    // Defense against a provider reporting cached_tokens > prompt_tokens or
@@ -0,0 +1,77 @@
+import { describe, expect, test } from "bun:test"
+import { Message, ToolCallPart, ToolResultPart } from "../src/schema/messages.js"
+import { normalizeToolHistory } from "../src/tool-history.js"
+
+const toolCall = (id: string, name = id) => ToolCallPart.make({ id, name, input: {} })
+const toolResult = (id: string, value: unknown, name = id, resultType?: "text" | "content" | "error") =>
+  Message.tool(ToolResultPart.make({ id, name, result: value, resultType }))
+
+describe("tool history normalization", () => {
+  test("fills missing local results before the next step", () => {
+    const normalized = normalizeToolHistory([
+      Message.assistant([toolCall("first"), toolCall("second")]),
+      toolResult("first", "done", "wrong", "text"),
+      Message.user("Continue."),
+      Message.assistant(toolCall("trailing")),
+    ])
+
+    expect(normalized.map((message) => message.role)).toEqual([
+      "assistant",
+      "tool",
+      "tool",
+      "user",
+      "assistant",
+    ])
+    expect(normalized[1]?.content[0]).toMatchObject({ type: "tool-result", id: "first", name: "first" })
+    expect(normalized[2]?.content).toEqual([
+      { type: "tool-result", id: "second", name: "second", result: { type: "error", value: "Tool result missing" } },
+    ])
+    expect(normalized[4]?.content).toEqual([toolCall("trailing")])
+  })
+
+  test("normalizes empty results without changing whitespace or media", () => {
+    const media = { type: "file" as const, uri: "data:image/png;base64,AQID", mime: "image/png" }
+    const normalized = normalizeToolHistory([
+      Message.assistant([
+        toolCall("text"),
+        toolCall("content"),
+        toolCall("error"),
+        toolCall("mixed"),
+        toolCall("whitespace"),
+      ]),
+      toolResult("text", "", "text", "text"),
+      toolResult("content", [], "content", "content"),
+      toolResult("error", "", "error", "error"),
+      toolResult("mixed", [{ type: "text", text: "" }, media], "mixed", "content"),
+      toolResult("whitespace", "   ", "whitespace", "text"),
+    ])
+
+    expect(normalized.slice(1).map((message) => message.content[0])).toEqual([
+      { type: "tool-result", id: "text", name: "text", result: { type: "text", value: "(no tool output)" } },
+      { type: "tool-result", id: "content", name: "content", result: { type: "text", value: "(no tool output)" } },
+      { type: "tool-result", id: "error", name: "error", result: { type: "error", value: "(no tool output)" } },
+      { type: "tool-result", id: "mixed", name: "mixed", result: { type: "content", value: [media] } },
+      { type: "tool-result", id: "whitespace", name: "whitespace", result: { type: "text", value: "   " } },
+    ])
+  })
+
+  test("leaves unmatched and provider-executed history unchanged", () => {
+    const hostedCall = ToolCallPart.make({
+      id: "hosted",
+      name: "web_search",
+      input: {},
+      providerExecuted: true,
+    })
+    const hostedResult = ToolResultPart.make({
+      id: "hosted",
+      name: "web_search",
+      result: "",
+      resultType: "text",
+      providerExecuted: true,
+    })
+    const hosted = Message.assistant([hostedCall, hostedResult])
+    const orphan = toolResult("orphan", "ignored", "orphan", "text")
+
+    expect(normalizeToolHistory([orphan, hosted])).toEqual([orphan, hosted])
+  })
+})
@@ -442,6 +442,33 @@ describe("LLMClient tools", () => {
    }),
  )

+  it.effect("projects malformed tagged dynamic output as opaque JSON", () =>
+    Effect.gen(function* () {
+      const malformed = { type: "content", value: [{ type: "text" }] }
+      const dynamic = Tool.make({
+        description: "Return caller-defined JSON.",
+        jsonSchema: { type: "object", properties: {} },
+        execute: () => Effect.succeed(malformed),
+      })
+
+      const dispatched = yield* ToolRuntime.dispatch(
+        { dynamic },
+        LLMEvent.toolCall({ id: "call_1", name: "dynamic", input: {} }),
+      )
+
+      expect(dispatched.result).toEqual({ type: "json", value: malformed })
+      expect(dispatched.output).toEqual({ structured: malformed, content: [] })
+      expect(dispatched.events).toEqual([
+        LLMEvent.toolResult({
+          id: "call_1",
+          name: "dynamic",
+          result: { type: "json", value: malformed },
+          output: { structured: malformed, content: [] },
+        }),
+      ])
+    }),
+  )
+
  it.effect("executes tool calls for one step without looping by default", () =>
    Effect.gen(function* () {
      const layer = scriptedResponses([
@@ -1,5 +1,6 @@
 src/assets/theme.css
 e2e/test-results
+e2e/performance/results/
 e2e/playwright-report
 component-tests/test-results
 component-tests/playwright-report
@@ -71,4 +71,25 @@ Environment options:

 ## Deployment

-You can deploy the `dist` folder to any static host provider (netlify, surge, now, etc.)
+The `deploy` GitHub Actions workflow uses SST to deploy the web app from these branches in `anomalyco/opencode`:
+
+| Branch       | Site                  |
+| ------------ | --------------------- |
+| `dev`        | `app.dev.opencode.ai` |
+| `production` | `app.opencode.ai`     |
+| `beta`       | `beta.opencode.ai`    |
+
+Changes merged into `v2` reach the beta site when they are promoted to `beta`. The beta SST stage deploys
+only the web app, using the same `WebApp` StaticSite definition as production. It sets the build channel
+and Sentry environment to `beta` without deploying the API, console, database, or billing infrastructure.
+
+The hosted app defaults to `http://localhost:49374`, matching the managed V2 service. Saved server selections
+override this default. Connecting still requires the service's credentials.
+
+The workflow reuses the repository's `CLOUDFLARE_API_TOKEN` and web Sentry settings. The Cloudflare token
+must cover SST's R2 state storage, KV assets, Workers, and custom-domain management in the account that
+owns `opencode.ai`. The beta GitHub environment must allow deployments from the `beta` branch; it does not
+need AWS credentials.
+
+SST manages the beta site's custom domain. The first deployment creates its DNS record and TLS certificate.
+Do not create a CNAME for `beta.opencode.ai` first, because it would conflict with the Workers custom domain.
@@ -1,5 +1,65 @@
 import { expect, story } from "../../storybook/playwright/story"

+for (const draft of ["empty-draft", "multiline-draft", "mixed-attachments"]) {
+  story(`select all stays inside the composer with ${draft}`, async ({ mount, page }) => {
+    const component = await mount(`opencode-composer-flow--${draft}`)
+    const input = component.getByRole("textbox", { name: "Prompt", exact: true })
+    const text = await input.textContent()
+
+    for (let count = 0; count < 2; count++) {
+      await input.press("ControlOrMeta+a")
+      expect(
+        await input.evaluate((editor) => {
+          const selection = window.getSelection()
+          return {
+            text: selection?.toString(),
+            inside: editor.contains(selection?.anchorNode ?? null) && editor.contains(selection?.focusNode ?? null),
+          }
+        }),
+      ).toEqual({ text, inside: true })
+    }
+
+    await page.keyboard.type("Replacement draft")
+    await expect(input).toHaveText("Replacement draft")
+    await expect(component.getByRole("status")).toHaveText("Ready")
+    if (draft === "mixed-attachments") {
+      await expect(component.getByAltText("layout.png")).toBeVisible()
+      await expect(component.getByText("Keep the normal flow flat", { exact: true })).toBeVisible()
+    }
+  })
+}
+
+story("renders a draft once and supports editing, caret restoration, and failure recovery", async ({ mount, page }) => {
+  await page.addInitScript(() => {
+    const replace = Element.prototype.replaceChildren
+    Element.prototype.replaceChildren = function (this: Element, ...nodes) {
+      // The ref can run before data-component is assigned, so count on every target.
+      this.setAttribute("data-test-replacements", String(Number(this.getAttribute("data-test-replacements")) + 1))
+      return replace.apply(this, nodes)
+    }
+  })
+  const component = await mount("opencode-composer-flow--failed-submission-restoration")
+  const input = component.getByRole("textbox", { name: "Prompt", exact: true })
+  await expect(input).toHaveText("Preserve this draft on failure")
+  await expect(input).toHaveAttribute("data-test-replacements", "1")
+
+  await input.press("Home")
+  await input.press("Shift+ArrowRight")
+  await input.pressSequentially("XY")
+  await expect(input).toHaveText("XYreserve this draft on failure")
+  await expect(input).toHaveAttribute("data-test-replacements", "1")
+
+  // Closing the model picker restores the controller's saved caret through its editor ref.
+  await component.locator('[data-action="composer-model"]').click()
+  await page.getByRole("menu").getByRole("textbox").press("Escape")
+  await expect(input).toBeFocused()
+  await input.pressSequentially("!")
+  await expect(input).toHaveText("XY!reserve this draft on failure")
+  await component.getByRole("button", { name: "Send", exact: true }).click()
+  await expect(component.getByRole("status")).toHaveText("Submission failed; draft restored")
+  await expect(input).toHaveText("Preserve this draft on failure")
+})
+
 // Moved from packages/app/e2e/regression/prompt-thinking-level.spec.ts
 story("shows the thinking level control while relevant", async ({ mount, page }) => {
  const component = await mount("opencode-composer-flow--model-and-variant")
@@ -0,0 +1,135 @@
+import { TimelineRow } from "@opencode-ai/session-ui/timeline/projection"
+import { onCleanup } from "solid-js"
+import { createStore } from "solid-js/store"
+import { render } from "solid-js/web"
+import { LanguageProvider } from "../src/runtime/i18n/language"
+import { createTimelineVirtualizer } from "../src/session/timeline/virtualizer"
+
+export function mountTimelineVirtualizer(input: { count: number; rowHeight: number; immediate?: boolean }) {
+  const host = document.createElement("main")
+  host.dataset.testid = "timeline-virtualizer-fixture"
+  host.dataset.scrolls = "0"
+  host.dataset.viewportResizes = "0"
+  host.style.cssText = "position:fixed;top:24px;right:24px;width:400px;z-index:1000"
+  document.body.appendChild(host)
+
+  function Fixture() {
+    const [state, setState] = createStore({ pinned: true, ready: false })
+    const rows = Array.from(
+      { length: input.count },
+      (_, index) => new TimelineRow.UserMessage({ userMessageID: `message-${index}` }),
+    )
+    const rowByKey = new Map(rows.map((row) => [TimelineRow.key(row), row]))
+    const indexes = new Map(rows.map((row, index) => [row.userMessageID, index]))
+    let viewport!: HTMLDivElement
+    let content!: HTMLDivElement
+    let container!: HTMLDivElement
+    const timeline = createTimelineVirtualizer({
+      sessionKey: () => "cold-reveal-fixture",
+      projection: {
+        rows: () => rows,
+        rowByKey: () => rowByKey,
+        activeMessageID: () => undefined,
+        messageRowIndex: () => indexes,
+        messageLastRowIndex: () => indexes,
+      },
+      showHeader: () => false,
+      pinned: () => state.pinned,
+      scroll: () => ({ overflow: false, jump: false }),
+      setScrollRef: (element) => {
+        if (!element) return
+        viewport = element
+        resize.observe(element, { box: "border-box" })
+      },
+      setContentRef: (element) => {
+        content = element
+        reveal.observe(element, { attributes: true, attributeFilter: ["style"] })
+      },
+      onPin: () => setState("pinned", true),
+      onUnpin: () => setState("pinned", false),
+      onScheduleScrollState: (element) => {
+        host.dataset.scrolls = String(Number(host.dataset.scrolls) + 1)
+        host.dataset.lastScrollTop = String(element.scrollTop)
+      },
+      onResumeScroll: () => {},
+      onSelectionInteraction: () => {},
+      onUserScroll: () => {},
+      onHistoryScroll: () => {},
+      canRenderImmediately: () => input.immediate ?? false,
+    })
+
+    const resize = new ResizeObserver((entries) => {
+      host.dataset.observedHeight = String(entries[0].borderBoxSize[0].blockSize)
+      host.dataset.viewportResizes = String(Number(host.dataset.viewportResizes) + 1)
+    })
+    const reveal = new MutationObserver(() => {
+      if (content.style.visibility === "hidden" || host.dataset.firstReveal) return
+      // Capture the first reveal, not a later frame after geometry has recovered.
+      const mounted = [...content.querySelectorAll<HTMLElement>("[data-timeline-key]")]
+      host.dataset.firstReveal = JSON.stringify({
+        rows: mounted.map((element) => Number(element.firstElementChild!.getAttribute("data-index"))),
+        pendingMarkdown: content.querySelectorAll('[data-component="markdown"]:not([data-markdown-ready])').length,
+        viewportHeight: viewport.clientHeight,
+        scrollTop: viewport.scrollTop,
+        clipped: mounted
+          .filter((element) => element.firstElementChild!.getBoundingClientRect().height > element.offsetHeight + 1)
+          .map((element) => element.dataset.timelineKey),
+      })
+    })
+    onCleanup(() => {
+      resize.disconnect()
+      reveal.disconnect()
+    })
+
+    return (
+      <div data-testid="timeline-controls" data-pinned={state.pinned}>
+        <button type="button" onClick={() => setState("ready", true)}>
+          Complete Markdown
+        </button>
+        <button type="button" onClick={() => (container.style.display = "none")}>
+          Hide viewport
+        </button>
+        <button
+          type="button"
+          onClick={() => {
+            const parent = viewport.parentElement!
+            host.dataset.scrolls = "0"
+            // Keep the same scroller and complete Markdown while it has no layout box.
+            viewport.remove()
+            viewport.scrollTop = 0
+            setState("ready", true)
+            parent.prepend(viewport)
+            container.style.removeProperty("display")
+          }}
+        >
+          Reconnect ready rows
+        </button>
+        <div ref={container} style={{ height: "180px", width: "400px" }}>
+          <timeline.View
+            header={null}
+            workspaceSession={() => false}
+            deferred={() => false}
+            renderRow={(row) => (
+              <div
+                data-component="markdown"
+                data-markdown-ready={state.ready ? "" : undefined}
+                style={{ height: `${input.rowHeight}px` }}
+              >
+                {row().userMessageID}
+              </div>
+            )}
+          />
+        </div>
+      </div>
+    )
+  }
+
+  render(
+    () => (
+      <LanguageProvider locale="en">
+        <Fixture />
+      </LanguageProvider>
+    ),
+    host,
+  )
+}
@@ -0,0 +1,79 @@
+import { fileURLToPath } from "node:url"
+import { expect, story } from "../../storybook/playwright/story"
+
+const fixture = `/@fs/${fileURLToPath(new URL("./timeline-virtualizer.fixture.tsx", import.meta.url)).replaceAll("\\", "/")}`
+
+story.beforeEach(async ({ mount }) => {
+  const component = await mount("opencode-composer-flow--mixed-attachments")
+  await expect(component.getByRole("textbox", { name: "Prompt", exact: true })).toBeVisible()
+})
+
+story("bounds the cheap suffix and reveals only ready measured rows", async ({ page }) => {
+  await page.evaluate(async (fixture) => {
+    const { mountTimelineVirtualizer } = await import(fixture)
+    mountTimelineVirtualizer({ count: 100, rowHeight: 60, immediate: true })
+  }, fixture)
+  const root = page.getByTestId("timeline-virtualizer-fixture")
+  const content = root.locator("[data-timeline-virtual-content]")
+  await expect(root).toHaveAttribute("data-observed-height", "180")
+  await expect(content).toHaveCSS("visibility", "hidden")
+  await expect(content.locator("[data-timeline-key]")).toHaveCount(4)
+  await root.getByRole("button", { name: "Complete Markdown", exact: true }).click()
+  await expect(content).toHaveCSS("visibility", "visible")
+  await expect(root).toHaveAttribute("data-first-reveal", /.+/)
+  expect(await root.evaluate((element) => JSON.parse(element.dataset.firstReveal!))).toMatchObject({
+    rows: [96, 97, 98, 99],
+    pendingMarkdown: 0,
+    clipped: [],
+    viewportHeight: 180,
+  })
+})
+
+for (const input of [
+  { name: "offset-only", count: 1, rowHeight: 600 },
+  { name: "zero-height", count: 4, rowHeight: 60 },
+]) {
+  story(`reveals ready measured rows after an ${input.name} reconnect`, async ({ page }) => {
+    await page.evaluate(
+      async ({ fixture, input }) => {
+        const { mountTimelineVirtualizer } = await import(fixture)
+        mountTimelineVirtualizer(input)
+      },
+      { fixture, input },
+    )
+    const root = page.getByTestId("timeline-virtualizer-fixture")
+    const content = root.locator("[data-timeline-virtual-content]")
+    await expect(root).toHaveAttribute("data-observed-height", "180")
+    await expect(content).toHaveCSS("visibility", "hidden")
+    await expect(content.locator("[data-timeline-key]")).toHaveCount(1)
+
+    if (input.name === "offset-only") {
+      await expect(root).toHaveAttribute("data-last-scroll-top", "484")
+      await root.locator("[data-scrollable]").dispatchEvent("wheel", { deltaY: -1 })
+      await expect(root.getByTestId("timeline-controls")).toHaveAttribute("data-pinned", "false")
+    }
+    if (input.name === "zero-height") {
+      await root.getByRole("button", { name: "Hide viewport", exact: true }).click()
+      // Wait for ResizeObserver to clear the actual range, not just for display:none.
+      await expect(root).toHaveAttribute("data-observed-height", "0")
+      await expect(content.locator("[data-timeline-key]")).toHaveCount(0)
+    }
+    await expect(root).not.toHaveAttribute("data-first-reveal")
+    const resizes = await root.getAttribute("data-viewport-resizes")
+    await root.getByRole("button", { name: "Reconnect ready rows", exact: true }).click()
+    await expect(content).toHaveCSS("visibility", "visible")
+    await expect(root).toHaveAttribute("data-first-reveal", /.+/)
+    expect(await root.evaluate((element) => JSON.parse(element.dataset.firstReveal!))).toMatchObject({
+      rows: input.count === 1 ? [0] : [0, 1, 2, 3],
+      pendingMarkdown: 0,
+      clipped: [],
+      viewportHeight: 180,
+      ...(input.name === "offset-only" ? { scrollTop: 0 } : {}),
+    })
+    if (input.name === "offset-only") {
+      // This repair must not depend on another native scroll or resize delivery.
+      await expect(root).toHaveAttribute("data-scrolls", "0")
+      await expect(root).toHaveAttribute("data-viewport-resizes", resizes!)
+    }
+  })
+}
@@ -65,7 +65,7 @@ The fixture requires every benchmark to call `report()`, automatically names and
 BENCHMARK {"name":"...","context":{"project":"chromium","platform":"darwin"},"metrics":{...}}
 ```

-Every observed page also emits `BENCHMARK_PAGE` with the same run ID, navigation history, and optional trace path before the final status-bearing `BENCHMARK` record. Chrome traces are browser-wide page-lifetime diagnostics; scenario metrics use narrower explicitly named observation windows.
+Every observed page also emits `BENCHMARK_PAGE` with the same run ID, navigation history, optional trace path, and trace scope before the final status-bearing `BENCHMARK` record. Chrome traces are browser-wide; the default window is page lifetime. Tab-switch traces begin after scenario setup and include explicit interaction markers. Scenario metrics use their own narrower observation windows.

 This follows the stack's own guidance: [Electron recommends repeated Chrome DevTools and Chrome Tracing measurement](https://www.electronjs.org/docs/latest/tutorial/performance), [Chrome DevTools recommends Performance recordings for runtime work](https://developer.chrome.com/docs/devtools/performance), and [Playwright uses traces for test debugging rather than renderer profiling](https://playwright.dev/docs/trace-viewer).

@@ -81,13 +81,60 @@ Committed smoke and regression tests continue to own correctness coverage for pa

 Tab-switch timing starts at `mousedown`, when mouse-selected tabs actually navigate, with a `click` fallback for keyboard activation. The probe excludes hidden/transparent content and intersects answers with their virtual-row clip and viewport. The tab workload requires the destination's final answer to be visible with Markdown ready. These results are not directly comparable to older click-start, geometry-only measurements. `stableObservedMs` includes confirmation across three correct samples; `firstCorrectObservedMs` is the first sample meeting all content and geometry checks. Neither is a compositor presentation timestamp.

-Each tab scenario reports one sample, including its raw observations. Use Playwright's `--repeat-each=5` for repeated measurements. Cached scenarios warm the destination at the same panel width before leaving it; a separate resized scenario validates reuse after opening the review pane changes that width.
+Each tab scenario reports one sample, including its raw observations. Use Playwright's `--repeat-each=20` for a baseline distribution. Warm scenarios prepare the destination at the same panel width before leaving it; a separate resized scenario validates reuse after opening the review pane changes that width.
+
+The tab-switch workload uses two equally long sessions: 200 user/assistant exchanges (400 messages) per tab. Every answer includes headings, emphasis, links, a blockquote, task and nested lists, an eight-row table, and four highlighted code fences (TSX, JSON, SQL, Bash), alongside the stress fixture's reasoning and tools. The mock API deliberately returns all 400 messages in one response so every scenario measures a long loaded history, not a short paginated tail. The viewport is fixed at 1440 x 900. Results include the fixture version, Markdown and serialized-message byte counts, and message-request count. These numbers are not directly comparable to the earlier 12-exchange source / 72-exchange destination fixture.
+
+Cold means the destination transcript has never rendered in that fresh browser context. Warm means its complex answer was rendered and ready before switching away and back. Both use the app's normal restored-tab data prefetch, which completes before measurement; neither includes app startup, the source session's Markdown engine initialization, or a cold backend fetch. The suite asserts no message fetch during either measured switch. Setup waits for mounted Markdown to finish and for the review-pane width transition to complete. Service workers are blocked to exclude the web build's background asset precache from this renderer benchmark. Screenshots are attached after measurement for the first repetition; Playwright video and trace recording are disabled for this workload, while opt-in Chrome profiling remains available. For a baseline distribution, use `--repeat-each=20 --retries=0`, keep profiling disabled, and report the median and p95 of `firstCorrectObservedMs` separately from the three-observation `stableObservedMs`.

 ```sh
 bunx playwright test --config e2e/performance/playwright.config.ts \
-  timeline/session-tab-switch-benchmark.spec.ts --repeat-each=5
+  timeline/session-tab-switch-benchmark.spec.ts --repeat-each=20 --retries=0
 ```

+**The tab-switch fixture is not an end-to-end cold-data benchmark.** It prefetches destination messages and returns full history. Measure cold API navigation, Home-row opening, and prefetched-but-unvisited tabs separately with normal pagination. Do not combine these entry paths or compare different transports and machine-load periods as one experiment.
+
+Keep one-off reports, recorded results, and traces outside git, in the ignored `e2e/performance/results/` directory or an external artifact directory. Preserve raw observations locally and publish anonymized summaries and charts in the PR description, not as committed experiment files.
+
+For a repeatable tab-switch summary, run from `packages/app`:
+
+```sh
+bun run bench:tabs
+```
+
+This runs only the tab-switch benchmark against the production build with 20 serial repetitions and no retries. It prints the median (mean of the two middle values for even sample counts) and nearest-rank p95 for `firstCorrectObservedMs` and `stableObservedMs` per scenario. Only records whose benchmark and Playwright statuses are passed and whose two metrics are finite enter the summary. Test and record statuses, missing records, and excluded samples are reported separately.
+
+For fresh entry paths, run `bun run bench:entry` from `packages/app`. It uses the same production, serial-repetition, and reporting defaults. The cases open an empty draft from the actual Home button, create a draft with the titlebar plus from an active session, and open a cold paginated session from Home. Draft readiness requires a focused editable composer, the expected model, project control, and new tab; typing and absence of backend mutations are checked afterward. Session readiness requires the latest group, ready answer Markdown, and bottom anchoring. These cases are separate from prefetched tab remounts.
+
+For milestone charts, rerun frozen builds with one workload and counterbalanced serial order. Do not connect historical medians from different transports, preparation, or machine-load periods. Show samples or ranges, name the checkpoints accurately, and distinguish experimental build snapshots from Git commits.
+
+Complete original `BENCHMARK` JSON records, including samples, context, and failed records, are saved as `tab-switch-benchmark.jsonl` in Playwright's configured output directory (default: `e2e/test-results/performance`). Standard Playwright flags can override defaults when appended:
+
+```sh
+bun run bench:tabs --repeat-each=3 --output=e2e/test-results/tabs-smoke
+```
+
+Set `OPENCODE_PERFORMANCE_MEMORY=1` for an opt-in renderer-main-isolate heap and DOM sample after mounted content is ready and an explicit GC completes. Probe DOM references are released before collection. This is not total desktop memory; do not mix these diagnostic runs with unprofiled latency samples. Set `OPENCODE_PERFORMANCE_TRACE_DIR` for a separate Chrome trace of each tab interaction, starting after preparation, with `session-switch:start`, `session-switch:ready`, and `session-switch:stable` markers.
+
+### Cache-Enabled HTTP Fixture
+
+The default tab harness uses Playwright routing for API responses. Playwright routing disables the browser HTTP cache, including for unrelated SVG assets. To measure with HTTP caching enabled, the same API handlers and tab data can run on a real loopback HTTP endpoint:
+
+```sh
+bun run build
+bun e2e/performance/tab-switch-server.ts --port 4639 --dist dist
+```
+
+With that fixture running, run the benchmark in a separate terminal from `packages/app`:
+
+```powershell
+$env:PLAYWRIGHT_BASE_URL = "http://127.0.0.1:4639"
+$env:OPENCODE_PERFORMANCE_HTTP_FIXTURE = "1"
+bun run bench:tabs
+```
+
+Use `--dist` to select a frozen production bundle when comparing revisions. An explicit `PLAYWRIGHT_BASE_URL` means the benchmark does not rebuild or start another preview. The fixture gives hashed assets immutable cache headers; it serves the deterministic read workload, not the live OpenCode service. Each test still gets a fresh browser context, and source-session setup still occurs before the measured switch. API responses use `no-store`, service workers remain blocked, and no destination Markdown is rendered before a cold switch. Records identify the transport as `http` or `playwright-route`; keep these series separate. Unset `OPENCODE_PERFORMANCE_HTTP_FIXTURE` when returning to the default routed harness.
+
 ## Retained renderer memory

 Run the catalog workload against the production app bundle:
@@ -111,7 +158,7 @@ bunx playwright test --config e2e/performance/playwright.config.ts \

 The emitted JSON is a standard Chrome trace and can be loaded directly into the Chrome DevTools Performance panel. `devtools-tracing` can optionally inspect it from the command line without adding package scripts or dependencies:

-Trace capture mirrors [Puppeteer's official tracing defaults and lifecycle](https://pptr.dev/api/puppeteer.tracing), using Chrome's `ReturnAsStream` transfer mode and failing when Chromium reports trace data loss.
+Trace capture follows [Puppeteer's tracing lifecycle](https://pptr.dev/api/puppeteer.tracing), using Chrome's `ReturnAsStream` transfer mode and failing when Chromium reports trace data loss. V8 CPU sample stacks support attribution through the frozen build's source maps. Set `OPENCODE_PERFORMANCE_STACK_TRACE=1` only when per-event timeline stacks are needed; they add substantial overhead. Keep profiled runs separate from latency distributions, including when comparing the stack-capture modes.

 ```sh
 bunx devtools-tracing stats <trace-path-from-BENCHMARK_PAGE>
@@ -5,16 +5,20 @@ type BenchmarkFixtures = {
  report: (metrics: Record<string, unknown>, context?: Record<string, unknown>) => void
  reportState: { payload?: { metrics: Record<string, unknown>; context: Record<string, unknown> } }
  benchmarkResult: void
+  traceScope: "page" | "interaction"
 }

 export type PerformancePageDiagnostics = {
  navigations: string[]
+  traceScope: "page" | "interaction"
+  startTrace: () => Promise<void>
  stop: () => Promise<string | undefined>
 }

 const pages = new WeakMap<Page, PerformancePageDiagnostics>()

 export const benchmark = base.extend<BenchmarkFixtures>({
+  traceScope: ["page", { option: true }],
  reportState: async ({}, use) => use({}),
  report: async ({ reportState }, use) => {
    await use((metrics, context = {}) => {
@@ -49,9 +53,9 @@ export const benchmark = base.extend<BenchmarkFixtures>({
    },
    { auto: true },
  ],
-  page: async ({ page }, use, testInfo) => {
+  page: async ({ page, traceScope }, use, testInfo) => {
    const name = benchmarkName(testInfo)
-    const diagnostics = await observePerformancePage(page, name)
+    const diagnostics = await observePerformancePage(page, name, traceScope)
    try {
      await use(page)
    } finally {
@@ -75,25 +79,30 @@ function benchmarkName(testInfo: TestInfo) {

 export { expect }

-async function observePerformancePage(page: Page, name: string) {
+async function observePerformancePage(page: Page, name: string, traceScope: "page" | "interaction" = "page") {
  const navigations: string[] = []
  const onNavigation = (frame: ReturnType<Page["mainFrame"]>) => {
    if (frame === page.mainFrame()) navigations.push(frame.url())
  }
  page.on("framenavigated", onNavigation)
-  const stopTrace = await startChromeTrace(page, name).catch((error) => {
-    page.off("framenavigated", onNavigation)
-    throw error
-  })
+  let stopTrace: Awaited<ReturnType<typeof startChromeTrace>>
  let stopping: Promise<string | undefined> | undefined
  const diagnostics: PerformancePageDiagnostics = {
    navigations,
+    traceScope,
+    async startTrace() {
+      stopTrace ??= await startChromeTrace(page, name).catch((error) => {
+        page.off("framenavigated", onNavigation)
+        throw error
+      })
+    },
    stop() {
      page.off("framenavigated", onNavigation)
      return (stopping ??= stopTrace?.() ?? Promise.resolve(undefined))
    },
  }
  pages.set(page, diagnostics)
+  if (traceScope === "page") await diagnostics.startTrace()
  return diagnostics
 }

@@ -130,6 +139,7 @@ async function reportPerformancePage(name: string, diagnostics: PerformancePageD
      context: {
        platform: process.platform,
        trace,
+        traceScope: diagnostics.traceScope,
        selectorTrace: process.env.OPENCODE_PERFORMANCE_SELECTOR_TRACE === "1",
      },
      navigations: diagnostics.navigations,
@@ -14,7 +14,6 @@ const categories = [
  "blink.console",
  "blink.user_timing",
  "latencyInfo",
-  "disabled-by-default-devtools.timeline.stack",
  "disabled-by-default-v8.cpu_profiler",
 ]

@@ -34,6 +33,9 @@ export async function startChromeTrace(page: Page, name: string): Promise<undefi
          .map((category) => category.slice(1)),
        includedCategories: [
          ...categories.filter((category) => !category.startsWith("-")),
+          ...(process.env.OPENCODE_PERFORMANCE_STACK_TRACE === "1"
+            ? ["disabled-by-default-devtools.timeline.stack"]
+            : []),
          ...(selectors
            ? ["disabled-by-default-blink.debug", "disabled-by-default-devtools.timeline.invalidationTracking"]
            : []),
@@ -0,0 +1,99 @@
+import type { FullConfig, FullResult, Reporter, Suite, TestCase, TestResult } from "@playwright/test/reporter"
+import { mkdir, writeFile } from "node:fs/promises"
+import path from "node:path"
+
+type BenchmarkRecord = {
+  status?: string
+  metrics?: { firstCorrectObservedMs?: unknown; stableObservedMs?: unknown } | null
+}
+
+export default class TabSwitchReporter implements Reporter {
+  private output = ""
+  private tests: TestCase[] = []
+  private results: { test: TestCase; status: TestResult["status"]; records: string[] }[] = []
+
+  onBegin(config: FullConfig, suite: Suite) {
+    this.output = config.projects[0].outputDir
+    this.tests = suite.allTests()
+  }
+
+  onTestEnd(test: TestCase, result: TestResult) {
+    this.results.push({
+      test,
+      status: result.status,
+      records: Buffer.concat(result.stdout.map((chunk) => (typeof chunk === "string" ? Buffer.from(chunk) : chunk)))
+        .toString("utf8")
+        .split(/\r?\n/)
+        .filter((line) => line.startsWith("BENCHMARK "))
+        .map((line) => line.slice("BENCHMARK ".length)),
+    })
+  }
+
+  async onEnd(result: FullResult) {
+    const file = path.join(this.output, "tab-switch-benchmark.jsonl")
+    try {
+      await mkdir(this.output, { recursive: true })
+      await writeFile(file, this.results.flatMap((entry) => entry.records.map((raw) => `${raw}\n`)).join(""), "utf8")
+    } catch (error) {
+      console.error("Could not save tab-switch benchmark records:", error)
+      return { status: "failed" as const }
+    }
+
+    console.log(`\nTab-switch benchmark: ${result.status}`)
+    Array.from(new Set(this.tests.map((test) => test.title))).forEach((name) => {
+      const results = this.results.filter((entry) => entry.test.title === name)
+      const unrun = this.tests.filter(
+        (test) => test.title === name && !results.some((entry) => entry.test.id === test.id),
+      ).length
+      const records = results.flatMap((entry) =>
+        entry.records.map((raw) => {
+          try {
+            return { status: entry.status, record: JSON.parse(raw) as BenchmarkRecord | null }
+          } catch {
+            return { status: entry.status, record: { status: "invalid JSON", metrics: null } }
+          }
+        }),
+      )
+      const passed = records.filter((entry) => entry.status === "passed" && entry.record?.status === "passed")
+      const valid = passed
+        .map((entry) => ({
+          firstCorrectObservedMs: entry.record?.metrics?.firstCorrectObservedMs,
+          stableObservedMs: entry.record?.metrics?.stableObservedMs,
+        }))
+        .filter(
+          (metrics): metrics is { firstCorrectObservedMs: number; stableObservedMs: number } =>
+            typeof metrics.firstCorrectObservedMs === "number" &&
+            Number.isFinite(metrics.firstCorrectObservedMs) &&
+            typeof metrics.stableObservedMs === "number" &&
+            Number.isFinite(metrics.stableObservedMs),
+        )
+
+      console.log(`\n${name}`)
+      console.log(`  Tests: ${counts(results.map((entry) => entry.status))}; unrun=${unrun}`)
+      console.log(
+        `  Records: ${counts(records.map((entry) => entry.record?.status ?? "missing status"))}; ` +
+          `missing=${results.filter((entry) => entry.records.length === 0).length + unrun}; ` +
+          `excluded=${records.length - valid.length}; invalid metrics=${passed.length - valid.length}`,
+      )
+      ;(["firstCorrectObservedMs", "stableObservedMs"] as const).forEach((metric) => {
+        const values = valid.map((entry) => entry[metric]).sort((a, b) => a - b)
+        if (values.length === 0) {
+          console.log(`  ${metric}: n=0, median=n/a, p95=n/a`)
+          return
+        }
+        const median = (values[Math.floor((values.length - 1) / 2)] + values[Math.floor(values.length / 2)]) / 2
+        const p95 = values[Math.ceil(values.length * 0.95) - 1]
+        console.log(`  ${metric}: n=${values.length}, median=${median.toFixed(2)} ms, p95=${p95.toFixed(2)} ms`)
+      })
+    })
+    console.log(`\nRaw BENCHMARK records: ${file}`)
+  }
+}
+
+function counts(statuses: string[]) {
+  return (
+    Array.from(new Set(statuses))
+      .map((status) => `${status}=${statuses.filter((value) => value === status).length}`)
+      .join(", ") || "none"
+  )
+}
@@ -0,0 +1,61 @@
+import path from "node:path"
+import { parseArgs } from "node:util"
+import { createMockServerHandler } from "../utils/mock-server"
+import { fixture } from "./timeline/session-timeline-stress.fixture"
+import { messages } from "./timeline/session-tab-switch.fixture"
+import { createReviewDiffs } from "./timeline/timeline-test-helpers"
+
+const args = parseArgs({
+  args: Bun.argv.slice(2),
+  options: { port: { type: "string", default: "4639" }, dist: { type: "string", default: "dist" } },
+})
+const directory = path.resolve(args.values.dist)
+const api = createMockServerHandler({
+  directory: fixture.directory,
+  project: fixture.project,
+  provider: fixture.provider,
+  sessions: fixture.sessions,
+  pageMessages: (sessionID) => ({ items: messages[sessionID] ?? [] }),
+  vcsDiff: createReviewDiffs(),
+})
+const server = Bun.serve({
+  hostname: "127.0.0.1",
+  port: Number(args.values.port),
+  idleTimeout: 0,
+  async fetch(request) {
+    const url = new URL(request.url)
+    if (url.pathname === "/api/event") {
+      return new Response(
+        new ReadableStream({
+          start(controller) {
+            controller.enqueue(
+              new TextEncoder().encode('data: {"id":"evt_fixture_connected","type":"server.connected","data":{}}\n\n'),
+            )
+          },
+        }),
+        { headers: { "content-type": "text/event-stream", "cache-control": "no-store" } },
+      )
+    }
+    if (url.pathname.startsWith("/api/")) {
+      const response = await api.handler(request)
+      response.headers.set("cache-control", "no-store")
+      return response
+    }
+    const file = Bun.file(path.join(directory, url.pathname))
+    if (!url.pathname.endsWith("/") && (await file.exists())) {
+      return new Response(file, {
+        headers: {
+          "cache-control": url.pathname.startsWith("/_assets/") ? "public, max-age=31536000, immutable" : "no-cache",
+        },
+      })
+    }
+    return new Response(Bun.file(path.join(directory, "index.html")), { headers: { "cache-control": "no-cache" } })
+  },
+})
+console.log(`Tab fixture: ${server.url} (${directory})`)
+const close = async () => {
+  await server.stop(true)
+  await api.dispose()
+}
+process.once("SIGINT", close)
+process.once("SIGTERM", close)
@@ -0,0 +1,33 @@
+import type { NavigationMilestoneSample } from "./navigation-milestones"
+import { measureNavigationMilestones } from "./navigation-milestones"
+import { benchmark, expect } from "../benchmark"
+
+benchmark(
+  "navigation milestones start at mousedown and wait for the expected ready controls",
+  async ({ page, report }) => {
+    await page.setContent('<button id="open">Open</button><input id="editor" disabled><span id="model">Loading</span>')
+    const result = await measureNavigationMilestones(page, {
+      triggerSelector: "#open",
+      milestones: { editor: { selector: "#editor:enabled:focus" }, model: { selector: "#model", text: "Ready model" } },
+      navigate: async () => {
+        await page.getByRole("button", { name: "Open", exact: true }).dispatchEvent("mousedown", { button: 0 })
+        await page.locator("#editor").evaluate((element: HTMLInputElement) => {
+          element.disabled = false
+          element.focus()
+        })
+        await page.waitForFunction(() => {
+          const samples = (window as Window & { __navigationMilestones?: { samples: NavigationMilestoneSample[] } })
+            .__navigationMilestones?.samples
+          return samples?.some((sample) => sample.milestones.editor && !sample.milestones.model)
+        })
+        await page.locator("#model").evaluate((element) => {
+          element.textContent = "Ready model"
+        })
+      },
+    })
+    expect(result.summary.all.firstObservedMs).not.toBeNull()
+    expect(result.summary.all.firstObservedMs).toBeGreaterThan(result.summary.milestones.editor.firstObservedMs!)
+    expect(await page.evaluate(() => "__navigationMilestones" in window)).toBe(false)
+    report(result)
+  },
+)
@@ -36,7 +36,7 @@ export async function measureNavigationMilestones(
  page: Page,
  input: {
    triggerSelector: string
-    milestones: Record<string, { selector: string; visible?: boolean }>
+    milestones: Record<string, { selector: string; visible?: boolean; text?: string }>
    navigate: () => Promise<void>
  },
 ) {
@@ -47,11 +47,19 @@ export async function measureNavigationMilestones(
      const marked = new Set<string>()
      let started: number | undefined
      let running = true
-      const visible = (selector: string) =>
+      const visible = (selector: string, text?: string) =>
        [...document.querySelectorAll<HTMLElement>(selector)].some((element) => {
+          if (!element.checkVisibility({ checkOpacity: true, checkVisibilityCSS: true })) return false
+          if (text !== undefined && element.textContent?.replace(/\s+/g, " ").trim() !== text) return false
          const rect = element.getBoundingClientRect()
-          const style = getComputedStyle(element)
-          return rect.width > 0 && rect.height > 0 && style.visibility !== "hidden" && style.display !== "none"
+          return (
+            rect.width > 0 &&
+            rect.height > 0 &&
+            rect.bottom > 0 &&
+            rect.top < innerHeight &&
+            rect.right > 0 &&
+            rect.left < innerWidth
+          )
        })
      const sample = () => {
        if (!running || started === undefined) return
@@ -61,7 +69,9 @@ export async function measureNavigationMilestones(
            const current = Object.fromEntries(
              Object.entries(milestones).map(([name, milestone]) => [
                name,
-                milestone.visible === false ? !document.querySelector(milestone.selector) : visible(milestone.selector),
+                milestone.visible === false
+                  ? !document.querySelector(milestone.selector)
+                  : visible(milestone.selector, milestone.text),
              ]),
            )
            samples.push({
@@ -93,36 +103,46 @@ export async function measureNavigationMilestones(
          }, 0)
        })
      }
-      document.addEventListener(
-        "click",
-        (event) => {
-          if (!(event.target instanceof Element) || !event.target.closest(triggerSelector)) return
-          started = performance.now()
-          performance.mark("opencode.navigation.click")
-          sample()
-        },
-        { capture: true, once: true },
-      )
+      const start = (event: MouseEvent) => {
+        if (started !== undefined || event.button !== 0) return
+        if (!(event.target instanceof Element) || !event.target.closest(triggerSelector)) return
+        started = performance.now()
+        performance.mark("opencode.navigation.start")
+        sample()
+      }
+      document.addEventListener("mousedown", start, true)
+      document.addEventListener("click", start, true)
      ;(window as Window & { __navigationMilestones?: NavigationMilestoneProbe }).__navigationMilestones = {
        samples,
        stop: () => {
          running = false
+          document.removeEventListener("mousedown", start, true)
+          document.removeEventListener("click", start, true)
        },
      }
    },
    { triggerSelector: input.triggerSelector, milestones: input.milestones },
  )
-  await input.navigate()
-  await page.waitForFunction(() => {
-    const samples = (window as Window & { __navigationMilestones?: NavigationMilestoneProbe }).__navigationMilestones
-      ?.samples
-    if (!samples || samples.length < 3) return false
-    return samples.slice(-3).every((sample) => Object.values(sample.milestones).every(Boolean))
-  })
-  const samples = await page.evaluate(() => {
-    const probe = (window as Window & { __navigationMilestones?: NavigationMilestoneProbe }).__navigationMilestones!
-    probe.stop()
-    return probe.samples
-  })
-  return { summary: summarizeNavigationMilestones(samples), samples }
+  try {
+    await input.navigate()
+    await page.waitForFunction(() => {
+      const samples = (window as Window & { __navigationMilestones?: NavigationMilestoneProbe }).__navigationMilestones
+        ?.samples
+      return (
+        samples &&
+        samples.length >= 3 &&
+        samples.slice(-3).every((sample) => Object.values(sample.milestones).every(Boolean))
+      )
+    })
+    const samples = await page.evaluate(
+      () => (window as Window & { __navigationMilestones?: NavigationMilestoneProbe }).__navigationMilestones!.samples,
+    )
+    return { summary: summarizeNavigationMilestones(samples), samples }
+  } finally {
+    await page.evaluate(() => {
+      const host = window as Window & { __navigationMilestones?: NavigationMilestoneProbe }
+      host.__navigationMilestones?.stop()
+      delete host.__navigationMilestones
+    })
+  }
 }
@@ -0,0 +1,91 @@
+import { benchmark, benchmarkDiagnostics, expect } from "../benchmark"
+import { measureNavigationMilestones } from "./navigation-milestones"
+import { fixture } from "./session-timeline-stress.fixture"
+import { measureSessionSwitch, waitForStableTimeline } from "./session-tab-switch-probe"
+import { installStressSessionTabs, mockStressTimeline, stressSessionHref } from "./timeline-test-helpers"
+
+benchmark.use({
+  viewport: { width: 1440, height: 900 },
+  serviceWorkers: "block",
+  traceScope: "interaction",
+  trace: "off",
+  video: "off",
+})
+
+for (const entry of ["home", "session"] as const) {
+  benchmark(`entry: new session from ${entry}`, async ({ page, report }) => {
+    await mockStressTimeline(page)
+    await installStressSessionTabs(page, { sessionIDs: entry === "home" ? [] : [fixture.sourceID] })
+    await page.goto(entry === "home" ? "/" : stressSessionHref(fixture.sourceID))
+    if (entry === "session") await waitForStableTimeline(page, fixture.expected.sourceMessageIDs.at(-1)!)
+    const trigger = entry === "home" ? '[data-action="home-new-session"]' : 'button[aria-label="New session"]'
+    await expect(page.locator(trigger)).toBeVisible()
+    await expect(page.locator('[data-component="new-session"]')).toHaveCount(0)
+    const writes: string[] = []
+    page.on("request", (request) => {
+      if (new URL(request.url()).pathname.startsWith("/api/") && !["GET", "HEAD", "OPTIONS"].includes(request.method()))
+        writes.push(request.method())
+    })
+    await benchmarkDiagnostics(page).startTrace()
+    const result = await measureNavigationMilestones(page, {
+      triggerSelector: trigger,
+      milestones: {
+        editor: {
+          selector: '[data-component="new-session"] [data-component="composer-editor"][contenteditable="true"]:focus',
+        },
+        model: { selector: '[data-component="new-session"] [data-action="composer-model"]', text: "Claude Opus 4.6" },
+        project: { selector: '[data-component="new-session"] [data-action="prompt-project"]' },
+        tab: { selector: '[data-titlebar-tab-link][href^="/new-session?draftId="]' },
+      },
+      navigate: () => page.locator(trigger).click(),
+    })
+    await benchmarkDiagnostics(page).stop()
+    const editor = page.locator('[data-component="new-session"] [data-component="composer-editor"]')
+    await expect(editor).toHaveText("")
+    await page.keyboard.type("Draft input")
+    await expect(editor).toHaveText("Draft input")
+    expect(writes).toEqual([])
+    report(
+      {
+        firstCorrectObservedMs: result.summary.all.firstObservedMs,
+        stableObservedMs: result.summary.all.stableObservedMs,
+        ...result,
+      },
+      { entry, data: "fixture", inputEvent: "mousedown" },
+    )
+  })
+}
+
+benchmark("entry: cold session from Home", async ({ page, report }) => {
+  const requests: string[] = []
+  await mockStressTimeline(page, {
+    onMessages: (request) => {
+      if (request.phase === "start") requests.push(request.sessionID)
+    },
+  })
+  await installStressSessionTabs(page, { sessionIDs: [] })
+  await page.goto("/")
+  const selector = `[data-component="home-session-row-container"][data-session-id="${fixture.targetID}"] [data-component="home-session-row"]`
+  await expect(page.locator(selector)).toBeVisible()
+  expect(requests).not.toContain(fixture.targetID)
+  const href = stressSessionHref(fixture.targetID)
+  await benchmarkDiagnostics(page).startTrace()
+  const result = await measureSessionSwitch(page, {
+    destinationIDs: fixture.messages[fixture.targetID].map((message) => message.id),
+    sourceIDs: [],
+    lastID: fixture.expected.targetMessageIDs.at(-1)!,
+    requiredPartID: fixture.expected.targetPartIDs.at(-1)!,
+    href,
+    triggerSelector: selector,
+    switch: async () => {
+      await page.locator(selector).click()
+      await waitForStableTimeline(page, fixture.expected.targetMessageIDs.at(-1)!)
+    },
+  })
+  await benchmarkDiagnostics(page).stop()
+  await expect(
+    page.locator(`[data-titlebar-tab-slot][data-active="true"] [data-titlebar-tab-link][href="${href}"]`),
+  ).toHaveCount(1)
+  expect(requests).toContain(fixture.targetID)
+  report(result, { entry: "home", data: "cold paginated fixture", inputEvent: "mousedown" })
+})
@@ -1,71 +1,129 @@
 import type { Page } from "@playwright/test"
+import { mockOpenCodeServer } from "../../utils/mock-server"
 import { expectSessionTitle } from "../../utils/waits"
-import { benchmark, expect, withBenchmarkPage } from "../benchmark"
+import { benchmark, benchmarkDiagnostics, expect } from "../benchmark"
 import { fixture } from "./session-timeline-stress.fixture"
+import { expected, messages, workload } from "./session-tab-switch.fixture"
 import {
  createReviewDiffs,
  installStressSessionTabs,
  installTimelineSettings,
-  mockStressTimeline,
  stressSessionHref,
 } from "./timeline-test-helpers"
 import { measureSessionSwitch, waitForStableTimeline } from "./session-tab-switch-probe"

 const scenarios = [
-  { cached: false, review: false, resized: false },
-  { cached: false, review: true, resized: false },
-  { cached: true, review: false, resized: false },
-  { cached: true, review: true, resized: false },
-  { cached: true, review: true, resized: true },
-]
+  { cache: "cold", review: "closed" },
+  { cache: "cold", review: "open" },
+  { cache: "warm", review: "closed" },
+  { cache: "warm", review: "open" },
+  { cache: "warm", review: "resized" },
+] as const
+
+const viewport = { width: 1440, height: 900 }
+const reviewDiffs = createReviewDiffs()
+benchmark.use({ viewport, video: "off", trace: "off", serviceWorkers: "block", traceScope: "interaction" })

 scenarios.forEach((scenario) => {
-  const name = `tab switch: ${scenario.cached ? "cached" : "unmounted"}, review ${scenario.review ? "open" : "closed"}${scenario.resized ? ", resized" : ""}`
-  benchmark(name, async ({ browser, report }, testInfo) => {
-    const result = await withBenchmarkPage(
-      browser,
-      name,
-      async (page) => {
-        await mockStressTimeline(page, { vcsDiff: createReviewDiffs() })
-        await installTimelineSettings(page)
-        await installStressSessionTabs(page)
-        await page.goto(stressSessionHref(fixture.sourceID))
-        await expectSessionTitle(page, fixture.expected.sourceTitle)
-        await waitForStableTimeline(page, fixture.expected.sourceMessageIDs.at(-1)!)
-        if (scenario.review && !scenario.resized) await openReviewPane(page)
-        if (scenario.cached) {
-          await switchSession(page, fixture.targetID, fixture.expected.targetTitle)
-          const answer = page.locator(`[data-timeline-part-id="${fixture.expected.targetPartIDs.at(-1)}"]`)
-          await expect(answer.locator('[data-component="markdown"]')).toHaveAttribute("data-markdown-ready", "")
-          await expect
-            .poll(() =>
-              answer.evaluate((element) => element.checkVisibility({ checkOpacity: true, checkVisibilityCSS: true })),
-            )
-            .toBe(true)
-          await waitForStableTimeline(page, fixture.expected.targetMessageIDs.at(-1)!)
-          await switchSession(page, fixture.sourceID, fixture.expected.sourceTitle)
-        }
-        if (scenario.resized) await openReviewPane(page)
-        await waitForStableTimeline(page, fixture.expected.sourceMessageIDs.at(-1)!)
+  benchmark(`tab switch: ${scenario.cache}, review ${scenario.review}`, async ({ page, report }, testInfo) => {
+    const requests = await prepareSessionTabs(page)
+    if (scenario.review === "open") await openReviewPane(page)
+    if (scenario.cache === "warm") {
+      await switchSession(page, fixture.targetID, fixture.expected.targetTitle)
+      await expectReadyTimeline(page, fixture.targetID)
+      await switchSession(page, fixture.sourceID, fixture.expected.sourceTitle)
+    }
+    if (scenario.review === "resized") await openReviewPane(page)
+    await expectReadyTimeline(page, fixture.sourceID)
+    await benchmarkDiagnostics(page).startTrace()
+    const requestsBefore = requests.length

-        return measureSessionSwitch(page, {
-          destinationIDs: fixture.messages[fixture.targetID].map((message) => message.id),
-          sourceIDs: fixture.messages[fixture.sourceID].map((message) => message.id),
-          lastID: fixture.expected.targetMessageIDs.at(-1)!,
-          requiredPartID: fixture.expected.targetPartIDs.at(-1),
-          href: stressSessionHref(fixture.targetID),
-          switch: () => switchSession(page, fixture.targetID, fixture.expected.targetTitle),
-        })
+    const result = await measureSessionSwitch(page, {
+      destinationIDs: messages[fixture.targetID].map((message) => message.id),
+      sourceIDs: messages[fixture.sourceID].map((message) => message.id),
+      lastID: expected[fixture.targetID].lastID,
+      requiredPartID: expected[fixture.targetID].answerID,
+      href: stressSessionHref(fixture.targetID),
+      switch: () => switchSession(page, fixture.targetID, fixture.expected.targetTitle),
+    })
+
+    expect(result.firstCorrectObservedMs).not.toBeNull()
+    expect(result.stableObservedMs).not.toBeNull()
+    expect(requests).toHaveLength(requestsBefore)
+    await expectReadyTimeline(page, fixture.targetID)
+    report(
+      {
+        ...result,
+        messageRequestsDuringSwitch: requests.length - requestsBefore,
+        rendererMemory:
+          process.env.OPENCODE_PERFORMANCE_MEMORY === "1" ? await retainedRendererMemory(page) : undefined,
+      },
+      {
+        ...scenario,
+        ...workload,
+        viewport,
+        browserVersion: page.context().browser()!.version(),
+        serviceWorkers: "blocked",
+        reviewFiles: scenario.review === "closed" ? 0 : reviewDiffs.length,
+        data: "prefetched",
+        transport: process.env.OPENCODE_PERFORMANCE_HTTP_FIXTURE === "1" ? "http" : "playwright-route",
+        inputEvent: "mousedown",
+        requireReadyAnswer: true,
      },
-      testInfo,
    )
-    expect(result.unknownSamples).toBe(0)
-    expect(result.wrongDestinationSamples).toBe(0)
-    if (scenario.cached) expect(result.blankSamples).toBe(0)
-    report(result, { ...scenario, inputEvent: "mousedown", requireReadyAnswer: true })
+    if (testInfo.repeatEachIndex === 0) {
+      await page.screenshot({ path: testInfo.outputPath("destination.png") })
+      await testInfo.attach("destination", { path: testInfo.outputPath("destination.png"), contentType: "image/png" })
+    }
  })
 })

+async function prepareSessionTabs(page: Page) {
+  const requests: string[] = []
+  page.on("request", (request) => {
+    if (request.method() !== "GET") return
+    const match = new URL(request.url()).pathname.match(/^\/api\/session\/([^/]+)\/message$/)
+    if (match) requests.push(decodeURIComponent(match[1]))
+  })
+  if (process.env.OPENCODE_PERFORMANCE_HTTP_FIXTURE !== "1")
+    await mockOpenCodeServer(page, {
+      sessions: fixture.sessions,
+      provider: fixture.provider,
+      directory: fixture.directory,
+      project: fixture.project,
+      // Return the full history so every scenario exercises a long loaded timeline.
+      pageMessages: (sessionID) => ({ items: messages[sessionID] ?? [] }),
+      vcsDiff: reviewDiffs,
+    })
+  await installTimelineSettings(page)
+  await installStressSessionTabs(page)
+  // Restored tabs prefetch their data even when their transcript has never rendered.
+  const prefetch = page.waitForResponse((response) =>
+    new URL(response.url()).pathname.endsWith(`/session/${fixture.targetID}/message`),
+  )
+  await page.goto(stressSessionHref(fixture.sourceID))
+  expect(await (await prefetch).finished()).toBeNull()
+  await expectSessionTitle(page, fixture.expected.sourceTitle)
+  await expectReadyTimeline(page, fixture.sourceID)
+  await expect(page.locator(`[data-timeline-part-id="${expected[fixture.targetID].answerID}"]`)).toHaveCount(0)
+  expect(requests.toSorted()).toEqual([fixture.sourceID, fixture.targetID].toSorted())
+  return requests
+}
+
+async function expectReadyTimeline(page: Page, sessionID: string) {
+  const answer = page.locator(`[data-timeline-part-id="${expected[sessionID].answerID}"]`)
+  await expect(answer.locator('[data-component="markdown"]')).toHaveAttribute("data-markdown-ready", "")
+  await expect(answer.getByRole("table")).toHaveCount(1)
+  await expect(answer.locator("pre")).toHaveCount(4)
+  await expect
+    .poll(() => answer.evaluate((element) => element.checkVisibility({ checkOpacity: true, checkVisibilityCSS: true })))
+    .toBe(true)
+  await waitForStableTimeline(page, expected[sessionID].lastID)
+  await expect(page.locator('[data-timeline-key] [data-component="markdown"]:not([data-markdown-ready])')).toHaveCount(
+    0,
+  )
+}
+
 async function switchSession(page: Page, sessionID: string, title: string) {
  const tab = page.locator(`[data-slot="titlebar-tabs"] a[href="${stressSessionHref(sessionID)}"]`)
  await expect(tab).toHaveCount(1)
@@ -80,4 +138,20 @@ async function openReviewPane(page: Page) {
    const text = document.querySelector("#review-panel")?.textContent ?? ""
    return text.includes("generated-000.ts") && text.includes("+3")
  })
+  await page.locator('[data-slot="session-chat-panel"]').evaluate(async (panel) => {
+    await Promise.all(panel.getAnimations().map((animation) => animation.finished))
+  })
+}
+
+async function retainedRendererMemory(page: Page) {
+  const cdp = await page.context().newCDPSession(page)
+  try {
+    await cdp.send("HeapProfiler.collectGarbage")
+    return {
+      heap: await cdp.send("Runtime.getHeapUsage"),
+      dom: await cdp.send("Memory.getDOMCounters"),
+    }
+  } finally {
+    await cdp.detach()
+  }
 }
@@ -2,66 +2,77 @@ import { benchmark, expect } from "../benchmark"
 import { measureSessionSwitch } from "./session-tab-switch-probe"
 import type { SessionSwitchSample } from "./session-tab-switch-metrics"

-benchmark("starts at mousedown and excludes hidden or unfinished destination content", async ({ page, report }) => {
-  await page.setContent(`
-    <a href="/session/destination">Destination</a>
+for (const tag of ["a", "button"] as const) {
+  benchmark(
+    `starts at ${tag} mousedown and excludes hidden or unfinished destination content`,
+    async ({ page, report }) => {
+      await page.setContent(`
+    <${tag} id="destination" ${tag === "a" ? 'href="/session/destination"' : 'type="button"'}>Destination</${tag}>
    <div class="scroll-view__viewport" style="height:200px;overflow:auto">
      <div data-timeline-row="message" data-timeline-key="row" data-message-id="source">
        <div data-timeline-part-id="answer"><div data-component="markdown">Destination answer</div></div>
      </div>
    </div>
  `)
-  await page.evaluate(() => {
-    document.querySelector("a")!.addEventListener("mousedown", () => {
-      const row = document.querySelector<HTMLElement>("[data-message-id]")!
-      row.dataset.messageId = "destination"
-      row.style.visibility = "hidden"
-    })
-  })
-  const result = await measureSessionSwitch(page, {
-    destinationIDs: ["destination"],
-    sourceIDs: ["source"],
-    lastID: "destination",
-    requiredPartID: "answer",
-    requireBottomAnchor: false,
-    href: "/session/destination",
-    switch: async () => {
-      // No click is dispatched: the probe must observe the event that activates tabs.
-      await page.getByRole("link", { name: "Destination" }).dispatchEvent("mousedown", { button: 0 })
-      await page.waitForFunction(() => {
-        const host = window as Window & { __sessionSwitchProbe?: { samples: SessionSwitchSample[] } }
-        return host.__sessionSwitchProbe?.samples.some((sample) => !sample.hasVisibleRows)
+      await page.evaluate(() => {
+        document.querySelector("#destination")!.addEventListener("mousedown", () => {
+          const row = document.querySelector<HTMLElement>("[data-message-id]")!
+          row.dataset.messageId = "destination"
+          row.style.visibility = "hidden"
+        })
      })
-      await page.locator("[data-message-id]").evaluate((row) => row.style.removeProperty("visibility"))
-      await page.waitForFunction(() => {
-        const host = window as Window & { __sessionSwitchProbe?: { samples: SessionSwitchSample[] } }
-        return host.__sessionSwitchProbe?.samples.some(
-          (sample) => sample.destination.length > 0 && sample.requiredPartVisible === false,
-        )
-      })
-      const beforeClip = await page.evaluate(() => {
-        const row = document.querySelector<HTMLElement>("[data-timeline-key]")!
-        row.style.cssText = "height:10px;position:relative;overflow:clip"
-        const answer = row.querySelector<HTMLElement>("[data-timeline-part-id]")!
-        answer.style.cssText = "position:absolute;top:30px;width:150px"
-        answer.querySelector('[data-component="markdown"]')!.setAttribute("data-markdown-ready", "")
-        return (
-          (window as Window & { __sessionSwitchProbe?: { samples: SessionSwitchSample[] } }).__sessionSwitchProbe
-            ?.samples.length ?? 0
-        )
-      })
-      await page.waitForFunction((count) => {
-        const host = window as Window & { __sessionSwitchProbe?: { samples: SessionSwitchSample[] } }
-        return host.__sessionSwitchProbe?.samples.slice(count).some((sample) => sample.requiredPartVisible === false)
-      }, beforeClip)
-      await page.locator("[data-timeline-key]").evaluate((row) => {
-        row.style.height = "100px"
+      const result = await measureSessionSwitch(page, {
+        destinationIDs: ["destination"],
+        sourceIDs: ["source"],
+        lastID: "destination",
+        requiredPartID: "answer",
+        requireBottomAnchor: false,
+        href: "/session/destination",
+        triggerSelector: tag === "button" ? "#destination" : undefined,
+        switch: async () => {
+          // No click is dispatched: the probe must observe the event that activates tabs.
+          await page
+            .getByRole(tag === "a" ? "link" : "button", { name: "Destination", exact: true })
+            .dispatchEvent("mousedown", { button: 0 })
+          await page.waitForFunction(() => {
+            const host = window as Window & { __sessionSwitchProbe?: { samples: SessionSwitchSample[] } }
+            return host.__sessionSwitchProbe?.samples.some((sample) => !sample.hasVisibleRows)
+          })
+          await page.locator("[data-message-id]").evaluate((row) => row.style.removeProperty("visibility"))
+          await page.waitForFunction(() => {
+            const host = window as Window & { __sessionSwitchProbe?: { samples: SessionSwitchSample[] } }
+            return host.__sessionSwitchProbe?.samples.some(
+              (sample) => sample.destination.length > 0 && sample.requiredPartVisible === false,
+            )
+          })
+          const beforeClip = await page.evaluate(() => {
+            const row = document.querySelector<HTMLElement>("[data-timeline-key]")!
+            row.style.cssText = "height:10px;position:relative;overflow:clip"
+            const answer = row.querySelector<HTMLElement>("[data-timeline-part-id]")!
+            answer.style.cssText = "position:absolute;top:30px;width:150px"
+            answer.querySelector('[data-component="markdown"]')!.setAttribute("data-markdown-ready", "")
+            return (
+              (window as Window & { __sessionSwitchProbe?: { samples: SessionSwitchSample[] } }).__sessionSwitchProbe
+                ?.samples.length ?? 0
+            )
+          })
+          await page.waitForFunction((count) => {
+            const host = window as Window & { __sessionSwitchProbe?: { samples: SessionSwitchSample[] } }
+            return host.__sessionSwitchProbe?.samples
+              .slice(count)
+              .some((sample) => sample.requiredPartVisible === false)
+          }, beforeClip)
+          await page.locator("[data-timeline-key]").evaluate((row) => {
+            row.style.height = "100px"
+          })
+        },
      })
+      expect(result.blankSamples).toBeGreaterThan(0)
+      expect(result.firstCorrectObservedMs).not.toBeNull()
+      expect(result.stableObservedMs).not.toBeNull()
+      expect(result.firstCorrectObservedMs).toBeGreaterThan(result.firstDestinationObservedMs!)
+      expect(await page.evaluate(() => "__sessionSwitchProbe" in window)).toBe(false)
+      report(result)
    },
-  })
-  expect(result.blankSamples).toBeGreaterThan(0)
-  expect(result.firstCorrectObservedMs).not.toBeNull()
-  expect(result.stableObservedMs).not.toBeNull()
-  expect(result.firstCorrectObservedMs).toBeGreaterThan(result.firstDestinationObservedMs!)
-  report(result)
-})
+  )
+}
@@ -14,124 +14,132 @@ async function installSessionSwitchProbe(
    lastID: string
    requiredPartID?: string
    requireBottomAnchor?: boolean
+    triggerSelector?: string
    href: string
  },
 ) {
-  await page.evaluate(({ destinationIDs, sourceIDs, lastID, requiredPartID, requireBottomAnchor, href }) => {
-    const destination = new Set(destinationIDs)
-    const source = new Set(sourceIDs)
-    const samples: SessionSwitchSample[] = []
-    let started: number | undefined
-    let running = true
-    const reviewLevels: Record<string, string> = {
-      panel: "#review-panel",
-      tabs: '#review-panel [data-component="tabs"]',
-      body: '#review-panel [data-slot="session-review-v2-body"]',
-      review: '#review-panel [data-component="session-review-v2"]',
-      preview: '#review-panel [data-slot="session-review-v2-preview"]',
-      scroll: '#review-panel [data-slot="session-review-v2-diff-scroll"]',
-      file: '#review-panel [data-component="file"][data-mode="diff"]',
-    }
-    const initialReviewNodes: Record<string, Element | null> = {}
-    const sample = () => {
-      if (!running || started === undefined) return
-      setTimeout(() => {
+  await page.evaluate(
+    ({ destinationIDs, sourceIDs, lastID, requiredPartID, requireBottomAnchor, triggerSelector, href }) => {
+      const destination = new Set(destinationIDs)
+      const source = new Set(sourceIDs)
+      const samples: SessionSwitchSample[] = []
+      let started: number | undefined
+      let running = true
+      const reviewLevels: Record<string, string> = {
+        panel: "#review-panel",
+        tabs: '#review-panel [data-component="tabs"]',
+        body: '#review-panel [data-slot="session-review-v2-body"]',
+        review: '#review-panel [data-component="session-review-v2"]',
+        preview: '#review-panel [data-slot="session-review-v2-preview"]',
+        scroll: '#review-panel [data-slot="session-review-v2-diff-scroll"]',
+        file: '#review-panel [data-component="file"][data-mode="diff"]',
+      }
+      const initialReviewNodes: Record<string, Element | null> = {}
+      const sample = () => {
        if (!running || started === undefined) return
-        const reviewPanel = document.querySelector<HTMLElement>("#review-panel")
-        const reviewFile = reviewPanel?.querySelector('[data-component="file"][data-mode="diff"]')
-        const initialReviewFile = initialReviewNodes.file
-        const replacedLevels = Object.entries(reviewLevels).flatMap(([name, selector]) => {
-          const initial = initialReviewNodes[name]
-          if (!initial) return []
-          const current = document.querySelector(selector)
-          return current && current !== initial ? [name] : []
-        })
-        const review = reviewPanel
-          ? {
-              fileHost: !!reviewFile,
-              fileHostReplaced: !!initialReviewFile && !!reviewFile && reviewFile !== initialReviewFile,
-              header:
-                reviewPanel
-                  .querySelector<HTMLElement>('[data-slot="session-review-v2-file-header"]')
-                  ?.textContent?.trim() ?? "",
-              replacedLevels,
-            }
-          : undefined
-        const root = [...document.querySelectorAll<HTMLElement>(".scroll-view__viewport")].find((element) =>
-          element.querySelector("[data-timeline-row]"),
-        )
-        if (root) {
-          const view = root.getBoundingClientRect()
-          const inViewport = (element: HTMLElement) => {
-            if (!element.checkVisibility({ checkOpacity: true, checkVisibilityCSS: true })) return false
-            const rect = element.getBoundingClientRect()
-            const clip = element.closest<HTMLElement>("[data-timeline-key]")?.getBoundingClientRect() ?? view
-            return (
-              Math.min(rect.bottom, clip.bottom, view.bottom) > Math.max(rect.top, clip.top, view.top) &&
-              Math.min(rect.right, clip.right, view.right) > Math.max(rect.left, clip.left, view.left)
-            )
-          }
-          const visible = [...root.querySelectorAll<HTMLElement>("[data-message-id]")]
-            .filter(inViewport)
-            .map((element) => element.dataset.messageId!)
-          const hasVisibleRows = [...root.querySelectorAll<HTMLElement>("[data-timeline-key]")].some(inViewport)
-          const requiredPartVisible = requiredPartID
-            ? [...root.querySelectorAll<HTMLElement>("[data-timeline-part-id]")].some((element) => {
-                if (element.dataset.timelinePartId !== requiredPartID) return false
-                if (!element.textContent?.trim()) return false
-                if (element.querySelector('[data-component="markdown"]:not([data-markdown-ready])')) return false
-                return inViewport(element)
-              })
+        setTimeout(() => {
+          if (!running || started === undefined) return
+          const reviewPanel = document.querySelector<HTMLElement>("#review-panel")
+          const reviewFile = reviewPanel?.querySelector('[data-component="file"][data-mode="diff"]')
+          const initialReviewFile = initialReviewNodes.file
+          const replacedLevels = Object.entries(reviewLevels).flatMap(([name, selector]) => {
+            const initial = initialReviewNodes[name]
+            if (!initial) return []
+            const current = document.querySelector(selector)
+            return current && current !== initial ? [name] : []
+          })
+          const review = reviewPanel
+            ? {
+                fileHost: !!reviewFile,
+                fileHostReplaced: !!initialReviewFile && !!reviewFile && reviewFile !== initialReviewFile,
+                header:
+                  reviewPanel
+                    .querySelector<HTMLElement>('[data-slot="session-review-v2-file-header"]')
+                    ?.textContent?.trim() ?? "",
+                replacedLevels,
+              }
            : undefined
-          const spacer = root.querySelector<HTMLElement>('[data-timeline-row="bottom-spacer"]')?.getBoundingClientRect()
-          samples.push({
-            observedAtMs: performance.now() - started,
-            destination: visible.filter((id) => destination.has(id)),
-            source: visible.filter((id) => source.has(id)),
-            hasVisibleRows,
-            last: visible.includes(lastID),
-            requiredPartVisible,
-            bottomAnchorRequired: requireBottomAnchor !== false,
-            bottomErrorPx: spacer ? spacer.bottom - view.bottom : undefined,
-            review,
-          })
-        } else {
-          samples.push({
-            observedAtMs: performance.now() - started,
-            destination: [],
-            source: [],
-            hasVisibleRows: false,
-            last: false,
-            requiredPartVisible: requiredPartID ? false : undefined,
-            bottomAnchorRequired: requireBottomAnchor !== false,
-            review,
-          })
+          const root = [...document.querySelectorAll<HTMLElement>(".scroll-view__viewport")].find((element) =>
+            element.querySelector("[data-timeline-row]"),
+          )
+          if (root) {
+            const view = root.getBoundingClientRect()
+            const inViewport = (element: HTMLElement) => {
+              if (!element.checkVisibility({ checkOpacity: true, checkVisibilityCSS: true })) return false
+              const rect = element.getBoundingClientRect()
+              const clip = element.closest<HTMLElement>("[data-timeline-key]")?.getBoundingClientRect() ?? view
+              return (
+                Math.min(rect.bottom, clip.bottom, view.bottom) > Math.max(rect.top, clip.top, view.top) &&
+                Math.min(rect.right, clip.right, view.right) > Math.max(rect.left, clip.left, view.left)
+              )
+            }
+            const visible = [...root.querySelectorAll<HTMLElement>("[data-message-id]")]
+              .filter(inViewport)
+              .map((element) => element.dataset.messageId!)
+            const hasVisibleRows = [...root.querySelectorAll<HTMLElement>("[data-timeline-key]")].some(inViewport)
+            const requiredPartVisible = requiredPartID
+              ? [...root.querySelectorAll<HTMLElement>("[data-timeline-part-id]")].some((element) => {
+                  if (element.dataset.timelinePartId !== requiredPartID) return false
+                  if (!element.textContent?.trim()) return false
+                  if (element.querySelector('[data-component="markdown"]:not([data-markdown-ready])')) return false
+                  return inViewport(element)
+                })
+              : undefined
+            const spacer = root
+              .querySelector<HTMLElement>('[data-timeline-row="bottom-spacer"]')
+              ?.getBoundingClientRect()
+            samples.push({
+              observedAtMs: performance.now() - started,
+              destination: visible.filter((id) => destination.has(id)),
+              source: visible.filter((id) => source.has(id)),
+              hasVisibleRows,
+              last: visible.includes(lastID),
+              requiredPartVisible,
+              bottomAnchorRequired: requireBottomAnchor !== false,
+              bottomErrorPx: spacer ? spacer.bottom - view.bottom : undefined,
+              review,
+            })
+          } else {
+            samples.push({
+              observedAtMs: performance.now() - started,
+              destination: [],
+              source: [],
+              hasVisibleRows: false,
+              last: false,
+              requiredPartVisible: requiredPartID ? false : undefined,
+              bottomAnchorRequired: requireBottomAnchor !== false,
+              review,
+            })
+          }
+          requestAnimationFrame(sample)
+        }, 0)
+      }
+      const start = (event: MouseEvent) => {
+        if (started !== undefined || event.button !== 0) return
+        const trigger = event.target instanceof Element ? event.target.closest(triggerSelector ?? "a") : undefined
+        if (!trigger || (!triggerSelector && trigger.getAttribute("href") !== href)) return
+        started = performance.now()
+        performance.mark("session-switch:start", { startTime: started })
+        for (const [name, selector] of Object.entries(reviewLevels)) {
+          initialReviewNodes[name] = document.querySelector(selector)
        }
        requestAnimationFrame(sample)
-      }, 0)
-    }
-    const start = (event: MouseEvent) => {
-      if (started !== undefined || event.button !== 0) return
-      const link = event.target instanceof Element ? event.target.closest("a") : undefined
-      if (link?.getAttribute("href") !== href) return
-      started = performance.now()
-      for (const [name, selector] of Object.entries(reviewLevels)) {
-        initialReviewNodes[name] = document.querySelector(selector)
      }
-      requestAnimationFrame(sample)
-    }
-    // Tabs activate on mousedown; click alone misses the synchronous navigation work.
-    document.addEventListener("mousedown", start, true)
-    document.addEventListener("click", start, true)
-    ;(window as Window & { __sessionSwitchProbe?: SessionSwitchProbe }).__sessionSwitchProbe = {
-      samples,
-      stop: () => {
-        running = false
-        document.removeEventListener("mousedown", start, true)
-        document.removeEventListener("click", start, true)
-      },
-    }
-  }, input)
+      // Tabs activate on mousedown; click alone misses the synchronous navigation work.
+      document.addEventListener("mousedown", start, true)
+      document.addEventListener("click", start, true)
+      ;(window as Window & { __sessionSwitchProbe?: SessionSwitchProbe }).__sessionSwitchProbe = {
+        samples,
+        stop: () => {
+          running = false
+          document.removeEventListener("mousedown", start, true)
+          document.removeEventListener("click", start, true)
+          Object.keys(initialReviewNodes).forEach((key) => (initialReviewNodes[key] = null))
+        },
+      }
+    },
+    input,
+  )
 }

 async function waitForStableSessionSwitch(page: Page) {
@@ -159,9 +167,17 @@ async function collectSessionSwitchResult(page: Page) {
  const samples = await page.evaluate(() => {
    const probe = (window as Window & { __sessionSwitchProbe?: SessionSwitchProbe }).__sessionSwitchProbe!
    probe.stop()
+    delete (window as Window & { __sessionSwitchProbe?: SessionSwitchProbe }).__sessionSwitchProbe
    return probe.samples
  })
-  return classifySessionSwitch(samples)
+  const result = classifySessionSwitch(samples)
+  await page.evaluate(({ firstCorrectObservedMs, stableObservedMs }) => {
+    const start = performance.getEntriesByName("session-switch:start").at(-1)!.startTime
+    if (firstCorrectObservedMs !== null)
+      performance.mark("session-switch:ready", { startTime: start + firstCorrectObservedMs })
+    if (stableObservedMs !== null) performance.mark("session-switch:stable", { startTime: start + stableObservedMs })
+  }, result)
+  return result
 }

 export async function measureSessionSwitch(
@@ -172,6 +188,7 @@ export async function measureSessionSwitch(
    lastID: string
    requiredPartID?: string
    requireBottomAnchor?: boolean
+    triggerSelector?: string
    href: string
    switch: () => Promise<void>
  },
@@ -185,6 +202,7 @@ export async function measureSessionSwitch(
  } finally {
    await page.evaluate(() => {
      ;(window as Window & { __sessionSwitchProbe?: SessionSwitchProbe }).__sessionSwitchProbe?.stop()
+      delete (window as Window & { __sessionSwitchProbe?: SessionSwitchProbe }).__sessionSwitchProbe
    })
  }
 }
@@ -0,0 +1,143 @@
+import type { SessionMessageInfo } from "@opencode-ai/client/promise"
+import { fixture } from "./session-timeline-stress.fixture"
+
+export const exchanges = 200
+
+export const messages: Record<string, SessionMessageInfo[]> = Object.fromEntries(
+  [fixture.sourceID, fixture.targetID].map((sessionID) => [
+    sessionID,
+    Array.from({ length: exchanges }, (_, index) => {
+      const seed = fixture.messages[fixture.targetID]
+      const user = seed[(index % (seed.length / 2)) * 2]!
+      const assistant = seed[(index % (seed.length / 2)) * 2 + 1]!
+      if (user.type !== "user" || assistant.type !== "assistant") throw new Error("Expected a user/assistant pair")
+      const suffix = `${sessionID}_${String(index).padStart(4, "0")}`
+      return [
+        {
+          ...user,
+          id: `msg_user_${suffix}`,
+          time: { created: 1700000000000 + index * 10_000 },
+        },
+        {
+          ...assistant,
+          id: `msg_assistant_${suffix}`,
+          time: { created: 1700000001000 + index * 10_000, completed: 1700000008000 + index * 10_000 },
+          content: [
+            ...assistant.content
+              .filter((part) => part.type !== "text")
+              .map((part) => (part.type === "tool" ? { ...part, id: `${part.id}_${suffix}` } : part)),
+            { type: "text", text: complexMarkdown(sessionID, index) },
+          ],
+        },
+      ] satisfies SessionMessageInfo[]
+    }).flat(),
+  ]),
+)
+
+export const expected = Object.fromEntries(
+  [fixture.sourceID, fixture.targetID].map((sessionID) => [
+    sessionID,
+    {
+      lastID: messages[sessionID].at(-2)!.id,
+      answerID: `${messages[sessionID].at(-1)!.id}:text:0`,
+    },
+  ]),
+)
+
+export const workload = {
+  fixture: "long-complex-markdown-v1",
+  exchangesPerSession: exchanges,
+  messagesPerSession: exchanges * 2,
+  history: "full fixture history in one response",
+  sessions: Object.fromEntries(
+    Object.entries(messages).map(([sessionID, items]) => [
+      sessionID,
+      {
+        payloadBytes: Buffer.byteLength(JSON.stringify(items)),
+        markdownBytes: items.reduce(
+          (total, message) =>
+            total +
+            (message.type === "assistant"
+              ? message.content.reduce(
+                  (size, part) => size + (part.type === "text" ? Buffer.byteLength(part.text) : 0),
+                  0,
+                )
+              : 0),
+          0,
+        ),
+      },
+    ]),
+  ),
+}
+
+function complexMarkdown(sessionID: string, index: number) {
+  return `## Renderer review ${sessionID} / ${index}
+
+Preserve **semantic identity**, *measured geometry*, and ~~obsolete estimates~~ when switching sessions. The \`measureElement(node)\` result must agree with the [rendering contract](https://example.com/rendering/${sessionID}/${index}).
+
+> A completed answer contains formatted prose, highlighted source, and structured results.
+> Keep the previous view until the destination is ready, rather than exposing partially formatted content.
+
+### Readiness checklist
+
+- [x] Resolve the destination session and its messages.
+- [x] Parse Markdown and highlight fenced code.
+- [ ] Verify a different panel width.
+  - Preserve the bottom anchor.
+  - Reuse the measured rows when their width matches.
+
+| Stage | Input | Expected result | Verification |
+| :--- | ---: | :--- | :--- |
+${Array.from({ length: 8 }, (_, row) => `| stage-${index}-${row} | ${index * 8 + row} | **ready** with \`row[${row}]\` | stable geometry and visible content |`).join("\n")}
+
+### Implementation
+
+\`\`\`tsx
+import { For, Show, createMemo } from "solid-js"
+
+type Row = { id: string; title: string; ready: boolean; height: number }
+
+export function SessionRows${index}(props: { rows: Row[]; selected: string }) {
+  const visible = createMemo(() => props.rows.filter((row) => row.ready))
+  return (
+    <section aria-label="${sessionID}-${index}">
+      <For each={visible()}>{(row) => (
+        <article data-selected={row.id === props.selected}>
+          <h3>{row.title}</h3>
+          <Show when={row.height > 0} fallback={<span>Measuring</span>}>
+            <output>{row.height.toFixed(2)} pixels</output>
+          </Show>
+        </article>
+      )}</For>
+    </section>
+  )
+}
+\`\`\`
+
+\`\`\`json
+${JSON.stringify({ session: sessionID, exchange: index, stages: ["hydrate", "parse", "highlight", "measure"], viewport: { width: 1440, height: 900 }, cache: { markdown: true, geometry: true } }, null, 2)}
+\`\`\`
+
+\`\`\`sql
+SELECT session_id, COUNT(*) AS messages, MAX(created_at) AS latest
+FROM session_message
+WHERE session_id = '${sessionID}' AND ordinal >= ${index}
+GROUP BY session_id
+ORDER BY latest DESC;
+\`\`\`
+
+### Verification
+
+1. Open the long source session and wait for its final answer.
+2. Select the destination tab, without changing the viewport.
+3. Confirm that **all Markdown is ready** and the bottom anchor is correct.
+
+\`\`\`bash
+bun typecheck
+bunx playwright test --config e2e/performance/playwright.config.ts
+git diff --check # ${sessionID}-${index}
+\`\`\`
+
+**Review complete: ${sessionID} / ${index}.**
+`
+}
@@ -0,0 +1,114 @@
+import { expect, spyOn, test } from "bun:test"
+import type { FullConfig, Suite, TestCase, TestResult } from "@playwright/test/reporter"
+import { mkdtemp, readFile, rm } from "node:fs/promises"
+import os from "node:os"
+import path from "node:path"
+import TabSwitchReporter from "../tab-switch-reporter"
+
+test("summarizes each scenario and saves complete records in the configured output directory", async () => {
+  const root = await mkdtemp(path.join(os.tmpdir(), "tab-switch-reporter-"))
+  const output = path.join(root, "configured-output")
+  const log = spyOn(console, "log").mockImplementation(() => {})
+  try {
+    const reporter = new TabSwitchReporter()
+    const cases = Array.from(
+      { length: 23 },
+      (_, index) => ({ id: String(index), title: index < 20 ? "cold" : "warm" }) as TestCase,
+    )
+    const records = cases.map((_, index) => {
+      const first = index < 20 ? 20 - index : [40, 0, 20][index - 20]
+      return JSON.stringify({
+        status: "passed",
+        metrics: {
+          firstCorrectObservedMs: first,
+          stableObservedMs: first * 2,
+          samples: [{ observedAtMs: first, destination: ["answer"], source: [] }],
+        },
+        extra: { preserved: "\u03b1" },
+      })
+    })
+    reporter.onBegin({ projects: [{ outputDir: output }] } as FullConfig, { allTests: () => cases } as Suite)
+    cases.forEach((item, index) => {
+      const bytes = Buffer.from(`BENCHMARK ${records[index]}\r\n`)
+      const split = bytes.indexOf(Buffer.from("\u03b1")) + 1
+      reporter.onTestEnd(item, {
+        status: "passed",
+        stdout:
+          index === 0
+            ? [bytes.subarray(0, split), bytes.subarray(split)]
+            : ["other output\nBENCHMARK_PAGE {}\nBENCH", "MARK ", records[index], "\n"],
+      } as TestResult)
+    })
+    await reporter.onEnd({ status: "passed", startTime: new Date(0), duration: 0 })
+
+    expect(await readFile(path.join(output, "tab-switch-benchmark.jsonl"), "utf8")).toBe(`${records.join("\n")}\n`)
+    const summary = log.mock.calls.map((call) => call.join(" ")).join("\n")
+    expect(summary).toContain("cold\n  Tests: passed=20; unrun=0")
+    expect(summary).toContain("warm\n  Tests: passed=3; unrun=0")
+    expect(summary).toContain("firstCorrectObservedMs: n=20, median=10.50 ms, p95=19.00 ms")
+    expect(summary).toContain("stableObservedMs: n=20, median=21.00 ms, p95=38.00 ms")
+    expect(summary).toContain("firstCorrectObservedMs: n=3, median=20.00 ms, p95=40.00 ms")
+    expect(summary).toContain("stableObservedMs: n=3, median=40.00 ms, p95=80.00 ms")
+  } finally {
+    log.mockRestore()
+    await rm(root, { recursive: true, force: true })
+  }
+})
+
+test("reports failures, missing records, and invalid metrics without discarding raw data", async () => {
+  const output = await mkdtemp(path.join(os.tmpdir(), "tab-switch-reporter-"))
+  const log = spyOn(console, "log").mockImplementation(() => {})
+  try {
+    const reporter = new TabSwitchReporter()
+    const entries = [
+      {
+        status: "passed",
+        raw: '{"status":"passed","metrics":{"firstCorrectObservedMs":12,"stableObservedMs":24}}',
+      },
+      {
+        status: "failed",
+        raw: '{"status":"passed","metrics":{"firstCorrectObservedMs":900,"stableObservedMs":950}}',
+      },
+      {
+        status: "passed",
+        raw: '{"status":"failed","metrics":{"firstCorrectObservedMs":900,"stableObservedMs":950}}',
+      },
+      { status: "timedOut", raw: '{"status":"failed","metrics":null,"error":"Benchmark did not report metrics"}' },
+      {
+        status: "passed",
+        raw: '{"status":"passed","metrics":{"firstCorrectObservedMs":null,"stableObservedMs":40}}',
+      },
+      { status: "failed", raw: '{"status":' },
+      { status: "skipped", raw: undefined },
+    ] as const
+    const cases = Array.from(
+      { length: entries.length + 2 },
+      (_, index) => ({ id: String(index), title: index <= entries.length ? "cold" : "empty" }) as TestCase,
+    )
+    reporter.onBegin({ projects: [{ outputDir: output }] } as FullConfig, { allTests: () => cases } as Suite)
+    entries.forEach((entry, index) => {
+      reporter.onTestEnd(cases[index], {
+        status: entry.status,
+        stdout: entry.raw === undefined ? [] : [`BENCHMARK ${entry.raw}\n`],
+      } as TestResult)
+    })
+    await reporter.onEnd({ status: "interrupted", startTime: new Date(0), duration: 0 })
+
+    expect(await readFile(path.join(output, "tab-switch-benchmark.jsonl"), "utf8")).toBe(
+      entries.flatMap((entry) => (entry.raw === undefined ? [] : [`${entry.raw}\n`])).join(""),
+    )
+    const summary = log.mock.calls.map((call) => call.join(" ")).join("\n")
+    expect(summary).toContain("Tab-switch benchmark: interrupted")
+    expect(summary).toContain("Tests: passed=3, failed=2, timedOut=1, skipped=1; unrun=1")
+    expect(summary).toContain("Records: passed=3, failed=2, invalid JSON=1; missing=2; excluded=5; invalid metrics=1")
+    expect(summary).toContain("firstCorrectObservedMs: n=1, median=12.00 ms, p95=12.00 ms")
+    expect(summary).toContain("stableObservedMs: n=1, median=24.00 ms, p95=24.00 ms")
+    expect(summary).toContain("empty\n  Tests: none; unrun=1")
+    expect(summary).toContain("Records: none; missing=1; excluded=0; invalid metrics=0")
+    expect(summary).toContain("firstCorrectObservedMs: n=0, median=n/a, p95=n/a")
+    expect(summary).toContain("stableObservedMs: n=0, median=n/a, p95=n/a")
+  } finally {
+    log.mockRestore()
+    await rm(output, { recursive: true, force: true })
+  }
+})
@@ -1,4 +1,4 @@
-import { expect, test } from "@playwright/test"
+import { expect, test, type Locator } from "@playwright/test"
 import { mockOpenCodeServer } from "../utils/mock-server"
 import { expectAppVisible } from "../utils/waits"

@@ -65,11 +65,64 @@ for (const lines of [6000, 25000]) {
    await expect.poll(async () => (await input.innerText()) === text).toBe(true)
    expect(await events.evaluate((events) => events.count)).toBe(1)
    await expect(input).toBeFocused()
+    await expectCaretVisible(input)
+    const scroll = page.locator('[data-component="composer-scroll"]')
+    await expect(scroll.locator(".scroll-view__viewport")).toHaveCSS("scrollbar-width", "none")
+    await expect(scroll.locator(".scroll-view__thumb")).toBeVisible()
    await page.keyboard.type("!")
    await expect.poll(async () => (await input.innerText()) === text + "!").toBe(true)
+    await expectCaretVisible(input)
+    const thumb = await scroll.locator(".scroll-view__thumb").boundingBox()
+    const bounds = await scroll.boundingBox()
+    if (!thumb || !bounds) throw new Error("Missing composer scrollbar bounds")
+    await page.mouse.move(thumb.x + thumb.width / 2, thumb.y + thumb.height / 2)
+    await page.mouse.down()
+    await page.mouse.move(thumb.x + thumb.width / 2, bounds.y + 8 + thumb.height / 2)
+    await page.mouse.up()
+    await expect(scroll.locator(".scroll-view__viewport")).toHaveJSProperty("scrollTop", 0)
+    await expect(input).toBeFocused()
+    await page.keyboard.press("ControlOrMeta+Home")
+    await page.keyboard.press("ControlOrMeta+End")
+    await expectCaretVisible(input)
  })
 }

+async function expectCaretVisible(input: Locator) {
+  await expect
+    .poll(() =>
+      input.evaluate((element) => {
+        const selection = window.getSelection()
+        if (!selection?.isCollapsed || !selection.rangeCount || !element.contains(selection.anchorNode)) return false
+        const caret = selection.getRangeAt(0).getBoundingClientRect()
+        const viewport = (element.closest("[data-scrollable]") ?? element).getBoundingClientRect()
+        return caret.height > 0 && caret.top >= viewport.top - 1 && caret.bottom <= viewport.bottom + 1
+      }),
+    )
+    .toBe(true)
+}
+
+for (const width of [390, 1280]) {
+  for (const direction of ["ltr", "rtl"]) {
+    test(`reveals a multiline paste in the middle at ${width}px in ${direction}`, async ({ page }) => {
+      await page.setViewportSize({ width, height: 800 })
+      await page.evaluate((direction) => (document.documentElement.dir = direction), direction)
+      const input = page.getByRole("textbox", { name: "Prompt", exact: true })
+      const suffix = "\nExisting trailing content".repeat(100)
+      await input.fill("Before " + suffix)
+      await input.press("ControlOrMeta+Home")
+      await input.press("ArrowRight")
+      const text = "Pasted line /tmp/example.ts 123 \u0645\u0631\u062d\u0628\u0627\n".repeat(100) + "End of paste"
+      await page.evaluate((text) => navigator.clipboard.writeText(text), text)
+      await page.keyboard.press("ControlOrMeta+V")
+      await expect.poll(() => input.innerText()).toBe("B" + text + "efore " + suffix)
+      await expectCaretVisible(input)
+      await page.keyboard.type("!")
+      await expect.poll(() => input.innerText()).toBe("B" + text + "!efore " + suffix)
+      await expectCaretVisible(input)
+    })
+  }
+}
+
 for (const text of [
  "single line <b> &amp;",
  "first\nsecond",
@@ -0,0 +1,92 @@
+import { expect, test } from "@playwright/test"
+import { fixture } from "../smoke/session-timeline.fixture"
+import { mockOpenCodeServer } from "../utils/mock-server"
+
+test.use({ viewport: { width: 390, height: 844 }, hasTouch: true, isMobile: true })
+
+test.beforeEach(async ({ page }) => {
+  await mockOpenCodeServer(page, {
+    directory: fixture.directory,
+    project: fixture.project,
+    provider: fixture.provider,
+    sessions: fixture.sessions,
+    pageMessages: () => ({ items: [] }),
+  })
+  await page.addInitScript(
+    ({ directory, server, sessions }) => {
+      localStorage.setItem(
+        "opencode.global.dat:server",
+        JSON.stringify({ projects: { local: [{ worktree: directory, expanded: true }] } }),
+      )
+      localStorage.setItem(
+        "opencode.window.browser.dat:tabs",
+        JSON.stringify(sessions.map((session) => ({ type: "session", server, sessionId: session.id }))),
+      )
+    },
+    { directory: fixture.directory, server: fixture.serverKey, sessions: fixture.sessions },
+  )
+  await page.goto("/")
+  await expect(page.locator('[data-component="home-session-row"]')).toHaveCount(fixture.sessions.length)
+})
+
+test("mobile project selection and drawer navigation preserve session identity", async ({ page }) => {
+  const projects = page.getByRole("button", { name: "Projects", exact: true })
+  await projects.click()
+  const picker = page.getByRole("dialog", { name: "Projects", exact: true })
+  await expect(picker.getByRole("button", { name: "All projects", exact: true })).toBeVisible()
+  await picker.getByRole("button", { name: new RegExp(` ${fixture.project.name}$`) }).click()
+  await expect(picker).toBeHidden()
+  await expect(projects).toContainText(fixture.project.name)
+
+  const trigger = page.locator('[data-slot="mobile-tabs-trigger"]')
+  const drawer = page.locator('[data-slot="mobile-tabs-drawer"]')
+  await trigger.click()
+  await expect(trigger).toHaveAttribute("aria-expanded", "true")
+  await expect(drawer.locator('[data-slot="tab-project"]')).toHaveText([fixture.project.name, fixture.project.name])
+  const settings = drawer.getByRole("button", { name: "Settings", exact: true })
+  const help = drawer.getByRole("button", { name: "Help", exact: true })
+  await expect(settings).toBeVisible()
+  await expect(help).toBeVisible()
+  expect((await settings.boundingBox())?.width).toBeGreaterThan((await help.boundingBox())?.width ?? 0)
+
+  await drawer.locator('[data-slot="tab-link"]').filter({ hasText: fixture.expected.sourceTitle }).click()
+  await expect(page).toHaveURL(new RegExp(`/session/${fixture.sourceID}$`))
+  await expect(drawer).toBeHidden()
+  await expect(trigger).toContainText(fixture.expected.sourceTitle)
+  await trigger.click()
+  await drawer.getByRole("button", { name: "Home", exact: true }).click()
+  await expect(page).toHaveURL("/")
+  await expect(drawer).toBeHidden()
+  await expect(page.locator('[data-component="home-session-row"]')).toHaveCount(fixture.sessions.length)
+})
+
+test("mobile settings section menu stays above a full-width panel", async ({ page }) => {
+  await page.locator('[data-slot="mobile-tabs-trigger"]').click()
+  await page.locator('[data-slot="mobile-tabs-drawer"]').getByRole("button", { name: "Settings", exact: true }).click()
+  const settings = page.getByTestId("settings-screen")
+  const menu = settings.getByRole("button", { name: "Preferences", exact: true })
+  const panel = settings.getByRole("tabpanel")
+  await expect(settings.getByRole("heading", { name: "General", exact: true })).toBeVisible()
+  await expect(menu).toBeVisible()
+  await menu.click()
+  await expect(page.getByRole("menuitemradio", { name: "Preferences", exact: true })).toBeChecked()
+  await page.keyboard.press("Escape")
+  await expect(page.getByRole("menuitemradio", { name: "Preferences", exact: true })).toBeHidden()
+  await expect(settings).toBeVisible()
+  await menu.click()
+  await page.getByRole("menuitemradio", { name: "Models", exact: true }).click()
+  await expect(settings.getByRole("heading", { name: "Models", exact: true })).toBeVisible()
+  await expect(settings.getByRole("button", { name: "Models", exact: true })).toBeVisible()
+  await expect(settings.getByRole("switch", { name: "Claude Opus 4.6", exact: true })).toBeAttached()
+  await expect
+    .poll(async () => {
+      const navigation = await settings.getByRole("button", { name: "Models", exact: true }).boundingBox()
+      const content = await panel.boundingBox()
+      return !!navigation && !!content && navigation.y + navigation.height <= content.y
+    })
+    .toBe(true)
+  await expect.poll(async () => (await panel.boundingBox())?.width ?? 0).toBeGreaterThan(350)
+  await settings.getByRole("button", { name: "Back to app", exact: true }).click()
+  await expect(settings).toBeHidden()
+  await expect(page.locator('[data-component="home-session-row"]')).toHaveCount(fixture.sessions.length)
+})
@@ -65,8 +65,8 @@ test("project Extensions stays inside settings while plugins load", async ({ pag
          data: (project ? ["shared-plugin", "project-plugin"] : ["shared-plugin"]).map((id) => ({
            id,
            source: { type: "package", package: id },
-            status: "active",
-            tui: false,
+            state: { status: "active" },
+            features: { server: true },
          })),
        },
      })
@@ -32,12 +32,13 @@ test("server dialog keeps focus above fullscreen settings", async ({ page }) =>

  const editor = page.getByRole("dialog", { name: "Add server" })
  await expect(editor.getByPlaceholder("http://localhost:4096")).toBeFocused()
-  const username = editor.getByPlaceholder("username")
+  await expect(editor.getByPlaceholder("username")).toHaveCount(0)
+  const name = editor.getByPlaceholder("Localhost", { exact: true })
  const password = editor.getByPlaceholder("password")
-  await username.click()
-  await expect(username).toBeFocused()
-  await username.fill("kit")
-  await expect(username).toHaveValue("kit")
+  await name.click()
+  await expect(name).toBeFocused()
+  await name.fill("Remote")
+  await expect(name).toHaveValue("Remote")
  await page.keyboard.press("Tab")
  await expect(password).toBeFocused()
  await password.fill("secret")
@@ -0,0 +1,193 @@
+import { expect, test } from "@playwright/test"
+import { base64Encode } from "@opencode-ai/util/encode"
+import type { OpenCodeEvent } from "@opencode-ai/client/promise"
+import { fixture } from "../smoke/session-timeline.fixture"
+import { mockOpenCodeServer } from "../utils/mock-server"
+
+test("keeps history visible and recovers the composer by choosing another directory", async ({ page }) => {
+  const directory = "/projects/deleted"
+  const destination = "/projects/restored"
+  const sessionID = "ses_missing_location"
+  const session = { id: sessionID, projectID: fixture.project.id, directory, title: "Missing location" }
+  const events: OpenCodeEvent[] = []
+  const moves: unknown[] = []
+  await mockOpenCodeServer(page, {
+    directory: destination,
+    project: { ...fixture.project, worktree: destination },
+    provider: fixture.provider,
+    sessions: [session],
+    events: () => events.splice(0),
+    fileList: () => [],
+    pageMessages: () => ({
+      items: [{ id: "msg_saved", type: "user", text: "Keep this session history", time: { created: 1 } }],
+    }),
+  })
+  await page.route("**/api/**", (route) => {
+    if (new URL(route.request().url()).searchParams.get("location[directory]") !== directory) return route.fallback()
+    return route.fulfill({ status: 500, body: "", headers: { "access-control-allow-origin": "*" } })
+  })
+  await page.goto(`/server/${base64Encode(fixture.serverKey)}/session/${sessionID}`)
+  await expect(page.getByText("Keep this session history", { exact: true })).toBeVisible()
+  await expect(page.getByRole("status")).toContainText("Session location unavailable")
+  await expect(page.getByRole("status")).toContainText(directory)
+  await expect(page.getByRole("textbox", { name: "Prompt", exact: true })).toHaveCount(0)
+  const choose = page.getByRole("button", { name: "Choose directory", exact: true })
+  await expect(choose).toBeEnabled()
+  await choose.click()
+  const dialog = page.getByRole("dialog", { name: "Choose directory", exact: true })
+  await expect(dialog.getByRole("combobox")).toBeFocused()
+  await dialog.getByRole("combobox").press("Escape")
+  await expect(dialog).toHaveCount(0)
+  await expect(choose).toBeFocused()
+  expect(moves).toEqual([])
+
+  // The server can still reject a destination after it was selected in the picker.
+  await page.route(`**/api/session/${sessionID}/move`, (route) => {
+    if (route.request().method() !== "POST") return route.fallback()
+    moves.push(route.request().postDataJSON())
+    return route.fulfill({
+      status: 400,
+      json: { _tag: "InvalidRequestError", message: "Destination is unavailable" },
+      headers: { "access-control-allow-origin": "*" },
+    })
+  })
+  await choose.press("Enter")
+  await expect(dialog.getByRole("combobox")).toBeFocused()
+  await dialog.getByRole("combobox").fill(destination)
+  await dialog.getByRole("combobox").press("Enter")
+  await expect(dialog.locator(".directory-picker-selection")).toHaveText(destination)
+  await dialog.getByRole("button", { name: "Select folder", exact: true }).click()
+  await expect(page.getByText("Failed to move session", { exact: true })).toBeVisible()
+  await expect(choose).toBeEnabled()
+  await expect(page.getByRole("textbox", { name: "Prompt", exact: true })).toHaveCount(0)
+  await expect(choose).toBeFocused()
+  expect(moves).toEqual([{ directory: destination }])
+
+  await page.route(`**/api/session/${sessionID}/move`, (route) => {
+    if (route.request().method() !== "POST") return route.fallback()
+    moves.push(route.request().postDataJSON())
+    session.directory = destination
+    events.push({
+      id: "evt_location_recovered",
+      type: "session.moved",
+      created: 2,
+      durable: { aggregateID: sessionID, seq: 1, version: 1 },
+      data: { sessionID, location: { directory: destination }, projectID: fixture.project.id },
+    })
+    return route.fulfill({ status: 204, headers: { "access-control-allow-origin": "*" } })
+  })
+  await choose.click()
+  await dialog.getByRole("combobox").fill(destination)
+  await dialog.getByRole("combobox").press("Enter")
+  await expect(dialog.locator(".directory-picker-selection")).toHaveText(destination)
+  await dialog.getByRole("button", { name: "Select folder", exact: true }).click()
+  await expect(page.getByText("Session location unavailable", { exact: true })).toHaveCount(0)
+  await expect(page.getByRole("textbox", { name: "Prompt", exact: true })).toBeEditable()
+  await expect(page.locator('[data-action="composer-model"]')).toBeVisible()
+  await expect(page.getByText("Keep this session history", { exact: true })).toBeVisible()
+  await expect(page).toHaveURL(new RegExp(`/session/${sessionID}$`))
+  expect(moves).toEqual([{ directory: destination }, { directory: destination }])
+})
+
+for (const create of [false, true]) {
+  test(`recovers into ${create ? "a new" : "an existing"} worktree without resolving the deleted location`, async ({
+    page,
+  }) => {
+    const canonical = "/projects/repository"
+    const directory = "/projects/deleted-worktree"
+    const destination = create ? "/projects/new-worktree" : "/projects/existing-worktree"
+    const sessionID = "ses_worktree_recovery"
+    const events: OpenCodeEvent[] = []
+    const requests: { operation: string; body: unknown }[] = []
+    let listing = Promise.withResolvers<void>()
+    let listingRequested = Promise.withResolvers<void>()
+    const moving = Promise.withResolvers<void>()
+    const moveRequested = Promise.withResolvers<void>()
+    await mockOpenCodeServer(page, {
+      directory: canonical,
+      project: { ...fixture.project, worktree: canonical },
+      provider: fixture.provider,
+      sessions: [{ id: sessionID, projectID: fixture.project.id, directory }],
+      pageMessages: () => ({
+        items: [{ id: "msg_worktree", type: "user", text: "Recover my worktree", time: { created: 1 } }],
+      }),
+      events: () => events.splice(0),
+    })
+    await page.route("**/api/**", async (route) => {
+      const url = new URL(route.request().url())
+      const headers = { "access-control-allow-origin": "*" }
+      if (url.searchParams.get("location[directory]") === directory)
+        return route.fulfill({ status: 500, body: "", headers })
+      if (url.pathname === `/api/worktree/${fixture.project.id}`) {
+        if (route.request().method() === "GET") {
+          listingRequested.resolve()
+          await listing.promise
+          return route.fulfill({
+            json: [
+              { directory: canonical, strategy: null },
+              { directory: "/projects/existing-worktree", strategy: "git" },
+              ...Array.from({ length: 20 }, (_, index) => ({
+                directory: `/projects/worktree-${index}`,
+                strategy: "git",
+              })),
+            ],
+            headers,
+          })
+        }
+        if (route.request().method() === "POST") {
+          requests.push({ operation: "create", body: route.request().postDataJSON() })
+          return route.fulfill({
+            json: { directory: destination, name: "new-worktree", branch: "new-worktree" },
+            headers,
+          })
+        }
+      }
+      if (url.pathname === `/api/session/${sessionID}/move` && route.request().method() === "POST") {
+        requests.push({ operation: "move", body: route.request().postDataJSON() })
+        moveRequested.resolve()
+        await moving.promise
+        events.push({
+          id: "evt_worktree_recovered",
+          type: "session.moved",
+          created: 2,
+          durable: { aggregateID: sessionID, seq: 1, version: 1 },
+          data: { sessionID, location: { directory: destination }, projectID: fixture.project.id },
+        })
+        return route.fulfill({ status: 204, headers })
+      }
+      return route.fallback()
+    })
+    await page.goto(`/server/${base64Encode(fixture.serverKey)}/session/${sessionID}`)
+    await page.getByRole("button", { name: "Choose worktree", exact: true }).click()
+    await listingRequested.promise
+    await expect(page.getByRole("menuitem", { name: "Loading", exact: true })).toBeVisible()
+    await expect(page.getByRole("menuitem", { name: "New workspace", exact: true })).toBeVisible()
+    await expect(page.getByText("Recover my worktree", { exact: true })).toBeVisible()
+    await expect(page.getByText("Session location unavailable", { exact: true })).toBeVisible()
+    listing.resolve()
+    const existing = page.getByRole("menuitem", { name: "existing-worktree", exact: true })
+    await expect(existing).toBeVisible()
+    await expect(existing).toHaveCSS("height", "28px")
+    await page.keyboard.press("Escape")
+    await expect(page.getByRole("menu")).toHaveCount(0)
+    listing = Promise.withResolvers<void>()
+    listingRequested = Promise.withResolvers<void>()
+    await page.getByRole("button", { name: "Choose worktree", exact: true }).click()
+    await listingRequested.promise
+    await expect(page.getByRole("menuitem", { name: "existing-worktree", exact: true })).toBeVisible()
+    await expect(page.getByText("Recover my worktree", { exact: true })).toBeVisible()
+    listing.resolve()
+    await page.getByRole("menuitem", { name: create ? "New workspace" : "existing-worktree", exact: true }).click()
+    await moveRequested.promise
+    await expect(page.getByRole("button", { name: "Moving session…", exact: true })).toBeDisabled()
+    await expect(page.getByRole("button", { name: "Choose worktree", exact: true })).toBeDisabled()
+    moving.resolve()
+    await expect(page.getByText("Session location unavailable", { exact: true })).toHaveCount(0)
+    await expect(page.getByRole("textbox", { name: "Prompt", exact: true })).toBeEditable()
+    await expect(page.getByText("Recover my worktree", { exact: true })).toBeVisible()
+    expect(requests).toEqual([
+      ...(create ? [{ operation: "create", body: { strategy: "git", from: canonical, directory: "/projects/" } }] : []),
+      { operation: "move", body: { directory: destination } },
+    ])
+  })
+}
@@ -0,0 +1,130 @@
+import { expect, test } from "@playwright/test"
+import { fixture, pageMessages } from "../smoke/session-timeline.fixture"
+import { mockOpenCodeServer } from "../utils/mock-server"
+
+test.beforeEach(async ({ page }) => {
+  const sessions = fixture.sessions.map((session) => ({ ...session }))
+  await mockOpenCodeServer(page, {
+    sessions,
+    provider: fixture.provider,
+    directory: fixture.directory,
+    project: fixture.project,
+    pageMessages,
+  })
+  await page.route("**/api/session/*/rename", async (route) => {
+    if (route.request().method() !== "POST") return route.fallback()
+    const id = new URL(route.request().url()).pathname.split("/").at(-2)
+    const session = sessions.find((item) => item.id === id)
+    const payload: unknown = route.request().postDataJSON()
+    if (
+      !session ||
+      !payload ||
+      typeof payload !== "object" ||
+      !("title" in payload) ||
+      typeof payload.title !== "string"
+    )
+      throw new Error("Invalid rename request")
+    session.title = payload.title
+    await route.fulfill({ status: 204, headers: { "access-control-allow-origin": "*" } })
+  })
+  await page.goto("/")
+  await page.locator('[data-component="home-session-row"]').filter({ hasText: fixture.expected.targetTitle }).click()
+  await expect(page.getByRole("heading", { name: fixture.expected.targetTitle, exact: true })).toBeVisible()
+})
+
+for (const commit of ["Enter", "blur", "click outside"]) {
+  test(`saves the session heading on ${commit}`, async ({ page }) => {
+    await page.getByRole("heading", { name: fixture.expected.targetTitle, exact: true }).click()
+    const input = page.locator('input[data-slot="session-title-child"]')
+    await expect(input).toBeFocused()
+    await input.fill("Renamed session")
+    if (commit === "Enter") await input.press("Enter")
+    if (commit === "blur") await input.press("Tab")
+    if (commit === "click outside") await page.locator('[data-component="composer-editor"]').click()
+    await expect(page.getByRole("heading", { name: "Renamed session", exact: true })).toBeVisible()
+    await expect(page.locator('[data-slot="titlebar-tabs"] a').filter({ hasText: "Renamed session" })).toBeVisible()
+    await page.reload()
+    await expect(page.getByRole("heading", { name: "Renamed session", exact: true })).toBeVisible()
+  })
+}
+
+test("cancels the session heading with Escape", async ({ page }) => {
+  await page.getByRole("heading", { name: fixture.expected.targetTitle, exact: true }).click()
+  const input = page.locator('input[data-slot="session-title-child"]')
+  await input.fill("Discard this title")
+  await input.press("Escape")
+  await expect(page.getByRole("heading", { name: fixture.expected.targetTitle, exact: true })).toBeVisible()
+  await page.reload()
+  await expect(page.getByRole("heading", { name: fixture.expected.targetTitle, exact: true })).toBeVisible()
+})
+
+test("keeps the draft when saving the session heading fails", async ({ page }) => {
+  await page.route("**/api/session/*/rename", (route) =>
+    route.fulfill({ status: 500, headers: { "access-control-allow-origin": "*" } }),
+  )
+  await page.getByRole("heading", { name: fixture.expected.targetTitle, exact: true }).click()
+  const input = page.locator('input[data-slot="session-title-child"]')
+  await input.fill("Retry this title")
+  await input.press("Tab")
+  await expect(page.getByText("Request failed", { exact: true })).toBeVisible()
+  await expect(input).toBeEnabled()
+  await expect(input).toHaveValue("Retry this title")
+  await expect(
+    page.locator('[data-slot="titlebar-tabs"] a').filter({ hasText: fixture.expected.targetTitle }),
+  ).toBeVisible()
+})
+
+test("does not save an empty session heading", async ({ page }) => {
+  await page.getByRole("heading", { name: fixture.expected.targetTitle, exact: true }).click()
+  const input = page.locator('input[data-slot="session-title-child"]')
+  await input.fill("   ")
+  await input.press("Tab")
+  await expect(page.getByRole("heading", { name: fixture.expected.targetTitle, exact: true })).toBeVisible()
+  await page.reload()
+  await expect(page.getByRole("heading", { name: fixture.expected.targetTitle, exact: true })).toBeVisible()
+})
+
+test("renames and closes the session tab from its context menu", async ({ page }) => {
+  const tab = page.locator('[data-slot="titlebar-tabs"] a').filter({ hasText: fixture.expected.targetTitle })
+  await tab.click({ button: "right" })
+  await expect(page.getByRole("menuitem", { name: "Rename", exact: true })).toBeVisible()
+  await page.keyboard.press("Escape")
+  await expect(page.getByRole("menuitem", { name: "Rename", exact: true })).toBeHidden()
+  await expect(tab).toBeFocused()
+  await tab.press("Shift+F10")
+  await page.getByRole("menuitem", { name: "Rename", exact: true }).click()
+  const input = page.locator('[data-slot="tab-title"][contenteditable="true"]')
+  await expect(input).toBeFocused()
+  await input.fill("Renamed from tab")
+  await input.press("Enter")
+  await expect(page.getByRole("heading", { name: "Renamed from tab", exact: true })).toBeVisible()
+  await page.reload()
+  await expect(page.getByRole("heading", { name: "Renamed from tab", exact: true })).toBeVisible()
+  const renamed = page.locator('[data-slot="titlebar-tabs"] a').filter({ hasText: "Renamed from tab" })
+  await renamed.click({ button: "right" })
+  await page.getByRole("menuitem", { name: "Close tab", exact: true }).click()
+  await expect(renamed).toBeHidden()
+  await page.getByRole("button", { name: "Home", exact: true }).click()
+  await expect(
+    page.locator('[data-component="home-session-row"]').filter({ hasText: "Renamed from tab" }),
+  ).toBeVisible()
+})
+
+test("renames an inactive tab without switching sessions", async ({ page }) => {
+  await page.getByRole("button", { name: "Home", exact: true }).click()
+  await page.locator('[data-component="home-session-row"]').filter({ hasText: fixture.expected.sourceTitle }).click()
+  await expect(page.getByRole("heading", { name: fixture.expected.sourceTitle, exact: true })).toBeVisible()
+  const tab = page.locator('[data-slot="titlebar-tabs"] a').filter({ hasText: fixture.expected.targetTitle })
+  await tab.click({ button: "right" })
+  await page.getByRole("menuitem", { name: "Rename", exact: true }).click()
+  const input = page.locator('[data-slot="tab-title"][contenteditable="true"]')
+  await expect(input).toBeFocused()
+  await input.fill("Inactive tab renamed")
+  await input.press("Tab")
+  await expect(page.getByRole("heading", { name: fixture.expected.sourceTitle, exact: true })).toBeVisible()
+  await expect(page).toHaveURL(new RegExp(`/session/${fixture.sourceID}$`))
+  await page.locator('[data-slot="titlebar-tabs"] a').filter({ hasText: "Inactive tab renamed" }).click()
+  await expect(page.getByRole("heading", { name: "Inactive tab renamed", exact: true })).toBeVisible()
+  await page.reload()
+  await expect(page.getByRole("heading", { name: "Inactive tab renamed", exact: true })).toBeVisible()
+})
@@ -0,0 +1,99 @@
+import type { SessionMessageInfo } from "@opencode-ai/client/promise"
+import { expect, test } from "@playwright/test"
+import { fixture } from "../performance/timeline/session-timeline-stress.fixture"
+import { stressSessionHref } from "../performance/timeline/timeline-test-helpers"
+import { mockOpenCodeServer } from "../utils/mock-server"
+
+test.use({ viewport: { width: 1440, height: 900 }, serviceWorkers: "block" })
+
+for (const window of ["assistant-only", "mixed"] as const) {
+  test(`renders the ${window} latest page before parent hydration and preserves it afterward`, async ({ page }) => {
+    const session = { ...fixture.sessions[0]!, id: `ses_hydration_${window}` }
+    // Both 20-message pages begin with an assistant; only page three supplies its parent.
+    const messages = Array.from({ length: 41 }, (_, index): SessionMessageInfo => {
+      const id = `msg_hydration_${index}`
+      const time = { created: 1700000000000 + index * 1_000 }
+      if (index === 0 || (window === "mixed" && index === 39))
+        return { id, type: "user", time, text: `Prompt ${index}` }
+      return {
+        id,
+        type: "assistant",
+        time: { ...time, completed: time.created + 500 },
+        model: { id: "claude-opus-4-6", providerID: "opencode" },
+        agent: "build",
+        content: [{ type: "text", text: index === 40 ? "## Hydrated tail\n\n**Ready.**" : `Answer ${index}` }],
+      }
+    })
+    const gates = [21, 1].map((index) => ({
+      before: messages[index]!.id,
+      parent: messages[index === 21 ? 1 : 0]!.id,
+      requested: Promise.withResolvers<void>(),
+      release: Promise.withResolvers<void>(),
+    }))
+    const requests: (string | undefined)[] = []
+    await mockOpenCodeServer(page, {
+      ...fixture,
+      sessions: [session],
+      beforeMessagesResponse: async ({ before }) => {
+        requests.push(before)
+        if (!before) return
+        const gate = gates.find((gate) => gate.before === before)
+        if (!gate) throw new Error(`Unexpected older-page boundary: ${before}`)
+        gate.requested.resolve()
+        await gate.release.promise
+      },
+      pageMessages: (_, limit, before) => {
+        expect(limit).toBe(20)
+        const end = before ? messages.findIndex((message) => message.id === before) : messages.length
+        const start = Math.max(0, end - limit)
+        return { items: messages.slice(start, end), cursor: start > 0 ? messages[start]!.id : undefined }
+      },
+    })
+    const tail = page.locator('[data-timeline-part-id="msg_hydration_40:text:0"]')
+    const markdown = tail.locator('[data-component="markdown"]')
+    const content = page.locator("[data-timeline-virtual-content]", { has: tail })
+    const viewport = page.locator(".scroll-view__viewport", { has: tail })
+    const orphan = page.locator('[data-timeline-row="AssistantPart"]', {
+      has: page.locator('[data-timeline-part-id="msg_hydration_38:text:0"]'),
+    })
+    const expectReadyTail = async () => {
+      await expect(content).toHaveCSS("visibility", "visible")
+      await expect(markdown).toHaveAttribute("data-markdown-ready", "")
+      await expect(markdown.getByRole("heading", { name: "Hydrated tail", exact: true })).toBeInViewport({ ratio: 1 })
+      await expect
+        .poll(() =>
+          viewport.evaluate((element) => Math.abs(element.scrollHeight - element.clientHeight - element.scrollTop)),
+        )
+        .toBeLessThanOrEqual(1)
+    }
+
+    try {
+      await page.goto(stressSessionHref(session.id))
+      await gates[0]!.requested.promise
+      // This must pass while the first older response is still held.
+      await expectReadyTail()
+      await expect(orphan).toHaveAttribute("data-message-id", "msg_hydration_21")
+      if (window === "mixed")
+        await expect(
+          page.locator('[data-timeline-row="UserMessage"][data-message-id="msg_hydration_39"]'),
+        ).toBeInViewport()
+      const original = await markdown.elementHandle()
+
+      for (const gate of gates) {
+        await gate.requested.promise
+        gate.release.resolve()
+        // Parent ownership proves the page reached the projection, not just the network.
+        await expect(orphan).toHaveAttribute("data-message-id", gate.parent)
+        await expectReadyTail()
+        expect(await markdown.evaluate((element, original) => element === original, original)).toBe(true)
+      }
+      expect(requests).toEqual([undefined, ...gates.map((gate) => gate.before)])
+      const ids = await content
+        .locator("[data-timeline-part-id]")
+        .evaluateAll((elements) => elements.map((element) => element.getAttribute("data-timeline-part-id")))
+      expect(new Set(ids).size).toBe(ids.length)
+    } finally {
+      gates.forEach((gate) => gate.release.resolve())
+    }
+  })
+}
@@ -0,0 +1,123 @@
+import { expect, test, type Page } from "@playwright/test"
+import type { SessionMessageInfo } from "@opencode-ai/client/promise"
+import { mockOpenCodeServer } from "../utils/mock-server"
+import { fixture } from "../performance/timeline/session-timeline-stress.fixture"
+import { expected, messages } from "../performance/timeline/session-tab-switch.fixture"
+import { installTimelineSettings, stressSessionHref } from "../performance/timeline/timeline-test-helpers"
+
+test.use({ viewport: { width: 1440, height: 900 }, serviceWorkers: "block" })
+
+type Reveal = { pending: number; clipped: string[]; bottomError: number; tables: number; codeBlocks: number }
+
+for (const width of [1440, 390]) {
+  test(`reveals measured Markdown after the worker completes at ${width}px`, async ({ page }) => {
+    await page.setViewportSize({ width, height: 900 })
+    const requested = Promise.withResolvers<void>()
+    const release = Promise.withResolvers<void>()
+    await page.route(/markdown\.worker(?:-[^/?]+\.js|\.ts)(?:\?.*)?$/, async (route) => {
+      requested.resolve()
+      await release.promise
+      await route.continue()
+    })
+    await page.addInitScript((partID) => {
+      const observer = new MutationObserver(() => {
+        const answer = document.querySelector<HTMLElement>(`[data-timeline-part-id="${partID}"]`)
+        const content = answer?.closest<HTMLElement>("[data-timeline-virtual-content]")
+        const root = content?.closest<HTMLElement>(".scroll-view__viewport")
+        if (!answer || !content || !root || !content.checkVisibility({ checkVisibilityCSS: true })) return
+        const spacer = content.querySelector('[data-timeline-row="bottom-spacer"]')
+        ;(window as Window & { __coldReveal?: Reveal }).__coldReveal = {
+          pending: content.querySelectorAll('[data-component="markdown"]:not([data-markdown-ready])').length,
+          clipped: [...content.querySelectorAll<HTMLElement>("[data-timeline-key]")].flatMap((row) =>
+            (row.firstElementChild?.getBoundingClientRect().height ?? 0) > row.getBoundingClientRect().height + 1
+              ? [row.dataset.timelineKey!]
+              : [],
+          ),
+          bottomError: (spacer?.getBoundingClientRect().bottom ?? Infinity) - root.getBoundingClientRect().bottom,
+          tables: answer.querySelectorAll("table").length,
+          codeBlocks: answer.querySelectorAll("pre").length,
+        }
+        observer.disconnect()
+      })
+      observer.observe(document, { childList: true, subtree: true, attributes: true, attributeFilter: ["style"] })
+    }, expected[fixture.sourceID].answerID)
+    await mockOpenCodeServer(page, {
+      sessions: fixture.sessions,
+      project: fixture.project,
+      provider: fixture.provider,
+      directory: fixture.directory,
+      pageMessages: () => ({ items: messages[fixture.sourceID] }),
+    })
+    await installTimelineSettings(page)
+    try {
+      await page.goto(stressSessionHref(fixture.sourceID), { waitUntil: "domcontentloaded" })
+      await requested.promise
+      await expect(page.locator("[data-timeline-virtual-content]")).toHaveCSS("visibility", "hidden")
+      release.resolve()
+      await expect(page.locator("[data-timeline-virtual-content]")).toHaveCSS("visibility", "visible")
+      const reveal = await page.evaluate(() => (window as Window & { __coldReveal?: Reveal }).__coldReveal)
+      expect(reveal).toMatchObject({ pending: 0, clipped: [], tables: 1, codeBlocks: 4 })
+      expect(Math.abs(reveal?.bottomError ?? Infinity)).toBeLessThanOrEqual(1)
+    } finally {
+      release.resolve()
+    }
+  })
+}
+
+test("scrolls within a long answer without mounting unrelated history", async ({ page }) => {
+  await openTimeline(page, messages[fixture.sourceID])
+  const answer = page.locator(`[data-timeline-part-id="${expected[fixture.sourceID].answerID}"]`)
+  await expect(answer.locator('[data-component="markdown"]')).toHaveAttribute("data-markdown-ready", "")
+  await expect(answer.getByRole("table")).toHaveCount(1)
+  const scroller = page.locator(".scroll-view__viewport", { has: answer })
+  await expect
+    .poll(() => scroller.evaluate((element) => element.scrollHeight - element.clientHeight - element.scrollTop))
+    .toBeLessThanOrEqual(1)
+  const rows = page.locator("[data-timeline-key]")
+  const keys = await rows.evaluateAll((elements) =>
+    elements.map((element) => element.getAttribute("data-timeline-key")),
+  )
+  const top = await answer.evaluate((element) => element.getBoundingClientRect().top)
+
+  await scroller.hover()
+  await page.mouse.wheel(0, -240)
+
+  await expect.poll(() => answer.evaluate((element) => element.getBoundingClientRect().top)).toBeCloseTo(top + 240, 0)
+  expect(
+    await rows.evaluateAll((elements) => elements.map((element) => element.getAttribute("data-timeline-key"))),
+  ).toEqual(keys)
+  await expect(answer.locator('[data-component="markdown"]')).toHaveAttribute("data-markdown-ready", "")
+})
+
+test("fills a short cold transcript before revealing it", async ({ page }) => {
+  const history = messages[fixture.sourceID].slice(-6).map((message, index) => {
+    if (message.type === "user") return { ...message, text: `Prompt ${index}`, metadata: undefined }
+    if (message.type === "assistant")
+      return { ...message, content: [{ type: "text" as const, text: `**Answer ${index}**` }] }
+    return message
+  })
+  await openTimeline(page, history)
+  for (const message of history) {
+    if (message.type === "user") {
+      await expect(page.locator(`[data-timeline-row="UserMessage"][data-message-id="${message.id}"]`)).toBeInViewport()
+    }
+    if (message.type === "assistant") {
+      const answer = page.locator(`[data-timeline-part-id="${message.id}:text:0"]`)
+      await expect(answer).toBeInViewport()
+      await expect(answer.locator('[data-component="markdown"]')).toHaveAttribute("data-markdown-ready", "")
+    }
+  }
+})
+
+async function openTimeline(page: Page, history: SessionMessageInfo[]) {
+  await mockOpenCodeServer(page, {
+    sessions: fixture.sessions,
+    project: fixture.project,
+    provider: fixture.provider,
+    directory: fixture.directory,
+    pageMessages: () => ({ items: history }),
+  })
+  await installTimelineSettings(page)
+  await page.goto(stressSessionHref(fixture.sourceID))
+  await expect(page.locator("[data-timeline-virtual-content]")).toHaveCSS("visibility", "visible")
+}
@@ -83,7 +83,12 @@ test("extensions opens without waiting for MCPs or plugins", async ({ page }) =>
      json: {
        location: { directory },
        data: [
-          { id: "demo-plugin", source: { type: "package", package: "demo-plugin" }, status: "active", tui: false },
+          {
+            id: "demo-plugin",
+            source: { type: "package", package: "demo-plugin" },
+            state: { status: "active" },
+            features: { server: true },
+          },
        ],
      },
    })
@@ -141,7 +141,7 @@ test("shows the not found fallback when the viewed session is deleted", async ({
  })

  await expect(page.getByText("This session cannot be found")).toBeVisible()
-  await expect(page.getByRole("button", { name: "Close Tab" })).toBeVisible()
+  await expect(page.getByRole("button", { name: "Close Tab", exact: true })).toBeVisible()
  await expect(page.getByRole("heading", { name: taskDescription })).toHaveCount(0)
 })

@@ -69,7 +69,7 @@ test("keyboard navigation follows the visible tab order", async ({ page }) => {
  await expect(page).toHaveURL(new RegExp(`${hrefC.replace(/[.*+?^${}()|[\]\\]/g, "\\$&")}$`))
 })

-test("cramped tabs only show the close button for the active tab", async ({ page }) => {
+test("mobile drawer exposes close controls and navigates between tabs", async ({ page }) => {
  await page.setViewportSize({ width: 360, height: 720 })
  await mockServer(page)
  await page.addInitScript(
@@ -89,19 +89,32 @@ test("cramped tabs only show the close button for the active tab", async ({ page
  const hrefA = `/server/${base64Encode(server)}/session/${sessionA.id}`
  const hrefB = `/server/${base64Encode(server)}/session/${sessionB.id}`
  await page.goto(hrefA)
+  await page.getByRole("button", { name: "Tabs", exact: true }).click()

  const tabA = page.locator(`[data-titlebar-tab-slot]:has(a[href="${hrefA}"])`)
  const tabB = page.locator(`[data-titlebar-tab-slot]:has(a[href="${hrefB}"])`)
  await expect(tabA).toHaveAttribute("data-active", "true")
  await expect(tabB).toBeVisible()
  await expect(tabA.locator('[data-slot="tab-close"]')).toBeVisible()
-  await expect(tabB.locator('[data-slot="tab-close"]')).toBeHidden()
+  await expect(tabB.locator('[data-slot="tab-close"]')).toBeVisible()

  await tabB.locator(`a[href="${hrefB}"]`).click()

  await expect(page).toHaveURL(new RegExp(`${hrefB.replace(/[.*+?^${}()|[\]\\]/g, "\\$&")}$`))
-  await expect(tabA.locator('[data-slot="tab-close"]')).toBeHidden()
+  await expect(page.getByRole("dialog", { name: "Tabs", exact: true })).toBeHidden()
+  await page.getByRole("button", { name: "Tabs", exact: true }).click()
+  await expect(tabA.locator('[data-slot="tab-close"]')).toBeVisible()
  await expect(tabB.locator('[data-slot="tab-close"]')).toBeVisible()
+
+  for (const direction of ["ltr", "rtl"]) {
+    await page.evaluate((direction) => document.documentElement.setAttribute("dir", direction), direction)
+    await page.setViewportSize({ width: 450, height: 720 })
+    await expect(tabA).toBeVisible()
+    await page.setViewportSize({ width: 1280, height: 720 })
+    await expect(tabA.locator("[data-titlebar-tab]")).toHaveAttribute("data-title-overflow", "false")
+    await page.setViewportSize({ width: 450, height: 720 })
+    await page.getByRole("button", { name: "Tabs", exact: true }).click()
+  }
 })

 test("vertical tabs show project details, resize, and navigate", async ({ page }) => {
@@ -194,28 +207,28 @@ test("appearance experimental setting switches tab orientation", async ({ page }

  await page.setViewportSize({ width: 920, height: 720 })
  await expect(page.locator('[data-slot="vertical-tabs-sidebar"]')).toHaveCSS("width", "260px")
-  await expect(settings.getByRole("tablist")).toHaveCSS("width", "160px")
+  await expect(settings.getByRole("tablist")).toBeHidden()
+  await expect(settings.getByRole("button", { name: "Appearance", exact: true })).toBeVisible()

  await page.setViewportSize({ width: 800, height: 720 })
-  await expect(settings.getByRole("tablist")).toHaveCSS("width", "160px")
-  await expect(version).toBeInViewport()
+  await expect(settings.getByRole("tablist")).toBeHidden()
+  await expect(settings.getByRole("button", { name: "Appearance", exact: true })).toBeVisible()

  await page.setViewportSize({ width: 390, height: 720 })
-  await expect(version).toBeInViewport()
+  await expect(settings.getByRole("button", { name: "Appearance", exact: true })).toBeVisible()
  await settings.evaluate((element) => element.setAttribute("dir", "rtl"))
-  await expect(version).toBeInViewport()
-  await expect(version).toHaveCSS("direction", "ltr")
+  await expect(settings.getByRole("button", { name: "Appearance", exact: true })).toBeInViewport()

  await page.setViewportSize({ width: 390, height: 360 })
-  await version.scrollIntoViewIfNeeded()
-  await expect(version).toBeInViewport()
+  await expect(settings.getByRole("button", { name: "Appearance", exact: true })).toBeInViewport()

  // Reload the UI-selected preference without seeding settings storage.
  await page.reload()
  const href = `/server/${base64Encode(server)}/session/${sessionA.id}`
+  await page.getByRole("button", { name: "Tabs", exact: true }).click()
  await expect(
    page
-      .locator('[data-slot="titlebar-tabs"]')
+      .locator('[data-slot="mobile-tabs-drawer"]')
      .locator(`[data-titlebar-tab-link][href="${href}"]`)
      .getByText(sessionA.title, { exact: true }),
  ).toBeVisible()
@@ -234,7 +247,7 @@ test("appearance experimental setting switches tab orientation", async ({ page }
  await expect(layout).toContainText("Vertical")
 })

-test("vertical tab preference falls back to horizontal on mobile", async ({ page }) => {
+test("vertical tab preference uses the drawer on mobile", async ({ page }) => {
  await page.setViewportSize({ width: 390, height: 720 })
  await mockServer(page)
  await page.addInitScript(
@@ -251,7 +264,8 @@ test("vertical tab preference falls back to horizontal on mobile", async ({ page
  const href = `/server/${base64Encode(server)}/session/${sessionA.id}`
  await page.goto(href)

-  const tabs = page.locator('[data-slot="titlebar-tabs"]')
+  await page.getByRole("button", { name: "Tabs", exact: true }).click()
+  const tabs = page.locator('[data-slot="mobile-tabs-drawer"]')
  await expect(tabs.locator(`[data-titlebar-tab-link][href="${href}"]`)).toContainText(sessionA.title)
  await expect(page.locator('[data-slot="vertical-tabs-sidebar"]')).toHaveCount(0)

@@ -288,11 +302,12 @@ async function mockServer(page: Page) {
    if (currentSessionInfo) return json(route, { data: currentSession(currentSessionInfo) })
    if (sessions.some((item) => url.pathname === `/api/session/${item.id}/message`))
      return json(route, { data: [], cursor: {} })
+    if (sessions.some((item) => url.pathname === `/api/session/${item.id}/inbox`)) return json(route, { data: [] })
    if (["/api/agent", "/api/provider", "/api/model", "/api/command", "/api/reference"].includes(url.pathname))
      return json(route, { location: { directory: sessionA.directory }, data: [] })
    if (url.pathname === "/api/model/default")
      return json(route, { location: { directory: sessionA.directory }, data: null })
-    if (url.pathname === "/api/permission/request" || url.pathname === "/api/question/request")
+    if (url.pathname === "/api/permission/request" || url.pathname === "/api/form/request")
      return json(route, { location: { directory: sessionA.directory }, data: [] })
    if (url.pathname === "/api/mcp") return json(route, { location: { directory: sessionA.directory }, data: [] })
    if (url.pathname === "/api/mcp/resource")
@@ -20,95 +20,134 @@ const inventory: WorktreeDirectory[] = [

 test.use({ serviceWorkers: "block" })

-for (const scenario of [
-  { name: "managed Git worktree", directory: workspace, accent: true },
-  { name: "linked Git worktree outside main", directory: "C:/OpenCode/LinkedWorkspace", accent: true },
-  {
-    name: "linked Git worktree on a narrow screen",
-    directory: "C:/OpenCode/LinkedWorkspace",
-    accent: true,
-    viewport: { width: 390, height: 844 },
-  },
-  { name: "main root with Windows case and separators", directory: "c:\\OPENCODE\\workspaceaccent\\", accent: false },
-  { name: "nested main directory", directory: `${root}/packages/app`, accent: false },
-  { name: "nested workspace inside main", directory: `${workspace}/packages/app`, accent: true },
-  {
-    name: "workspace with Windows case and separators",
-    directory: "c:\\opencode\\WORKSPACEACCENT\\.worktrees\\FEATURE\\src\\",
-    accent: true,
-  },
-  { name: "unregistered sibling with the same prefix", directory: `${workspace}-unregistered`, accent: false },
-  { name: "workspace using another strategy", directory: "C:/OpenCode/WorkspaceCopy", accent: true },
-  { name: "registered directory without a strategy", directory: "C:/OpenCode/RegisteredDirectory", accent: true },
-]) {
-  test(`existing session send button: ${scenario.name}`, async ({ page }, testInfo) => {
-    if (scenario.viewport) await page.setViewportSize(scenario.viewport)
-    const view = await openSession(page, scenario.directory)
-    await view.input.fill("Inspect this fixture workspace.")
-    await expect(view.send).toBeEnabled()
+for (const theme of ["light", "dark"] as const) {
+  test.describe(theme, () => {
+    test.beforeEach(async ({ page }) => {
+      await page.addInitScript((theme) => {
+        localStorage.setItem("opencode-theme-id", "oc-2")
+        localStorage.setItem("opencode-color-scheme", theme)
+      }, theme)
+    })

-    if (scenario.name === "managed Git worktree") {
-      // Capture before the color assertion so both red and green runs have evidence.
-      const path = testInfo.outputPath("workspace-accent.png")
-      await view.composer.screenshot({ path })
-      await testInfo.attach("workspace-accent", { path, contentType: "image/png" })
+    for (const scenario of [
+      { name: "managed Git worktree", directory: workspace, accent: true },
+      { name: "linked Git worktree outside main", directory: "C:/OpenCode/LinkedWorkspace", accent: true },
+      {
+        name: "linked Git worktree on a narrow screen",
+        directory: "C:/OpenCode/LinkedWorkspace",
+        accent: true,
+        viewport: { width: 390, height: 844 },
+      },
+      {
+        name: "main root with Windows case and separators",
+        directory: "c:\\OPENCODE\\workspaceaccent\\",
+        accent: false,
+      },
+      { name: "nested main directory", directory: `${root}/packages/app`, accent: false },
+      { name: "nested workspace inside main", directory: `${workspace}/packages/app`, accent: true },
+      {
+        name: "workspace with Windows case and separators",
+        directory: "c:\\opencode\\WORKSPACEACCENT\\.worktrees\\FEATURE\\src\\",
+        accent: true,
+      },
+      { name: "unregistered sibling with the same prefix", directory: `${workspace}-unregistered`, accent: false },
+      { name: "workspace using another strategy", directory: "C:/OpenCode/WorkspaceCopy", accent: true },
+      { name: "registered directory without a strategy", directory: "C:/OpenCode/RegisteredDirectory", accent: true },
+    ]) {
+      test(`existing session send button: ${scenario.name}`, async ({ page }, testInfo) => {
+        if (scenario.viewport) await page.setViewportSize(scenario.viewport)
+        const view = await openSession(page, scenario.directory)
+        await view.input.fill("Inspect this fixture workspace.")
+        await expect(view.send).toBeEnabled()
+
+        if (scenario.name === "managed Git worktree") {
+          // Capture before the color assertion so both red and green runs have evidence.
+          const path = testInfo.outputPath("workspace-accent.png")
+          await view.composer.screenshot({ path })
+          await testInfo.attach("workspace-accent", { path, contentType: "image/png" })
+        }
+
+        await expectBackground(view.send, "contrast")
+        await view.send.hover()
+        await expectBackground(view.send, "contrast")
+        await view.composer.locator('[data-action="composer-model"]').press("Tab")
+        await expect(view.send).toBeFocused()
+        await expectBackground(view.send, "contrast")
+        const message = page.locator('[data-slot="user-message-text"]')
+        await expect(message).toHaveText("Check this fixture workspace.")
+        await expectBackground(
+          message,
+          scenario.accent ? "accent" : theme === "light" ? "layer-02" : "layer-01",
+          "background-color",
+        )
+      })
    }

-    await expectBackground(view.send, scenario.accent ? "accent" : "contrast")
-    const message = page.locator('[data-slot="user-message-text"]')
-    await expect(message).toHaveText("Check this fixture workspace.")
-    await expectBackground(message, scenario.accent ? "accent" : "layer-02", "background-color")
+    test("inventory updates leave send neutral; disabled and stop stay neutral", async ({ page }) => {
+      const view = await openSession(page, workspace, [{ directory: root }])
+      await view.input.fill("Keep this draft while the inventory changes.")
+      await expect(view.send).toBeEnabled()
+      await expectBackground(view.send, "contrast")
+      const url = page.url()
+
+      const refreshed = page.waitForResponse(
+        (response) =>
+          new URL(response.url()).pathname === `/api/worktree/${projectID}` && response.request().method() === "GET",
+      )
+      view.worktrees.push({ directory: workspace, strategy: "git" })
+      view.events.push({
+        id: "evt_workspace_accent_inventory",
+        created: 1700000001000,
+        type: "worktree.updated",
+        data: { projectID },
+      })
+      expect((await refreshed).ok()).toBe(true)
+      await expectBackground(view.send, "contrast")
+      await expect(page).toHaveURL(url)
+      await expect(view.input).toHaveText("Keep this draft while the inventory changes.")
+      await expect(view.send).toBeEnabled()
+
+      await view.input.fill("")
+      await expect(view.send).toBeDisabled()
+      await expectBackground(view.send, "contrast")
+
+      view.events.push({
+        id: "evt_workspace_accent_running",
+        created: 1700000002000,
+        type: "session.execution.started",
+        durable: { aggregateID: sessionID, seq: 1, version: 1 },
+        data: { sessionID },
+      })
+      const stop = view.composer.getByRole("button", { name: "Stop", exact: true })
+      await expect(stop).toBeEnabled()
+      await expectBackground(stop, "contrast")
+
+      await view.input.fill("Send a follow-up instead of stopping.")
+      await expect(view.send).toBeEnabled()
+      await expectBackground(view.send, "contrast")
+      await expect(page).toHaveURL(url)
+    })
+
+    test("new workspace send button stays neutral", async ({ page }) => {
+      const view = await openSession(page, root, [...inventory], true)
+      await expect(view.send).toBeDisabled()
+      await expectBackground(view.send, "contrast")
+      await page.getByRole("button", { name: "Local", exact: true }).click()
+      await page.getByRole("menuitem", { name: "New workspace", exact: true }).click()
+      await expect(page.getByRole("button", { name: "New workspace", exact: true })).toBeVisible()
+      await view.input.fill("Inspect this fixture workspace.")
+      await expect(view.send).toBeEnabled()
+      await expectBackground(view.send, "contrast")
+      await view.send.hover()
+      await expectBackground(view.send, "contrast")
+      await view.composer.locator('[data-action="composer-model"]').press("Tab")
+      await expect(view.send).toBeFocused()
+      await expectBackground(view.send, "contrast")
+    })
  })
 }

-test("inventory updates recolor the send button without navigation; disabled and stop stay neutral", async ({
-  page,
-}) => {
-  const view = await openSession(page, workspace, [{ directory: root }])
-  await view.input.fill("Keep this draft while the inventory changes.")
-  await expect(view.send).toBeEnabled()
-  await expectBackground(view.send, "contrast")
-  const url = page.url()
-
-  const refreshed = page.waitForResponse(
-    (response) =>
-      new URL(response.url()).pathname === `/api/worktree/${projectID}` && response.request().method() === "GET",
-  )
-  view.worktrees.push({ directory: workspace, strategy: "git" })
-  view.events.push({
-    id: "evt_workspace_accent_inventory",
-    created: 1700000001000,
-    type: "worktree.updated",
-    data: { projectID },
-  })
-  expect((await refreshed).ok()).toBe(true)
-  await expectBackground(view.send, "accent")
-  await expect(page).toHaveURL(url)
-  await expect(view.input).toHaveText("Keep this draft while the inventory changes.")
-  await expect(view.send).toBeEnabled()
-
-  await view.input.fill("")
-  await expect(view.send).toBeDisabled()
-  await expectBackground(view.send, "contrast")
-
-  view.events.push({
-    id: "evt_workspace_accent_running",
-    created: 1700000002000,
-    type: "session.execution.started",
-    durable: { aggregateID: sessionID, seq: 1, version: 1 },
-    data: { sessionID },
-  })
-  const stop = view.composer.getByRole("button", { name: "Stop", exact: true })
-  await expect(stop).toBeEnabled()
-  await expectBackground(stop, "contrast")
-
-  await view.input.fill("Send a follow-up instead of stopping.")
-  await expect(view.send).toBeEnabled()
-  await expectBackground(view.send, "accent")
-  await expect(page).toHaveURL(url)
-})
-
-async function openSession(page: Page, directory: string, worktrees = [...inventory]) {
+async function openSession(page: Page, directory: string, worktrees = [...inventory], draft = false) {
  const events: OpenCodeEvent[] = []
  await mockOpenCodeServer(page, {
    directory,
@@ -159,18 +198,32 @@ async function openSession(page: Page, directory: string, worktrees = [...invent
    if (route.request().method() !== "GET") return route.fallback()
    return route.fulfill({ json: worktrees, headers: { "access-control-allow-origin": "*" } })
  })
-  await page.addInitScript(() => {
-    localStorage.setItem("opencode-theme-id", "oc-2")
-    localStorage.setItem("opencode-color-scheme", "light")
-  })
+  if (draft)
+    await page.addInitScript(
+      ({ root, server }) => {
+        localStorage.setItem(
+          "opencode.global.dat:server",
+          JSON.stringify({
+            projects: { local: [{ worktree: root, expanded: true }] },
+            lastProject: { local: root },
+          }),
+        )
+        localStorage.setItem(
+          "opencode.window.browser.dat:tabs",
+          JSON.stringify([{ type: "draft", draftID: "draft_workspace_accent", server, directory: root }]),
+        )
+      },
+      { root, server },
+    )
  const loaded = page.waitForResponse(
    (response) =>
      new URL(response.url()).pathname === `/api/worktree/${projectID}` && response.request().method() === "GET",
  )
-  await page.goto(`/server/${base64Encode(server)}/session/${sessionID}`)
+  await page.goto(
+    draft ? "/new-session?draftId=draft_workspace_accent" : `/server/${base64Encode(server)}/session/${sessionID}`,
+  )
  expect((await loaded).ok()).toBe(true)
-  await expectSessionReady(page, { server, sessionID, title })
-  await expect(page.locator("html")).toHaveAttribute("data-color-scheme", "light")
+  if (!draft) await expectSessionReady(page, { server, sessionID, title })
  const composer = page.locator('[data-component="composer"]')
  await expectAppVisible(composer)
  const input = composer.getByRole("textbox", { name: "Prompt", exact: true })
@@ -355,6 +355,14 @@ test.describe("smoke: session timeline", () => {
    await expectCanScrollToStart(page, expectedPartIDs, expectedMessageIDs, errors)

    const shell = page.locator(`[data-timeline-part-id="${fixture.expected.expandedShellPartID}"]`)
+    // The shell is below a long diff; reveal it rather than depending on offscreen overscan.
+    while ((await shell.count()) === 0) {
+      const before = await timelineState(page)
+      await timelineScroller(page).press("PageDown")
+      await expect.poll(async () => (await timelineState(page)).signature).not.toBe(before.signature)
+    }
+    await shell.scrollIntoViewIfNeeded()
+    await expect(shell).toBeInViewport()
    const shellTrigger = shell.locator('[data-slot="collapsible-trigger"]')
    const shellSubtitle = shell.locator('[data-slot="basic-tool-tool-subtitle"]')
    await expect(shellSubtitle).toHaveCount(0)
@@ -694,6 +702,7 @@ async function expectSessionTimelineReady(
  expectedMessageIDs: string[],
  errors: string[],
 ) {
+  await expect(page.locator("[data-timeline-virtual-content]")).toHaveCSS("visibility", "visible")
  await waitForTimelineStable(page)
  for (const text of forbiddenText) await expect(page.getByText(text)).toHaveCount(0)
  const currentState = await timelineState(page)
@@ -1,10 +1,16 @@
 import { expect, test } from "@playwright/test"
+import { fixture } from "../performance/timeline/session-timeline-stress.fixture"
+import { installStressSessionTabs, stressSessionHref } from "../performance/timeline/timeline-test-helpers"
 import { mockOpenCodeServer } from "../utils/mock-server"
 import { expectAppVisible } from "../utils/waits"

 const directory = "C:/OpenCode/NewProject"

 test("creates a session in a new project and selects its model", async ({ page }) => {
+  // An empty draft must remain usable when the file viewer is unavailable.
+  await page.route(/(?:\/_assets\/file-(?!icon-)[^/]+\.js|\/session-ui\/src\/components\/file\.tsx)(?:\?|$)/, (route) =>
+    route.abort(),
+  )
  await mockOpenCodeServer(page, {
    directory,
    project: {
@@ -60,7 +66,7 @@ test("creates a session in a new project and selects its model", async ({ page }
          { providerID: "opencode", modelID: "free-model", visibility: "show" },
          { providerID: "opencode-go", modelID: "go-model-1", visibility: "show" },
        ],
-        recent: [],
+        recent: [{ providerID: "opencode-go", modelID: "go-model-1" }],
        variant: {},
      }),
    )
@@ -81,11 +87,73 @@ test("creates a session in a new project and selects its model", async ({ page }
  await expectAppVisible(page.locator('[data-component="composer"]'))

  const modelControl = page.locator('[data-action="composer-model"]')
+  await expect(modelControl).toContainText("Go Model 1")
+  await modelControl.click()
+  await page.locator('[data-option-key="opencode:free-model"]').click()
+  await expect(modelControl).toContainText("Free Model")
+
  await modelControl.click()
-  await expect(page.locator('[data-option-key="opencode:free-model"]')).toBeVisible()
  const goModel = page.locator('[data-option-key="opencode-go:go-model-1"]')
  await expect(goModel).toBeVisible()
  await goModel.click()

  await expect(modelControl).toContainText("Go Model 1")
 })
+
+test("restores each existing session's model and variant when switching tabs", async ({ page }) => {
+  const sessions = ["A", "B"].map((name) => ({
+    ...fixture.sessions[0],
+    id: `ses_model_${name}`,
+    title: `Model ${name}`,
+    model: { id: `model-${name}`, providerID: "opencode", variant: "balanced" },
+  }))
+  await mockOpenCodeServer(page, {
+    ...fixture,
+    sessions,
+    provider: {
+      all: [
+        {
+          id: "opencode",
+          name: "OpenCode",
+          models: Object.fromEntries(
+            sessions.map((session) => [
+              session.model.id,
+              {
+                id: session.model.id,
+                name: session.title,
+                limit: { context: 200_000 },
+                variants: { balanced: {}, high: {} },
+              },
+            ]),
+          ),
+        },
+      ],
+      connected: ["opencode"],
+      default: { providerID: "opencode", modelID: sessions[0]!.model.id },
+    },
+    pageMessages: () => ({ items: [] }),
+  })
+  await installStressSessionTabs(page, { sessionIDs: sessions.map((session) => session.id) })
+
+  const hrefA = stressSessionHref(sessions[0]!.id)
+  const hrefB = stressSessionHref(sessions[1]!.id)
+  await page.goto(hrefA)
+  const composer = page.locator('[data-component="composer"]')
+  const modelControl = composer.locator('[data-action="composer-model"]')
+  const variant = composer.getByRole("button", { name: "Choose model variant", exact: true })
+  await expect(modelControl).toHaveText("Model A")
+  await expect(variant).toHaveText("balanced")
+  await variant.click()
+  await page.getByRole("menuitemradio", { name: "high", exact: true }).click()
+  await expect(variant).toHaveText("high")
+
+  await page.locator(`[data-titlebar-tab-link][href="${hrefB}"]`).click()
+  await expect(page).toHaveURL(hrefB)
+  await expect(modelControl).toHaveText("Model B")
+  await expect(variant).toHaveText("balanced")
+
+  await page.locator(`[data-titlebar-tab-link][href="${hrefA}"]`).click()
+  await expect(page).toHaveURL(hrefA)
+  await expect(modelControl).toHaveText("Model A")
+  await expect(variant).toHaveText("high")
+})
@@ -47,7 +47,6 @@ type MockStreamWindow = Window & {
 }

 export async function mockOpenCodeServer(page: Page, config: MockServerConfig) {
-  const state = { cursors: new Map<string, string>(), nextCursor: 0 }
  const server = `http://${process.env.PLAYWRIGHT_SERVER_HOST ?? "127.0.0.1"}:${process.env.PLAYWRIGHT_SERVER_PORT ?? "4096"}`

  await page.addInitScript(
@@ -135,13 +134,7 @@ export async function mockOpenCodeServer(page: Page, config: MockServerConfig) {
    }, 50)
    page.on("close", () => clearInterval(timer))
  }
-  const transport = HttpRouter.toWebHandler(
-    HttpApiBuilder.layer(MockApi).pipe(
-      Layer.provide(mockHandlers(config, state)),
-      Layer.provide(HttpServer.layerServices),
-    ),
-    { disableLogger: true },
-  )
+  const transport = createMockServerHandler(config)
  page.on("close", () => void transport.dispose())

  await page.route("**/api/**", async (route) => {
@@ -173,6 +166,16 @@ export async function mockOpenCodeServer(page: Page, config: MockServerConfig) {
  })
 }

+export function createMockServerHandler(config: MockServerConfig) {
+  return HttpRouter.toWebHandler(
+    HttpApiBuilder.layer(MockApi).pipe(
+      Layer.provide(mockHandlers(config, { cursors: new Map<string, string>(), nextCursor: 0 })),
+      Layer.provide(HttpServer.layerServices),
+    ),
+    { disableLogger: true },
+  )
+}
+
 const corsHeaders = {
  "access-control-allow-origin": "*",
  "access-control-allow-headers": "*",
@@ -34,6 +34,8 @@
    "test:service-worker": "bun run build && playwright test --config e2e/service-worker/playwright.config.ts",
    "test:stability": "bun test ./e2e/performance/unit/visual-stability.test.ts && playwright test --config e2e/performance/timeline-stability/playwright.config.ts",
    "test:bench": "bun test ./e2e/performance/unit && playwright test --config e2e/performance/playwright.config.ts",
+    "bench:tabs": "PLAYWRIGHT_BUILD=1 playwright test --config e2e/performance/playwright.config.ts timeline/session-tab-switch-benchmark.spec.ts --repeat-each=20 --workers=1 --retries=0 --reporter=line,./e2e/performance/tab-switch-reporter.ts",
+    "bench:entry": "PLAYWRIGHT_BUILD=1 playwright test --config e2e/performance/playwright.config.ts timeline/session-entry-benchmark.spec.ts --repeat-each=20 --workers=1 --retries=0 --reporter=line,./e2e/performance/tab-switch-reporter.ts",
    "test:bench:devex": "bun test ./e2e/performance/unit/desktop-startup.test.ts && playwright test --config e2e/performance/devex/playwright.config.ts"
  },
  "license": "MIT",
@@ -54,6 +56,7 @@
    "vite-plugin-solid": "2.11.14"
  },
  "dependencies": {
+    "@corvu/drawer": "catalog:",
    "@dnd-kit/abstract": "0.5.0",
    "@dnd-kit/dom": "0.5.0",
    "@dnd-kit/helpers": "0.5.0",
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
Shoubhit Dash	48aedea579	refactor(ai): simplify response fragment guards	2026-08-31 19:30:04 +05:30
Shoubhit Dash	96c0f0d8c3	fix(ai): enforce exclusive checkpoint representations	2026-08-31 19:29:23 +05:30
Shoubhit Dash	4b7d49c9a4	fix(ai): preserve compacted image detail on replay	2026-08-31 19:27:34 +05:30
Shoubhit Dash	0e1081b0a5	fix(ai): apply cache policy to bedrock messages	2026-08-31 19:26:21 +05:30
Shoubhit Dash	ce96bbe8a7	refactor(ai): share request preparation across operations	2026-08-31 19:25:28 +05:30
Shoubhit Dash	027f0e7cc9	fix(ai): reject order-unsafe checkpoint recovery	2026-08-31 19:23:58 +05:30
Shoubhit Dash	0a442fa2f8	chore: merge v2 into provider-compaction	2026-08-31 19:08:10 +05:30
Brendan Allan	5a4914c670	feat(app): refine mobile shell home and settings (#46388 )	2026-08-31 21:20:17 +08:00
Kit Langton	3a797bf6e4	fix(ai): validate canonical tool results (#46062 )	2026-08-31 09:14:31 -04:00
Shoubhit Dash	59d1ab783c	refactor(ai): make compaction conversion synchronous	2026-08-31 17:29:38 +05:30
Shoubhit Dash	6b4426bbd3	docs(ai): describe typed compaction history	2026-08-31 17:19:57 +05:30
Shoubhit Dash	1b9f762ea5	refactor(ai): model compaction as typed conversation parts	2026-08-31 17:19:45 +05:30
opencode-agent[bot]andvimtor	33536da231	fix(core): commit undo before compaction (#46383 ) Co-authored-by: vimtor <36263538+vimtor@users.noreply.github.com>	2026-08-31 17:03:51 +05:30
Shoubhit Dash	c8b4963d98	refactor(ai): simplify compaction control flow	2026-08-31 16:24:01 +05:30
Shoubhit Dash	14a90331bc	docs(ai): explain provider compaction and replay ownership	2026-08-31 16:11:38 +05:30
Shoubhit Dash	cd2880075a	test(ai): cover compaction across http and websocket flows	2026-08-31 16:11:17 +05:30
Shoubhit Dash	af22c16249	fix(ai): validate compaction boundaries and incomplete blocks	2026-08-31 16:09:00 +05:30
Shoubhit Dash	bc4db825a4	fix(core): reject unsupported compaction replay in ai sdk routes	2026-08-31 16:08:32 +05:30
Shoubhit Dash	4f0ba4f06e	feat(ai): add bedrock messages route for claude compaction	2026-08-31 16:02:30 +05:30
Shoubhit Dash	8dd92e2d95	feat(ai): add explicit responses compaction calls	2026-08-31 15:58:29 +05:30
Shoubhit Dash	cb5a6c7db4	feat(ai): support anthropic compaction and iteration usage	2026-08-31 15:53:57 +05:30
Shoubhit Dash	eaa3dfe04e	feat(ai): support automatic responses compaction	2026-08-31 15:51:01 +05:30
Shoubhit Dash	5fd800d169	feat(ai): preserve provider compaction in messages and events	2026-08-31 15:48:42 +05:30
Shoubhit Dash	e56ceed32b	Revert "fix(tui): surface subagent permissions and questions" (#46376 )	2026-08-31 15:41:46 +05:30
Shoubhit Dash	b2c7246134	fix(core): refresh git references on daily activity (#45575 )	2026-08-31 15:36:37 +05:30
Shoubhit Dash	9c39e75ce2	fix(tui): surface subagent permissions and questions (#44976 )	2026-08-31 14:38:05 +05:30
opencode-agent[bot]andBrendonovich	0dad76e618	fix(ui): prevent menu items from shrinking (#46353 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-31 16:00:55 +08:00
Luke Parker	90fb6562ce	fix(shell): bound post-exit pipe draining on all platforms (#46085 )	2026-08-31 16:53:35 +10:00
opencode-agent[bot]andBrendonovich	174d263890	fix(app): save session titles on blur and add tab context menu (#46113 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-31 13:07:32 +08:00
opencode-agent[bot]andBrendonovich	5ec29e7a87	refactor(desktop): use password-only server authentication (#45958 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-31 13:06:35 +08:00
Luke Parker	3c6b85acf7	fix(app): reveal pasted composer content with custom scrollbar (#46339 )	2026-08-31 14:38:28 +10:00
Luke Parker	50e77f66fd	fix(app): keep composer select all scoped to the editor (#46338 )	2026-08-31 14:38:00 +10:00
Aiden Cline	d484f070d1	fix(core): recover reads with non-breaking spaces (#45807 )	2026-08-30 23:13:26 -05:00
Luke Parker	8890294bf0	fix(desktop): preserve Windows editing shortcuts (#46336 )	2026-08-31 04:07:42 +00:00
Luke Parker	a1925de0c1	fix(core): flush trailing stream chunks while providers pause (#46326 )	2026-08-31 14:03:17 +10:00
Aiden Cline	24e826d06b	fix(ai): validate Bedrock media data (#46333 )	2026-08-30 22:58:58 -05:00
Aiden Cline	19625400c1	fix(ai): normalize tool result history (#46309 )	2026-08-30 22:19:48 -05:00
opencode-agent[bot]	52cbe7b3ee	chore: update nix node_modules hashes	2026-08-31 02:04:13 +00:00
Dax	6a2c3e91c7	feat(plugin): add typed rpc and custom events (#46105 )	2026-08-30 21:28:19 -04:00
Aiden Cline	4afd8e81be	fix(ai): deduplicate request tools (#46306 )	2026-08-30 16:24:46 -05:00
Aiden Cline	485bdc9c4e	fix(ai): omit empty Responses user messages (#46297 )	2026-08-30 16:03:41 -05:00
Aiden Cline	1a3aee39de	fix(ai): sanitize empty Bedrock tool input keys (#46296 )	2026-08-30 16:02:31 -05:00
Aiden Cline	2ddf257c6f	fix(ai): filter empty Anthropic messages (#46291 )	2026-08-30 14:39:46 -05:00
Aiden Cline	ef50e0b6d8	fix(ai): omit empty Bedrock system blocks (#46294 )	2026-08-30 14:38:51 -05:00
Aiden Cline	583a1a2b6f	fix(ai): omit empty Anthropic system blocks (#46289 )	2026-08-30 14:21:05 -05:00
Aiden Cline	68c1207b52	fix(ai): close Gemini text before tools (#46286 )	2026-08-30 14:08:25 -05:00
Aiden Cline	f77647ad12	fix(ai): accumulate bedrock redacted content (#46283 )	2026-08-30 13:45:39 -05:00
Aiden Cline	c746ea3210	fix(ai): use unique Gemini block ids (#46279 )	2026-08-30 13:21:35 -05:00
Aiden Cline	d323b34826	feat(ai): add native Mistral provider (#46278 )	2026-08-30 12:56:01 -05:00
Aiden Cline	bf50194f99	fix(ai): reject truncated bedrock frames (#46281 )	2026-08-30 12:47:33 -05:00
Kit Langton	4a977b2b31	refactor(core): bind standalone skill activation to Session (#46077 ) Move standalone skill activation into ID-bound Session handles and delegate from the public service. Preserve current-placement lookup, raw skill content, ambient publication context, validation order, and host-scoped detached resume behavior. Cover ownership and lifecycle contracts with focused regressions.	2026-08-30 10:22:03 -04:00
Aiden Cline	b1e3a7b222	fix(ai): preserve forced reasoning signature (#46218 )	2026-08-29 23:09:03 -05:00
opencode-agent[bot]andBrendonovich	b12d43698e	fix(app): recover sessions with unavailable locations (#46215 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-30 11:51:48 +08:00
opencode-agent[bot]	5cd2c22b27	chore: update nix node_modules hashes	2026-08-30 03:43:36 +00:00
Aiden Cline	c2ef80a287	fix(ai): fail malformed converse output (#46193 )	2026-08-29 22:29:53 -05:00
opencode-agent[bot]andBrendonovich	4df3029536	fix(desktop): scope library validation exception to CLI (#46212 ) Co-authored-by: Brendonovich <14191578+Brendonovich@users.noreply.github.com>	2026-08-30 03:29:31 +00:00
Aiden Cline	e70d667a9f	fix(ai): preserve Anthropic finish across usage deltas (#46171 )	2026-08-29 15:45:27 -05:00
Kit Langton	8ba434b597	refactor(core): move projected Session reads into Store (#46075 ) Move existing Session list and message queries into SessionStore. Preserve public response wrapping, Session existence checks, pagination, ordering, and typed message decoding errors.	2026-08-29 09:31:04 -04:00
Kit Langton	171947787c	test(tui): wait for diff base search focus (#46083 ) Wait for the diff-base search input to receive focus before typing. Preserve existing assertions and timeouts while removing the render-versus-focus test race.	2026-08-29 08:37:32 -04:00
Luke Parker	106629aa11	feat(infra): deploy beta web app with SST (#46086 )	2026-08-29 14:51:22 +10:00
Luke Parker	3ee2e482ce	fix(app): preserve Windows panel top outlines (#46090 )	2026-08-29 04:40:55 +00:00
Kit Langton	849824efd2	refactor(core): merge defaults for selected MCP servers (#46072 )	2026-08-28 23:48:29 -04:00
Kit Langton	cf2c3a536d	refactor(core): reuse catalog response digest (#46071 )	2026-08-28 23:41:49 -04:00
opencode-agent[bot]	7852cecd72	chore: update nix node_modules hashes	2026-08-29 03:33:15 +00:00
Kit Langton	6cfffeb031	refactor(core): avoid encoding rejected image candidates (#46073 )	2026-08-28 23:32:58 -04:00
Kit Langton	6e954f75ee	refactor(core): isolate Session admission and controls (#46019 ) Separate ID-bound Session policy from host routing. Bind Inbox and Location preparation dependencies at construction, preserve admission and execution semantics, and cover the extracted ownership contracts directly.	2026-08-28 23:22:25 -04:00
Kit Langton	0116a98371	refactor(tui): share app lifecycle test fixture Reuse a file-local fixture for renderer, HTTP server, event stream, app startup, and teardown across lifecycle tests. Preserve scenario-specific handlers, configuration, deferred responses, and assertions while removing 187 lines of repeated setup.	2026-08-28 23:19:49 -04:00
Kit Langton	a38cbd42aa	refactor(core): isolate shell tool preparation Name the tool-owned pre-spawn preparation boundary while preserving hook edits, permission ordering, directory validation, and effective timeout reporting. Strengthen the existing regression assertions.	2026-08-28 23:17:48 -04:00
Luke Parker	4ab31867c4	fix(app): reduce session-switch latency (#46044 )	2026-08-29 13:16:53 +10:00
Luke Parker	51a082cea3	fix(core): release exited shell execution state (#46058 )	2026-08-29 03:08:50 +00:00
Kit Langton	e3bda5e2d0	feat(tui): add branch review scopes	2026-08-28 22:46:23 -04:00
Kit Langton	ad9117b107	fix(session): run user shells immediately in the background Run user shell commands concurrently with model execution, retain their output in one shell entry, and admit non-waking completion messages. Share result capture and notification formatting across shell callers.	2026-08-28 22:46:20 -04:00
Kit Langton	6b7a1d419c	fix(core): preserve Responses tool identities (#45658 )	2026-08-28 22:14:13 -04:00
Brendan Allan	c0d0f5f4bc	feat(cli): configure server CORS origins (#45544 )	2026-08-29 09:41:52 +08:00
opencode-agent[bot]andHona	095ed63ea0	fix(app): keep workspace submit buttons neutral (#46052 ) Co-authored-by: Hona <10430890+Hona@users.noreply.github.com>	2026-08-29 11:20:41 +10:00
Kit Langton	fa5ccac707	refactor(core): share subagent completion delivery (#46054 )	2026-08-29 01:20:26 +00:00
Kit Langton	6809be2d0a	fix(core): reuse permission policy for pending approvals (#46050 )	2026-08-28 21:03:52 -04:00
Aiden Cline	a5f8869b35	fix(ai): reject unknown chat finishes (#46036 )	2026-08-28 18:22:47 -05:00
Aiden Cline	1c9c5305a6	fix(ai): classify retryable server failures (#46038 )	2026-08-28 18:11:52 -05:00
Aiden Cline	cf014bf2c1	fix(ai): avoid truncated tool calls (#46040 )	2026-08-28 18:11:40 -05:00
Aiden Cline	8e7190f795	fix(ai): respect completed response item text (#45854 )	2026-08-28 17:32:19 -05:00
Kit Langton	82b6e0e316	chore: enforce mapError simplifications (#46028 )	2026-08-28 17:59:18 -04:00
Kit Langton	5b39f5184f	test(core): share selector sentinels (#45748 )	2026-08-28 17:59:12 -04:00
Kit Langton	1da25727b2	refactor(core): remove unused insert state (#45738 )	2026-08-28 17:59:07 -04:00
Kit Langton	51d53f45c1	test(ci): register omitted V2 unit suites (#45970 )	2026-08-28 21:58:54 +00:00
Kit Langton	f779b2748a	test(core): align runner service style (#45753 )	2026-08-28 17:47:29 -04:00
Kit Langton	f2fb191f53	test(core): inject vertex auth transport (#45751 )	2026-08-28 17:47:24 -04:00
Kit Langton	0b32bdf1e5	test(core): use disposable session fixtures (#45745 )	2026-08-28 17:47:18 -04:00
Kit Langton	2751813454	test(core): simplify provider test fixtures (#45743 )	2026-08-28 17:47:13 -04:00
Aiden Cline	f61858e683	fix(ai): avoid filtered tool calls (#46029 )	2026-08-28 16:44:05 -05:00
Kit Langton	87525e00b9	test(core): synchronize retry tests on scheduled events (#46027 )	2026-08-28 21:37:04 +00:00
Kit Langton	803b7718b8	refactor(core): import canonical schema contracts (#45746 )	2026-08-28 17:29:07 -04:00
Kit Langton	8a24a01bff	test(core): repair runner fixture claims (#45737 )	2026-08-28 17:29:03 -04:00
Kit Langton	9e39a4fbdf	test(core): simplify skill test setup (#45721 )	2026-08-28 17:29:00 -04:00
Kit Langton	d0baff184b	fix(core): make provider publication cancellation-safe (#46026 )	2026-08-28 17:24:27 -04:00
Kit Langton	d82a0b28a9	docs: clarify release workflow (#46025 )	2026-08-28 17:18:47 -04:00
Kit Langton	bd379e13cb	refactor(core): remove unused instruction import (#45747 )	2026-08-28 17:18:44 -04:00
Kit Langton	6e1f783aec	refactor(core): expose pty layer directly (#45733 )	2026-08-28 17:18:41 -04:00
Kit Langton	edef6a4b15	refactor(core): flatten git runner arguments (#45731 )	2026-08-28 17:18:38 -04:00
Kit Langton	5990679ebd	refactor: use typed error mapping operators (#45727 )	2026-08-28 17:18:35 -04:00
Kit Langton	b0c8a8c827	fix(util): refresh flock heartbeat time (#45720 )	2026-08-28 17:18:32 -04:00
Aiden Cline	d837ffe70f	feat(plugin): add session retry hook (#45999 )	2026-08-28 16:16:19 -05:00
Kit Langton	b1d7dd82fc	refactor(core): use nonempty array guards (#45718 )	2026-08-28 17:12:03 -04:00
Kit Langton	e409567428	refactor(core): express operation policies as transforms (#45713 )	2026-08-28 17:11:59 -04:00
Kit Langton	9538c2171f	refactor(core): derive config recognition fields (#45701 )	2026-08-28 17:11:55 -04:00
Kit Langton	0a718be0d9	refactor(plugin): normalize web search API type (#45717 )	2026-08-28 16:55:32 -04:00
Kit Langton	67845091ba	fix(util): share streamed stderr consumption (#45716 )	2026-08-28 16:55:28 -04:00
Kit Langton	fe788b7842	fix(core): use runtime update columns (#45700 )	2026-08-28 16:55:24 -04:00
Kit Langton	ee42eb3ca3	fix(core): render built-in command arguments (#45697 )	2026-08-28 16:55:21 -04:00
Kit Langton	80323a4deb	test(core): name acquired test services (#45696 )	2026-08-28 16:55:17 -04:00
Kit Langton	a35f96f427	fix(core): preserve late opaque reasoning (#45694 )	2026-08-28 16:55:13 -04:00
Kit Langton	d354c3d640	fix(server): avoid loading locations for pending reads (#45994 )	2026-08-28 20:44:48 +00:00
Kit Langton	ce005ce002	refactor(core): remove unused result plumbing (#45695 )	2026-08-28 16:44:27 -04:00
				`@@ -0,0 +1 @@`
				`export { messagesModel as model } from "../amazon-bedrock.js"`