fix(core): isolate shell jobs by shell ID (#45985 )

Co-authored-by: rekram1-node <rekram1-node@users.noreply.github.com>
fix(tui): admit compaction requests optimistically (#45973 )
2026-08-28 20:46:14 +00:00 · 2026-08-28 13:13:16 -05:00 · 2026-08-28 14:00:09 -04:00 · 2026-08-28 13:47:50 -04:00 · 2026-08-28 13:37:42 -04:00 · 2026-08-28 13:37:11 -04:00
309 changed files with 10108 additions and 4037 deletions
@@ -1,5 +0,0 @@
---
-"@opencode-ai/theme": patch
---
-
-Use the unread accent color by default for question and permission status indicators, while preserving explicit theme overrides.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Align the built-in review and system prompts with available documentation and code-search tools and the current Plan mode reminder.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Preserve prompt cache prefixes when sessions move between locations with unchanged instructions.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Correct the published Newtype documentation to recommend `Schema.decodeUnknownSync` for schema-based synchronous decoding.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Fix OpenCode Console device authorization URLs when the server returns an origin-rooted verification path.
@@ -0,0 +1,5 @@
+---
+"@opencode-ai/core": patch
+---
+
+Correct directory page headings when the read offset is zero.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Propagate fatal Git ignore-check errors during index refresh so best-effort snapshot capture returns unavailable instead of treating the check as having no matches.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Prompt and synthetic inbox ID reuse is now idempotent: reusing an ID within the same Session succeeds and returns the first admission, ignoring the retried payload, metadata, and delivery mode. Previously reuse with a differing payload failed with a conflict. Cross-Session and cross-type reuse still fail, and control items keep their operation-specific conflict behavior.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": major
---
-
-Treat MCP as a word in current Core namespace exports: rename `MCP` to `Mcp`, `MCPClient` to `McpClient`, `MCPStdio` to `McpStdio`, `MCPOAuth` to `McpOAuth`, `ConfigMCPPlugin` to `ConfigMcpPlugin`, and `MCPCodeModeExclusionPlugin` to `McpCodeModeExclusionPlugin`. Direct consumers must update their imports. Module paths, Schema contracts, runtime service keys, error tags, and behavior are unchanged.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Rename the write, patch, and question tool formatting helpers from `toModelOutput` to `toModelContent` to match the result field they populate. Direct imports of these helpers must use the new name; generated content and declared machine output are unchanged.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": minor
---
-
-Rename ModelResolver.supported to ModelResolver.hasPackage. Consumers of the old export must update the name; the predicate remains Boolean(model.package), checking only whether a catalog model declares a provider package, not whether it can be loaded. Default-model selection behavior is unchanged.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Nested AGENTS.md instructions are re-injected after compaction. Previously the in-memory dedup claim outlived the synthetic message that compaction dropped from model-visible history, so nested instructions were silently lost for the rest of the process lifetime. The claim now only guards in-flight loads; the synthetic message metadata in durable history is the sole lasting ledger, so any history truncation (compaction, revert) self-heals on the next read in that subtree.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/plugin": patch
---
-
-Derive Promise plugin API request and response conversion from the canonical protocol schemas.
@@ -1,8 +0,0 @@
---
-"@opencode-ai/core": minor
-"@opencode-ai/schema": minor
-"@opencode-ai/protocol": minor
-"@opencode-ai/client": minor
---
-
-Remove the unused question request API and use session forms for question tool interactions.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Wake idle sessions when recovering background shell outcomes after a server restart. Admit shell outcomes before resuming background child sessions, while preserving restart retry budgets and completion-driven parent notifications.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Include both paths of renamed files in new snapshot change lists so undo restores the original file instead of only deleting the renamed file.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Apply shared Session model-request preparation to transient generation.
@@ -1,9 +0,0 @@
---
-"@opencode-ai/core": minor
-"@opencode-ai/schema": minor
-"@opencode-ai/protocol": minor
-"@opencode-ai/client": minor
---
-
-Add an opt-in portable shell permission scanner. Opaque commands use normal shell authorization without inferring
-external directories, while the default tree-sitter path remains unchanged.
@@ -1,10 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Make the experimental portable shell scanner authoritative, with no Tree-sitter
-fallback. Scan common Bash and PowerShell control flow, heredocs, functions,
-expressions, quoting, and substitutions natively. Preserve existing redirect and
-declaration permission matching, and make PowerShell saved approvals cover the
-original command spelling. Parser failures remain visible without changing the
-permission engine. The default Tree-sitter path is unchanged.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Use parent directory names for SKILL.md IDs so individually downloaded skills do not overwrite each other.
@@ -1,8 +0,0 @@
---
-"@opencode-ai/core": patch
-"@opencode-ai/app": patch
---
-
-Keep project labels stable when opening multiple clones of the same repository, while still refreshing the canonical path when its directory is renamed or removed.
-
-Worktree setup scripts receive the selected source directory as `OPENCODE_WORKTREE_BASE` rather than another clone's shared project path.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Refresh expired OAuth credentials resolved during batched plugin activation. Integration connection resolution now materializes registrations deferred by the activation batch before consulting refresh implementations, so a just-registered OAuth method is no longer skipped and an expired token no longer used as-is.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Simplify interrupt continuation: the steer-scoped resume decision now lives in SessionExecution as a post-cleanup inbox check, and the run coordinator drops its continuation state machine. Wakes arriving during cancellation cleanup now restart a normal full drain, and interrupting an idle session with continue now resumes pending steering input. Recovery-applied moves now end with the same full wake as inbox-admitted moves, retrying any stranded inbox work at the new location. Interrupting with continue now also resumes a next-in-line control item: between-turn manual compaction and moves run under any drain scope, while queued prompts remain parked.
@@ -1,6 +0,0 @@
---
-"@opencode-ai/client": patch
-"@opencode-ai/plugin": patch
---
-
-Add form reply and cancellation operations that reconcile terminal forms in the local TUI projection.
@@ -1,5 +0,0 @@
---
-"@opencode-ai/core": patch
---
-
-Title generation and compaction summaries now build their model requests through the shared session request boundary, gaining unsupported-media filtering and image bounds while explicitly opting out of session context hooks: plugins that shape the agent conversation do not observe title or compaction requests. Title requests gain the fork-aware session prompt cache key, and compaction summaries in forked sessions reuse the fork root's prompt cache key instead of the fork's own.
@@ -1,6 +0,0 @@
---
-"@opencode-ai/plugin": patch
-"@opencode-ai/core": patch
---
-
-Add transport-neutral Session model request hooks and provider-scoped hook registration so eligible OpenAI Responses requests can prefer WebSocket without bypassing HTTP-only middleware.
@@ -185,7 +185,7 @@
        "solid-js": "catalog:",
      },
      "peerDependencies": {
-        "effect": "4.0.0-rc.111",
+        "effect": "4.0.0-rc.112",
        "solid-js": ">=1.9.0",
      },
      "optionalPeers": [
@@ -426,7 +426,7 @@
        "@sentry/vite-plugin": "catalog:",
        "@solid-primitives/storage": "catalog:",
        "@solidjs/meta": "catalog:",
-        "@solidjs/router": "0.15.4",
+        "@solidjs/router": "catalog:",
        "@types/bun": "catalog:",
        "@types/node": "catalog:",
        "@typescript/native-preview": "catalog:",
@@ -506,7 +506,7 @@
      "name": "@opencode-ai/http-recorder",
      "version": "1.18.15",
      "dependencies": {
-        "@effect/platform-node-shared": "4.0.0-rc.111",
+        "@effect/platform-node-shared": "4.0.0-rc.112",
      },
      "devDependencies": {
        "@effect/platform-node": "catalog:",
@@ -1043,7 +1043,7 @@
    "@ai-sdk/mistral@3.0.51": "patches/@ai-sdk%2Fmistral@3.0.51.patch",
    "@npmcli/agent@4.0.2": "patches/@npmcli%2Fagent@4.0.2.patch",
    "@silvia-odwyer/photon-node@0.3.4": "patches/@silvia-odwyer%2Fphoton-node@0.3.4.patch",
-    "solid-js@1.9.10": "patches/solid-js@1.9.10.patch",
+    "solid-js@1.9.15": "patches/solid-js@1.9.15.patch",
    "@ff-labs/fff-bun@0.10.5": "patches/@ff-labs%2Ffff-bun@0.10.5.patch",
    "@ai-sdk/google@3.0.73": "patches/@ai-sdk%2Fgoogle@3.0.73.patch",
    "@dnd-kit/dom@0.5.0": "patches/@dnd-kit%2Fdom@0.5.0.patch",
@@ -1057,14 +1057,15 @@
    "@types/bun": "catalog:",
    "@types/node": "catalog:",
    "effect": "catalog:",
+    "solid-js": "catalog:",
  },
  "catalog": {
    "@cloudflare/workers-types": "4.20251008.0",
    "@corvu/drawer": "0.2.4",
-    "@effect/opentelemetry": "4.0.0-rc.111",
-    "@effect/platform-node": "4.0.0-rc.111",
-    "@effect/platform-node-shared": "4.0.0-rc.111",
-    "@effect/sql-sqlite-bun": "4.0.0-rc.111",
+    "@effect/opentelemetry": "4.0.0-rc.112",
+    "@effect/platform-node": "4.0.0-rc.112",
+    "@effect/platform-node-shared": "4.0.0-rc.112",
+    "@effect/sql-sqlite-bun": "4.0.0-rc.112",
    "@hono/standard-validator": "0.2.0",
    "@hono/zod-validator": "0.4.2",
    "@kobalte/core": "0.13.13",
@@ -1085,7 +1086,7 @@
    "@solid-primitives/resize-observer": "2.2.0",
    "@solid-primitives/storage": "4.4.0",
    "@solidjs/meta": "0.29.4",
-    "@solidjs/router": "0.15.4",
+    "@solidjs/router": "1.0.0",
    "@solidjs/start": "https://pkg.pr.new/@solidjs/start@dfb2020",
    "@standard-schema/spec": "1.1.0",
    "@tailwindcss/vite": "4.1.11",
@@ -1104,7 +1105,7 @@
    "dompurify": "3.4.14",
    "drizzle-kit": "1.0.0-rc.5-ab785fc",
    "drizzle-orm": "1.0.0-rc.5-169397b",
-    "effect": "4.0.0-rc.111",
+    "effect": "4.0.0-rc.112",
    "fuzzysort": "3.1.0",
    "get-east-asian-width": "1.6.0",
    "hono": "4.10.7",
@@ -1118,7 +1119,7 @@
    "resolve.exports": "2.0.3",
    "semver": "7.7.4",
    "shiki": "4.4.3",
-    "solid-js": "1.9.10",
+    "solid-js": "1.9.15",
    "solid-list": "0.3.0",
    "solid-sonner": "0.3.1",
    "sst": "4.13.1",
@@ -1663,13 +1664,13 @@

    "@drizzle-team/brocli": ["@drizzle-team/brocli@0.12.0", "", {}, "sha512-mlUE+rZ8CatQekLhnaiN91Iemdd+e2gFKooGlnRB3oPTL3VghLfX24dx7HrzMNeC1JrIB/0kpsfyty3f5HNfxQ=="],

-    "@effect/opentelemetry": ["@effect/opentelemetry@4.0.0-rc.111", "", { "peerDependencies": { "@opentelemetry/api": ">=1.9.0 <2.0.0", "@opentelemetry/api-logs": ">=0.203.0 <0.300.0", "@opentelemetry/resources": ">=2.0.0 <3.0.0", "@opentelemetry/sdk-logs": ">=0.203.0 <0.300.0", "@opentelemetry/sdk-metrics": ">=2.0.0 <3.0.0", "@opentelemetry/sdk-trace-base": ">=2.0.0 <3.0.0", "@opentelemetry/sdk-trace-node": ">=2.0.0 <3.0.0", "@opentelemetry/sdk-trace-web": ">=2.0.0 <3.0.0", "@opentelemetry/semantic-conventions": ">=1.33.0 <2.0.0", "effect": "^4.0.0-rc.111" }, "optionalPeers": ["@opentelemetry/api", "@opentelemetry/api-logs", "@opentelemetry/resources", "@opentelemetry/sdk-logs", "@opentelemetry/sdk-metrics", "@opentelemetry/sdk-trace-base", "@opentelemetry/sdk-trace-node", "@opentelemetry/sdk-trace-web"] }, "sha512-bztAYpWoipn/a4tFjECMNvtWLkf2TS+Qn23PBI+f4q5DP9cBKfV00uhX8cwEVcfAtZd+lflSbGURHa4xlSewOA=="],
+    "@effect/opentelemetry": ["@effect/opentelemetry@4.0.0-rc.112", "", { "peerDependencies": { "@opentelemetry/api": ">=1.9.0 <2.0.0", "@opentelemetry/api-logs": ">=0.203.0 <0.300.0", "@opentelemetry/resources": ">=2.0.0 <3.0.0", "@opentelemetry/sdk-logs": ">=0.203.0 <0.300.0", "@opentelemetry/sdk-metrics": ">=2.0.0 <3.0.0", "@opentelemetry/sdk-trace-base": ">=2.0.0 <3.0.0", "@opentelemetry/sdk-trace-node": ">=2.0.0 <3.0.0", "@opentelemetry/sdk-trace-web": ">=2.0.0 <3.0.0", "@opentelemetry/semantic-conventions": ">=1.33.0 <2.0.0", "effect": "^4.0.0-rc.112" }, "optionalPeers": ["@opentelemetry/api", "@opentelemetry/api-logs", "@opentelemetry/resources", "@opentelemetry/sdk-logs", "@opentelemetry/sdk-metrics", "@opentelemetry/sdk-trace-base", "@opentelemetry/sdk-trace-node", "@opentelemetry/sdk-trace-web"] }, "sha512-OTRv1DxTHUmnakgJ6XVM8wVgF1KgZH4UXnOemwSLUwUjXO+RCikzF8oR/rlVmOGq81KtQzj1URM4M4nchlQOuQ=="],

-    "@effect/platform-node": ["@effect/platform-node@4.0.0-rc.111", "", { "dependencies": { "@effect/platform-node-shared": "^4.0.0-rc.111", "mime": "^4.1.0", "undici": "^8.10.0" }, "peerDependencies": { "effect": "^4.0.0-rc.111", "redis": ">=5.0.0 <7.0.0" } }, "sha512-oy1i7HsOGg/5r+DuBe5+ddmnUhnXmyZFPFPXZCBdO/RQpHK3PIFe1/2UMGxZE+ngDymrSksuQTSgQLF6P+MLqw=="],
+    "@effect/platform-node": ["@effect/platform-node@4.0.0-rc.112", "", { "dependencies": { "@effect/platform-node-shared": "^4.0.0-rc.112", "mime": "^4.1.0", "undici": "^8.10.0" }, "peerDependencies": { "effect": "^4.0.0-rc.112", "redis": ">=5.0.0 <7.0.0" } }, "sha512-/BMAcdNGQQskLmI0Zoa95KfTZkr9HV9N4NSxaSrusG6GeW6Ulp9KvZ+Rlaiw8lnOt43CXjFLdfll5/k5rxL4hQ=="],

-    "@effect/platform-node-shared": ["@effect/platform-node-shared@4.0.0-rc.111", "", { "dependencies": { "@types/ws": "^8.18.1", "ws": "^8.21.3" }, "peerDependencies": { "effect": "^4.0.0-rc.111" } }, "sha512-iES0Q9vmjhaUKqeW9ceonuD45MUg/Ouk08LzRSptZ+B5qB0w9WlRjDmUz5TJmY2betNop5FRI5k4AD4mtQt3Bw=="],
+    "@effect/platform-node-shared": ["@effect/platform-node-shared@4.0.0-rc.112", "", { "dependencies": { "@types/ws": "^8.18.1", "ws": "^8.21.3" }, "peerDependencies": { "effect": "^4.0.0-rc.112" } }, "sha512-ttjz0xKamFN7vL8pNDYVwddJLjZvqKePc05djlz2VcdaKbLsnYbtMnL1rbOfHgEnIUSHGh7FkjaN4DM1Ov81sQ=="],

-    "@effect/sql-sqlite-bun": ["@effect/sql-sqlite-bun@4.0.0-rc.111", "", { "peerDependencies": { "effect": "^4.0.0-rc.111" } }, "sha512-u5HqWLYISTH5ydsCr45nOGmkspmAeNzK4q+plMrJntDoG8sQttwizC4akGfbQDgoSxyYfyBzi3YyVW/NvhvSQQ=="],
+    "@effect/sql-sqlite-bun": ["@effect/sql-sqlite-bun@4.0.0-rc.112", "", { "peerDependencies": { "effect": "^4.0.0-rc.112" } }, "sha512-EqR8pWZo3VzedRvP6qo5HcfqNmFvSa+sQLWQl3oBMx2KY1FCc8NYjep79wYjNkW0qABnffrPUXCMTOncd1A+5g=="],

    "@electron-internal/extract-zip": ["@electron-internal/extract-zip@1.0.5", "", {}, "sha512-+bqFCP98pLI0Tt0XQo1TmlXtwjWchISndDOxCkEcIuUgXWpBnLyRI+2DU+mesvnMMX6L1XDqYNA0lXNDHd/yiA=="],

@@ -2893,7 +2894,7 @@

    "@solidjs/meta": ["@solidjs/meta@0.29.4", "", { "peerDependencies": { "solid-js": ">=1.8.4" } }, "sha512-zdIWBGpR9zGx1p1bzIPqF5Gs+Ks/BH8R6fWhmUa/dcK1L2rUC8BAcZJzNRYBQv74kScf1TSOs0EY//Vd/I0V8g=="],

-    "@solidjs/router": ["@solidjs/router@0.15.4", "", { "peerDependencies": { "solid-js": "^1.8.6" } }, "sha512-WOpgg9a9T638cR+5FGbFi/IV4l2FpmBs1GpIMSPa0Ce9vyJN7Wts+X2PqMf9IYn0zUj2MlSJtm1gp7/HI/n5TQ=="],
+    "@solidjs/router": ["@solidjs/router@1.0.0", "", { "peerDependencies": { "solid-js": "^1.8.6" } }, "sha512-cCSk1hvgCowiMa9bzzYWHiLu1U4E22+DfJe6/rOwAyECKrxc3jrd5QnoW3sDDJtW+e077cz/M67bPl3DqOBw1Q=="],

    "@solidjs/start": ["@solidjs/start@https://pkg.pr.new/@solidjs/start@dfb2020", { "dependencies": { "@babel/core": "^7.28.3", "@babel/traverse": "^7.28.3", "@babel/types": "^7.28.5", "@solidjs/meta": "^0.29.4", "@tanstack/server-functions-plugin": "1.134.5", "@types/babel__traverse": "^7.28.0", "@types/micromatch": "^4.0.9", "cookie-es": "^2.0.0", "defu": "^6.1.4", "error-stack-parser": "^2.1.4", "es-module-lexer": "^1.7.0", "esbuild": "^0.25.3", "fast-glob": "^3.3.3", "h3": "npm:h3@2.0.1-rc.4", "html-to-image": "^1.11.13", "micromatch": "^4.0.8", "path-to-regexp": "^8.2.0", "pathe": "^2.0.3", "radix3": "^1.1.2", "seroval": "^1.3.2", "seroval-plugins": "^1.2.1", "shiki": "^1.26.1", "solid-js": "^1.9.9", "source-map-js": "^1.2.1", "srvx": "^0.9.1", "terracotta": "^1.0.6", "vite": "7.1.10", "vite-plugin-solid": "^2.11.9", "vitest": "^4.0.10" } }, "sha512-7JjjA49VGNOsMRI8QRUhVudZmv0CnJ18SliSgK1ojszs/c3ijftgVkzvXdkSLN4miDTzbkXewf65D6ZBo6W+GQ=="],

@@ -3833,7 +3834,7 @@

    "ee-first": ["ee-first@1.1.1", "", {}, "sha512-WMwm9LhRUo+WUaRN+vRuETqG89IgZphVSNkdFgeb6sS/E4OrDIN7t48CAewSHXc6C8lefD8KKfr5vY61brQlow=="],

-    "effect": ["effect@4.0.0-rc.111", "", { "dependencies": { "@standard-schema/spec": "^1.1.0", "fast-check": "^4.9.0", "msgpackr": "^2.0.5" } }, "sha512-ASd5L58EIR0CUNueZNKKjSsyOCd+2alxOAIaTcHaqkJkPsaYSsw5Cg/cfANk5K4Jr2YsX756xvX11shzqsreWA=="],
+    "effect": ["effect@4.0.0-rc.112", "", { "dependencies": { "fast-check": "^4.9.0", "msgpackr": "^2.0.5" } }, "sha512-wXxwuh1Ywnv4cPRM3Wfa0vDwuOHnZ1TsTgHJkG9XgzND6inhBH9n1vBxhg3iIXOia/OrpmvVmd3lrD4vq6bF3A=="],

    "ejs": ["ejs@3.1.10", "", { "dependencies": { "jake": "^10.8.5" }, "bin": { "ejs": "bin/cli.js" } }, "sha512-UeJmFfOrAQS8OJWPZ4qtgHyWExa088/MtK5UEyoJGFH67cDEXkZSviOiKRCZ4Xij0zxI3JECgYs3oKx+AizQBA=="],

@@ -5235,9 +5236,9 @@

    "serialize-javascript": ["serialize-javascript@7.1.0", "", {}, "sha512-RNEqWOyhhUQYN9V1GfHwu9AR/g+NTciH6Z5u3/no6X3/w+04J2lVDL+svFQVXgXrEGBMG2puMVN3gq2SNGuTGw=="],

-    "seroval": ["seroval@1.3.2", "", {}, "sha512-RbcPH1n5cfwKrru7v7+zrZvjLurgHhGyso3HTyGtRivGWgYjbOmGuivCQaORNELjNONoK35nj28EoWul9sb1zQ=="],
+    "seroval": ["seroval@1.5.6", "", {}, "sha512-rVQVWjjSvlINzaQPZH5JFqsqEsIWdTxY3iJZCnTL/5gQbXIRooVZKI60tVCkOVfzcRPejboxO2t0P89dg5mQaA=="],

-    "seroval-plugins": ["seroval-plugins@1.3.3", "", { "peerDependencies": { "seroval": "^1.0" } }, "sha512-16OL3NnUBw8JG1jBLUoZJsLnQq0n5Ua6aHalhJK4fMQkz1lqR7Osz1sA30trBtd9VUDc2NgkuRCn8+/pBwqZ+w=="],
+    "seroval-plugins": ["seroval-plugins@1.5.6", "", { "peerDependencies": { "seroval": "^1.0" } }, "sha512-HXuLAX2pu/UByPpaeo/TaMfvMIi+1QqIoPJYCcAtU8QkVNwgR6MPlGuCQTErV1JwraaMbYaWVIBX7mppzGLATQ=="],

    "serve-static": ["serve-static@2.2.1", "", { "dependencies": { "encodeurl": "^2.0.0", "escape-html": "^1.0.3", "parseurl": "^1.3.3", "send": "^1.2.0" } }, "sha512-xRXBn0pPqQTVQiC8wyQrKs2MOlX24zQ0POGaj0kultvoOCstBQM5yvOhAVSUwOMjQtTvsPWoNCHfPGwaaQJhTw=="],

@@ -5299,7 +5300,7 @@

    "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-js": ["solid-js@1.9.10", "", { "dependencies": { "csstype": "^3.1.0", "seroval": "~1.3.0", "seroval-plugins": "~1.3.0" } }, "sha512-Coz956cos/EPDlhs6+jsdTxKuJDPT7B5SVIWgABwROyxjY7Xbr8wkzD68Et+NxnV7DLJ3nJdAC2r9InuV/4Jew=="],
+    "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=="],

@@ -6307,10 +6308,18 @@

    "@solidjs/start/path-to-regexp": ["path-to-regexp@8.4.2", "", {}, "sha512-qRcuIdP69NPm4qbACK+aDogI5CBDMi1jKe0ry5rSQJz8JVLsC7jV8XpiJjGRLLol3N+R5ihGYcrPLTno6pAdBA=="],

+    "@solidjs/start/seroval": ["seroval@1.3.2", "", {}, "sha512-RbcPH1n5cfwKrru7v7+zrZvjLurgHhGyso3HTyGtRivGWgYjbOmGuivCQaORNELjNONoK35nj28EoWul9sb1zQ=="],
+
+    "@solidjs/start/seroval-plugins": ["seroval-plugins@1.3.3", "", { "peerDependencies": { "seroval": "^1.0" } }, "sha512-16OL3NnUBw8JG1jBLUoZJsLnQq0n5Ua6aHalhJK4fMQkz1lqR7Osz1sA30trBtd9VUDc2NgkuRCn8+/pBwqZ+w=="],
+
    "@solidjs/start/shiki": ["shiki@1.29.2", "", { "dependencies": { "@shikijs/core": "1.29.2", "@shikijs/engine-javascript": "1.29.2", "@shikijs/engine-oniguruma": "1.29.2", "@shikijs/langs": "1.29.2", "@shikijs/themes": "1.29.2", "@shikijs/types": "1.29.2", "@shikijs/vscode-textmate": "^10.0.1", "@types/hast": "^3.0.4" } }, "sha512-njXuliz/cP+67jU2hukkxCNuH1yUi4QfdZZY+sMr5PPrIyXSu5iTb/qYC4BiWWB0vZ+7TbdvYUCeL23zpwCfbg=="],

    "@solidjs/start/vite": ["vite@7.1.10", "", { "dependencies": { "esbuild": "^0.25.0", "fdir": "^6.5.0", "picomatch": "^4.0.3", "postcss": "^8.5.6", "rollup": "^4.43.0", "tinyglobby": "^0.2.15" }, "optionalDependencies": { "fsevents": "~2.3.3" }, "peerDependencies": { "@types/node": "^20.19.0 || >=22.12.0", "jiti": ">=1.21.0", "less": "^4.0.0", "lightningcss": "^1.21.0", "sass": "^1.70.0", "sass-embedded": "^1.70.0", "stylus": ">=0.54.8", "sugarss": "^5.0.0", "terser": "^5.16.0", "tsx": "^4.8.1", "yaml": "^2.4.2" }, "optionalPeers": ["@types/node", "jiti", "less", "lightningcss", "sass", "sass-embedded", "stylus", "sugarss", "terser", "tsx", "yaml"], "bin": { "vite": "bin/vite.js" } }, "sha512-CmuvUBzVJ/e3HGxhg6cYk88NGgTnBoOo7ogtfJJ0fefUWAxN/WDSUa50o+oVBxuIhO8FoEZW0j2eW7sfjs5EtA=="],

+    "@standard-community/standard-json/effect": ["effect@4.0.0-rc.111", "", { "dependencies": { "@standard-schema/spec": "^1.1.0", "fast-check": "^4.9.0", "msgpackr": "^2.0.5" } }, "sha512-ASd5L58EIR0CUNueZNKKjSsyOCd+2alxOAIaTcHaqkJkPsaYSsw5Cg/cfANk5K4Jr2YsX756xvX11shzqsreWA=="],
+
+    "@standard-community/standard-openapi/effect": ["effect@4.0.0-rc.111", "", { "dependencies": { "@standard-schema/spec": "^1.1.0", "fast-check": "^4.9.0", "msgpackr": "^2.0.5" } }, "sha512-ASd5L58EIR0CUNueZNKKjSsyOCd+2alxOAIaTcHaqkJkPsaYSsw5Cg/cfANk5K4Jr2YsX756xvX11shzqsreWA=="],
+
    "@tailwindcss/node/lightningcss": ["lightningcss@1.32.0", "", { "dependencies": { "detect-libc": "^2.0.3" }, "optionalDependencies": { "lightningcss-android-arm64": "1.32.0", "lightningcss-darwin-arm64": "1.32.0", "lightningcss-darwin-x64": "1.32.0", "lightningcss-freebsd-x64": "1.32.0", "lightningcss-linux-arm-gnueabihf": "1.32.0", "lightningcss-linux-arm64-gnu": "1.32.0", "lightningcss-linux-arm64-musl": "1.32.0", "lightningcss-linux-x64-gnu": "1.32.0", "lightningcss-linux-x64-musl": "1.32.0", "lightningcss-win32-arm64-msvc": "1.32.0", "lightningcss-win32-x64-msvc": "1.32.0" } }, "sha512-NXYBzinNrblfraPGyrbPoD19C1h9lfI/1mzgWYvXUTe414Gz/X1FD2XBZSZM7rRTrMA8JL3OtAaGifrIKhQ5yQ=="],

    "@tailwindcss/oxide-wasm32-wasi/@emnapi/core": ["@emnapi/core@1.11.3", "", { "dependencies": { "@emnapi/wasi-threads": "1.2.3", "tslib": "^2.4.0" }, "bundled": true }, "sha512-zLpS5asjEb7lq8jYLq37N6XKaE41DIexlY1rF/z4/tIl3wo13Sqm28fRyfIsKZD+NZ8mM5RoKkpW/rBcuoSZSg=="],
@@ -6891,6 +6900,8 @@

    "@brendonovich/vite-plugin-opencode/@opencode-ai/client/@opencode-ai/schema": ["@opencode-ai/schema@0.0.0-beta-18050", "", { "dependencies": { "@standard-schema/spec": "1.1.0", "effect": "4.0.0-rc.111" } }, "sha512-/D6VXaWlytTXR3IOiMLIKuPcfp7FQNUzRPm9z3K7UBFd1Bw4q/WZksaf5RVcBGz+0YRxYMc1V4D7MFlceSgtyg=="],

+    "@brendonovich/vite-plugin-opencode/@opencode-ai/client/effect": ["effect@4.0.0-rc.111", "", { "dependencies": { "@standard-schema/spec": "^1.1.0", "fast-check": "^4.9.0", "msgpackr": "^2.0.5" } }, "sha512-ASd5L58EIR0CUNueZNKKjSsyOCd+2alxOAIaTcHaqkJkPsaYSsw5Cg/cfANk5K4Jr2YsX756xvX11shzqsreWA=="],
+
    "@bruits/satteri-wasm32-wasi/@emnapi/core/@emnapi/wasi-threads": ["@emnapi/wasi-threads@1.2.2", "", { "dependencies": { "tslib": "^2.4.0" } }, "sha512-c95qOXkHdydNKhscBTebqEC1CVAZpyqOfVfBzQ1qgzyl3gfeldUjIggDbIZgDKsHLgnsM+igH7TJ/eAasaVuMA=="],

    "@electron/asar/minimatch/brace-expansion": ["brace-expansion@1.1.18", "", { "dependencies": { "balanced-match": "^1.0.0", "concat-map": "0.0.1" } }, "sha512-Edep/X9fGqVNmzKBVsDYIOtD+z1tuezV70LBjdCst9Tqu76lsnvRiZ6oTic1n+/BIwX6QDGAO94PN4N2SADvtw=="],
@@ -1,8 +1,8 @@
 {
  "nodeModules": {
-    "x86_64-linux": "sha256-vn4haT0uuD5tCbMD2F5IuDj+/XS/U/sVQ0bk5xJCbIM=",
-    "aarch64-linux": "sha256-XUGJ6hE7xvIyT53k5uCqaF/eGU4d4tIbuVuGrHU12Fg=",
-    "aarch64-darwin": "sha256-o3Os6rO8nGVVYs1aJBJAdYQaLbDsYNtGub4Gg1gHRQU=",
-    "x86_64-darwin": "sha256-YB00nH7WmgqRXvzA+zYEbaaik/vKheu3n/vJ6Ombx9k="
+    "x86_64-linux": "sha256-EtUp4pHl9TyPtRrLGvk/X7kd2LuIxNxCpUwF5aLtzN4=",
+    "aarch64-linux": "sha256-m0j/pMZCguclR3/T9JmzCfi11YzmIvBFyR2bVhIO37Y=",
+    "aarch64-darwin": "sha256-nqefk68ZTUfNU15q1WkXaGsFzPNwOjCtMHpp6WrpNqM=",
+    "x86_64-darwin": "sha256-syD7hX62E4yCDV/wux1QKw4q/zZr24f99Y2mmzMJo6o="
  }
 }
@@ -39,10 +39,10 @@
      "packages/stats/*"
    ],
    "catalog": {
-      "@effect/opentelemetry": "4.0.0-rc.111",
-      "@effect/platform-node": "4.0.0-rc.111",
-      "@effect/platform-node-shared": "4.0.0-rc.111",
-      "@effect/sql-sqlite-bun": "4.0.0-rc.111",
+      "@effect/opentelemetry": "4.0.0-rc.112",
+      "@effect/platform-node": "4.0.0-rc.112",
+      "@effect/platform-node-shared": "4.0.0-rc.112",
+      "@effect/sql-sqlite-bun": "4.0.0-rc.112",
      "@npmcli/arborist": "9.4.0",
      "@types/bun": "1.3.13",
      "@types/cross-spawn": "6.0.6",
@@ -76,7 +76,7 @@
      "dompurify": "3.4.14",
      "drizzle-kit": "1.0.0-rc.5-ab785fc",
      "drizzle-orm": "1.0.0-rc.5-169397b",
-      "effect": "4.0.0-rc.111",
+      "effect": "4.0.0-rc.112",
      "ai": "6.0.168",
      "cross-spawn": "7.0.6",
      "hono": "4.10.7",
@@ -101,11 +101,11 @@
      "tailwindcss": "4.1.11",
      "vite": "7.3.6",
      "@solidjs/meta": "0.29.4",
-      "@solidjs/router": "0.15.4",
+      "@solidjs/router": "1.0.0",
      "@solidjs/start": "https://pkg.pr.new/@solidjs/start@dfb2020",
      "@sentry/solid": "10.71.0",
      "@sentry/vite-plugin": "5.4.0",
-      "solid-js": "1.9.10",
+      "solid-js": "1.9.15",
      "solid-sonner": "0.3.1",
      "vite-plugin-solid": "2.11.10",
      "@lydell/node-pty": "1.2.0-beta.12"
@@ -161,7 +161,8 @@
    "@effect/platform-node-shared": "catalog:",
    "@types/bun": "catalog:",
    "@types/node": "catalog:",
-    "effect": "catalog:"
+    "effect": "catalog:",
+    "solid-js": "catalog:"
  },
  "patchedDependencies": {
    "@ai-sdk/openai-compatible@2.0.41": "patches/@ai-sdk%2Fopenai-compatible@2.0.41.patch",
@@ -169,7 +170,7 @@
    "@npmcli/agent@4.0.2": "patches/@npmcli%2Fagent@4.0.2.patch",
    "@silvia-odwyer/photon-node@0.3.4": "patches/@silvia-odwyer%2Fphoton-node@0.3.4.patch",
    "@standard-community/standard-openapi@0.2.9": "patches/@standard-community%2Fstandard-openapi@0.2.9.patch",
-    "solid-js@1.9.10": "patches/solid-js@1.9.10.patch",
+    "solid-js@1.9.15": "patches/solid-js@1.9.15.patch",
    "@ai-sdk/mistral@3.0.51": "patches/@ai-sdk%2Fmistral@3.0.51.patch",
    "gcp-metadata@8.1.2": "patches/gcp-metadata@8.1.2.patch",
    "pacote@21.5.0": "patches/pacote@21.5.0.patch",
@@ -214,22 +214,40 @@ the requests sent by code under test:
 import { Effect } from "effect"
 import { TestLLM } from "@opencode-ai/ai/testing"

-const testLLM = TestLLM.layer({
-  fallback: TestLLM.text("Hello from the test model", "text-1"),
-})
-
-// TestLLM.clientLayer provides LLMClient.Service and consumes TestLLM.Service.
 const programWithTestClient = Effect.gen(function* () {
+  const test = yield* TestLLM.Test
+  yield* test.push(TestLLM.text("Hello from the test model", "text-1"))
  const result = yield* program
-  const test = yield* TestLLM.Service
-  console.log(test.requests)
+  console.log(yield* test.requests())
  return result
-}).pipe(Effect.provide(TestLLM.clientLayer), Effect.provide(testLLM))
+}).pipe(Effect.provide(TestLLM.testLayer()))
 ```

-`TestLLM.push(...)` scripts one-shot responses, `TestLLM.always(...)` changes the fallback, and
-`TestLLM.wait(...)` lets concurrent tests wait until a request has arrived. Every received canonical request is
-available on the yielded `TestLLM.Service`.
+`testLayer()` provides the same object under `LLMClient.Service` and `TestLLM.Test`. Production consumes the
+normal client; tests use the additional controls. Each layer build has fresh state.
+
+- `test.push(...)` queues one-shot responses in execution order. Each argument is one response.
+- `test.always(response)` installs a repeatable fallback. The layer's `fallback` option sets its initial value.
+- `test.serve(request => response)` installs a request-dependent fallback. `always` and `serve` replace each
+  other without changing queued replies; queued replies take precedence.
+- `test.requests()` returns an array snapshot. `transformRequest` changes only the recorded observation;
+  `serve` receives the original canonical request.
+- `test.wait(count)` waits for request arrivals, not output or completion, and supports concurrent waiters.
+- `test.gate()` returns a scoped gate with countable `started` notifications and a `release` Effect. Release
+  unblocks all requests captured by that gate; closing its scope also releases it. Effect-aware test runners
+  already provide Scope.
+
+Constructing `stream()` or `generate()` does not record a request, invoke a responder, or consume a script.
+Each execution does. An exhausted queue without a fallback defects immediately rather than waiting for a
+future reply.
+
+Responses remain canonical event arrays or arbitrary `Stream<LLMEvent, AIError>` values. The client consumes
+supplied streams directly, preserving failure identity, finalizers, incomplete output, and post-finish tails;
+it does not repair or truncate them.
+
+The published legacy `Service`, `layer`, `clientLayer`, and module-level controls remain available as adapters
+over the same implementation, including the legacy live `requests` array. New tests should use `Test` and
+`testLayer`.

 ## Caching

@@ -496,6 +496,7 @@ const mapUsage = (usage: BedrockUsageSchema | undefined, providerMetadataKey: st
 interface ParserState {
  readonly providerMetadataKey: string
  readonly tools: ToolStream.State<number>
+  readonly finishedTools: ReadonlySet<number>
  // Bedrock splits the finish into `messageStop` (carries `stopReason`) and
  // `metadata` (carries usage). Hold the terminal event in state so `onHalt`
  // can emit exactly one finish after both chunks have had a chance to arrive.
@@ -574,6 +575,7 @@ const step = (state: ParserState, event: BedrockEvent) =>

    if (event.contentBlockDelta?.delta?.toolUse) {
      const index = event.contentBlockDelta.contentBlockIndex
+      if (state.finishedTools.has(index)) return [state, []] as const
      const result = ToolStream.appendExisting(
        ADAPTER,
        state.tools,
@@ -612,6 +614,7 @@ const step = (state: ParserState, event: BedrockEvent) =>
            state.hasToolCalls,
          lifecycle,
          tools: result.tools,
+          finishedTools: resultEvents.length > 0 ? new Set([...state.finishedTools, index]) : state.finishedTools,
          reasoningSignatures: Object.fromEntries(
            Object.entries(state.reasoningSignatures).filter(([key]) => key !== String(index)),
          ),
@@ -703,6 +706,7 @@ export const protocol = Protocol.make({
    initial: (request) => ({
      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
      tools: ToolStream.empty<number>(),
+      finishedTools: new Set<number>(),
      pendingFinish: undefined,
      hasToolCalls: false,
      lifecycle: Lifecycle.initial(),
@@ -391,17 +391,19 @@ export interface ParserState {
  readonly name: string
  readonly providerMetadataKey: string
  readonly tools: ToolStream.State<string>
+  // Call ids stay independent of item ids, which may be omitted or reused.
+  readonly completedTools: ReadonlySet<string>
  readonly hasFunctionCall: boolean
  readonly lifecycle: Lifecycle.State
  readonly outputItems: Readonly<Record<number, string>>
-  readonly messageItems: ReadonlySet<string>
-  readonly messagePhases: Readonly<Record<string, MessagePhase | null>>
+  readonly message: { readonly id: string; readonly phase: MessagePhase | null | undefined } | undefined
  readonly reasoningItems: Readonly<Record<string, ReasoningStreamItem>>
 }

 type ReasoningSummaryStatus = "active" | "can-conclude" | "concluded"

 interface ReasoningStreamItem {
+  readonly open: boolean
  readonly encryptedContent: string | null | undefined
  // Keyed by the wire protocol's numeric `summary_index`. JS object keys coerce to
  // strings, but typing the map as `Record<number, ...>` documents intent
@@ -826,16 +828,16 @@ const TERMINAL_TYPES = new Set(["error", "response.completed", "response.incompl
 export const terminal = (event: Event) => TERMINAL_TYPES.has(event.type)

 const onOutputTextDelta = (state: ParserState, event: Event, id: string): StepResult => {
-  if (!event.delta || !state.messageItems.has(id)) return [state, NO_EVENTS]
+  if (!event.delta || state.message?.id !== id) return [state, NO_EVENTS]
  const events: LLMEvent[] = []
-  const phase = state.messagePhases[id]
+  const phase = state.message.phase
  const metadata = providerMetadata(state, { itemId: id, ...(phase === undefined ? {} : { phase }) })
  const lifecycle = Lifecycle.textStart(state.lifecycle, events, id, metadata)
  return [{ ...state, lifecycle: Lifecycle.textDelta(lifecycle, events, id, event.delta) }, events]
 }

 const onOutputTextDone = (state: ParserState, event: Event, id: string): StepResult => {
-  if (state.messageItems.has(id)) {
+  if (state.message?.id === id) {
    if (state.lifecycle.text.has(id) || event.text === undefined) return [state, NO_EVENTS]
    return onOutputTextDelta(state, { ...event, delta: event.text }, id)
  }
@@ -846,18 +848,62 @@ const onOutputTextDone = (state: ParserState, event: Event, id: string): StepRes
 export const outputItemID = (state: ParserState, event: Event) =>
  event.output_index === undefined ? event.item_id : (state.outputItems[event.output_index] ?? event.item_id)

-export const onReasoningDelta = (state: ParserState, event: Event, itemID: string): StepResult => {
+const startReasoningSummaryPart = (state: ParserState, itemID: string, index: number): StepResult => {
  const item = state.reasoningItems[itemID]
-  if (!event.delta || !item) return [state, NO_EVENTS]
-  const index = event.summary_index ?? 0
+  if (!item?.open || index === 0 || item.summaryParts[index] !== undefined) return [state, NO_EVENTS]
+
  const events: LLMEvent[] = []
+  const lifecycle = Object.entries(item.summaryParts)
+    .filter((entry) => entry[1] !== "concluded")
+    .reduce(
+      (lifecycle, entry) =>
+        Lifecycle.reasoningEnd(lifecycle, events, `${itemID}:${entry[0]}`, providerMetadata(state, { itemId: itemID })),
+      state.lifecycle,
+    )
  return [
    {
      ...state,
-      lifecycle: Lifecycle.reasoningDelta(state.lifecycle, events, `${itemID}:${index}`, event.delta),
+      lifecycle: Lifecycle.reasoningStart(
+        lifecycle,
+        events,
+        `${itemID}:${index}`,
+        providerMetadata(state, { itemId: itemID, reasoningEncryptedContent: item.encryptedContent ?? null }),
+      ),
      reasoningItems: {
        ...state.reasoningItems,
-        [itemID]: { ...item, deltaIndexes: new Set([...item.deltaIndexes, index]) },
+        [itemID]: {
+          ...item,
+          summaryParts: {
+            ...Object.fromEntries(
+              Object.entries(item.summaryParts).map((entry) =>
+                entry[1] === "concluded" ? entry : [entry[0], "concluded" as const],
+              ),
+            ),
+            [index]: "active",
+          },
+        },
+      },
+    },
+    events,
+  ]
+}
+
+export const onReasoningDelta = (state: ParserState, event: Event, itemID: string): StepResult => {
+  const item = state.reasoningItems[itemID]
+  if (!event.delta || !item?.open) return [state, NO_EVENTS]
+  const index = event.summary_index ?? 0
+  if (item.summaryParts[index] === "concluded") return [state, NO_EVENTS]
+  const [started, emitted] = startReasoningSummaryPart(state, itemID, index)
+  const current = started.reasoningItems[itemID]
+  if (!current) return [started, emitted]
+  const events: LLMEvent[] = [...emitted]
+  return [
+    {
+      ...started,
+      lifecycle: Lifecycle.reasoningDelta(started.lifecycle, events, `${itemID}:${index}`, event.delta),
+      reasoningItems: {
+        ...started.reasoningItems,
+        [itemID]: { ...current, deltaIndexes: new Set([...current.deltaIndexes, index]) },
      },
    },
    events,
@@ -869,7 +915,7 @@ export const onReasoningDelta = (state: ParserState, event: Event, itemID: strin
 // as a single delta unless that summary index already streamed one.
 export const onReasoningDone = (state: ParserState, event: Event, itemID: string): StepResult => {
  const item = state.reasoningItems[itemID]
-  if (!item || typeof event.text !== "string") return [state, NO_EVENTS]
+  if (!item?.open || typeof event.text !== "string") return [state, NO_EVENTS]
  const index = event.summary_index ?? 0
  if (item.deltaIndexes.has(index)) return [state, NO_EVENTS]
  return onReasoningDelta(state, { ...event, delta: event.text }, itemID)
@@ -878,32 +924,48 @@ export const onReasoningDone = (state: ParserState, event: Event, itemID: string
 const reasoningMetadata = (state: ParserState, item: StreamItem & { id: string }) =>
  providerMetadata(state, { itemId: item.id, reasoningEncryptedContent: item.encrypted_content ?? null })

-// Responses APIs stream reasoning items in a stable order:
+// Responses APIs normally stream reasoning items in this order:
 //   `output_item.added` (reasoning) →
 //     `reasoning_summary_part.added` (index=0) →
 //     `reasoning_summary_text.delta` →
 //     `reasoning_summary_part.done` (index=0) →
 //     (repeat for index>0) →
 //   `output_item.done` (reasoning).
-// The handlers below rely on this ordering: `onOutputItemAdded` seeds the
-// per-item entry, `onReasoningSummaryPartAdded` for `summary_index === 0`
-// short-circuits when the entry already exists, and higher-index handlers
-// fold against the same entry. Behaviour for out-of-order events is
-// best-effort, not guaranteed.
+// `onOutputItemAdded` seeds the per-item entry, while each later part start is
+// also an implicit boundary for the previous part. This keeps the common event
+// lifecycle ordered when a compatible provider omits or delays a part-done event.
 const onOutputItemAdded = (state: ParserState, event: Event): StepResult => {
  const item = event.item
  if (item?.type === "message" && item.id !== undefined) {
+    const itemID = item.id
    const phase = messagePhase(item.phase)
+    // A new message closes earlier messages, including ones that never streamed.
+    const events: LLMEvent[] = []
+    const lifecycle = [...state.lifecycle.text]
+      .filter((id) => id !== itemID)
+      .reduce((lifecycle, id) => {
+        const openPhase = state.message?.id === id ? state.message.phase : undefined
+        return Lifecycle.textEnd(
+          lifecycle,
+          events,
+          id,
+          providerMetadata(state, { itemId: id, ...(openPhase === undefined ? {} : { phase: openPhase }) }),
+        )
+      }, state.lifecycle)
    return [
      {
        ...state,
-        messageItems: new Set([...state.messageItems, item.id]),
-        messagePhases: phase === undefined ? state.messagePhases : { ...state.messagePhases, [item.id]: phase },
+        lifecycle,
+        message: {
+          id: itemID,
+          phase: phase === undefined && state.message?.id === itemID ? state.message.phase : phase,
+        },
      },
-      NO_EVENTS,
+      events,
    ]
  }
  if (item && isReasoningItem(item)) {
+    if (state.reasoningItems[item.id] !== undefined) return [state, NO_EVENTS]
    const events: LLMEvent[] = []
    return [
      {
@@ -912,6 +974,7 @@ const onOutputItemAdded = (state: ParserState, event: Event): StepResult => {
        reasoningItems: {
          ...state.reasoningItems,
          [item.id]: {
+            open: true,
            encryptedContent: item.encrypted_content,
            summaryParts: { 0: "active" },
            deltaIndexes: new Set(),
@@ -923,6 +986,8 @@ const onOutputItemAdded = (state: ParserState, event: Event): StepResult => {
  }
  if (item?.type !== "function_call" || !item.call_id) return [state, NO_EVENTS]
  const id = item.id ?? item.call_id
+  if (Object.values(state.tools).some((tool) => tool?.id === item.call_id) || state.completedTools.has(item.call_id))
+    return [state, NO_EVENTS]
  const metadata = item.id !== undefined ? providerMetadata(state, { itemId: item.id }) : undefined
  const events: LLMEvent[] = []
  const lifecycle = Lifecycle.stepStart(state.lifecycle, events)
@@ -943,55 +1008,14 @@ const onOutputItemAdded = (state: ParserState, event: Event): StepResult => {

 const onReasoningSummaryPartAdded = (state: ParserState, event: Event): StepResult => {
  if (event.item_id === undefined || event.summary_index === undefined) return [state, NO_EVENTS]
-  const item = state.reasoningItems[event.item_id]
-  if (!item) return [state, NO_EVENTS]
-  if (event.summary_index === 0) return [state, NO_EVENTS]
-
-  const events: LLMEvent[] = []
-  const closed = Object.entries(item.summaryParts)
-    .filter((entry) => entry[1] === "can-conclude")
-    .reduce(
-      (lifecycle, entry) =>
-        Lifecycle.reasoningEnd(
-          lifecycle,
-          events,
-          `${event.item_id}:${entry[0]}`,
-          providerMetadata(state, { itemId: event.item_id }),
-        ),
-      state.lifecycle,
-    )
-  return [
-    {
-      ...state,
-      lifecycle: Lifecycle.reasoningStart(
-        closed,
-        events,
-        `${event.item_id}:${event.summary_index}`,
-        providerMetadata(state, { itemId: event.item_id, reasoningEncryptedContent: item.encryptedContent ?? null }),
-      ),
-      reasoningItems: {
-        ...state.reasoningItems,
-        [event.item_id]: {
-          ...item,
-          summaryParts: {
-            ...Object.fromEntries(
-              Object.entries(item.summaryParts).map((entry) =>
-                entry[1] === "can-conclude" ? [entry[0], "concluded" as const] : entry,
-              ),
-            ),
-            [event.summary_index]: "active",
-          },
-        },
-      },
-    },
-    events,
-  ]
+  return startReasoningSummaryPart(state, event.item_id, event.summary_index)
 }

 const onReasoningSummaryPartDone = (state: ParserState, event: Event): StepResult => {
  if (event.item_id === undefined || event.summary_index === undefined) return [state, NO_EVENTS]
  const item = state.reasoningItems[event.item_id]
-  if (!item) return [state, NO_EVENTS]
+  if (!item?.open) return [state, NO_EVENTS]
+  if (item.summaryParts[event.summary_index] !== "active") return [state, NO_EVENTS]
  return [
    {
      ...state,
@@ -1047,11 +1071,8 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (

  if (item.type === "message" && item.id !== undefined) {
    const itemPhase = messagePhase(item.phase)
-    const phase = itemPhase === undefined ? state.messagePhases[item.id] : itemPhase
+    const phase = itemPhase === undefined && state.message?.id === item.id ? state.message.phase : itemPhase
    const events: LLMEvent[] = []
-    const messageItems = new Set(state.messageItems)
-    messageItems.delete(item.id)
-    const { [item.id]: _phase, ...messagePhases } = state.messagePhases
    return [
      {
        ...state,
@@ -1061,8 +1082,7 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
          item.id,
          providerMetadata(state, { itemId: item.id, ...(phase === undefined ? {} : { phase }) }),
        ),
-        messageItems,
-        messagePhases,
+        message: state.message?.id === item.id ? undefined : state.message,
      },
      events,
    ] satisfies StepResult
@@ -1070,20 +1090,36 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (

  if (item.type === "function_call") {
    if (!item.call_id || !item.name) return [state, NO_EVENTS] satisfies StepResult
-    const id = item.id ?? item.call_id
-    const tools = state.tools[id]
-      ? state.tools
-      : ToolStream.start(state.tools, id, {
-          id: item.call_id,
-          name: item.name,
-          providerMetadata: item.id !== undefined ? providerMetadata(state, { itemId: item.id }) : undefined,
-        })
+    const callID = item.call_id
+    if (state.completedTools.has(callID)) return [state, NO_EVENTS] satisfies StepResult
+    const metadata = item.id !== undefined ? providerMetadata(state, { itemId: item.id }) : undefined
+    const fallback = item.id ?? callID
+    // Match the pending tool by call id so item events that disagree on
+    // whether `item.id` is present still resolve the same call.
+    const registered =
+      state.tools[fallback] !== undefined
+        ? fallback
+        : Object.keys(state.tools).find((key) => state.tools[key]?.id === callID)
+    const id = registered ?? fallback
+    const tools =
+      registered !== undefined
+        ? state.tools
+        : ToolStream.start(state.tools, id, {
+            id: callID,
+            name: item.name,
+            providerMetadata: metadata,
+          })
    const result =
      item.arguments === undefined
        ? yield* ToolStream.finish(state.id, tools, id)
        : yield* ToolStream.finishWithInput(state.id, tools, id, item.arguments)
    const events: LLMEvent[] = []
-    const resultEvents = result.events ?? []
+    const finished = result.events ?? []
+    // A done-only call never streamed a start event, so open its lifecycle here.
+    const resultEvents =
+      registered !== undefined || finished.length === 0
+        ? finished
+        : [LLMEvent.toolInputStart({ id: callID, name: item.name, providerMetadata: metadata }), ...finished]
    const lifecycle = resultEvents.length ? Lifecycle.stepStart(state.lifecycle, events) : state.lifecycle
    events.push(...resultEvents)
    return [
@@ -1094,6 +1130,7 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
          resultEvents.some((event) => LLMEvent.is.toolCall(event) || LLMEvent.is.toolInputError(event)) ||
          state.hasFunctionCall,
        tools: result.tools,
+        completedTools: new Set([...state.completedTools, callID]),
      },
      events,
    ] satisfies StepResult
@@ -1104,20 +1141,49 @@ const onOutputItemDone = Effect.fn("OpenResponses.onOutputItemDone")(function* (
    const metadata = reasoningMetadata(state, item)
    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 { [item.id]: _removed, ...reasoningItems } = state.reasoningItems
-      return [{ ...state, lifecycle, reasoningItems }, events] satisfies StepResult
+      return [
+        {
+          ...state,
+          lifecycle,
+          reasoningItems: {
+            ...state.reasoningItems,
+            [item.id]: {
+              ...reasoningItem,
+              open: false,
+              encryptedContent: item.encrypted_content ?? reasoningItem.encryptedContent,
+            },
+          },
+        },
+        events,
+      ] satisfies StepResult
    }
    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 }))
-      return [{ ...state, lifecycle }, events] satisfies StepResult
+      return [
+        {
+          ...state,
+          lifecycle,
+          reasoningItems: {
+            ...state.reasoningItems,
+            [item.id]: {
+              open: false,
+              encryptedContent: item.encrypted_content,
+              summaryParts: { 0: "concluded" },
+              deltaIndexes: new Set(),
+            },
+          },
+        },
+        events,
+      ] satisfies StepResult
    }
    return [
      { ...state, lifecycle: Lifecycle.reasoningEnd(state.lifecycle, events, item.id, metadata) },
@@ -1139,7 +1205,7 @@ const onResponseFinish = Effect.fn("OpenResponses.onResponseFinish")(function* (
            if (
              id === undefined ||
              ((item.type !== "function_call" || !current.tools[id]) &&
-                (item.type !== "reasoning" || !current.reasoningItems[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(
@@ -1256,6 +1322,13 @@ export const step = (state: ParserState, input: Event) => {
  if (event.type === "response.output_item.added") {
    if (event.item?.type === "message" && event.item.id === undefined)
      return ProviderShared.eventError(state.id, `${event.type} message is missing id`)
+    if (
+      event.item &&
+      isReasoningItem(event.item) &&
+      state.reasoningItems[event.item.id] === undefined &&
+      state.lifecycle.reasoning.size > 0
+    )
+      return ProviderShared.eventError(state.id, `${event.type} started reasoning before the previous item ended`)
    const id = event.item?.id ?? (event.item?.type === "function_call" ? event.item.call_id : undefined)
    return Effect.succeed(
      onOutputItemAdded(
@@ -1305,10 +1378,10 @@ export const initial = (request: LLMRequest, extension: Extension = BASE): Parse
  providerMetadataKey: request.model.route.providerMetadataKey ?? "openresponses",
  hasFunctionCall: false,
  tools: ToolStream.empty<string>(),
+  completedTools: new Set<string>(),
  lifecycle: Lifecycle.initial(),
  outputItems: {},
-  messageItems: new Set<string>(),
-  messagePhases: {},
+  message: undefined,
  reasoningItems: {},
 })

@@ -3,6 +3,7 @@ import { Tool } from "@opencode-ai/schema/tool"
 import { Route } from "../route/client.js"
 import { Auth } from "../route/auth.js"
 import { Endpoint } from "../route/endpoint.js"
+import { Framing } from "../route/framing.js"
 import { HttpTransport } from "../route/transport/index.js"
 import { Protocol } from "../route/protocol.js"
 import {
@@ -245,6 +246,8 @@ export const OpenAIChatEvent = Schema.StructWithRest(
  [Schema.Record(Schema.String, Schema.Unknown)],
 )
 export type OpenAIChatEvent = Schema.Schema.Type<typeof OpenAIChatEvent>
+const DONE = "[DONE]" as const
+const OpenAIChatStreamEvent = Schema.Union([Schema.Literal(DONE), Protocol.jsonEvent(OpenAIChatEvent)])
 type OpenAIChatRequestMessage = LLMRequest["messages"][number]

 interface PendingToolDelta {
@@ -997,33 +1000,12 @@ const step = (state: ParserState, event: OpenAIChatEvent) =>
      lifecycle = Lifecycle.reasoningStart(lifecycle, events, "reasoning-0", deltaMetadata)
    const reasoningEmitted = state.reasoningEmitted || lifecycle.reasoning.has("reasoning-0")

-    if (delta?.content) {
-      lifecycle = Lifecycle.reasoningEnd(
-        lifecycle,
-        events,
-        "reasoning-0",
-        reasoningMetadata(
-          state.providerMetadataKey,
-          reasoningField,
-          reasoningDetailsObserved ? state.reasoningDetails : undefined,
-        ),
-      )
-      lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.content)
-    }
+    // Reasoning is one response-wide channel: it stays open alongside text and
+    // refusal output so late reasoning deltas and details join the same block,
+    // and `finishEvents` closes it once with the complete metadata.
+    if (delta?.content) lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.content)

-    if (delta?.refusal) {
-      lifecycle = Lifecycle.reasoningEnd(
-        lifecycle,
-        events,
-        "reasoning-0",
-        reasoningMetadata(
-          state.providerMetadataKey,
-          reasoningField,
-          reasoningDetailsObserved ? state.reasoningDetails : undefined,
-        ),
-      )
-      lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.refusal)
-    }
+    if (delta?.refusal) lifecycle = Lifecycle.textDelta(lifecycle, events, "text-0", delta.refusal)

    // Compatible providers may omit indexes. Prefer durable identity, then use
    // batch position for parallel deltas or the latest call for sparse chunks.
@@ -1129,10 +1111,12 @@ const finishEvents = Effect.fn("OpenAIChat.finishEvents")(function* (state: Pars
          state.finishReason.normalized === "stop" && hasToolCalls ? "tool-calls" : state.finishReason.normalized,
      }
    : { normalized: hasToolCalls ? ("tool-calls" as const) : ("stop" as const) }
+  // Snapshot details at publish time so the emitted event never observes later
+  // mutation of the accumulated `reasoningDetails` array.
  const metadata = reasoningMetadata(
    state.providerMetadataKey,
    state.reasoningField,
-    state.reasoningDetailsObserved ? state.reasoningDetails : undefined,
+    state.reasoningDetailsObserved ? [...state.reasoningDetails] : undefined,
  )
  const started =
    state.reasoningDetailsObserved && !state.reasoningEmitted
@@ -1166,7 +1150,7 @@ export const protocol = Protocol.make({
    from: fromRequest,
  },
  stream: {
-    event: Protocol.jsonEvent(OpenAIChatEvent),
+    event: OpenAIChatStreamEvent,
    initial: (request) => ({
      providerMetadataKey: request.model.route.providerMetadataKey ?? String(request.model.provider),
      tools: ToolStream.empty<number>(),
@@ -1180,12 +1164,14 @@ export const protocol = Protocol.make({
      nextToolIndex: 0,
      requireFinishReason: request.model.compatibility?.requireFinishReason ?? true,
    }),
-    step,
+    step: (state: ParserState, event) => (event === DONE ? Effect.succeed([state, []] as const) : step(state, event)),
+    terminal: (event) => event === DONE,
    onHalt: finishEvents,
  },
 })

-export const httpTransport = HttpTransport.sseJson.with<OpenAIChatBody>()
+export const framing = Framing.sseWithDone
+export const httpTransport = HttpTransport.sseJson.with<OpenAIChatBody>().with({ framing })

 export const route = Route.make({
  id: ADAPTER,
@@ -1,6 +1,5 @@
 import { Route, type RouteRoutedLanguageModelInput } from "../route/client.js"
 import { Endpoint } from "../route/endpoint.js"
-import { Framing } from "../route/framing.js"
 import * as OpenAIChat from "./openai-chat.js"

 const ADAPTER = "openai-compatible-chat"
@@ -19,7 +18,7 @@ export const route = Route.make({
  providerMetadataKey: "openai",
  protocol: OpenAIChat.protocol,
  endpoint: Endpoint.path("/chat/completions"),
-  framing: Framing.sse,
+  framing: OpenAIChat.framing,
 })

 export * as OpenAICompatibleChat from "./openai-compatible-chat.js"
@@ -207,15 +207,16 @@ export const errorText = (error: unknown) => {

 /**
 * `framing` step for Server-Sent Events. Decodes UTF-8, runs the SSE channel
- * decoder, optionally filters named events, and drops empty / `[DONE]`
- * keep-alive events so the protocol event schema sees one JSON string per
- * element. Retry control events are ignored without interrupting the stream.
+ * decoder, optionally filters named events, and drops empty events. `[DONE]`
+ * is dropped by default or retained for protocols that use it as their stream
+ * boundary. Retry control events are ignored without interrupting the stream.
 * Decoder failures become provider output errors so the public error channel
 * stays `AIError`.
 */
 export const sseFraming = (
  bytes: Stream.Stream<Uint8Array, AIError>,
  events?: ReadonlySet<string>,
+  includeDone = false,
 ): Stream.Stream<string, AIError> =>
  bytes.pipe(
    Stream.decodeText(),
@@ -240,7 +241,7 @@ export const sseFraming = (
      (event) =>
        (events === undefined || events.has(event.event)) &&
        event.data.length > 0 &&
-        (event.data !== "[DONE]" || (events !== undefined && event.event !== "message")),
+        (event.data !== "[DONE]" || includeDone || (events !== undefined && event.event !== "message")),
    ),
    Stream.map((event) => event.data),
  )
@@ -62,22 +62,31 @@ export const reasoningEnd = (
  events: LLMEvent[],
  id: string,
  providerMetadata?: ProviderMetadata,
+  /** Authoritative complete value; replaces accumulated deltas when present. */
+  text?: string,
 ): State => {
  if (!state.reasoning.has(id)) return state
  const stepped = stepStart(state, events)
-  events.push(LLMEvent.reasoningEnd({ id, providerMetadata }))
+  events.push(LLMEvent.reasoningEnd({ id, text, providerMetadata }))
  const reasoning = new Set(stepped.reasoning)
  reasoning.delete(id)
  return { ...stepped, reasoning }
 }

-export const textEnd = (state: State, events: LLMEvent[], id: string, providerMetadata?: ProviderMetadata): State => {
+export const textEnd = (
+  state: State,
+  events: LLMEvent[],
+  id: string,
+  providerMetadata?: ProviderMetadata,
+  /** Authoritative complete value; replaces accumulated deltas when present. */
+  text?: string,
+): State => {
  if (!state.text.has(id)) return state
  const stepped = stepStart(state, events)
-  events.push(LLMEvent.textEnd({ id, providerMetadata }))
-  const text = new Set(stepped.text)
-  text.delete(id)
-  return { ...stepped, text }
+  events.push(LLMEvent.textEnd({ id, text, providerMetadata }))
+  const open = new Set(stepped.text)
+  open.delete(id)
+  return { ...stepped, text: open }
 }

 const closeOpenBlocks = (state: State, events: LLMEvent[]): State => {
@@ -37,6 +37,7 @@ const patterns = [
  /too large for model with \d+ maximum context length/i,
  /prompt has [\d,]+ tokens?, but the configured context size is [\d,]+ tokens?/i,
  /model_context_window_exceeded/i,
+  /range of input length should be/i,
  /too many tokens/i,
  /token limit exceeded/i,
  /request_too_large/i,
@@ -59,6 +60,7 @@ export const isContextOverflowFailure = (failure: unknown) =>

 const decodeJson = Schema.decodeUnknownOption(Schema.fromJsonString(Schema.Unknown))
 const QUOTA_CODES = new Set(["insufficient_quota", "usage_not_included", "billing_error"])
+const AUTH_CODES = new Set(["authentication_error", "permission_error"])
 const SERVER_CODES = new Set([
  "api_error",
  "internal_error",
@@ -74,7 +76,6 @@ 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 NETWORK_ERROR_TEXT = /network[-_\s]error/i

 export interface ProviderFailure {
  readonly message: string
@@ -90,8 +91,12 @@ export interface ProviderFailure {
  readonly rateLimit?: HttpRateLimitDetails | undefined
 }

-// Keep HTTP failures and provider-reported stream failures on one typed path so
-// session retry policy never needs provider-specific string matching.
+// Classification records affirmative evidence about a failure. Deterministic
+// failures need positive identification (a 4xx status, quota/auth/policy
+// signals); anything unrecognized stays UnknownProvider, which the session
+// retry policy treats as retry-eligible because transient failures arrive in
+// unpredictable shapes while deterministic rejections almost always carry a
+// status or known code.
 export function classifyProviderFailure(input: ProviderFailure): AIError["reason"] {
  const details = { message: input.message, body: input.rawBody, http: input.http, cause: input.cause }
  const body = input.rawBody ?? ""
@@ -116,46 +121,33 @@ export function classifyProviderFailure(input: ProviderFailure): AIError["reason
  if (CONTENT_POLICY_TEXT.test(text)) return new ContentPolicyError(details)
  if (codes.some((code) => QUOTA_CODES.has(code)) || (input.status === 429 && QUOTA_TEXT.test(text)))
    return new QuotaExceededError(details)
-  if (input.status === 401) return new AuthenticationError({ ...details, kind: "invalid" })
-  if (input.status === 403) return new AuthenticationError({ ...details, kind: "insufficient-permissions" })
-  if (codes.includes("authentication_error")) return new AuthenticationError({ ...details, kind: "invalid" })
-  if (codes.includes("permission_error"))
-    return new AuthenticationError({ ...details, kind: "insufficient-permissions" })
+  if (input.status === 401 || input.status === 403 || codes.some((code) => AUTH_CODES.has(code)))
+    return new AuthenticationError(details)
  if (
-    codes.some((code) => code.includes("rate_limit") || code === "too_many_requests" || code === "throttlingexception")
+    input.status === 429 ||
+    codes.some(
+      (code) => code.includes("rate_limit") || code === "too_many_requests" || code === "throttlingexception",
+    ) ||
+    RATE_LIMIT_TEXT.test(text)
  )
    return new RateLimitError({
      ...details,
      retryAfterMs: input.retryAfterMs,
      rateLimit: input.rateLimit,
    })
-  if (RATE_LIMIT_TEXT.test(text))
-    return new RateLimitError({
-      ...details,
-      retryAfterMs: input.retryAfterMs,
-      rateLimit: input.rateLimit,
-    })
-  if (NETWORK_ERROR_TEXT.test(text)) return new ProviderInternalError(details)
-  if (codes.some((code) => SERVER_CODES.has(code) || code.includes("exhausted") || code.includes("unavailable")))
-    return new ProviderInternalError({
-      ...details,
-      retryAfterMs: input.retryAfterMs,
-    })
-  if (input.status === 429) {
-    return new RateLimitError({
-      ...details,
-      retryAfterMs: input.retryAfterMs,
-      rateLimit: input.rateLimit,
-    })
-  }
-  if (input.status === 408 || input.status === 409 || (input.status !== undefined && input.status >= 500))
+  if (
+    input.status === 408 ||
+    input.status === 409 ||
+    (input.status !== undefined && input.status >= 500) ||
+    codes.some((code) => SERVER_CODES.has(code) || code.includes("exhausted") || code.includes("unavailable"))
+  )
    return new ProviderInternalError({
      ...details,
      retryAfterMs: input.retryAfterMs,
    })
  if (codes.some((code) => INVALID_REQUEST_CODES.has(code))) return new InvalidRequestError(details)
-  if (input.status === 400 || input.status === 404 || input.status === 413 || input.status === 422)
-    return new InvalidRequestError(details)
+  // Any remaining 4xx is a deterministic rejection of this request.
+  if (input.status !== undefined && input.status >= 400 && input.status < 500) return new InvalidRequestError(details)
  return new UnknownProviderError(details)
 }

@@ -5,7 +5,6 @@ import { ProviderShared } from "../protocols/shared.js"
 import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
 import { Route, type RouteDefaultsInput } from "../route/client.js"
 import { Endpoint } from "../route/endpoint.js"
-import { Framing } from "../route/framing.js"
 import { Protocol } from "../route/protocol.js"
 import { ProviderID, type ModelID, type LLMRequest } from "../schema/index.js"
 import { profiles } from "./openai-compatible-profile.js"
@@ -75,7 +74,7 @@ export const route = Route.make({
  providerMetadataKey: "openai",
  protocol,
  endpoint: Endpoint.path("/chat/completions", { baseURL: profiles.groq.baseURL }),
-  framing: Framing.sse,
+  framing: OpenAIChat.framing,
 })

 export const configure = (input: LanguageModelOptions = {}) => {
@@ -1,7 +1,6 @@
 import { Effect, Schema } from "effect"
 import { Route, type RouteDefaultsInput } from "../route/client.js"
 import { Endpoint } from "../route/endpoint.js"
-import { Framing } from "../route/framing.js"
 import { Protocol } from "../route/protocol.js"
 import { AuthOptions, type ProviderAuthOption } from "../route/auth-options.js"
 import { ProviderID, type CacheHint, type ModelID } from "../schema/index.js"
@@ -167,7 +166,7 @@ export const route = Route.make({
  providerMetadataKey: "openrouter",
  protocol,
  endpoint: Endpoint.path("/chat/completions", { baseURL: profile.baseURL }),
-  framing: Framing.sse,
+  framing: OpenAIChat.framing,
 })

 export const routes = [route]
@@ -139,7 +139,7 @@ const toAIError = (error: AuthError): AIError => {
    return new AIError({
      reason:
        error instanceof MissingCredentialError
-          ? new AuthenticationError({ message: error.message, cause: error, kind: "missing" })
+          ? new AuthenticationError({ message: error.message, cause: error })
          : new InvalidRequestError({ message: `Failed to resolve auth config: ${error.message}`, cause: error }),
    })
  }
@@ -8,8 +8,8 @@ import type { AIError } from "../schema/index.js"
 * `Framing` is the byte-stream-shaped seam between transport and protocol:
 *
 * - SSE (`Framing.sse`) — UTF-8 decode the body, run the SSE channel decoder,
- *   drop empty / `[DONE]` keep-alives. Each emitted frame is the JSON `data:`
- *   payload of one event.
+ *   and emit the `data:` payload of each non-empty event. The default drops
+ *   `[DONE]`; protocols that use it as a terminal select `sseWithDone`.
 * - AWS event stream — length-prefixed binary frames with CRC checksums.
 *   Each emitted frame is one parsed binary event record.
 *
@@ -26,6 +26,12 @@ export interface Definition<Frame> {
 /** Server-Sent Events framing. Used by every JSON-streaming HTTP provider. */
 export const sse: Definition<string> = { id: "sse", frame: ProviderShared.sseFraming }

+/** Server-Sent Events framing that retains the conventional `[DONE]` sentinel. */
+export const sseWithDone: Definition<string> = {
+  id: "sse",
+  frame: (bytes) => ProviderShared.sseFraming(bytes, undefined, true),
+}
+
 /** SSE framing restricted to protocol-recognized event names. */
 export const sseEvents = (events: ReadonlySet<string>): Definition<string> => ({
  id: "sse",
@@ -44,10 +44,7 @@ export class NoRouteError extends Schema.TaggedError<NoRouteError>("AI.Error.NoR

 export class AuthenticationError extends Schema.TaggedError<AuthenticationError>("AI.Error.Authentication")(
  "Authentication",
-  {
-    ...ReasonFields,
-    kind: Schema.Literals(["missing", "invalid", "expired", "insufficient-permissions", "unknown"]),
-  },
+  ReasonFields,
 ) {}

 export class RateLimitError extends Schema.TaggedError<RateLimitError>("AI.Error.RateLimit")("RateLimit", {
@@ -112,6 +112,8 @@ export type TextDelta = Schema.Schema.Type<typeof TextDelta>
 export const TextEnd = Schema.Struct({
  type: Schema.tag("text-end"),
  id: ContentBlockID,
+  /** Authoritative complete value; replaces accumulated deltas when present. */
+  text: Schema.optional(Schema.String),
  providerMetadata: Schema.optional(ProviderMetadata),
 }).annotate({ identifier: "LLM.Event.TextEnd" })
 export type TextEnd = Schema.Schema.Type<typeof TextEnd>
@@ -134,6 +136,8 @@ export type ReasoningDelta = Schema.Schema.Type<typeof ReasoningDelta>
 export const ReasoningEnd = Schema.Struct({
  type: Schema.tag("reasoning-end"),
  id: ContentBlockID,
+  /** Authoritative complete value; replaces accumulated deltas when present. */
+  text: Schema.optional(Schema.String),
  providerMetadata: Schema.optional(ProviderMetadata),
 }).annotate({ identifier: "LLM.Event.ReasoningEnd" })
 export type ReasoningEnd = Schema.Schema.Type<typeof ReasoningEnd>
@@ -328,17 +332,32 @@ 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 }>(
+  events: ReadonlyArray<LLMEvent>,
+  isDelta: (event: LLMEvent) => event is Extract<LLMEvent, Delta>,
+  isEnd: (event: LLMEvent) => event is Extract<LLMEvent, End>,
+) => {
+  const order: string[] = []
+  const parts = new Map<string, string>()
+  for (const event of events) {
+    if (isDelta(event)) {
+      if (!parts.has(event.id)) order.push(event.id)
+      parts.set(event.id, (parts.get(event.id) ?? "") + event.text)
+    }
+    if (isEnd(event) && event.text !== undefined) {
+      if (!parts.has(event.id)) order.push(event.id)
+      parts.set(event.id, event.text)
+    }
+  }
+  return order.map((id) => parts.get(id)).join("")
+}
+
 const responseText = (events: ReadonlyArray<LLMEvent>) =>
-  events
-    .filter(LLMEvent.is.textDelta)
-    .map((event) => event.text)
-    .join("")
+  joinFragments(events, LLMEvent.is.textDelta, LLMEvent.is.textEnd)

 const responseReasoning = (events: ReadonlyArray<LLMEvent>) =>
-  events
-    .filter(LLMEvent.is.reasoningDelta)
-    .map((event) => event.text)
-    .join("")
+  joinFragments(events, LLMEvent.is.reasoningDelta, LLMEvent.is.reasoningEnd)

 const responseUsage = (events: ReadonlyArray<LLMEvent>) =>
  events.reduce<Usage | undefined>(
@@ -445,10 +464,11 @@ const reduceTextDelta = (state: ResponseState, event: TextDelta): ResponseState
 const reduceTextEnd = (state: ResponseState, event: TextEnd): ResponseState => {
  const current = state.textParts[event.id]
  if (!current) return state
+  const text = event.text ?? current.text
  const providerMetadata = event.providerMetadata ?? current.providerMetadata
  return {
-    ...replaceContent(state, current.contentIndex, textContent(current.text, providerMetadata)),
-    textParts: { ...state.textParts, [event.id]: { ...current, providerMetadata } },
+    ...replaceContent(state, current.contentIndex, textContent(text, providerMetadata)),
+    textParts: { ...state.textParts, [event.id]: { ...current, text, providerMetadata } },
  }
 }

@@ -478,10 +498,11 @@ const reduceReasoningDelta = (state: ResponseState, event: ReasoningDelta): Resp
 const reduceReasoningEnd = (state: ResponseState, event: ReasoningEnd): ResponseState => {
  const current = state.reasoningParts[event.id]
  if (!current) return state
+  const text = event.text ?? current.text
  const providerMetadata = event.providerMetadata ?? current.providerMetadata
  return {
-    ...replaceContent(state, current.contentIndex, reasoningContent(current.text, providerMetadata)),
-    reasoningParts: { ...state.reasoningParts, [event.id]: { ...current, providerMetadata } },
+    ...replaceContent(state, current.contentIndex, reasoningContent(text, providerMetadata)),
+    reasoningParts: { ...state.reasoningParts, [event.id]: { ...current, text, providerMetadata } },
  }
 }

@@ -579,12 +600,12 @@ export class LLMResponse extends Schema.Class<LLMResponse>("LLM.Response")({
  usage: Schema.optional(Usage),
  finishReason: FinishReasonDetails,
 }) {
-  /** Concatenated assistant text assembled from streamed `text-delta` events. */
+  /** Concatenated assistant text; each fragment's `text-end` value replaces its accumulated deltas when present. */
  get text() {
    return responseText(this.events)
  }

-  /** Concatenated reasoning text assembled from streamed `reasoning-delta` events. */
+  /** Concatenated reasoning text; each fragment's `reasoning-end` value replaces its accumulated deltas when present. */
  get reasoning() {
    return responseReasoning(this.events)
  }
@@ -1,6 +1,6 @@
 export * as TestLLM from "./testing.js"

-import { LLMClient, type Interface as LLMClientShape } from "./route/client.js"
+import { LLMClient } from "./route/client.js"
 import {
  LLMEvent,
  LLMResponse,
@@ -16,13 +16,33 @@ export type Response = readonly LLMEvent[] | Stream.Stream<LLMEvent, AIError>

 export type Gate = Readonly<{ started: Effect.Effect<void>; release: Effect.Effect<void> }>

+type ClientInterface = Context.Service.Shape<typeof LLMClient.Service>
+
+export type Responder = (request: LLMRequest) => Response
+
+export interface TestInterface extends ClientInterface {
+  /** Returns a snapshot of requests observed at execution time. */
+  readonly requests: () => Effect.Effect<readonly LLMRequest[]>
+  readonly push: (...responses: readonly Response[]) => Effect.Effect<void>
+  /** Replaces the fallback without changing queued responses. */
+  readonly always: (response: Response) => Effect.Effect<void>
+  /** Answers requests after the one-shot queue is exhausted; receives the original request. */
+  readonly serve: (responder: Responder) => Effect.Effect<void>
+  /** Waits for request arrivals, not output or completion. */
+  readonly wait: (count: number) => Effect.Effect<void>
+  readonly gate: () => Effect.Effect<Gate, never, Scope.Scope>
+}
+
+export class Test extends Context.Service<Test, TestInterface>()("@opencode/ai/TestLLM/Test") {}
+
+/** @deprecated Use TestInterface through Test and testLayer. */
 export interface Interface {
  readonly requests: LLMRequest[]
  readonly push: (...responses: readonly Response[]) => Effect.Effect<void>
  readonly always: (response: Response) => Effect.Effect<void>
  readonly wait: (count: number) => Effect.Effect<void>
  readonly gate: Effect.Effect<Gate, never, Scope.Scope>
-  readonly client: LLMClientShape
+  readonly client: ClientInterface
 }

 export interface LayerOptions {
@@ -31,6 +51,7 @@ export interface LayerOptions {
  readonly fallback?: Response
 }

+/** @deprecated Use Test and testLayer for normal client methods and test controls. */
 export class Service extends Context.Service<Service, Interface>()("@opencode/ai/TestLLM") {}

 export const complete = (
@@ -80,59 +101,64 @@ export const hangAfter = (...events: readonly LLMEvent[]) => Stream.concat(Strea

 const toStream = (response: Response) => (Stream.isStream(response) ? response : Stream.fromIterable(response))

-export const layer = (options: LayerOptions = {}) =>
-  Layer.effect(
-    Service,
-    Effect.gen(function* () {
-      const requests: LLMRequest[] = []
-      const responses: Response[] = []
-      let started = Deferred.makeUnsafe<void>()
-      let fallback = options.fallback
-      let activeGate: { readonly started: Queue.Queue<void>; readonly release: Latch.Latch } | undefined
-      const wait = (count: number): Effect.Effect<void> =>
-        Effect.suspend(() =>
-          requests.length >= count ? Effect.void : Deferred.await(started).pipe(Effect.andThen(wait(count))),
-        )
+const make = (options: LayerOptions) =>
+  Effect.sync(() => {
+    const requests: LLMRequest[] = []
+    const responses: Response[] = []
+    let started = Deferred.makeUnsafe<void>()
+    let fallback: Response | Responder | undefined = options.fallback
+    let activeGate: { readonly started: Queue.Queue<void>; readonly release: Latch.Latch } | undefined
+    const wait = (count: number): Effect.Effect<void> =>
+      Effect.suspend(() =>
+        requests.length >= count ? Effect.void : Deferred.await(started).pipe(Effect.andThen(wait(count))),
+      )

-      const stream = ((request: LLMRequest) => {
-        requests.push(options.transformRequest?.(request) ?? request)
+    const stream: ClientInterface["stream"] = (request) =>
+      Stream.suspend(() => {
+        const count = requests.push(options.transformRequest?.(request) ?? request)
        const waiting = started
        started = Deferred.makeUnsafe()
-        Deferred.doneUnsafe(waiting, Effect.void)
-        const response = responses.shift() ?? fallback
-        if (!response) return Stream.die(new Error(`TestLLM has no response for request ${requests.length}`))
-        const streamed = toStream(response)
        const gate = activeGate
-        if (!gate) return streamed
-        return Stream.unwrap(
-          Queue.offer(gate.started, undefined).pipe(Effect.andThen(gate.release.await), Effect.as(streamed)),
-        )
-      }) as LLMClientShape["stream"]
-      const client = LLMClient.Service.of({
-        stream,
-        generate: (request) =>
-          stream(request).pipe(
-            Stream.runFold(LLMResponse.empty, LLMResponse.reduce),
-            Effect.flatMap((state) => {
-              const response = LLMResponse.complete(state)
-              if (response) return Effect.succeed(response)
-              return Effect.die("TestLLM response ended without a terminal finish event")
-            }),
-          ),
+        try {
+          const response = responses.shift() ?? (typeof fallback === "function" ? fallback(request) : fallback)
+          if (!response) return Stream.die(new Error(`TestLLM has no response for request ${count}`))
+          const streamed = toStream(response)
+          if (!gate) return streamed
+          return Stream.unwrap(
+            Queue.offer(gate.started, undefined).pipe(Effect.andThen(gate.release.await), Effect.as(streamed)),
+          )
+        } finally {
+          // Waiters can resume synchronously; assign the reply and gate before notifying them.
+          Deferred.doneUnsafe(waiting, Effect.void)
+        }
      })
-
-      return Service.of({
-        requests,
-        push: (...input) =>
-          Effect.sync(() => {
-            responses.push(...input)
+    const test = Test.of({
+      stream,
+      generate: (request) =>
+        stream(request).pipe(
+          Stream.runFold(LLMResponse.empty, LLMResponse.reduce),
+          Effect.flatMap((state) => {
+            const response = LLMResponse.complete(state)
+            if (response) return Effect.succeed(response)
+            return Effect.die("TestLLM response ended without a terminal finish event")
          }),
-        always: (response) =>
-          Effect.sync(() => {
-            fallback = response
-          }),
-        wait,
-        gate: Effect.gen(function* () {
+        ),
+      requests: () => Effect.sync(() => [...requests]),
+      push: (...input) =>
+        Effect.sync(() => {
+          responses.push(...input)
+        }),
+      always: (response) =>
+        Effect.sync(() => {
+          fallback = response
+        }),
+      serve: (responder) =>
+        Effect.sync(() => {
+          fallback = responder
+        }),
+      wait,
+      gate: () =>
+        Effect.gen(function* () {
          const gate = {
            started: yield* Effect.acquireRelease(Queue.unbounded<void>(), Queue.shutdown),
            release: yield* Latch.make(),
@@ -147,11 +173,36 @@ export const layer = (options: LayerOptions = {}) =>
            release,
          }
        }),
-        client,
-      })
-    }),
+    })
+
+    return { test, requests }
+  })
+
+/** Provides one shared implementation under the normal client and test-control tags. */
+export const testLayer = (options: LayerOptions = {}) =>
+  Layer.effectContext(
+    Effect.map(make(options), (implementation) =>
+      Context.make(LLMClient.Service, implementation.test).pipe(Context.add(Test, implementation.test)),
+    ),
  )

+/** @deprecated Use testLayer; retained for published callers of the legacy control interface. */
+export const layer = (options: LayerOptions = {}) =>
+  Layer.effect(
+    Service,
+    Effect.map(make(options), (implementation) =>
+      Service.of({
+        requests: implementation.requests,
+        push: implementation.test.push,
+        always: implementation.test.always,
+        wait: implementation.test.wait,
+        gate: implementation.test.gate(),
+        client: implementation.test,
+      }),
+    ),
+  )
+
+/** @deprecated testLayer provides LLMClient.Service directly. */
 export const clientLayer = Layer.effect(
  LLMClient.Service,
  Effect.map(Service, (service) => service.client),
@@ -574,6 +574,7 @@ describe("WebSocket channel execution", () => {
  const model = configure({ baseURL: "https://api.openai.test/v1/", apiKey: "test" }).responses("gpt-4.1-mini")
  const request = LLM.request({ model, prompt: "Say hello." })
  const frames = [
+    JSON.stringify({ type: "response.output_item.added", item: { type: "message", id: "msg_1" } }),
    JSON.stringify({ type: "response.output_text.delta", item_id: "msg_1", delta: "Hi" }),
    JSON.stringify({ type: "response.completed", response: { id: "resp_1" } }),
  ]
@@ -32,6 +32,8 @@ describe("public exports", () => {
    expect(Provider.make).toBeFunction()
    expect(ProviderSubpath.make).toBe(Provider.make)
    expect(TestLLM.layer).toBeFunction()
+    expect(TestLLM.testLayer).toBeFunction()
+    expect(TestLLM.Test.of).toBeFunction()
  })

  test("route barrel exposes route-authoring APIs", () => {
@@ -10,7 +10,7 @@
      "transport": "http",
      "request": {
        "method": "POST",
-        "url": "https://api.anthropic.com/v1/messages",
+        "url": "https://api.anthropic.com/v1/messages?beta=true",
        "headers": {
          "anthropic-version": "2023-06-01",
          "content-type": "application/json"
@@ -29,7 +29,7 @@
      "transport": "http",
      "request": {
        "method": "POST",
-        "url": "https://api.anthropic.com/v1/messages",
+        "url": "https://api.anthropic.com/v1/messages?beta=true",
        "headers": {
          "anthropic-version": "2023-06-01",
          "content-type": "application/json"
@@ -10,7 +10,7 @@
      "transport": "http",
      "request": {
        "method": "POST",
-        "url": "https://api.anthropic.com/v1/messages",
+        "url": "https://api.anthropic.com/v1/messages?beta=true",
        "headers": {
          "anthropic-version": "2023-06-01",
          "content-type": "application/json"
@@ -29,7 +29,7 @@
      "transport": "http",
      "request": {
        "method": "POST",
-        "url": "https://api.anthropic.com/v1/messages",
+        "url": "https://api.anthropic.com/v1/messages?beta=true",
        "headers": {
          "anthropic-version": "2023-06-01",
          "content-type": "application/json"
@@ -24,7 +24,7 @@
      "transport": "http",
      "request": {
        "method": "POST",
-        "url": "https://api.anthropic.com/v1/messages",
+        "url": "https://api.anthropic.com/v1/messages?beta=true",
        "headers": {
          "anthropic-version": "2023-06-01",
          "content-type": "application/json"
@@ -23,7 +23,7 @@
        "headers": {
          "content-type": "application/json"
        },
-        "body": "{\"model\":\"anthropic/claude-sonnet-4.6\",\"messages\":[{\"role\":\"system\",\"content\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":\"What is the weather in Paris?\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false}}}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":1536,\"reasoning\":{\"max_tokens\":1024}}"
+        "body": "{\"model\":\"anthropic/claude-sonnet-4.6\",\"messages\":[{\"role\":\"system\",\"content\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":\"What is the weather in Paris?\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"reasoning\":{\"max_tokens\":1024},\"max_completion_tokens\":1536,\"store\":false,\"usage\":{\"include\":true}}"
      },
      "response": {
        "status": 200,
@@ -41,7 +41,7 @@
        "headers": {
          "content-type": "application/json"
        },
-        "body": "{\"model\":\"anthropic/claude-sonnet-4.6\",\"messages\":[{\"role\":\"system\",\"content\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":\"What is the weather in Paris?\"},{\"role\":\"assistant\",\"content\":\"Sure! Let me check the weather in Paris for you right now!\",\"tool_calls\":[{\"id\":\"toolu_01PaChhcyw3yu2P2bDS2bgAA\",\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"}}],\"reasoning\":\"The user wants to know the weather in Paris. I'll use the get_weather tool.\",\"reasoning_details\":[{\"type\":\"reasoning.text\",\"text\":\"The user wants to know the weather in Paris. I'll use the get_weather tool.\",\"format\":\"anthropic-claude-v1\",\"index\":0,\"signature\":\"ErkCCosBCA8YAipAjKnRKpxkZ4eHrMPJ63IWEOYPSzb+XSHyG+vLK+2ks2O9T4N9M37Xn2kausQSH1rfsrdmKxgUlBg6yUFRgMVR7DIRY2xhdWRlLXNvbm5ldC00LTY4AEIIdGhpbmtpbmdaJDRjMGYwNDZmLTI1ZmQtNDVmYi1iZmIzLWEwOGE4ZTI0OWNhNxIMNFb5O6pb4nX0HojdGgyL5h+CAIpsxpdM1QgiMGm/i3ST6F5mAhxB+Uez0Cm95ra9yvQkrzHaA/AmWoXpdmPlczSn1S1RDk2IqeA57Spbf7JT44jygtLQt6yZmGzoTBHn3VkwaNZsuuAtbdo4B5QJXooa/AoKKs54QZ2kfS640vsv5flQVCg7CoQCFuLKjIeLMO7MnxVyuskXJr1DgesTa7I0ScF53U9JGhgB\"}]},{\"role\":\"tool\",\"tool_call_id\":\"toolu_01PaChhcyw3yu2P2bDS2bgAA\",\"content\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false}}}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":1536,\"reasoning\":{\"max_tokens\":1024}}"
+        "body": "{\"model\":\"anthropic/claude-sonnet-4.6\",\"messages\":[{\"role\":\"system\",\"content\":\"Use the get_weather tool exactly once. After the tool result, reply exactly: Paris is sunny.\"},{\"role\":\"user\",\"content\":\"What is the weather in Paris?\"},{\"role\":\"assistant\",\"content\":\"Sure! Let me check the weather in Paris for you right now!\",\"tool_calls\":[{\"id\":\"toolu_01PaChhcyw3yu2P2bDS2bgAA\",\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"arguments\":\"{\\\"city\\\":\\\"Paris\\\"}\"}}],\"reasoning\":\"The user wants to know the weather in Paris. I'll use the get_weather tool.\",\"reasoning_details\":[{\"type\":\"reasoning.text\",\"text\":\"The user wants to know the weather in Paris. I'll use the get_weather tool.\",\"format\":\"anthropic-claude-v1\",\"index\":0,\"signature\":\"ErkCCosBCA8YAipAjKnRKpxkZ4eHrMPJ63IWEOYPSzb+XSHyG+vLK+2ks2O9T4N9M37Xn2kausQSH1rfsrdmKxgUlBg6yUFRgMVR7DIRY2xhdWRlLXNvbm5ldC00LTY4AEIIdGhpbmtpbmdaJDRjMGYwNDZmLTI1ZmQtNDVmYi1iZmIzLWEwOGE4ZTI0OWNhNxIMNFb5O6pb4nX0HojdGgyL5h+CAIpsxpdM1QgiMGm/i3ST6F5mAhxB+Uez0Cm95ra9yvQkrzHaA/AmWoXpdmPlczSn1S1RDk2IqeA57Spbf7JT44jygtLQt6yZmGzoTBHn3VkwaNZsuuAtbdo4B5QJXooa/AoKKs54QZ2kfS640vsv5flQVCg7CoQCFuLKjIeLMO7MnxVyuskXJr1DgesTa7I0ScF53U9JGhgB\"}]},{\"role\":\"tool\",\"tool_call_id\":\"toolu_01PaChhcyw3yu2P2bDS2bgAA\",\"content\":\"{\\\"temperature\\\":22,\\\"condition\\\":\\\"sunny\\\"}\"}],\"tools\":[{\"type\":\"function\",\"function\":{\"name\":\"get_weather\",\"description\":\"Get current weather for a city.\",\"parameters\":{\"type\":\"object\",\"properties\":{\"city\":{\"type\":\"string\"}},\"required\":[\"city\"],\"additionalProperties\":false},\"strict\":false}}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"reasoning\":{\"max_tokens\":1024},\"max_completion_tokens\":1536,\"store\":false,\"usage\":{\"include\":true}}"
      },
      "response": {
        "status": 200,
@@ -15,7 +15,7 @@
        "headers": {
          "content-type": "application/json"
        },
-        "body": "{\"model\":\"anthropic/claude-sonnet-4.6\",\"messages\":[{\"role\":\"system\",\"content\":\"Think through the arithmetic, then reply with only the final integer.\"},{\"role\":\"user\",\"content\":\"What is 173 multiplied by 219?\"}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"max_tokens\":1536,\"temperature\":0,\"reasoning\":{\"max_tokens\":1024}}"
+        "body": "{\"model\":\"anthropic/claude-sonnet-4.6\",\"messages\":[{\"role\":\"system\",\"content\":\"Think through the arithmetic, then reply with only the final integer.\"},{\"role\":\"user\",\"content\":\"What is 173 multiplied by 219?\"}],\"stream\":true,\"stream_options\":{\"include_usage\":true},\"temperature\":0,\"reasoning\":{\"max_tokens\":1024},\"max_completion_tokens\":1536,\"store\":false,\"usage\":{\"include\":true}}"
      },
      "response": {
        "status": 200,
@@ -10,11 +10,11 @@
      "transport": "http",
      "request": {
        "method": "POST",
-        "url": "https://api.anthropic.com/v1/messages",
+        "url": "https://api.anthropic.com/v1/messages?beta=true",
        "headers": {
          "content-type": "application/json"
        },
-        "body": "{\"model\":\"claude-haiku-4-5-20251001\",\"system\":[{\"type\":\"text\",\"text\":\"Read the PDF returned by the tool and follow the user's response format exactly.\"}],\"messages\":[{\"role\":\"user\",\"content\":[{\"type\":\"text\",\"text\":\"Return only the verification code from the PDF.\"}]},{\"role\":\"assistant\",\"content\":[{\"type\":\"tool_use\",\"id\":\"call_pdf_1\",\"name\":\"read_pdf\",\"input\":{}}]},{\"role\":\"user\",\"content\":[{\"type\":\"tool_result\",\"tool_use_id\":\"call_pdf_1\",\"content\":[{\"type\":\"text\",\"text\":\"PDF read successfully\"},{\"type\":\"document\",\"source\":{\"type\":\"base64\",\"media_type\":\"application/pdf\",\"data\":\"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\"}}]}]}],\"tools\":[{\"name\":\"read_pdf\",\"description\":\"Read the attached PDF.\",\"input_schema\":{\"type\":\"object\",\"properties\":{},\"additionalProperties\":false}}],\"stream\":true,\"max_tokens\":40,\"temperature\":0}"
+        "body": "{\"model\":\"claude-haiku-4-5-20251001\",\"system\":[{\"type\":\"text\",\"text\":\"Read the PDF returned by the tool and follow the user's response format exactly.\"}],\"messages\":[{\"role\":\"user\",\"content\":[{\"type\":\"text\",\"text\":\"Return only the verification code from the PDF.\"}]},{\"role\":\"assistant\",\"content\":[{\"type\":\"tool_use\",\"id\":\"call_pdf_1\",\"name\":\"read_pdf\",\"input\":{}}]},{\"role\":\"user\",\"content\":[{\"type\":\"tool_result\",\"tool_use_id\":\"call_pdf_1\",\"content\":[{\"type\":\"text\",\"text\":\"PDF read successfully\"},{\"type\":\"document\",\"source\":{\"type\":\"base64\",\"media_type\":\"application/pdf\",\"data\":\"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\"},\"title\":\"verification.pdf\"}]}]}],\"tools\":[{\"name\":\"read_pdf\",\"description\":\"Read the attached PDF.\",\"input_schema\":{\"type\":\"object\",\"properties\":{},\"additionalProperties\":false}}],\"stream\":true,\"max_tokens\":40,\"temperature\":0}"
      },
      "response": {
        "status": 200,
@@ -10,11 +10,11 @@
      "transport": "http",
      "request": {
        "method": "POST",
-        "url": "https://api.anthropic.com/v1/messages",
+        "url": "https://api.anthropic.com/v1/messages?beta=true",
        "headers": {
          "content-type": "application/json"
        },
-        "body": "{\"model\":\"claude-haiku-4-5-20251001\",\"messages\":[{\"role\":\"user\",\"content\":[{\"type\":\"document\",\"source\":{\"type\":\"base64\",\"media_type\":\"application/pdf\",\"data\":\"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\"}},{\"type\":\"text\",\"text\":\"Return only the verification code from the PDF.\"}]}],\"stream\":true,\"max_tokens\":40,\"temperature\":0}"
+        "body": "{\"model\":\"claude-haiku-4-5-20251001\",\"messages\":[{\"role\":\"user\",\"content\":[{\"type\":\"document\",\"source\":{\"type\":\"base64\",\"media_type\":\"application/pdf\",\"data\":\"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\"},\"title\":\"verification.pdf\"},{\"type\":\"text\",\"text\":\"Return only the verification code from the PDF.\"}]}],\"stream\":true,\"max_tokens\":40,\"temperature\":0}"
      },
      "response": {
        "status": 200,
@@ -11,6 +11,7 @@ describe("provider error classification", () => {
      "Input length 131393 exceeds the maximum allowed input length of 131040 tokens.",
      "The input (516368 tokens) is longer than the model's context length (262144 tokens).",
      "Prompt has 5,958,968 tokens, but the configured context size is 256,000 tokens",
+      "Range of input length should be [1, 129024]",
      "Too many tokens",
      "Token limit exceeded",
    ]
@@ -87,10 +88,12 @@ describe("provider error classification", () => {
    ])
  })

-  test("classifies network error text as provider internal", () => {
+  test("classifies any remaining 4xx status as an invalid request", () => {
    expect(
-      ["network error", "network-error", "network_error"].map((message) => classifyProviderFailure({ message })._tag),
-    ).toEqual(["ProviderInternal", "ProviderInternal", "ProviderInternal"])
+      [400, 402, 404, 418, 422, 451].map(
+        (status) => classifyProviderFailure({ message: `HTTP ${status}`, status })._tag,
+      ),
+    ).toEqual(Array(6).fill("InvalidRequest"))
  })

  test("classifies nested provider codes when a top-level code is also present", () => {
@@ -103,10 +106,12 @@ describe("provider error classification", () => {
    ).toEqual(["QuotaExceeded", "ProviderInternal", "InvalidRequest"])
  })

-  test("keeps unknown and malformed provider payloads non-retryable", () => {
+  test("leaves unrecognized failures unclassified for the retry default", () => {
    expect(classifyProviderFailure({ message: '{"error":{"message":"no_kv_space"}}' })._tag).toBe("UnknownProvider")
    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")
  })
 })

@@ -1,5 +1,7 @@
 import { describe, expect, test } from "bun:test"
 import { model } from "@opencode-ai/ai/providers/openai"
+import { LLM } from "../src/index.js"
+import { Endpoint } from "../src/route/endpoint.js"

 describe("provider package entrypoints", () => {
  test("semantic API aliases expose the same contract", async () => {
@@ -36,7 +38,8 @@ describe("provider package entrypoints", () => {
    expect(modules[0].model).toBe(modules[1].model)
    expect(modules[8].model).toBe(modules[9].model)
    expect(modules[12].model).toBe(modules[13].model)
-    expect(modules[19].model).toBe(modules[20].model)
+    expect(modules[19].model).toBe(modules[21].model)
+    expect(modules[19].model).not.toBe(modules[20].model)
  })

  test("maps DeepInfra package settings onto its native executable model", async () => {
@@ -139,8 +142,10 @@ describe("provider package entrypoints", () => {
    expect(selected.route.id).toBe("anthropic-messages")
    expect(selected.route.endpoint).toMatchObject({
      baseURL: "https://messages.example.test/v1",
-      path: "/messages",
    })
+    expect(
+      Endpoint.render(selected.route.endpoint, { request: LLM.request({ model: selected }), body: {} }).toString(),
+    ).toBe("https://messages.example.test/v1/messages")
    expect(selected.route.defaults.headers).toEqual({ "x-application": "opencode" })
    expect(selected.route.defaults.http?.body).toEqual({ metadata: { user_id: "user_1" } })
    expect(selected.route.defaults.providerOptions).toEqual({ effort: "low" })
@@ -491,6 +491,59 @@ describe("Bedrock Converse route", () => {
    }),
  )

+  it.effect("ignores late tool deltas after contentBlockStop", () =>
+    Effect.gen(function* () {
+      const body = eventStreamBody(
+        [
+          "contentBlockStart",
+          {
+            contentBlockIndex: 0,
+            start: { toolUse: { toolUseId: "tool_1", name: "lookup" } },
+          },
+        ],
+        ["contentBlockDelta", { contentBlockIndex: 0, delta: { toolUse: { input: '{"query":"weather"}' } } }],
+        ["contentBlockStop", { contentBlockIndex: 0 }],
+        ["contentBlockDelta", { contentBlockIndex: 0, delta: { toolUse: { input: '{"late":true}' } } }],
+        ["messageStop", { stopReason: "tool_use" }],
+      )
+      const response = yield* LLMClient.generate(baseRequest).pipe(Effect.provide(fixedBytes(body)))
+
+      expect(response.toolCalls).toEqual([
+        { type: "tool-call", id: "tool_1", name: "lookup", input: { query: "weather" } },
+      ])
+      expect(response.events.filter((event) => event.type === "tool-input-delta")).toEqual([
+        {
+          type: "tool-input-delta",
+          id: "tool_1",
+          name: "lookup",
+          text: '{"query":"weather"}',
+          input: { query: "weather" },
+        },
+      ])
+    }),
+  )
+
+  it.effect("rejects tool deltas without contentBlockStart", () =>
+    Effect.gen(function* () {
+      const error = yield* LLMClient.generate(baseRequest).pipe(
+        Effect.provide(
+          fixedBytes(
+            eventStreamBody(
+              ["contentBlockDelta", { contentBlockIndex: 0, delta: { toolUse: { input: "{}" } } }],
+              ["messageStop", { stopReason: "tool_use" }],
+            ),
+          ),
+        ),
+        Effect.flip,
+      )
+
+      expect(error).toMatchObject({
+        reason: { _tag: "InvalidProviderOutput" },
+        message: "Bedrock Converse tool delta is missing its tool call",
+      })
+    }),
+  )
+
  it.effect("recovers incomplete tool input at finalization", () =>
    Effect.gen(function* () {
      const body = eventStreamBody(
@@ -52,7 +52,7 @@ describe("provider error retention", () => {
          Effect.flip,
        )
        expect(error.message).toContain("Slow down")
-        expect(error.reason._tag).toBe(entry.name === "Gemini" ? "ProviderInternal" : "RateLimit")
+        expect(error.reason._tag).toBe("RateLimit")
        expect(error.reason.body).toBe(body)
        expect(error.reason.http).toMatchObject({ status: 200, headers: { "x-provider-trace": "trace-1" } })
        expect(error.reason.http?.url).toStartWith("https://provider.test/")
@@ -0,0 +1,467 @@
+import { describe, expect } from "bun:test"
+import { Effect, Stream } 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 collect = (...input: OpenResponses.Event[]) =>
+  Effect.gen(function* () {
+    const events = yield* LLMClient.stream(request).pipe(
+      Stream.runCollect,
+      Effect.provide(fixedResponse(sseEvents(...input))),
+    )
+    expectLifecycle(
+      events,
+      input.some((event) => event.type === "response.completed"),
+    )
+    return events
+  })
+
+// Deliberately local to these basic-item fixtures, not a general stream validator.
+function expectLifecycle(events: ReadonlyArray<LLMEvent>, completed: boolean) {
+  const active = { text: new Set<string>(), reasoning: new Set<string>() }
+  const tools = new Map<string, "started" | "ended" | "called">()
+  events.forEach((event) => {
+    if (event.type === "text-start" || event.type === "reasoning-start") {
+      const blocks = event.type === "text-start" ? active.text : active.reasoning
+      expect(blocks.size).toBe(0)
+      blocks.add(event.id)
+    }
+    if (event.type === "text-delta" || event.type === "reasoning-delta") {
+      expect((event.type === "text-delta" ? active.text : active.reasoning).has(event.id)).toBe(true)
+    }
+    if (event.type === "text-end" || event.type === "reasoning-end") {
+      expect((event.type === "text-end" ? active.text : active.reasoning).delete(event.id)).toBe(true)
+    }
+    if (event.type === "tool-input-start") {
+      expect(tools.has(event.id)).toBe(false)
+      tools.set(event.id, "started")
+    }
+    if (event.type === "tool-input-delta") expect(tools.get(event.id)).toBe("started")
+    if (event.type === "tool-input-end") {
+      expect(tools.get(event.id)).toBe("started")
+      tools.set(event.id, "ended")
+    }
+    if (event.type === "tool-call") {
+      expect(tools.get(event.id)).toBe("ended")
+      tools.set(event.id, "called")
+    }
+    // Incomplete responses may leave pending tool inputs without a call.
+    if (event.type === "finish" && completed) {
+      expect(active.text.size).toBe(0)
+      expect(active.reasoning.size).toBe(0)
+      expect([...tools.values()].every((status) => status === "called")).toBe(true)
+    }
+  })
+  expect(events.filter(LLMEvent.is.stepStart)).toHaveLength(1)
+  expect(events[0]?.type).toBe("step-start")
+  expect(events.filter(LLMEvent.is.stepFinish)).toHaveLength(1)
+  expect(events.filter(LLMEvent.is.finish)).toHaveLength(1)
+  expect(events.slice(-2).map((event) => event.type)).toEqual(["step-finish", "finish"])
+}
+
+describe("Open Responses basic-item lifecycles", () => {
+  it.effect("closes implicit summary boundaries and ignores late events for completed reasoning", () =>
+    Effect.gen(function* () {
+      const item = { type: "reasoning", id: "rs_1", encrypted_content: "encrypted-state" }
+      const events = yield* collect(
+        { type: "response.output_item.added", output_index: 0, item: { ...item, encrypted_content: null } },
+        { type: "response.output_item.added", item: { ...item, encrypted_content: null } },
+        { 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_part.done", item_id: "rs_1", summary_index: 0 },
+        { type: "response.reasoning_summary_text.done", item_id: "rs_1", summary_index: 1, text: "Second" },
+        // The third part omits both explicit summary boundaries.
+        {
+          type: "response.reasoning_summary_text.delta",
+          output_index: 0,
+          item_id: "wrong",
+          summary_index: 2,
+          delta: "Third",
+        },
+        { type: "response.output_item.done", item },
+        { type: "response.output_item.done", item },
+        { type: "response.output_item.added", item },
+        { type: "response.reasoning_summary_part.added", item_id: "rs_1", summary_index: 3 },
+        { type: "response.reasoning_summary_text.delta", item_id: "rs_1", summary_index: 3, delta: "late" },
+        { type: "response.reasoning_summary_text.done", item_id: "rs_1", summary_index: 2, text: "late final" },
+        { type: "response.reasoning_summary_part.done", item_id: "rs_1", summary_index: 3 },
+        completed,
+      )
+
+      expect(events.filter((event) => event.type.startsWith("reasoning-"))).toEqual([
+        {
+          type: "reasoning-start",
+          id: "rs_1:0",
+          providerMetadata: { "openai-compatible": { itemId: "rs_1", reasoningEncryptedContent: null } },
+        },
+        { type: "reasoning-delta", id: "rs_1:0", text: "First" },
+        { type: "reasoning-end", id: "rs_1:0", providerMetadata: { "openai-compatible": { itemId: "rs_1" } } },
+        {
+          type: "reasoning-start",
+          id: "rs_1:1",
+          providerMetadata: { "openai-compatible": { itemId: "rs_1", reasoningEncryptedContent: null } },
+        },
+        { type: "reasoning-delta", id: "rs_1:1", text: "Second" },
+        { type: "reasoning-end", id: "rs_1:1", providerMetadata: { "openai-compatible": { itemId: "rs_1" } } },
+        {
+          type: "reasoning-start",
+          id: "rs_1:2",
+          providerMetadata: { "openai-compatible": { itemId: "rs_1", reasoningEncryptedContent: null } },
+        },
+        { type: "reasoning-delta", id: "rs_1:2", text: "Third" },
+        {
+          type: "reasoning-end",
+          id: "rs_1:2",
+          providerMetadata: { "openai-compatible": { itemId: "rs_1", reasoningEncryptedContent: "encrypted-state" } },
+        },
+      ])
+    }),
+  )
+
+  it.effect("preserves done-only encrypted reasoning without replaying its summary or late events", () =>
+    Effect.gen(function* () {
+      const item = {
+        type: "reasoning",
+        id: "rs_1",
+        encrypted_content: "encrypted-state",
+        summary: [{ type: "summary_text", text: "Not streamed" }],
+      }
+      const events = yield* collect(
+        { type: "response.output_item.done", item },
+        { type: "response.output_item.done", item },
+        { type: "response.output_item.added", item },
+        { type: "response.reasoning_summary_text.delta", item_id: "rs_1", delta: "late" },
+        { type: "response.reasoning_summary_part.added", item_id: "rs_1", summary_index: 1 },
+        { type: "response.reasoning_summary_text.done", item_id: "rs_1", summary_index: 1, text: "late final" },
+        completed,
+        // Route termination must also prevent events after response completion.
+        { type: "response.output_item.added", item: { type: "reasoning", id: "rs_after" } },
+      )
+      expect(events.filter((event) => event.type.startsWith("reasoning-"))).toEqual([
+        {
+          type: "reasoning-start",
+          id: "rs_1",
+          providerMetadata: { "openai-compatible": { itemId: "rs_1", reasoningEncryptedContent: "encrypted-state" } },
+        },
+        {
+          type: "reasoning-end",
+          id: "rs_1",
+          providerMetadata: { "openai-compatible": { itemId: "rs_1", reasoningEncryptedContent: "encrypted-state" } },
+        },
+      ])
+    }),
+  )
+
+  it.effect("forgets never-streamed messages at implicit boundaries and preserves refusal phases", () =>
+    Effect.gen(function* () {
+      const events = yield* collect(
+        { type: "response.output_item.added", item: { type: "message", id: "msg_empty" } },
+        { type: "response.output_item.added", item: { type: "message", id: "msg_1", phase: "commentary" } },
+        { type: "response.output_text.done", item_id: "msg_1", text: "Checking" },
+        { type: "response.output_text.done", item_id: "msg_1", text: "Duplicate" },
+        { type: "response.output_item.added", item: { type: "message", id: "msg_2", phase: null } },
+        { type: "response.output_text.delta", item_id: "msg_empty", delta: "stale" },
+        { type: "response.output_text.done", item_id: "msg_empty", text: "stale final" },
+        { type: "response.output_text.delta", item_id: "msg_1", delta: "late" },
+        { type: "response.refusal.delta", item_id: "msg_2", delta: "Cannot help." },
+        { type: "response.refusal.done", item_id: "msg_2", refusal: "Cannot help." },
+        { type: "response.output_item.done", item: { type: "message", id: "msg_2", phase: "final_answer" } },
+        { type: "response.output_item.added", item: { type: "message", id: "msg_3", phase: null } },
+        { type: "response.refusal.done", item_id: "msg_3", refusal: "Done-only refusal." },
+        { type: "response.output_item.done", item: { type: "message", id: "msg_3" } },
+        completed,
+      )
+      expect(events.filter((event) => event.type.startsWith("text-"))).toEqual([
+        {
+          type: "text-start",
+          id: "msg_1",
+          providerMetadata: { "openai-compatible": { itemId: "msg_1", phase: "commentary" } },
+        },
+        { type: "text-delta", id: "msg_1", text: "Checking" },
+        {
+          type: "text-end",
+          id: "msg_1",
+          providerMetadata: { "openai-compatible": { itemId: "msg_1", phase: "commentary" } },
+        },
+        {
+          type: "text-start",
+          id: "msg_2",
+          providerMetadata: { "openai-compatible": { itemId: "msg_2", phase: null } },
+        },
+        { type: "text-delta", id: "msg_2", text: "Cannot help." },
+        {
+          type: "text-end",
+          id: "msg_2",
+          providerMetadata: { "openai-compatible": { itemId: "msg_2", phase: "final_answer" } },
+        },
+        {
+          type: "text-start",
+          id: "msg_3",
+          providerMetadata: { "openai-compatible": { itemId: "msg_3", phase: null } },
+        },
+        { type: "text-delta", id: "msg_3", text: "Done-only refusal." },
+        { type: "text-end", id: "msg_3", providerMetadata: { "openai-compatible": { itemId: "msg_3", phase: null } } },
+      ])
+    }),
+  )
+  it.effect("allows a message to be registered again without inheriting its previous phase", () =>
+    Effect.gen(function* () {
+      const events = yield* collect(
+        { type: "response.output_item.added", item: { type: "message", id: "msg_1", phase: "commentary" } },
+        { type: "response.output_text.delta", item_id: "msg_1", delta: "First" },
+        { type: "response.output_item.done", item: { type: "message", id: "msg_1" } },
+        { type: "response.output_item.added", item: { type: "message", id: "msg_1" } },
+        { type: "response.output_text.delta", item_id: "msg_1", delta: "Second" },
+        { type: "response.output_item.done", item: { type: "message", id: "msg_1" } },
+        completed,
+      )
+      expect(events.filter(LLMEvent.is.textEnd)).toEqual([
+        {
+          type: "text-end",
+          id: "msg_1",
+          providerMetadata: { "openai-compatible": { itemId: "msg_1", phase: "commentary" } },
+        },
+        { type: "text-end", id: "msg_1", providerMetadata: { "openai-compatible": { itemId: "msg_1" } } },
+      ])
+      expect(events.filter(LLMEvent.is.textDelta).map((event) => event.text)).toEqual(["First", "Second"])
+    }),
+  )
+  ;[undefined, "fc_1"].forEach((id) => {
+    it.effect(`opens and closes a done-only tool ${id === undefined ? "without" : "with"} an item id`, () =>
+      Effect.gen(function* () {
+        const item = {
+          type: "function_call",
+          ...(id === undefined ? {} : { id }),
+          call_id: "call_1",
+          name: "lookup",
+          arguments: '{"query":"weather"}',
+        }
+        const events = yield* collect(
+          { type: "response.output_item.done", item },
+          { type: "response.output_item.done", item: { ...item, id: "fc_1" } },
+          { type: "response.output_item.added", item },
+          completed,
+        )
+        const providerMetadata = id === undefined ? undefined : { "openai-compatible": { itemId: id } }
+        expect(events.filter((event) => event.type.startsWith("tool-"))).toEqual([
+          { type: "tool-input-start", id: "call_1", name: "lookup", providerMetadata },
+          { type: "tool-input-end", id: "call_1", name: "lookup", providerMetadata },
+          { type: "tool-call", id: "call_1", name: "lookup", input: { query: "weather" }, providerMetadata },
+        ])
+        expect(events.filter(LLMEvent.is.finish)).toEqual([
+          {
+            type: "finish",
+            reason: { normalized: "tool-calls", raw: undefined },
+            providerMetadata: { "openai-compatible": { responseId: "resp_1", serviceTier: undefined } },
+          },
+        ])
+      }),
+    )
+
+    it.effect(`deduplicates a pending call whose item id is ${id === undefined ? "introduced" : "omitted"} later`, () =>
+      Effect.gen(function* () {
+        const item = { type: "function_call", call_id: "call_1", name: "lookup" }
+        const first = { ...item, ...(id === undefined ? {} : { id }) }
+        const duplicate = { ...item, ...(id === undefined ? { id: "fc_1" } : {}) }
+        const events = yield* collect(
+          { type: "response.output_item.added", item: first },
+          { type: "response.function_call_arguments.delta", item_id: id ?? "call_1", delta: '{"query":"weather"}' },
+          { type: "response.output_item.added", item: duplicate },
+          { type: "response.output_item.done", item: duplicate },
+          { type: "response.output_item.done", item: first },
+          { type: "response.output_item.added", item: duplicate },
+          completed,
+        )
+        // Identity metadata comes from the first admission, not the duplicate.
+        const providerMetadata = id === undefined ? undefined : { "openai-compatible": { itemId: id } }
+        expect(events.filter((event) => event.type.startsWith("tool-"))).toEqual([
+          { type: "tool-input-start", id: "call_1", name: "lookup", providerMetadata },
+          {
+            type: "tool-input-delta",
+            id: "call_1",
+            name: "lookup",
+            text: '{"query":"weather"}',
+            input: { query: "weather" },
+          },
+          { type: "tool-input-end", id: "call_1", name: "lookup", providerMetadata },
+          { type: "tool-call", id: "call_1", name: "lookup", input: { query: "weather" }, providerMetadata },
+        ])
+      }),
+    )
+  })
+
+  it.effect("recovers pending items in completed output order with terminal encrypted metadata", () =>
+    Effect.gen(function* () {
+      const events = yield* collect(
+        {
+          type: "response.output_item.added",
+          item: { type: "function_call", id: "fc_1", call_id: "call_1", name: "lookup" },
+        },
+        { type: "response.function_call_arguments.delta", item_id: "fc_1", delta: '{"query":"draft"}' },
+        { type: "response.output_item.added", item: { type: "reasoning", id: "rs_1", encrypted_content: null } },
+        { type: "response.reasoning_summary_text.delta", item_id: "rs_1", delta: "Thinking" },
+        { type: "response.reasoning_summary_part.done", item_id: "rs_1", summary_index: 0 },
+        {
+          type: "response.completed",
+          response: {
+            id: "resp_1",
+            output: [
+              { type: "reasoning", id: "rs_1", encrypted_content: "terminal-state" },
+              { type: "function_call", id: "fc_1", call_id: "call_1", name: "lookup", arguments: '{"query":"final"}' },
+              { type: "function_call", id: "fc_unseen", call_id: "call_unseen", name: "lookup", arguments: "{}" },
+            ],
+          },
+        },
+      )
+      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",
+          name: "lookup",
+          providerMetadata: { "openai-compatible": { itemId: "fc_1" } },
+        },
+        {
+          type: "tool-call",
+          id: "call_1",
+          name: "lookup",
+          input: { query: "final" },
+          providerMetadata: { "openai-compatible": { itemId: "fc_1" } },
+        },
+      ])
+    }),
+  )
+
+  it.effect("preserves call identity and pending order when an item id is reused", () =>
+    Effect.gen(function* () {
+      const first = { type: "function_call", id: "fc_1", call_id: "call_1", name: "lookup", arguments: "{}" }
+      const events = yield* collect(
+        { type: "response.output_item.added", item: first },
+        { type: "response.output_item.added", item: { ...first, id: "fc_2", call_id: "call_2" } },
+        { type: "response.output_item.done", item: first },
+        { type: "response.output_item.added", item: { ...first, call_id: "call_3" } },
+        { type: "response.output_item.done", item: first },
+        completed,
+      )
+      expect(events.filter(LLMEvent.is.toolCall).map((event) => event.id)).toEqual(["call_1", "call_2", "call_3"])
+    }),
+  )
+
+  it.effect("keeps text and reasoning identities separate even with empty item ids", () =>
+    Effect.gen(function* () {
+      const events = yield* collect(
+        { type: "response.output_item.added", item: { type: "reasoning", id: "" } },
+        { type: "response.output_item.added", item: { type: "message", id: "" } },
+        { type: "response.output_item.added", item: { type: "reasoning", id: "" } },
+        { type: "response.reasoning_summary_text.delta", item_id: "", delta: "Thinking" },
+        { type: "response.output_text.delta", item_id: "", delta: "Answer" },
+        { type: "response.output_item.done", item: { type: "reasoning", id: "", encrypted_content: "state" } },
+        { type: "response.output_item.done", item: { type: "message", id: "" } },
+        completed,
+      )
+      expect(events.filter(LLMEvent.is.reasoningDelta).map((event) => event.text)).toEqual(["Thinking"])
+      expect(events.filter(LLMEvent.is.textDelta).map((event) => event.text)).toEqual(["Answer"])
+      expect(events.filter(LLMEvent.is.reasoningEnd)).toEqual([
+        {
+          type: "reasoning-end",
+          id: ":0",
+          providerMetadata: { "openai-compatible": { itemId: "", reasoningEncryptedContent: "state" } },
+        },
+      ])
+    }),
+  )
+
+  it.effect("does not recover a completed tool from a tracked message with the same id", () =>
+    Effect.gen(function* () {
+      const events = yield* collect(
+        { type: "response.output_item.added", item: { type: "message", id: "item_1" } },
+        { type: "response.output_text.delta", item_id: "item_1", delta: "Answer" },
+        {
+          type: "response.completed",
+          response: {
+            id: "resp_1",
+            output: [{ type: "function_call", id: "item_1", call_id: "call_1", name: "lookup", arguments: "{}" }],
+          },
+        },
+      )
+      expect(events.filter(LLMEvent.is.toolCall)).toEqual([])
+      expect(events.filter(LLMEvent.is.finish).map((event) => event.reason.normalized)).toEqual(["stop"])
+    }),
+  )
+
+  it.effect("flushes pending calls and open text when completed output is absent", () =>
+    Effect.gen(function* () {
+      const events = yield* collect(
+        { type: "response.output_item.added", item: { type: "message", id: "msg_1", phase: "final_answer" } },
+        { type: "response.output_text.delta", item_id: "msg_1", delta: "Answer" },
+        {
+          type: "response.output_item.added",
+          item: { type: "function_call", call_id: "call_1", name: "lookup", arguments: "{}" },
+        },
+        completed,
+      )
+      // Generic terminal closure does not repeat the message's phase metadata.
+      expect(events.slice(4, -2)).toEqual([
+        { type: "tool-input-end", id: "call_1", name: "lookup" },
+        { type: "tool-call", id: "call_1", name: "lookup", input: {} },
+        { type: "text-end", id: "msg_1" },
+      ])
+    }),
+  )
+
+  it.effect("does not reconcile pending calls or terminal reasoning metadata on incomplete responses", () =>
+    Effect.gen(function* () {
+      const events = yield* collect(
+        { type: "response.output_item.added", item: { type: "reasoning", id: "rs_1", encrypted_content: null } },
+        { type: "response.reasoning_summary_text.delta", item_id: "rs_1", delta: "Partial" },
+        {
+          type: "response.output_item.added",
+          item: { type: "function_call", id: "fc_1", call_id: "call_1", name: "lookup" },
+        },
+        { type: "response.function_call_arguments.delta", item_id: "fc_1", delta: '{"query":' },
+        {
+          type: "response.incomplete",
+          response: {
+            id: "resp_1",
+            incomplete_details: { reason: "max_output_tokens" },
+            output: [
+              { type: "reasoning", id: "rs_1", encrypted_content: "not-reconciled" },
+              {
+                type: "function_call",
+                id: "fc_1",
+                call_id: "call_1",
+                name: "lookup",
+                arguments: '{"query":"not-reconciled"}',
+              },
+            ],
+          },
+        },
+      )
+      expect(events.filter(LLMEvent.is.toolInputEnd)).toEqual([])
+      expect(events.filter(LLMEvent.is.toolCall)).toEqual([])
+      expect(events.filter(LLMEvent.is.reasoningEnd)).toEqual([{ type: "reasoning-end", id: "rs_1:0" }])
+      expect(events.filter(LLMEvent.is.finish)).toEqual([
+        {
+          type: "finish",
+          reason: { normalized: "length", raw: "max_output_tokens" },
+          providerMetadata: { "openai-compatible": { responseId: "resp_1", serviceTier: undefined } },
+        },
+      ])
+    }),
+  )
+})
@@ -7,6 +7,7 @@ import {
  AIError,
  LLMEvent,
  LLMRequest,
+  LLMResponse,
  Message,
  LanguageModel,
  ToolCallPart,
@@ -807,6 +808,28 @@ describe("OpenAI Chat route", () => {
    }),
  )

+  it.effect("finishes at the done sentinel without waiting for response EOF", () =>
+    Effect.gen(function* () {
+      const stream = new ReadableStream<Uint8Array>({
+        start(controller) {
+          controller.enqueue(
+            new TextEncoder().encode(sseEvents(deltaChunk({ content: "Hello" }), deltaChunk({}, "stop"))),
+          )
+        },
+      })
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(stream, {
+            headers: { "content-type": "text/event-stream" },
+          }),
+        ),
+      )
+
+      expect(response.text).toBe("Hello")
+      expect(response.events.at(-1)?.type).toBe("finish")
+    }),
+  )
+
  it.effect("preserves streamed refusals as ordinary assistant text", () =>
    Effect.gen(function* () {
      const response = yield* LLMClient.generate(request).pipe(
@@ -1127,7 +1150,7 @@ describe("OpenAI Chat route", () => {
    }),
  )

-  it.effect("preserves scalar reasoning after content starts", () =>
+  it.effect("preserves scalar reasoning after content starts in one lifecycle", () =>
    Effect.gen(function* () {
      const details = [{ type: "reasoning.text", text: "detail", format: "unknown", index: 0 }]
      const response = yield* LLMClient.generate(request).pipe(
@@ -1144,8 +1167,35 @@ describe("OpenAI Chat route", () => {
      )

      expect(response.reasoning).toBe("detailscalar")
-      expect(response.events.filter(LLMEvent.is.reasoningStart)).toHaveLength(2)
-      expect(response.events.filter(LLMEvent.is.reasoningEnd)).toHaveLength(2)
+      expect(response.events.filter(LLMEvent.is.reasoningStart)).toHaveLength(1)
+      expect(response.events.filter(LLMEvent.is.reasoningEnd)).toHaveLength(1)
+      expect(response.message.content.filter((part) => part.type === "reasoning")).toHaveLength(1)
+      expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
+        openai: { reasoningField: "reasoning", reasoningDetails: details },
+      })
+    }),
+  )
+
+  it.effect("keeps one reasoning lifecycle across many content chunks", () =>
+    Effect.gen(function* () {
+      const details = [{ type: "reasoning.text", text: "thinking", format: "anthropic-claude-v1", index: 0 }]
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { choices: [{ delta: { reasoning: "thinking", reasoning_details: details } }] },
+              ...Array.from({ length: 25 }, (_, index) => deltaChunk({ content: `chunk-${index} ` })),
+              deltaChunk({}, "stop"),
+            ),
+          ),
+        ),
+      )
+
+      expect(response.reasoning).toBe("thinking")
+      expect(response.text).toBe(Array.from({ length: 25 }, (_, index) => `chunk-${index} `).join(""))
+      expect(response.events.filter(LLMEvent.is.reasoningStart)).toHaveLength(1)
+      expect(response.events.filter(LLMEvent.is.reasoningEnd)).toHaveLength(1)
+      expect(response.message.content.filter((part) => part.type === "reasoning")).toHaveLength(1)
      expect(response.message.content.find((part) => part.type === "reasoning")?.providerMetadata).toEqual({
        openai: { reasoningField: "reasoning", reasoningDetails: details },
      })
@@ -1191,7 +1241,18 @@ describe("OpenAI Chat route", () => {
          index: 0,
        },
      ]
-      const response = yield* LLMClient.generate(request).pipe(
+      // Snapshot reasoning-end metadata as each event is published so the
+      // assertion cannot pass through later mutation of a shared array.
+      const publishedEndMetadata: unknown[] = []
+      const response = yield* LLMClient.stream(request).pipe(
+        Stream.tap((event) =>
+          Effect.sync(() => {
+            if (LLMEvent.is.reasoningEnd(event))
+              publishedEndMetadata.push(decodeJson(encodeJson(event.providerMetadata)))
+          }),
+        ),
+        Stream.runFold(LLMResponse.empty, LLMResponse.reduce),
+        Effect.map((state) => LLMResponse.complete(state)!),
        Effect.provide(
          fixedResponse(
            sseEvents(
@@ -1212,10 +1273,12 @@ describe("OpenAI Chat route", () => {
      expect(response.events.filter(LLMEvent.is.reasoningStart)).toHaveLength(1)
      expect(response.events.filter(LLMEvent.is.reasoningDelta)).toHaveLength(1)
      expect(response.events.filter(LLMEvent.is.reasoningEnd)).toHaveLength(1)
-      expect(response.events.filter(LLMEvent.is.reasoningEnd).at(-1)?.providerMetadata).toEqual({
-        openai: { reasoningField: "reasoning", reasoningDetails: merged },
-      })
-      expect(response.events.findIndex(LLMEvent.is.reasoningEnd)).toBeLessThan(
+      expect(publishedEndMetadata).toEqual([{ openai: { reasoningField: "reasoning", reasoningDetails: merged } }])
+      expect(response.events.findIndex(LLMEvent.is.reasoningStart)).toBeLessThan(
+        response.events.findIndex(LLMEvent.is.textStart),
+      )
+      // Reasoning stays open alongside text and closes once during finalization.
+      expect(response.events.findIndex(LLMEvent.is.reasoningEnd)).toBeGreaterThan(
        response.events.findIndex(LLMEvent.is.textStart),
      )

@@ -405,6 +405,19 @@ describe("OpenAI-compatible Chat route", () => {
    }),
  )

+  it.effect("ignores events after the done sentinel", () =>
+    Effect.gen(function* () {
+      const body = `${sseEvents(
+        deltaChunk({ content: "Hello" }),
+        deltaChunk({}, "stop"),
+      )}data: ${JSON.stringify(deltaChunk({ content: " late" }))}\n\n`
+      const response = yield* LLMClient.generate(request).pipe(Effect.provide(fixedResponse(body)))
+
+      expect(response.text).toBe("Hello")
+      expect(response.finishReason).toEqual({ normalized: "stop", raw: "stop" })
+    }),
+  )
+
  it.effect("accepts nullable usage and preserves provider fields", () =>
    Effect.gen(function* () {
      const response = yield* LLMClient.generate(request).pipe(
@@ -2750,6 +2750,227 @@ describe("OpenAI Responses route", () => {
    }),
  )

+  it.effect("concludes reasoning at implicit summary boundaries", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(
+        LLMRequest.update(request, { providerOptions: { store: false } }),
+      ).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              {
+                type: "response.output_item.added",
+                item: { type: "reasoning", id: "rs_1", encrypted_content: null },
+              },
+              { type: "response.reasoning_summary_part.added", item_id: "rs_1", summary_index: 0 },
+              { type: "response.reasoning_summary_text.delta", item_id: "rs_1", summary_index: 0, delta: "First" },
+              // The next part is enough to conclude the previous one even when
+              // its done event is delayed.
+              { type: "response.reasoning_summary_part.added", item_id: "rs_1", summary_index: 1 },
+              { type: "response.reasoning_summary_part.done", item_id: "rs_1", summary_index: 0 },
+              { type: "response.reasoning_summary_text.delta", item_id: "rs_1", summary_index: 1, delta: "Second" },
+              { type: "response.reasoning_summary_part.done", item_id: "rs_1", summary_index: 1 },
+              // Some compatible providers begin the next part with its first delta.
+              { type: "response.reasoning_summary_text.delta", item_id: "rs_1", summary_index: 2, delta: "Third" },
+              {
+                type: "response.output_item.done",
+                item: { type: "reasoning", id: "rs_1", encrypted_content: "encrypted-state" },
+              },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.reasoning).toBe("FirstSecondThird")
+      expect(response.events.filter((event) => event.type.startsWith("reasoning-"))).toEqual([
+        {
+          type: "reasoning-start",
+          id: "rs_1:0",
+          providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: null } },
+        },
+        { type: "reasoning-delta", id: "rs_1:0", text: "First", providerMetadata: undefined },
+        { type: "reasoning-end", id: "rs_1:0", providerMetadata: { openai: { itemId: "rs_1" } } },
+        {
+          type: "reasoning-start",
+          id: "rs_1:1",
+          providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: null } },
+        },
+        { type: "reasoning-delta", id: "rs_1:1", text: "Second", providerMetadata: undefined },
+        {
+          type: "reasoning-end",
+          id: "rs_1:1",
+          providerMetadata: { openai: { itemId: "rs_1" } },
+        },
+        {
+          type: "reasoning-start",
+          id: "rs_1:2",
+          providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: null } },
+        },
+        { type: "reasoning-delta", id: "rs_1:2", text: "Third", providerMetadata: undefined },
+        {
+          type: "reasoning-end",
+          id: "rs_1:2",
+          providerMetadata: { openai: { itemId: "rs_1", reasoningEncryptedContent: "encrypted-state" } },
+        },
+      ])
+    }),
+  )
+
+  it.effect("rejects a reasoning item that starts before the previous item ends", () =>
+    Effect.gen(function* () {
+      const error = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { 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.output_item.added", item: { type: "reasoning", id: "rs_2" } },
+            ),
+          ),
+        ),
+        Effect.flip,
+      )
+
+      expect(error.reason._tag).toBe("InvalidProviderOutput")
+      expect(error.message).toContain("started reasoning before the previous item ended")
+    }),
+  )
+
+  it.effect("concludes text at implicit message boundaries", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              // An item that never streams text is untracked at the boundary too.
+              { type: "response.output_item.added", item: { type: "message", id: "msg_0" } },
+              { type: "response.output_item.added", item: { type: "message", id: "msg_1" } },
+              { type: "response.output_text.delta", item_id: "msg_1", delta: "First" },
+              // The previous message's done event is missing; the next message
+              // item is the boundary for its open text.
+              { type: "response.output_item.added", item: { type: "message", id: "msg_2" } },
+              // Late deltas for concluded or untracked messages must stay no-ops.
+              { type: "response.output_text.delta", item_id: "msg_1", delta: " late" },
+              { type: "response.output_text.delta", item_id: "msg_0", delta: " stale" },
+              { type: "response.output_text.delta", item_id: "msg_2", delta: "Second" },
+              { type: "response.output_item.done", item: { type: "message", id: "msg_2" } },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.text).toBe("FirstSecond")
+      expect(response.events.filter((event) => event.type.startsWith("text-"))).toMatchObject([
+        { type: "text-start", id: "msg_1" },
+        { type: "text-delta", id: "msg_1", text: "First" },
+        { type: "text-end", id: "msg_1" },
+        { type: "text-start", id: "msg_2" },
+        { type: "text-delta", id: "msg_2", text: "Second" },
+        { type: "text-end", id: "msg_2" },
+      ])
+    }),
+  )
+
+  it.effect("opens the tool lifecycle for a done-only function call", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              // No output_item.added: the call arrives only as a completed item.
+              {
+                type: "response.output_item.done",
+                item: {
+                  type: "function_call",
+                  id: "fc_1",
+                  call_id: "call_1",
+                  name: "lookup",
+                  arguments: '{"query":"weather"}',
+                },
+              },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.events.filter((event) => event.type.startsWith("tool-"))).toMatchObject([
+        { type: "tool-input-start", id: "call_1", name: "lookup" },
+        { type: "tool-input-end", id: "call_1", name: "lookup" },
+        { type: "tool-call", id: "call_1", name: "lookup", input: { query: "weather" } },
+      ])
+      expect(response.finishReason.normalized).toBe("tool-calls")
+    }),
+  )
+
+  it.effect("ignores duplicate item boundary events", () =>
+    Effect.gen(function* () {
+      const response = yield* LLMClient.generate(request).pipe(
+        Effect.provide(
+          fixedResponse(
+            sseEvents(
+              { type: "response.output_item.added", item: { type: "reasoning", id: "rs_1" } },
+              // Duplicate added for a known item is not overlap and must no-op.
+              { 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: "Think" },
+              { type: "response.output_item.done", item: { type: "reasoning", id: "rs_1" } },
+              { type: "response.output_item.done", item: { type: "reasoning", id: "rs_1" } },
+              {
+                type: "response.output_item.added",
+                item: { type: "function_call", id: "fc_1", call_id: "call_1", name: "lookup", arguments: "" },
+              },
+              {
+                type: "response.output_item.added",
+                item: { type: "function_call", id: "fc_1", call_id: "call_1", name: "lookup", arguments: "" },
+              },
+              { type: "response.function_call_arguments.delta", item_id: "fc_1", delta: '{"query":"weather"}' },
+              {
+                type: "response.output_item.done",
+                item: {
+                  type: "function_call",
+                  id: "fc_1",
+                  call_id: "call_1",
+                  name: "lookup",
+                  arguments: '{"query":"weather"}',
+                },
+              },
+              {
+                type: "response.output_item.done",
+                item: {
+                  type: "function_call",
+                  id: "fc_1",
+                  call_id: "call_1",
+                  name: "lookup",
+                  arguments: '{"query":"weather"}',
+                },
+              },
+              // Duplicates that drop the item id still resolve the same call.
+              {
+                type: "response.output_item.done",
+                item: { type: "function_call", call_id: "call_1", name: "lookup", arguments: '{"query":"weather"}' },
+              },
+              {
+                type: "response.output_item.added",
+                item: { type: "function_call", call_id: "call_1", name: "lookup", arguments: "" },
+              },
+              { type: "response.completed", response: { id: "resp_1" } },
+            ),
+          ),
+        ),
+      )
+
+      expect(response.reasoning).toBe("Think")
+      expect(response.events.filter((event) => event.type === "reasoning-start")).toHaveLength(1)
+      expect(response.events.filter((event) => event.type === "reasoning-end")).toHaveLength(1)
+      expect(response.events.filter((event) => event.type === "tool-input-start")).toHaveLength(1)
+      expect(response.events.filter(LLMEvent.is.toolCall)).toEqual([
+        expect.objectContaining({ id: "call_1", input: { query: "weather" } }),
+      ])
+    }),
+  )
+
  it.effect("reconciles reasoning summaries that arrive only as finals", () =>
    Effect.gen(function* () {
      const response = yield* LLMClient.generate(
@@ -305,9 +305,11 @@ describe("OpenRouter", () => {
      expect(prepared.body.messages).toEqual([
        {
          role: "assistant",
-          content: null,
+          content: "",
          reasoning: "Thinking",
+          reasoning_content: undefined,
          reasoning_details: details,
+          reasoning_text: undefined,
        },
      ])
    }),
@@ -335,7 +337,14 @@ describe("OpenRouter", () => {
      )

      expect(prepared.body.messages).toEqual([
-        { role: "assistant", content: null, reasoning: "Thinking", reasoning_details: details },
+        {
+          role: "assistant",
+          content: "",
+          reasoning: "Thinking",
+          reasoning_content: undefined,
+          reasoning_details: details,
+          reasoning_text: undefined,
+        },
      ])
    }),
  )
@@ -361,7 +370,14 @@ describe("OpenRouter", () => {
      )

      expect(prepared.body.messages).toEqual([
-        { role: "assistant", content: null, reasoning: "AB", reasoning_details: details },
+        {
+          role: "assistant",
+          content: "",
+          reasoning: "AB",
+          reasoning_content: undefined,
+          reasoning_details: details,
+          reasoning_text: undefined,
+        },
      ])
    }),
  )
@@ -376,7 +392,16 @@ describe("OpenRouter", () => {
        }),
      )

-      expect(prepared.body.messages).toEqual([{ role: "assistant", content: null }])
+      expect(prepared.body.messages).toEqual([
+        {
+          role: "assistant",
+          content: "",
+          reasoning: undefined,
+          reasoning_content: undefined,
+          reasoning_details: undefined,
+          reasoning_text: undefined,
+        },
+      ])
    }),
  )
 })
@@ -104,6 +104,60 @@ describe("LLMResponse reducer", () => {
    ])
  })

+  test("authoritative text-end value replaces accumulated deltas", () => {
+    const response = LLMResponse.fromEvents([
+      LLMEvent.textStart({ id: "t1" }),
+      LLMEvent.textDelta({ id: "t1", text: "Hel" }),
+      LLMEvent.textEnd({ id: "t1", text: "Hello!" }),
+      LLMEvent.finish({ reason: { normalized: "stop" } }),
+    ])
+
+    expect(response?.message.content).toEqual([{ type: "text", text: "Hello!" }])
+    expect(response?.text).toBe("Hello!")
+  })
+
+  test("text-end without value keeps joined deltas", () => {
+    const response = LLMResponse.fromEvents([
+      LLMEvent.textStart({ id: "t1" }),
+      LLMEvent.textDelta({ id: "t1", text: "Hel" }),
+      LLMEvent.textDelta({ id: "t1", text: "lo" }),
+      LLMEvent.textEnd({ id: "t1" }),
+      LLMEvent.finish({ reason: { normalized: "stop" } }),
+    ])
+
+    expect(response?.message.content).toEqual([{ type: "text", text: "Hello" }])
+    expect(response?.text).toBe("Hello")
+  })
+
+  test("authoritative reasoning-end value replaces only its own fragment", () => {
+    const response = LLMResponse.fromEvents([
+      LLMEvent.reasoningStart({ id: "r1:0" }),
+      LLMEvent.reasoningDelta({ id: "r1:0", text: "First summ" }),
+      LLMEvent.reasoningEnd({ id: "r1:0", text: "First summary." }),
+      LLMEvent.reasoningStart({ id: "r1:1" }),
+      LLMEvent.reasoningDelta({ id: "r1:1", text: "Second summary." }),
+      LLMEvent.reasoningEnd({ id: "r1:1" }),
+      LLMEvent.finish({ reason: { normalized: "stop" } }),
+    ])
+
+    expect(response?.message.content).toEqual([
+      { type: "reasoning", text: "First summary." },
+      { type: "reasoning", text: "Second summary." },
+    ])
+    expect(response?.reasoning).toBe("First summary.Second summary.")
+  })
+
+  test("end value recovers a fragment that streamed no deltas", () => {
+    const response = LLMResponse.fromEvents([
+      LLMEvent.textStart({ id: "t1" }),
+      LLMEvent.textEnd({ id: "t1", text: "Hello!" }),
+      LLMEvent.finish({ reason: { normalized: "stop" } }),
+    ])
+
+    expect(response?.message.content).toEqual([{ type: "text", text: "Hello!" }])
+    expect(response?.text).toBe("Hello!")
+  })
+
  test("clears malformed tool input without appending an executable call", () => {
    const state = reduce([
      LLMEvent.toolInputStart({ id: "call_1", name: "lookup" }),
@@ -250,7 +250,7 @@ test("AI error reasons are tagged Errors with required messages", () => {
      provider: model.provider,
      model: model.id,
    }),
-    new AuthenticationError({ message: "Missing credentials", kind: "missing" }),
+    new AuthenticationError({ message: "Missing credentials" }),
    new RateLimitError({ message: "Rate limited" }),
    new QuotaExceededError({ message: "Quota exceeded" }),
    new ContentPolicyError({ message: "Content blocked" }),
@@ -0,0 +1,331 @@
+import { describe, expect } from "bun:test"
+import { AIError, LanguageModel, LLM, LLMClient, LLMEvent, LLMRequest, RateLimitError } from "../src/index.js"
+import { OpenAIChat } from "../src/protocols/openai-chat.js"
+import { TestLLM } from "../src/testing.js"
+import { Effect, Fiber, Latch, Stream } from "effect"
+import { testEffect } from "./lib/effect.js"
+
+const request = LLM.request({
+  model: LanguageModel.make({ id: "fictional-model", provider: "fixture", route: OpenAIChat.route }),
+  prompt: "Say hello",
+})
+const legacy = testEffect(TestLLM.layer())
+const it = testEffect(TestLLM.testLayer())
+
+describe("TestLLM legacy client", () => {
+  legacy.effect("does not observe requests or consume responses until execution", () =>
+    Effect.gen(function* () {
+      const llm = yield* TestLLM.Service
+      yield* llm.push(TestLLM.text("first", "first"), TestLLM.text("second", "second"))
+
+      llm.client.stream(request)
+      llm.client.generate(request)
+      expect(llm.requests).toEqual([])
+
+      expect((yield* llm.client.generate(request)).text).toBe("first")
+      expect((yield* llm.client.generate(request)).text).toBe("second")
+      expect(llm.requests).toEqual([request, request])
+    }),
+  )
+
+  legacy.effect("assigns and records a fresh response for each execution", () =>
+    Effect.gen(function* () {
+      const llm = yield* TestLLM.Service
+      yield* llm.push(
+        TestLLM.text("first", "first"),
+        TestLLM.text("second", "second"),
+        TestLLM.text("third", "third"),
+        TestLLM.text("fourth", "fourth"),
+      )
+      const stream = llm.client.stream(request)
+      const generate = llm.client.generate(request)
+
+      expect(yield* Stream.runCollect(stream)).toEqual(TestLLM.text("first", "first"))
+      expect(yield* Stream.runCollect(stream)).toEqual(TestLLM.text("second", "second"))
+      expect((yield* generate).text).toBe("third")
+      expect((yield* generate).text).toBe("fourth")
+      expect(llm.requests).toEqual([request, request, request, request])
+    }),
+  )
+
+  legacy.effect("keeps module-level controls and clientLayer on the same backing state", () =>
+    Effect.gen(function* () {
+      const llm = yield* TestLLM.Service
+      const requests = llm.requests
+      yield* TestLLM.push(TestLLM.text("queued", "queued"))
+      yield* TestLLM.always(TestLLM.text("fallback", "fallback"))
+      expect((yield* LLMClient.generate(request).pipe(Effect.provide(TestLLM.clientLayer))).text).toBe("queued")
+      yield* TestLLM.wait(1)
+      expect(requests).toEqual([request])
+      requests.length = 0
+      expect((yield* llm.client.generate(request)).text).toBe("fallback")
+      expect(llm.requests).toBe(requests)
+      expect(requests).toEqual([request])
+    }),
+  )
+})
+
+describe("TestLLM first-class client", () => {
+  it.effect("provides the same object under normal and test tags with snapshot observations", () =>
+    Effect.gen(function* () {
+      const llm = yield* TestLLM.Test
+      const client = yield* LLMClient.Service
+      expect(client).toBe(llm)
+      const before = yield* llm.requests()
+      yield* llm.push(TestLLM.text("hello", "answer"))
+      const generate = client.generate(request)
+      client.stream(request)
+      expect(yield* llm.requests()).toEqual([])
+
+      expect((yield* generate).text).toBe("hello")
+      expect(before).toEqual([])
+      expect(yield* llm.requests()).toEqual([request])
+      expect(yield* llm.requests()).not.toBe(yield* llm.requests())
+    }),
+  )
+
+  it.effect("prioritizes queued replies over request-dependent and constant fallbacks", () =>
+    Effect.gen(function* () {
+      const llm = yield* TestLLM.Test
+      const served: LLMRequest[] = []
+      yield* llm.always(TestLLM.text("old fallback", "old"))
+      yield* llm.push(TestLLM.text("first", "first"), TestLLM.text("second", "second"))
+      yield* llm.serve((request) => {
+        served.push(request)
+        return TestLLM.text(request.promptCacheKey ?? "default", "served")
+      })
+
+      expect((yield* LLMClient.generate(request)).text).toBe("first")
+      expect((yield* LLMClient.generate(request)).text).toBe("second")
+      expect(served).toEqual([])
+      const selected = LLMRequest.update(request, { promptCacheKey: "selected" })
+      expect((yield* LLMClient.generate(selected)).text).toBe("selected")
+      expect((yield* LLMClient.generate(request)).text).toBe("default")
+      expect(served).toEqual([selected, request])
+
+      yield* llm.push(TestLLM.text("queued again", "queued"))
+      yield* llm.always(TestLLM.text("constant", "constant"))
+      expect((yield* LLMClient.generate(request)).text).toBe("queued again")
+      expect((yield* LLMClient.generate(request)).text).toBe("constant")
+      expect((yield* LLMClient.generate(request)).text).toBe("constant")
+      expect(served).toEqual([selected, request])
+    }),
+  )
+
+  testEffect(
+    TestLLM.testLayer({
+      transformRequest: (request) => LLMRequest.update(request, { promptCacheKey: "observation" }),
+    }),
+  ).effect("transforms observations without changing the request passed to the responder", () =>
+    Effect.gen(function* () {
+      const llm = yield* TestLLM.Test
+      yield* llm.serve((input) => {
+        expect(input).toBe(request)
+        return TestLLM.text("original", "answer")
+      })
+      const generate = llm.generate(request)
+      expect(yield* llm.requests()).toEqual([])
+      expect((yield* generate).text).toBe("original")
+      expect(yield* llm.requests()).toEqual([LLMRequest.update(request, { promptCacheKey: "observation" })])
+    }),
+  )
+
+  it.effect("broadcasts request-arrival waits and satisfies waits registered afterward", () =>
+    Effect.gen(function* () {
+      const llm = yield* TestLLM.Test
+      yield* llm.always(TestLLM.stop())
+      const first = yield* llm.wait(2).pipe(Effect.forkChild({ startImmediately: true }))
+      const second = yield* llm.wait(2).pipe(Effect.forkChild({ startImmediately: true }))
+      yield* llm.generate(request)
+      expect(first.pollUnsafe()).toBeUndefined()
+      expect(second.pollUnsafe()).toBeUndefined()
+      yield* llm.generate(request)
+      yield* Fiber.join(first)
+      yield* Fiber.join(second)
+      yield* llm.wait(2)
+      expect(yield* llm.requests()).toHaveLength(2)
+    }),
+  )
+  ;(["queued", "served"] as const).forEach((mode) => {
+    it.effect(`assigns ${mode} replies before resuming request-arrival continuations`, () =>
+      Effect.gen(function* () {
+        const llm = yield* TestLLM.Test
+        const responses = [TestLLM.text("first", "first"), TestLLM.text("second", "second")]
+        yield* mode === "queued" ? llm.push(...responses) : llm.serve(() => responses.shift() ?? [])
+        const later = yield* llm
+          .wait(1)
+          .pipe(Effect.andThen(llm.generate(request)), Effect.forkChild({ startImmediately: true }))
+
+        expect((yield* llm.generate(request)).text).toBe("first")
+        expect((yield* Fiber.join(later)).text).toBe("second")
+        expect(yield* llm.requests()).toEqual([request, request])
+      }),
+    )
+  })
+
+  it.effect("notifies arrival waiters even when the responder defects", () =>
+    Effect.gen(function* () {
+      const llm = yield* TestLLM.Test
+      const defect = new Error("Broken fixture responder")
+      yield* llm.serve(() => {
+        throw defect
+      })
+      const waiter = yield* llm.wait(1).pipe(Effect.forkChild({ startImmediately: true }))
+      expect(yield* llm.generate(request).pipe(Effect.catchDefect(Effect.succeed))).toBe(defect)
+      yield* Fiber.join(waiter)
+    }),
+  )
+
+  it.effect("builds independent state even when the same layer is provided concurrently", () => {
+    const layer = TestLLM.testLayer()
+    const run = Effect.gen(function* () {
+      const llm = yield* TestLLM.Test
+      expect(yield* llm.requests()).toEqual([])
+      yield* llm.push(TestLLM.text("one", "answer"))
+      expect((yield* LLMClient.generate(request)).text).toBe("one")
+      return yield* llm.requests()
+    }).pipe(Effect.provide(layer))
+    return Effect.gen(function* () {
+      expect(yield* Effect.all([run, run], { concurrency: "unbounded" })).toEqual([[request], [request]])
+    })
+  })
+
+  it.effect("counts concurrent starts on one gate without serializing their response assignment", () =>
+    Effect.gen(function* () {
+      const llm = yield* TestLLM.Test
+      yield* llm.push(TestLLM.text("first", "first"), TestLLM.text("second", "second"))
+      const generate = llm.generate(request)
+      const gate = yield* llm.gate()
+      const first = yield* generate.pipe(Effect.forkChild({ startImmediately: true }))
+      yield* gate.started
+      const second = yield* generate.pipe(Effect.forkChild({ startImmediately: true }))
+      yield* gate.started
+      yield* llm.wait(2)
+      expect(first.pollUnsafe()).toBeUndefined()
+      expect(second.pollUnsafe()).toBeUndefined()
+      yield* gate.release
+      expect((yield* Fiber.join(first)).text).toBe("first")
+      expect((yield* Fiber.join(second)).text).toBe("second")
+    }),
+  )
+
+  it.effect("does not clear a replacement gate when the previous gate is released", () =>
+    Effect.gen(function* () {
+      const llm = yield* TestLLM.Test
+      yield* llm.always(TestLLM.stop())
+      const previous = yield* llm.gate()
+      const first = yield* llm.generate(request).pipe(Effect.forkChild({ startImmediately: true }))
+      yield* previous.started
+      const next = yield* llm.gate()
+      yield* previous.release
+      yield* Fiber.join(first)
+
+      const second = yield* llm.generate(request).pipe(Effect.forkChild({ startImmediately: true }))
+      yield* next.started
+      expect(second.pollUnsafe()).toBeUndefined()
+      yield* next.release
+      yield* Fiber.join(second)
+    }),
+  )
+
+  it.effect("releases a gate when its deliberately narrower scope closes", () =>
+    Effect.gen(function* () {
+      const llm = yield* TestLLM.Test
+      yield* llm.always(TestLLM.stop())
+      // Only the gate is scoped here; its release must happen before the test ends.
+      const run = yield* Effect.scoped(
+        Effect.gen(function* () {
+          const gate = yield* llm.gate()
+          const run = yield* llm.generate(request).pipe(Effect.forkChild({ startImmediately: true }))
+          yield* gate.started
+          return run
+        }),
+      )
+      yield* Fiber.join(run)
+      yield* llm.generate(request)
+      expect(yield* llm.requests()).toHaveLength(2)
+    }),
+  )
+
+  it.effect("keeps an executed response consumed after interruption and permits later requests", () =>
+    Effect.gen(function* () {
+      const llm = yield* TestLLM.Test
+      yield* llm.push(TestLLM.text("interrupted", "first"), TestLLM.text("next", "second"))
+      const gate = yield* llm.gate()
+      const run = yield* llm.generate(request).pipe(Effect.forkChild({ startImmediately: true }))
+      yield* gate.started
+      yield* Fiber.interrupt(run)
+      yield* gate.release
+      expect((yield* llm.generate(request)).text).toBe("next")
+      expect(yield* llm.requests()).toHaveLength(2)
+    }),
+  )
+
+  it.effect("consumes a supplied stream's post-finish tail and runs its finalizer", () =>
+    Effect.gen(function* () {
+      const llm = yield* TestLLM.Test
+      const tail = yield* Latch.make()
+      const release = yield* Latch.make()
+      const finalized = yield* Latch.make()
+      yield* llm.push(
+        Stream.unwrap(
+          Effect.gen(function* () {
+            yield* Effect.addFinalizer(() => finalized.open)
+            return Stream.fromIterable(TestLLM.text("complete", "answer")).pipe(
+              Stream.concat(Stream.fromEffect(tail.open.pipe(Effect.andThen(release.await))).pipe(Stream.drain)),
+            )
+          }),
+        ),
+      )
+      const run = yield* llm.generate(request).pipe(Effect.forkChild({ startImmediately: true }))
+      yield* tail.await
+      expect(run.pollUnsafe()).toBeUndefined()
+      yield* release.open
+      expect((yield* Fiber.join(run)).text).toBe("complete")
+      yield* finalized.await
+    }),
+  )
+
+  it.effect("preserves irregular events, ordinary EOF, typed failures, and responder defects", () =>
+    Effect.gen(function* () {
+      const llm = yield* TestLLM.Test
+      const events = [LLMEvent.textDelta({ id: "without-start", text: "partial" })]
+      yield* llm.push(events, [])
+      expect(yield* Stream.runCollect(llm.stream(request))).toEqual(events)
+      expect(yield* Stream.runCollect(llm.stream(request))).toEqual([])
+
+      const failure = new AIError({ reason: new RateLimitError({ message: "Try later" }) })
+      const observed: LLMEvent[] = []
+      yield* llm.serve(() => TestLLM.failAfter(failure, ...events))
+      expect(
+        yield* llm.stream(request).pipe(
+          Stream.runForEach((event) => Effect.sync(() => observed.push(event))),
+          Effect.flip,
+        ),
+      ).toBe(failure)
+      expect(observed).toEqual(events)
+      expect(yield* llm.generate(request).pipe(Effect.flip)).toBe(failure)
+
+      const defect = new Error("Broken fixture responder")
+      yield* llm.serve(() => {
+        throw defect
+      })
+      expect(yield* llm.generate(request).pipe(Effect.catchDefect(Effect.succeed))).toBe(defect)
+      yield* llm.push(TestLLM.text("recovered", "answer"))
+      expect((yield* llm.generate(request)).text).toBe("recovered")
+    }),
+  )
+
+  it.effect("defects on unexpected requests instead of waiting for a late script", () =>
+    Effect.gen(function* () {
+      const llm = yield* TestLLM.Test
+      const defect = yield* llm.generate(request).pipe(Effect.catchDefect(Effect.succeed))
+      expect(defect).toBeInstanceOf(Error)
+      if (!(defect instanceof Error)) return
+      expect(defect.message).toBe("TestLLM has no response for request 1")
+      expect(yield* llm.requests()).toEqual([request])
+      yield* llm.push(TestLLM.stop())
+      yield* llm.generate(request)
+    }),
+  )
+})
@@ -5,5 +5,5 @@
    "noEmit": true,
    "rootDir": "."
  },
-  "include": ["test/**/*.types.ts"]
+  "include": ["test/**/*.types.ts", "test/testing.test.ts"]
 }
@@ -80,8 +80,8 @@ export function createSessionRequestModel() {
        if (message.type !== "synthetic") return []
        if (message.metadata?.source === "subagent" && typeof message.metadata.childID === "string")
          return [message.metadata.childID]
-        if (message.metadata?.source === "shell" && typeof message.metadata.jobID === "string")
-          return [message.metadata.jobID]
+        if (message.metadata?.source === "shell")
+          return [message.metadata.shellID, message.metadata.jobID].filter((id): id is string => typeof id === "string")
        return []
      }),
    )
@@ -121,6 +121,7 @@ export function createSessionRequestModel() {
        if (part.type !== "tool" || part.name !== "shell" || completed.has(part.id)) return []
        if (part.state.status !== "completed" || part.state.metadata?.status !== "running") return []
        const shellID = part.state.metadata.shellID
+        if (typeof shellID === "string" && completed.has(shellID)) return []
        const command = part.state.input.command
        return [
          {
@@ -0,0 +1,85 @@
+import { expect, test } from "bun:test"
+import { MemoryRouter, Route, createMemoryHistory, useNavigate, useParams } from "@solidjs/router"
+import { createEffect, onCleanup } from "solid-js"
+import { createComponent, render } from "solid-js/web"
+import { createSessionOwnership } from "@/session/session-ownership"
+
+test("session ownership follows navigation, back, forward, replacement, and route disposal", async () => {
+  const host = document.createElement("div")
+  const history = createMemoryHistory()
+  history.set({ value: "/session/A", replace: true, scroll: false })
+  const captured: ReturnType<ReturnType<typeof createSessionOwnership>["capture"]>[] = []
+  const navigate: ReturnType<typeof useNavigate>[] = []
+  const cleaned: string[] = []
+  const settle = () => new Promise((resolve) => setTimeout(resolve, 0))
+
+  const Session = () => {
+    const params = useParams<{ id: string }>()
+    const ownership = createSessionOwnership(() => params.id)
+    const label = document.createElement("div")
+    navigate.push(useNavigate())
+    createEffect(() => {
+      label.textContent = params.id
+      captured.push(ownership.capture())
+    })
+    onCleanup(() => cleaned.push("session"))
+    return label
+  }
+
+  const dispose = render(
+    () =>
+      createComponent(MemoryRouter, {
+        history,
+        get children() {
+          return [
+            createComponent(Route, { path: "/", component: () => "Home" }),
+            createComponent(Route, { path: "/session/:id", component: Session }),
+          ]
+        },
+      }),
+    host,
+  )
+
+  try {
+    expect(host.textContent).toBe("A")
+    expect(captured[0].current()).toBe(true)
+
+    navigate[0]("/session/B")
+    await settle()
+    expect(host.textContent).toBe("B")
+    expect(captured[0].current()).toBe(false)
+    expect(captured[1].current()).toBe(true)
+
+    history.back()
+    await settle()
+    expect(host.textContent).toBe("A")
+    expect(captured[0].current()).toBe(false)
+    expect(captured[1].current()).toBe(false)
+    expect(captured[2].current()).toBe(true)
+
+    history.forward()
+    await settle()
+    expect(host.textContent).toBe("B")
+    expect(captured[2].current()).toBe(false)
+    expect(captured[3].current()).toBe(true)
+    expect(navigate).toHaveLength(1)
+    expect(cleaned).toEqual([])
+
+    navigate[0]("/", { replace: true })
+    await settle()
+    expect(host.textContent).toBe("Home")
+    expect(captured.every((owner) => !owner.current())).toBe(true)
+    expect(cleaned).toEqual(["session"])
+
+    history.back()
+    await settle()
+    expect(host.textContent).toBe("A")
+    expect(captured[4].current()).toBe(true)
+    expect(navigate).toHaveLength(2)
+  } finally {
+    dispose()
+  }
+
+  expect(captured.every((owner) => !owner.current())).toBe(true)
+  expect(cleaned).toEqual(["session", "session"])
+})
@@ -0,0 +1,89 @@
+import { expect, test } from "bun:test"
+import {
+  createComputed,
+  createMemo,
+  createResource,
+  createRoot,
+  createSignal,
+  onCleanup,
+  startTransition,
+  Suspense,
+  type Accessor,
+  type JSX,
+} from "solid-js"
+import { createComponent, render } from "solid-js/web"
+
+// Solid 1.9.15 includes the former transition patch for solidjs/solid#2046.
+test("a memo created during a paused transition has a committed value", async () => {
+  const host = document.createElement("div")
+  const loaded = Promise.withResolvers<string>()
+  const started = Promise.withResolvers<void>()
+  const memos: Accessor<string>[] = []
+  const [session, setSession] = createSignal(false)
+  const dispose = render(
+    () =>
+      createComponent(Suspense, {
+        fallback: "Loading",
+        get children() {
+          return createMemo(() => {
+            if (!session()) return "Home"
+            memos.push(createMemo(() => "Session"))
+            const [data] = createResource(() => loaded.promise)
+            started.resolve()
+            return createMemo(() => data() ?? "")
+          }) as unknown as JSX.Element
+        },
+      }),
+    host,
+  )
+
+  try {
+    expect(host.textContent).toBe("Home")
+    const pending = startTransition(() => setSession(true))
+    await started.promise
+    expect(host.textContent).toBe("Home")
+    expect(memos).toHaveLength(1)
+    expect(memos[0]()).toBe("Session")
+
+    loaded.resolve("Session loaded")
+    await pending
+    expect(host.textContent).toBe("Session loaded")
+    expect(memos[0]()).toBe("Session")
+  } finally {
+    loaded.resolve("Session loaded")
+    dispose()
+  }
+})
+
+// Exercise each patched bundle independently; importing both dev formats emits Solid's duplicate-runtime warning.
+test.each(["solid.js", "solid.cjs", "dev.js", "dev.cjs", "server.js", "server.cjs"])(
+  "%s reentrant disposal cleans owned computations and callbacks exactly once",
+  async (file) => {
+    const runtime: {
+      createRoot: typeof createRoot
+      createComputed: typeof createComputed
+      onCleanup: typeof onCleanup
+    } = await import(`solid-js/dist/${file}`)
+    const cleaned: string[] = []
+    const dispose = runtime.createRoot((dispose) => {
+      runtime.onCleanup(() => cleaned.push("root"))
+      runtime.createComputed(() => {
+        runtime.onCleanup(() => cleaned.push("first"))
+      })
+      runtime.createComputed(() => {
+        runtime.onCleanup(() => cleaned.push("second"))
+        runtime.onCleanup(() => {
+          cleaned.push("reentrant")
+          // Bound the regression so an unpatched runtime fails without overflowing the stack.
+          if (cleaned.filter((item) => item === "reentrant").length === 1) dispose()
+        })
+      })
+      return dispose
+    })
+
+    expect(dispose).not.toThrow()
+    expect(cleaned.toSorted()).toEqual(["first", "reentrant", "root", "second"])
+    dispose()
+    expect(cleaned).toHaveLength(4)
+  },
+)
@@ -1,58 +0,0 @@
-import { NodeFileSystem } from "@effect/platform-node"
-import { Global } from "@opencode-ai/util/global"
-import { AppProcess } from "@opencode-ai/util/process"
-import { Effect, FileSystem, Layer, Stream } from "effect"
-import { ChildProcessSpawner } from "effect/unstable/process"
-import assert from "node:assert/strict"
-import path from "node:path"
-import { Updater } from "../updater"
-
-const latest = { version: "0.0.0-beta-17498" }
-const installs: string[] = []
-
-// This fixture runs in its own process; no real update requests or installs occur.
-globalThis.fetch = Object.assign(async () => Response.json(latest), { preconnect() {} })
-
-await Effect.runPromise(
-  Effect.gen(function* () {
-    const fs = yield* FileSystem.FileSystem
-    const directory = yield* fs.makeTempDirectoryScoped({ prefix: "opencode-updater-" })
-    const dependencies = Layer.mergeAll(
-      Layer.succeed(FileSystem.FileSystem, fs),
-      Layer.succeed(
-        Global.Service,
-        Global.make({ home: directory, config: directory, cache: path.join(directory, "cache") }),
-      ),
-      Layer.succeed(AppProcess.Service, {
-        ...ChildProcessSpawner.make(() => Effect.die("Unexpected process spawn")),
-        runStream: () => Stream.die("Unexpected streaming process"),
-        run: (command) => {
-          assert.equal(command._tag, "StandardCommand")
-          if (command.command === "npm" && command.args[0] === "install") {
-            assert.ok(command.args.includes("--global"))
-            installs.push(command.args.at(-1)!)
-          }
-          return Effect.succeed({
-            command: command.command,
-            exitCode: 0,
-            stdout: Buffer.from(command.command === "npm" ? "@opencode-ai/cli" : ""),
-            stderr: Buffer.alloc(0),
-            stdoutTruncated: false,
-            stderrTruncated: false,
-          })
-        },
-      }),
-    )
-    yield* Effect.gen(function* () {
-      const updater = yield* Updater.Service
-      yield* updater.check()
-      assert.deepEqual(installs, ["@opencode-ai/cli@0.0.0-beta-17498"])
-      yield* updater.check()
-      yield* updater.check()
-      assert.deepEqual(installs, ["@opencode-ai/cli@0.0.0-beta-17498"])
-      latest.version = "0.0.0-beta-17499"
-      yield* updater.check()
-      assert.deepEqual(installs, ["@opencode-ai/cli@0.0.0-beta-17498", "@opencode-ai/cli@0.0.0-beta-17499"])
-    }).pipe(Effect.provide(Updater.layer.pipe(Layer.provide(dependencies))))
-  }).pipe(Effect.scoped, Effect.provide(NodeFileSystem.layer)),
-)
@@ -5,9 +5,9 @@ const maximumComponent = "9007199254740991"
 const versionPattern =
  /^v?([0-9]+)\.([0-9]+)\.([0-9]+)(?:-([0-9A-Za-z-]+(?:\.[0-9A-Za-z-]+)*))?(?:\+[0-9A-Za-z-]+(?:\.[0-9A-Za-z-]+)*)?$/

-export function action(current: string, latest: string, policy: Policy, installed = current): Action {
+export function action(current: string, latest: string, policy: Policy): Action {
  if (policy === false) return "none"
-  const currentVersion = parseReleaseVersion(installed)
+  const currentVersion = parseReleaseVersion(current)
  const latestVersion = parseReleaseVersion(latest)
  if (!currentVersion || !latestVersion || sameRelease(currentVersion, latestVersion)) return "none"
  // Major upgrades are never installed automatically.
@@ -1,30 +1,8 @@
 import { describe, expect, test } from "bun:test"
-import path from "node:path"
 import { action } from "./updater-action"
 import { decodePolicy } from "./updater"

 describe("updater", () => {
-  test("remembers successful installs across checks and accepts the next release", async () => {
-    // Isolate compiled version constants and the update endpoint from other tests.
-    const child = Bun.spawn(
-      [
-        process.execPath,
-        "--define",
-        'OPENCODE_VERSION="0.0.0-next-16473"',
-        "--define",
-        'OPENCODE_CHANNEL="beta"',
-        path.join(import.meta.dir, "fixtures/updater.ts"),
-      ],
-      { env: { ...process.env, OPENCODE_DISABLE_AUTOUPDATE: "" }, stdout: "pipe", stderr: "pipe" },
-    )
-    const [code, stdout, stderr] = await Promise.all([
-      child.exited,
-      new Response(child.stdout).text(),
-      new Response(child.stderr).text(),
-    ])
-    expect(code, stdout + stderr).toBe(0)
-  })
-
  test("reads autoupdate from JSONC", () => {
    expect(decodePolicy('{ // preference\n "autoupdate": "notify",\n}')).toBe("notify")
    expect(decodePolicy('{ "autoupdate": false }')).toBe(false)
@@ -54,14 +32,6 @@ describe("updater", () => {
    expect(action("1.2.3", "1.2.3", true)).toBe("none")
  })

-  test("skips an installed update but still accepts the next release", () => {
-    const current = "0.0.0-next-16473"
-    const installed = "0.0.0-beta-17498"
-    expect(action(current, installed, true)).toBe("upgrade")
-    expect(action(current, installed, true, installed)).toBe("none")
-    expect(action(current, "0.0.0-beta-17499", true, installed)).toBe("upgrade")
-  })
-
  test("upgrades when latest is lower (rollback)", () => {
    expect(action("1.2.4", "1.2.3", true)).toBe("upgrade")
  })
@@ -1,7 +1,7 @@
 import { Global } from "@opencode-ai/util/global"
 import { AppProcess } from "@opencode-ai/util/process"
 import { OPENCODE_CHANNEL, OPENCODE_LOCAL, OPENCODE_VERSION } from "../version"
-import { Context, Duration, Effect, FileSystem, Layer, Ref } from "effect"
+import { Context, Duration, Effect, FileSystem, Layer } from "effect"
 import { ChildProcess } from "effect/unstable/process"
 import { parse, type ParseError } from "jsonc-parser"
 import path from "node:path"
@@ -38,7 +38,6 @@ export const layer = Layer.effect(
    const fs = yield* FileSystem.FileSystem
    const global = yield* Global.Service
    const appProcess = yield* AppProcess.Service
-    const installed = yield* Ref.make(OPENCODE_VERSION)
    const channel = OPENCODE_CHANNEL.replace(/[^a-zA-Z0-9._-]/g, "-")

    const readPolicy = Effect.fnUntraced(function* () {
@@ -113,20 +112,18 @@ export const layer = Layer.effect(

    const upgrade = Effect.fnUntraced(function* (method: Method, version: string) {
      const target = `${packageName}@${version}`
-      const commands: Record<Exclude<Method, "bun" | "curl" | "npm">, string[]> = {
+      const commands: Record<Exclude<Method, "bun" | "curl">, string[]> = {
+        npm: ["npm", "install", "--global", target],
        pnpm: ["pnpm", "add", "--global", `--allow-build=${packageName}`, target],
        yarn: ["yarn", "global", "add", target],
      }
      const result = yield* Effect.scoped(
        Effect.gen(function* () {
-          if (method === "bun" || method === "npm") {
+          if (method === "bun") {
+            // Bun does not prune old versions from its shared package cache.
            yield* fs.makeDirectory(global.cache, { recursive: true })
            const cache = yield* fs.makeTempDirectoryScoped({ directory: global.cache, prefix: "update-" })
-            const command =
-              method === "bun"
-                ? ["bun", "install", "--global", "--trust", "--cache-dir", cache, target]
-                : ["npm", "install", "--global", "--cache", cache, target]
-            return yield* run(command, "5 minutes")
+            return yield* run(["bun", "install", "--global", "--trust", "--cache-dir", cache, target], "5 minutes")
          }
          if (method === "curl") {
            yield* fs.makeDirectory(global.cache, { recursive: true })
@@ -159,19 +156,18 @@ export const layer = Layer.effect(

        return yield* Effect.gen(function* () {
          const version = yield* latest()
-          const current = yield* Ref.get(installed)
          yield* Effect.logInfo("update check", {
-            current,
+            current: OPENCODE_VERSION,
            latest: version,
          })
-          const next = action(OPENCODE_VERSION, version, policy, current)
+          const next = action(OPENCODE_VERSION, version, policy)
          if (next === "none") return yield* Effect.logInfo("update check done", { action: "up-to-date" })
-          if (next === "notify") return yield* Effect.logInfo("OpenCode update available", { current, latest: version })
+          if (next === "notify")
+            return yield* Effect.logInfo("OpenCode update available", { current: OPENCODE_VERSION, latest: version })
          const detected = yield* method()
          if (!detected) return yield* Effect.logWarning("automatic update skipped: installation method not found")
          yield* upgrade(detected, version)
-          yield* Ref.set(installed, version)
-          yield* Effect.logInfo("updated OpenCode", { from: current, to: version, method: detected })
+          yield* Effect.logInfo("updated OpenCode", { from: OPENCODE_VERSION, to: version, method: detected })
        })
      },
      Effect.catchCause((cause) => Effect.logWarning("automatic update failed", { cause })),
@@ -37,7 +37,7 @@
    "@opencode-ai/protocol": "workspace:*"
  },
  "peerDependencies": {
-    "effect": "4.0.0-rc.111",
+    "effect": "4.0.0-rc.112",
    "solid-js": ">=1.9.0"
  },
  "peerDependenciesMeta": {
@@ -1639,6 +1639,23 @@ export interface PtyApi<E = never> {
  readonly connect: { readonly token: PtyConnectTokenOperation<E> }
 }

+export type ExperimentalPersistentPtyReadInput = { readonly sessionID: Session.ID; readonly lines?: number | undefined }
+export type ExperimentalPersistentPtyReadOutput = {
+  readonly ptyID: Pty.ID
+  readonly title: string
+  readonly cwd: string
+  readonly foregroundProcess: string | null
+  readonly screen: {
+    readonly text: string
+    readonly cols: number
+    readonly rows: number
+    readonly cursor: { readonly x: number; readonly y: number }
+  }
+} | null
+export type ExperimentalPersistentPtyReadOperation<E = never> = (
+  input: ExperimentalPersistentPtyReadInput,
+) => Effect.Effect<ExperimentalPersistentPtyReadOutput, E>
+
 export type ExperimentalPersistentPtyListInput = { readonly sessionID: Session.ID }
 export type ExperimentalPersistentPtyListOutput = ReadonlyArray<{
  readonly id: Pty.ID
@@ -1787,6 +1804,7 @@ export type ExperimentalPersistentPtyConnectTokenOperation<E = never> = (

 export interface ExperimentalApi<E = never> {
  readonly persistentPty: {
+    readonly read: ExperimentalPersistentPtyReadOperation<E>
    readonly list: ExperimentalPersistentPtyListOperation<E>
    readonly create: ExperimentalPersistentPtyCreateOperation<E>
    readonly shutdown: ExperimentalPersistentPtyShutdownOperation<E>
@@ -198,6 +198,8 @@ import type {
  PtyRemoveOutput,
  PtyConnectTokenInput,
  PtyConnectTokenOutput,
+  ExperimentalPersistentPtyReadInput,
+  ExperimentalPersistentPtyReadOutput,
  ExperimentalPersistentPtyListInput,
  ExperimentalPersistentPtyListOutput,
  ExperimentalPersistentPtyCreateInput,
@@ -1234,6 +1236,15 @@ const adaptGroupPty = (raw: RawClient["server.pty"]) => ({
  connect: { token: EndpointPtyConnectToken(raw) },
 })

+const EndpointExperimentalPersistentPtyRead =
+  (raw: RawClient["server.experimental"]) => (input: ExperimentalPersistentPtyReadInput) =>
+    preserveEffect<ExperimentalPersistentPtyReadOutput>()(
+      raw["persistentPty.read"]({ params: { sessionID: input["sessionID"] }, query: { lines: input["lines"] } }).pipe(
+        Effect.mapError(mapClientError),
+        Effect.map((value) => value.data),
+      ),
+    )
+
 const EndpointExperimentalPersistentPtyList =
  (raw: RawClient["server.experimental"]) => (input: ExperimentalPersistentPtyListInput) =>
    preserveEffect<ExperimentalPersistentPtyListOutput>()(
@@ -1322,6 +1333,7 @@ const EndpointExperimentalPersistentPtyConnectToken =

 const adaptGroupExperimental = (raw: RawClient["server.experimental"]) => ({
  persistentPty: {
+    read: EndpointExperimentalPersistentPtyRead(raw),
    list: EndpointExperimentalPersistentPtyList(raw),
    create: EndpointExperimentalPersistentPtyCreate(raw),
    shutdown: EndpointExperimentalPersistentPtyShutdown(raw),
@@ -194,6 +194,8 @@ import type {
  PtyRemoveOutput,
  PtyConnectTokenInput,
  PtyConnectTokenOutput,
+  ExperimentalPersistentPtyReadInput,
+  ExperimentalPersistentPtyReadOutput,
  ExperimentalPersistentPtyListInput,
  ExperimentalPersistentPtyListOutput,
  ExperimentalPersistentPtyCreateInput,
@@ -1684,6 +1686,18 @@ export function make(options: ClientOptions) {
    },
    experimental: {
      persistentPty: {
+        read: (input: ExperimentalPersistentPtyReadInput, requestOptions?: RequestOptions) =>
+          request<{ readonly data: ExperimentalPersistentPtyReadOutput }>(
+            {
+              method: "GET",
+              path: `/api/experimental/session/${encodeURIComponent(input.sessionID)}/terminal/read`,
+              query: { lines: input["lines"] },
+              successStatus: 200,
+              declaredStatuses: [503, 401, 400],
+              empty: false,
+            },
+            requestOptions,
+          ).then((value) => value.data),
        list: (input: ExperimentalPersistentPtyListInput, requestOptions?: RequestOptions) =>
          request<{ readonly data: ExperimentalPersistentPtyListOutput }>(
            {
@@ -378,6 +378,14 @@ export type SessionStatus =

 export type PtyTicketConnectToken = { ticket: string; expires_in: number }

+export type PersistentPtyReadResult = {
+  ptyID: string
+  title: string
+  cwd: string
+  foregroundProcess: string | null
+  screen: { text: string; cols: number; rows: number; cursor: { x: number; y: number } }
+}
+
 export type PersistentPtyHandoff = { directory: string; instanceID: string; ticket: string; expiresAt: number }

 export type ShellInfo1 = {
@@ -5769,6 +5777,13 @@ export type PtyConnectTokenOutput = {
  data: PtyTicketConnectToken
 }

+export type ExperimentalPersistentPtyReadInput = {
+  readonly sessionID: { readonly sessionID: string }["sessionID"]
+  readonly lines?: { readonly lines?: number | undefined }["lines"]
+}
+
+export type ExperimentalPersistentPtyReadOutput = { data: PersistentPtyReadResult | null }["data"]
+
 export type ExperimentalPersistentPtyListInput = { readonly sessionID: { readonly sessionID: string }["sessionID"] }

 export type ExperimentalPersistentPtyListOutput = { data: Array<PersistentPtyInfo> }["data"]
@@ -29,6 +29,7 @@ import type {
  SessionMessageAssistantTool,
  SessionInfo,
  SessionInboxInfo,
+  SessionInboxCompaction,
  ShellInfo,
  SkillInfo,
  VcsInfo,
@@ -215,8 +216,10 @@ export function createData(config: CreateDataInput) {
  )
  const messageIndex = new Map<string, Map<string, number>>()
  const sync = createSync()
+  let activeUpdates: Map<string, DataSessionStatus | undefined> | undefined

  function setSessionActive(sessionID: string, status: DataSessionStatus) {
+    activeUpdates?.set(sessionID, status)
    setStore("session", "active", sessionID, status)
  }

@@ -282,12 +285,11 @@ export function createData(config: CreateDataInput) {
    setStore("session", "pending", sessionID, index, { ...item, delivery })
  }

-  // Inbox IDs of optimistic prompt admissions still awaiting their durable
-  // echo. This is the one deliberate piece of in-flight bookkeeping in this
-  // layer: it exists so a rejection only rolls back rows the server never
-  // acknowledged, and so a concurrent pending re-fetch cannot wipe a row the
-  // server does not know about yet. Entries clear on the enqueued echo or on
-  // rollback — not on POST success, which typically precedes the echo.
+  // Inbox IDs of optimistic admissions awaiting acknowledgement, so rejection
+  // only rolls back unacknowledged rows and a pending re-fetch cannot wipe a
+  // row the server does not know about yet. Prompts clear on their durable
+  // echo, positive pending read, or rollback; compactions also reconcile the
+  // POST's canonical ID.
  const outbox = new Set<string>()

  // Session IDs of optimistic create admissions still awaiting acknowledgement
@@ -301,11 +303,12 @@ export function createData(config: CreateDataInput) {
  // to exist server-side instead of failing with "not found".
  const creating = new Map<string, Promise<unknown>>()

-  // Per-session send chain: prompts must be admitted in submission order,
-  // and HTTP gives no ordering across concurrent POSTs. Each prompt waits
-  // for the previous prompt's POST (settled, so one failure does not block
-  // the next) before sending its own.
+  // Per-session send chain: prompts and compactions must be admitted in
+  // submission order. Each waits for the previous POST to settle, so one
+  // failure does not block the next.
  const sending = new Map<string, Promise<unknown>>()
+  const compacting = new Map<string, { id: string; observed: Set<string>; request: Promise<SessionInboxCompaction> }>()
+  onCleanup(() => compacting.clear())

  // Register `promise` under `key` until it settles. A later registration
  // replaces an earlier one; settlement only clears its own entry.
@@ -317,9 +320,24 @@ export function createData(config: CreateDataInput) {
    void promise.then(settle, settle)
  }

+  // Capture creation before settlement clears its entry, so dependent RPCs still see a failed create.
+  function sendAdmission<Value>(sessionID: string, send: () => Promise<Value>, gate?: Promise<unknown>) {
+    const created = creating.get(sessionID)
+    const previous = sending.get(sessionID)
+    const request = Promise.resolve()
+      .then(() => Promise.all([gate, created, previous]))
+      .then(send)
+    track(
+      sending,
+      sessionID,
+      request.catch(() => undefined),
+    )
+    return request
+  }
+
  // Upsert an admitted inbox item into pending, input, and (for user and
  // synthetic items) the visible transcript. Used by the inbox.enqueued
-  // handler and by optimistic prompt admission; the upsert is what reconciles
+  // handler and by optimistic admission; the upsert is what reconciles
  // the durable echo with an optimistic placeholder — the durable payload and
  // times replace the client's guess.
  function admitLocal(item: SessionInboxInfo) {
@@ -332,6 +350,7 @@ export function createData(config: CreateDataInput) {
        item.sessionID,
        at < 0 ? [...pending, item] : pending.map((entry, index) => (index === at ? item : entry)),
      )
+      if (item.type === "compaction") return
      const input = store.session.input[item.sessionID] ?? []
      if (!input.includes(item.id)) setStore("session", "input", item.sessionID, [...input, item.id])
      materializeInboxMessage(item)
@@ -473,6 +492,7 @@ export function createData(config: CreateDataInput) {
  }

  function removeSession(sessionID: string) {
+    activeUpdates?.set(sessionID, undefined)
    store.session.pending[sessionID]?.forEach((item) => outbox.delete(item.id))
    messageIndex.delete(sessionID)
    sync.invalidate(`session:${sessionID}`)
@@ -504,17 +524,25 @@ export function createData(config: CreateDataInput) {

  function handleEvent(event: OpenCodeEvent) {
    switch (event.type) {
-      case "server.connected":
+      case "server.connected": {
+        const updates = new Map<string, DataSessionStatus | undefined>()
+        activeUpdates = updates
        void api()
          .session.active()
          .then((active) => {
-            setStore(
-              "session",
-              "active",
-              reconcile(Object.fromEntries(Object.keys(active).map((sessionID) => [sessionID, "running" as const]))),
-            )
+            if (activeUpdates !== updates) return
+            // Lifecycle events received during hydration supersede the snapshot.
+            const snapshot = new Map<string, DataSessionStatus>(Object.keys(active).map((id) => [id, "running"]))
+            updates.forEach((status, id) => {
+              if (status === undefined) return snapshot.delete(id)
+              snapshot.set(id, status)
+            })
+            activeUpdates = undefined
+            setStore("session", "active", reconcile(Object.fromEntries(snapshot)))
+          })
+          .catch(() => {
+            if (activeUpdates === updates) activeUpdates = undefined
          })
-          .catch(() => undefined)
        void api()
          .location.get({ location: locationQuery(defaultLocation()) })
          .then((location) => {
@@ -525,6 +553,7 @@ export function createData(config: CreateDataInput) {
        void result.location.vcs.sync().catch((error) => console.error("Failed to preload VCS info", error))
        void result.project.sync().catch((error) => console.error("Failed to preload projects", error))
        return
+      }
      case "project.updated":
        setStore("project", "info", event.data.id, reconcile(event.data))
        return
@@ -656,6 +685,7 @@ export function createData(config: CreateDataInput) {
          draft.push(existing)
          message.reindex(draft, index, position)
        })
+        compacting.get(event.data.sessionID)?.observed.add(event.data.inboxID)
        return
      }
      case "session.inbox.delivery.changed":
@@ -663,6 +693,7 @@ export function createData(config: CreateDataInput) {
        return
      case "session.inbox.cancelled": {
        retractLocal(event.data.sessionID, event.data.inboxID)
+        compacting.get(event.data.sessionID)?.observed.add(event.data.inboxID)
        return
      }
      case "session.inbox.enqueued": {
@@ -673,6 +704,12 @@ export function createData(config: CreateDataInput) {
          timeCreated: event.created,
          ...event.data.item,
        })
+        if (event.data.item.type === "compaction") {
+          const active = compacting.get(event.data.sessionID)
+          active?.observed.add(event.data.inboxID)
+          if (active && active.id !== event.data.inboxID && outbox.delete(active.id))
+            removePending(event.data.sessionID, active.id)
+        }
        return
      }
      case "session.instructions.updated":
@@ -971,6 +1008,7 @@ export function createData(config: CreateDataInput) {
            time: { created: event.created },
          })
        })
+        if (event.data.inputID) compacting.get(event.data.sessionID)?.observed.add(event.data.inputID)
        return
      case "session.execution.succeeded":
      case "session.execution.failed":
@@ -1068,6 +1106,7 @@ export function createData(config: CreateDataInput) {
          }
          message.append(draft, index, failed)
        })
+        if (event.data.inputID) compacting.get(event.data.sessionID)?.observed.add(event.data.inputID)
        return
      case "permission.asked":
        if (store.session.permission[event.data.sessionID]?.some((request) => request.id === event.data.id)) return
@@ -1254,12 +1293,17 @@ export function createData(config: CreateDataInput) {
        sync(sessionID: string) {
          return sync.run(`session.pending:${sessionID}`, async () => {
            const pending = await api().session.inbox.list({ sessionID })
+            // A positive read acknowledges admission even when its SSE echo is delayed.
+            pending.forEach((item) => outbox.delete(item.id))
+            // Compactions also coalesce by Session, not just by the proposed ID.
+            if (pending.some((item) => item.type === "compaction"))
+              store.session.pending[sessionID]
+                ?.filter((item) => item.type === "compaction")
+                .forEach((item) => outbox.delete(item.id))
            // Keep optimistic rows still awaiting their echo: this fetch may
            // have raced ahead of an in-flight admission the server does not
            // know about yet.
-            const inflight = (store.session.pending[sessionID] ?? []).filter(
-              (item) => outbox.has(item.id) && !pending.some((row) => row.id === item.id),
-            )
+            const inflight = (store.session.pending[sessionID] ?? []).filter((item) => outbox.has(item.id))
            const merged = inflight.length === 0 ? pending : [...pending, ...inflight]
            batch(() => {
              setStore("session", "pending", sessionID, reconcile(merged))
@@ -1333,13 +1377,56 @@ export function createData(config: CreateDataInput) {
        if (fresh) track(creating, id, request)
        return { id, request }
      },
+      compact(input: { sessionID: string; model?: ModelRef }) {
+        const active = compacting.get(input.sessionID)
+        if (active) return active.request
+        // A known pending control ID may be consumed while setup waits. Propose
+        // a fresh ID and let the server coalesce, without duplicating its row.
+        const id = SessionMessage.ID.create()
+        if (!store.session.pending[input.sessionID]?.some((item) => item.type === "compaction")) {
+          outbox.add(id)
+          admitLocal({
+            id,
+            sessionID: input.sessionID,
+            timeCreated: Date.now(),
+            type: "compaction",
+            delivery: "steer",
+            payload: {},
+          })
+        }
+        // Compaction admission can coalesce onto a different ID. Retire the
+        // speculative row on an echo, and remember consumed IDs until the POST
+        // settles so its older response cannot resurrect a queued row.
+        const observed = new Set<string>()
+        const request = sendAdmission(input.sessionID, async () => {
+          if (input.model) await api().session.switchModel({ sessionID: input.sessionID, model: input.model })
+          return api().session.compact({ sessionID: input.sessionID, id })
+        })
+          .then((item) => {
+            batch(() => {
+              outbox.delete(id)
+              if (item.id !== id) removePending(input.sessionID, id)
+              if (!observed.has(item.id) && !messageIndex.get(input.sessionID)?.has(item.id)) admitLocal(item)
+            })
+            return item
+          })
+          .catch((error) => {
+            if (outbox.delete(id)) removePending(input.sessionID, id)
+            throw error
+          })
+          .finally(() => {
+            if (compacting.get(input.sessionID)?.request === request) compacting.delete(input.sessionID)
+          })
+        compacting.set(input.sessionID, { id, observed, request })
+        return request
+      },
      // Optimistic prompt admission: render the prompt immediately under a
      // client-minted ID, send it, and let the durable inbox.enqueued echo
      // upsert that same ID with the server's payload. Server admission is
      // idempotent per ID, so retrying with the identical payload cannot
      // double-admit.
-      prompt(input: SessionPromptInput & { gate?: Promise<unknown> }) {
-        const { gate, ...request } = input
+      prompt(input: SessionPromptInput & { gate?: Promise<unknown>; prepare?: () => Promise<unknown> }) {
+        const { gate, prepare, ...request } = input
        const id = request.id ?? SessionMessage.ID.create()
        // A retry may reuse an ID that is already rendered — and possibly
        // already durable. Admit optimistically only for new IDs so a failed
@@ -1365,25 +1452,15 @@ export function createData(config: CreateDataInput) {
            },
          })
        }
-        // Wrapped so even a synchronous client failure reaches the rollback.
-        // The POST additionally waits for the caller's gate, for any
-        // in-flight optimistic create of this session, and for the previous
-        // prompt's POST: the row renders now, the send happens once the
-        // session exists server-side and earlier prompts are admitted.
-        const previous = sending.get(request.sessionID)
-        const send = Promise.resolve()
-          .then(() => Promise.all([gate, creating.get(request.sessionID), previous]))
-          .then(() => api().session.prompt({ ...request, id }))
-        track(
-          sending,
+        return sendAdmission(
          request.sessionID,
-          send.then(
-            () => undefined,
-            () => undefined,
-          ),
-        )
-        return send.catch((error) => {
-          // Roll back only rows this call admitted and the echo has not
+          async () => {
+            await prepare?.()
+            return api().session.prompt({ ...request, id })
+          },
+          gate,
+        ).catch((error) => {
+          // Roll back only rows this call admitted and the server has not
          // acknowledged: anything else is server state.
          if (fresh && outbox.delete(id)) retractLocal(request.sessionID, id)
          throw error
@@ -50,6 +50,8 @@ test("exposes every standard HTTP API group", () => {
  expect(Object.keys(client.vcs)).toEqual(["get", "status", "branches", "diff"])
  expect(Object.keys(client.pty)).toEqual(["list", "create", "get", "update", "remove", "connect"])
  expect(Object.keys(client.pty.connect)).toEqual(["token"])
+  expect(Object.keys(client.experimental)).toEqual(["persistentPty"])
+  expect(client.experimental.persistentPty.read).toBeFunction()
  expect(Object.keys(client.shell)).toEqual(["list", "create", "get", "timeout", "output", "remove"])
  expect(Object.keys(client.project)).toEqual(["list", "update", "current"])
  expect(Object.keys(client.worktree)).toEqual(["list", "create", "remove", "refresh"])
@@ -0,0 +1,400 @@
+import { expect, test } from "bun:test"
+import { createRoot } from "solid-js"
+import { createData, type CreateDataInput } from "../src/solid"
+import { OpenCode, type OpenCodeEvent, type SessionInboxCompaction, type SessionInboxInfo } from "../src/promise"
+
+test("admits compaction before model setup and serializes the following prompt", async () => {
+  using fixture = setup()
+  const compact = fixture.data.session.compact({ sessionID, model: { providerID: "demo", id: "model" } })
+  const proposed = fixture.data.session.pending.list(sessionID)[0]
+  expect(proposed).toMatchObject({ type: "compaction", sessionID })
+  expect(fixture.calls).toEqual([])
+  expect(fixture.data.session.message.list(sessionID)).toEqual([])
+  expect(fixture.data.session.status(sessionID)).toBe("idle")
+
+  const prompt = fixture.data.session.prompt({ sessionID, text: "Follow up" })
+  expect(fixture.data.session.message.list(sessionID)).toMatchObject([{ type: "user", text: "Follow up" }])
+  await wait(() => fixture.calls.length === 1)
+  expect(fixture.calls).toEqual(["model"])
+  fixture.model.resolve()
+  await wait(() => fixture.calls.length === 2)
+  expect(fixture.calls).toEqual(["model", "compact"])
+  fixture.response.resolve(Response.json({ data: item(proposed.id) }))
+  await Promise.all([compact, prompt])
+  expect(fixture.calls).toEqual(["model", "compact", "prompt"])
+  expect(fixture.proposals).toEqual([proposed.id])
+})
+
+test("coalesces duplicate gestures until the admission request settles", async () => {
+  using fixture = setup()
+  const first = fixture.data.session.compact({ sessionID })
+  expect(fixture.data.session.compact({ sessionID })).toBe(first)
+  expect(fixture.data.session.pending.list(sessionID)).toHaveLength(1)
+  await wait(() => fixture.calls.length === 1)
+  fixture.response.resolve(Response.json({ data: item("msg_canonical") }))
+  await first
+  expect(fixture.calls).toEqual(["compact"])
+  const next = fixture.data.session.compact({ sessionID })
+  expect(next).not.toBe(first)
+  await next
+  expect(fixture.calls).toEqual(["compact", "compact"])
+  expect(new Set(fixture.proposals).size).toBe(2)
+  expect(fixture.proposals).not.toContain("msg_canonical")
+})
+
+test("substitutes the canonical response ID and reconciles its later echo", async () => {
+  using fixture = setup()
+  const request = fixture.data.session.compact({ sessionID })
+  const proposed = fixture.data.session.pending.list(sessionID)[0].id
+  await fixture.data.session.pending.sync(sessionID)
+  expect(fixture.data.session.pending.list(sessionID).map((row) => row.id)).toEqual([proposed])
+  fixture.response.resolve(Response.json({ data: item("msg_canonical") }))
+  await request
+  expect(fixture.data.session.pending.list(sessionID)).toEqual([item("msg_canonical")])
+  fixture.enqueue("msg_canonical", 20)
+  expect(fixture.data.session.pending.list(sessionID)).toEqual([item("msg_canonical", 20)])
+  expect(fixture.data.session.input.list(sessionID)).toEqual([])
+})
+
+test.each(["proposed", "canonical"])("adopts the %s echo before the response without duplicating it", async (kind) => {
+  using fixture = setup()
+  const request = fixture.data.session.compact({ sessionID })
+  const id = kind === "proposed" ? fixture.data.session.pending.list(sessionID)[0].id : "msg_canonical"
+  fixture.enqueue(id, 20)
+  expect(fixture.data.session.pending.list(sessionID)).toEqual([item(id, 20)])
+  fixture.response.resolve(Response.json({ data: item(id) }))
+  await request
+  expect(fixture.data.session.pending.list(sessionID)).toEqual([item(id, 20)])
+})
+
+test.each(["started", "cancelled", "failed"])(
+  "does not resurrect a canonical item already %s before the response",
+  async (kind) => {
+    using fixture = setup()
+    const request = fixture.data.session.compact({ sessionID })
+    fixture.enqueue("msg_canonical")
+    if (kind === "started")
+      fixture.emit({
+        ...event,
+        type: "session.compaction.started",
+        data: { sessionID, inputID: "msg_canonical", reason: "manual" },
+      })
+    if (kind === "cancelled")
+      fixture.emit({ ...event, type: "session.inbox.cancelled", data: { sessionID, inboxID: "msg_canonical" } })
+    if (kind === "failed")
+      fixture.emit({
+        ...event,
+        type: "session.compaction.failed",
+        data: {
+          sessionID,
+          inputID: "msg_canonical",
+          reason: "manual",
+          error: { type: "aborted", message: "Cancelled" },
+        },
+      })
+    expect(fixture.data.session.pending.list(sessionID)).toEqual([])
+    fixture.response.resolve(Response.json({ data: item("msg_canonical") }))
+    await request
+    expect(fixture.data.session.pending.list(sessionID)).toEqual([])
+    if (kind === "started") {
+      expect(fixture.data.session.message.list(sessionID)).toMatchObject([{ type: "compaction", status: "running" }])
+      fixture.emit({
+        ...event,
+        type: "session.compaction.ended",
+        data: { sessionID, reason: "manual", text: "Summary", recent: "Recent" },
+      })
+      expect(fixture.data.session.message.list(sessionID)).toMatchObject([
+        { type: "compaction", status: "completed", summary: "Summary" },
+      ])
+    }
+  },
+)
+
+test.each(["model", "compact"])("rolls back a rejected %s RPC and releases the following prompt", async (rpc) => {
+  using fixture = setup()
+  const request = fixture.data.session.compact({ sessionID, model: { providerID: "demo", id: "model" } })
+  const failed = request.catch((error: unknown) => error)
+  const prompt = fixture.data.session.prompt({ sessionID, text: "Follow up" })
+  if (rpc === "model") fixture.model.reject(new Error("Model setup failed"))
+  if (rpc === "compact") {
+    fixture.model.resolve()
+    fixture.response.resolve(new Response("Admission failed", { status: 500 }))
+  }
+  expect(await failed).toBeInstanceOf(Error)
+  await prompt
+  expect(fixture.data.session.pending.list(sessionID).map((row) => row.type)).toEqual(["user"])
+  expect(fixture.data.session.message.list(sessionID)).toMatchObject([{ type: "user", text: "Follow up" }])
+})
+
+test.each(["proposed", "canonical", "existing"])(
+  "preserves acknowledged %s compaction after an HTTP error",
+  async (kind) => {
+    using fixture = setup()
+    if (kind === "existing") fixture.enqueue("msg_canonical")
+    const request = fixture.data.session.compact({ sessionID })
+    const failed = request.catch((error: unknown) => error)
+    const id = kind === "proposed" ? fixture.data.session.pending.list(sessionID)[0].id : "msg_canonical"
+    if (kind !== "existing") fixture.enqueue(id)
+    expect(fixture.data.session.pending.list(sessionID)).toEqual([item(id)])
+    fixture.response.resolve(new Response("Lost response", { status: 500 }))
+    expect(await failed).toBeInstanceOf(Error)
+    expect(fixture.data.session.pending.list(sessionID)).toEqual([item(id)])
+    expect(fixture.listeners.size).toBe(1)
+  },
+)
+
+test("uses a fresh control ID when the known pending compaction starts during model setup", async () => {
+  const proposed = Promise.withResolvers<string>()
+  using fixture = setup(async (request) => {
+    if (!request.url.endsWith("/compact")) return undefined
+    const body = await request.json()
+    proposed.resolve(body.id)
+    if (body.id === "msg_existing") return Response.json({ message: "Control ID already consumed" }, { status: 409 })
+    return Response.json({ data: item(body.id) })
+  })
+  fixture.enqueue("msg_existing")
+  const request = fixture.data.session.compact({ sessionID, model: { providerID: "demo", id: "model" } })
+  const result = request.catch((error: unknown) => error)
+  expect(fixture.data.session.pending.list(sessionID)).toEqual([item("msg_existing")])
+  await wait(() => fixture.calls.includes("model"))
+  fixture.emit({
+    ...event,
+    type: "session.compaction.started",
+    data: { sessionID, inputID: "msg_existing", reason: "manual" },
+  })
+  fixture.model.resolve()
+  expect(await proposed.promise).not.toBe("msg_existing")
+  expect(await result).toEqual(item(await proposed.promise))
+  expect(fixture.data.session.pending.list(sessionID)).toEqual([item(await proposed.promise)])
+  expect(fixture.data.session.message.list(sessionID)).toMatchObject([
+    { id: "msg_existing", type: "compaction", status: "running" },
+  ])
+})
+
+test.each(["compaction", "canonical compaction", "user"])(
+  "preserves a fetched durable %s when SSE is delayed and HTTP fails",
+  async (type) => {
+    using fixture = setup(async (request) => {
+      if (request.url.endsWith("/prompt")) return fixture.response.promise
+      return undefined
+    })
+    const request =
+      type === "user"
+        ? fixture.data.session.prompt({ sessionID, text: "Follow up" })
+        : fixture.data.session.compact({ sessionID })
+    const result = request.catch((error: unknown) => error)
+    const id = type === "canonical compaction" ? "msg_canonical" : fixture.data.session.pending.list(sessionID)[0].id
+    const durable: SessionInboxInfo =
+      type === "user" ? { ...item(id, 20), type: "user", payload: { text: "Follow up" } } : item(id, 20)
+    fixture.pending.push(durable)
+    await fixture.data.session.pending.sync(sessionID)
+    expect(fixture.data.session.pending.list(sessionID)).toEqual([durable])
+    fixture.response.resolve(new Response("Lost response", { status: 500 }))
+    expect(await result).toBeInstanceOf(Error)
+    expect(fixture.data.session.pending.list(sessionID)).toEqual([durable])
+    if (type === "user")
+      expect(fixture.data.session.message.list(sessionID)).toMatchObject([{ id, type: "user", text: "Follow up" }])
+  },
+)
+
+test("keeps one event listener and removes it when the data owner is disposed during a gate", async () => {
+  using fixture = setup()
+  const gate = Promise.withResolvers<void>()
+  const first = fixture.data.session.prompt({ sessionID, text: "First", gate: gate.promise })
+  const compact = fixture.data.session.compact({ sessionID })
+  expect(fixture.listeners.size).toBe(1)
+  fixture.dispose()
+  expect(fixture.listeners.size).toBe(0)
+  gate.resolve()
+  fixture.response.resolve(Response.json({ data: item("msg_canonical") }))
+  await Promise.all([first, compact])
+  expect(fixture.listeners.size).toBe(0)
+})
+
+test("routes concurrent compaction observations by session through one listener", async () => {
+  const firstResponse = Promise.withResolvers<Response>()
+  const secondResponse = Promise.withResolvers<Response>()
+  using fixture = setup(async (request) => {
+    if (!request.url.endsWith("/compact")) return undefined
+    return request.url.includes(`/session/${sessionID}/`) ? firstResponse.promise : secondResponse.promise
+  })
+  const first = fixture.data.session.compact({ sessionID })
+  const second = fixture.data.session.compact({ sessionID: "ses_other" })
+  const firstID = fixture.data.session.pending.list(sessionID)[0].id
+  const secondID = fixture.data.session.pending.list("ses_other")[0].id
+  expect(fixture.listeners.size).toBe(1)
+  fixture.emit({ ...event, type: "session.inbox.cancelled", data: { sessionID, inboxID: firstID } })
+  expect(fixture.data.session.pending.list(sessionID)).toEqual([])
+  expect(fixture.data.session.pending.list("ses_other").map((row) => row.id)).toEqual([secondID])
+
+  firstResponse.resolve(Response.json({ data: item(firstID) }))
+  secondResponse.resolve(Response.json({ data: { ...item(secondID), sessionID: "ses_other" } }))
+  await Promise.all([first, second])
+  expect(fixture.data.session.pending.list(sessionID)).toEqual([])
+  expect(fixture.data.session.pending.list("ses_other")).toEqual([{ ...item(secondID), sessionID: "ses_other" }])
+  expect(fixture.listeners.size).toBe(1)
+})
+
+test.each(["gate", "prepare"])(
+  "a preceding prompt's failed %s does not block compaction or following model preparation",
+  async (kind) => {
+    using fixture = setup()
+    const gate = Promise.withResolvers<void>()
+    const prepared: string[] = []
+    const first = fixture.data.session
+      .prompt({
+        sessionID,
+        id: "msg_first",
+        text: "First",
+        gate: kind === "gate" ? gate.promise : undefined,
+        prepare: () => {
+          prepared.push("first")
+          return gate.promise
+        },
+      })
+      .catch((error: unknown) => error)
+    const compact = fixture.data.session.compact({ sessionID, model: { providerID: "demo", id: "first" } })
+    const following = fixture.data.session.prompt({
+      sessionID,
+      text: "Follow up",
+      prepare: () => {
+        prepared.push("following")
+        return fixture.api.session.switchModel({ sessionID, model: { providerID: "demo", id: "second" } })
+      },
+    })
+    if (kind === "prepare") await wait(() => prepared.includes("first"))
+    gate.reject(new Error("Preparation failed"))
+    expect(await first).toBeInstanceOf(Error)
+    await wait(() => fixture.calls.includes("model"))
+    expect(prepared).toEqual(kind === "prepare" ? ["first"] : [])
+    fixture.model.resolve()
+    fixture.response.resolve(Response.json({ data: item("msg_canonical") }))
+    await Promise.all([compact, following])
+    expect(fixture.calls).toEqual(["model", "compact", "model", "prompt"])
+    expect(prepared.at(-1)).toBe("following")
+    expect(fixture.data.session.message.list(sessionID)).toMatchObject([{ type: "user", text: "Follow up" }])
+  },
+)
+
+test("creation failure rejects gated prompt, compaction, and following preparation without sending their RPCs", async () => {
+  const creation = Promise.withResolvers<Response>()
+  const requested = Promise.withResolvers<void>()
+  using fixture = setup(async (request) => {
+    if (!request.url.endsWith("/api/session")) return undefined
+    requested.resolve()
+    return creation.promise
+  })
+  const gate = Promise.withResolvers<void>()
+  const prepared: string[] = []
+  const created = fixture.data.session.create({ id: sessionID })
+  const first = fixture.data.session.prompt({ sessionID, text: "First", gate: gate.promise })
+  const compact = fixture.data.session.compact({ sessionID, model: { providerID: "demo", id: "model" } })
+  const following = fixture.data.session.prompt({
+    sessionID,
+    text: "Follow up",
+    prepare: async () => {
+      prepared.push("following")
+    },
+  })
+  const results = Promise.allSettled([created.request, first, compact, following])
+  await requested.promise
+  creation.resolve(new Response("Creation failed", { status: 500 }))
+  expect((await results).map((result) => result.status)).toEqual(["rejected", "rejected", "rejected", "rejected"])
+  expect(fixture.calls).toEqual([])
+  expect(prepared).toEqual([])
+  expect(fixture.data.session.get(sessionID)).toBeUndefined()
+  expect(fixture.data.session.pending.list(sessionID)).toEqual([])
+  expect(fixture.listeners.size).toBe(1)
+  gate.resolve()
+})
+
+const sessionID = "ses_compact"
+const event = { id: "evt_compact", created: 10, durable: { aggregateID: sessionID, seq: 1, version: 1 } }
+const item = (id: string, timeCreated = 10): SessionInboxCompaction => ({
+  id,
+  sessionID,
+  timeCreated,
+  type: "compaction",
+  delivery: "steer",
+  payload: {},
+})
+
+function setup(override?: (request: Request) => Promise<Response | undefined>) {
+  const model = Promise.withResolvers<void>()
+  const response = Promise.withResolvers<Response>()
+  const calls: string[] = []
+  const proposals: string[] = []
+  const pending: SessionInboxInfo[] = []
+  const listeners = new Set<Parameters<CreateDataInput["event"]["listen"]>[0]>()
+  const api = OpenCode.make({
+    baseUrl: "http://opencode.local",
+    fetch: async (input, init) => {
+      const request = input instanceof Request ? input : new Request(input, init)
+      const overridden = await override?.(request)
+      if (overridden) return overridden
+      const rpc = new URL(request.url).pathname.split("/").at(-1)
+      if (rpc === "inbox") return Response.json({ data: pending })
+      if (rpc === "model") {
+        calls.push(rpc)
+        await model.promise
+        return new Response(null, { status: 204 })
+      }
+      if (rpc === "compact") {
+        calls.push(rpc)
+        proposals.push((await request.json()).id)
+        return (await response.promise).clone()
+      }
+      if (rpc === "prompt") {
+        calls.push(rpc)
+        return Response.json({
+          data: { ...item((await request.json()).id), type: "user", payload: { text: "Follow up" } },
+        })
+      }
+      throw new Error(`Unexpected request: ${request.url}`)
+    },
+  })
+  const root = createRoot((dispose) => ({
+    data: createData({
+      api: () => api,
+      directory: "/project",
+      event: {
+        on: () => () => {},
+        listen(handler) {
+          listeners.add(handler)
+          return () => listeners.delete(handler)
+        },
+      },
+    }),
+    dispose,
+  }))
+  const emit = (details: OpenCodeEvent) => listeners.forEach((listener) => listener({ name: details.type, details }))
+  return {
+    data: root.data,
+    api,
+    dispose: root.dispose,
+    [Symbol.dispose]: root.dispose,
+    model,
+    response,
+    calls,
+    proposals,
+    pending,
+    listeners,
+    emit,
+    enqueue(id: string, created = 10) {
+      emit({
+        ...event,
+        created,
+        type: "session.inbox.enqueued",
+        data: { sessionID, inboxID: id, item: { type: "compaction", delivery: "steer", payload: {} } },
+      })
+    },
+  }
+}
+
+async function wait(predicate: () => boolean) {
+  for (let attempt = 0; attempt < 100; attempt++) {
+    if (predicate()) return
+    await Bun.sleep(5)
+  }
+  throw new Error("Timed out waiting for request")
+}
@@ -485,6 +485,102 @@ test("preserves assistant content replacement events across an active message re
  }
 })

+test.each([
+  "session.execution.succeeded",
+  "session.execution.failed",
+  "session.execution.interrupted",
+  "session.execution.started",
+  "session.deleted",
+] as const)("preserves %s activity when an older snapshot arrives", async (type) => {
+  const release = Promise.withResolvers<void>()
+  const requested = Promise.withResolvers<void>()
+  const setup = activityFixture(async () => {
+    requested.resolve()
+    await release.promise
+    return Response.json({
+      data: {
+        ...(type === "session.execution.started" ? {} : { ses_refresh: { type: "running" } }),
+        ses_hydrated: { type: "running" },
+      },
+    })
+  })
+
+  try {
+    if (type !== "session.execution.started") setup.data.session.setStatus("ses_refresh", "running")
+    setup.emit({ type: "server.connected", data: {} })
+    await requested.promise
+    setup.emit({
+      id: "evt_activity",
+      created: 2,
+      type,
+      durable: { aggregateID: "ses_refresh", seq: 2, version: 1 },
+      data: { sessionID: "ses_refresh", reason: "user" },
+    })
+    expect(setup.data.session.status("ses_refresh")).toBe(type === "session.execution.started" ? "running" : "idle")
+    release.resolve()
+    await wait(() => setup.data.session.status("ses_hydrated") === "running")
+    expect(setup.data.session.status("ses_refresh")).toBe(type === "session.execution.started" ? "running" : "idle")
+  } finally {
+    release.resolve()
+    setup.dispose()
+  }
+})
+
+test("ignores activity snapshots from an older connection", async () => {
+  const reads: ReturnType<typeof Promise.withResolvers<Response>>[] = []
+  const setup = activityFixture(() => {
+    const read = Promise.withResolvers<Response>()
+    reads.push(read)
+    return read.promise
+  })
+
+  try {
+    setup.emit({ type: "server.connected", data: {} })
+    await wait(() => reads.length === 1)
+    setup.emit({ type: "server.connected", data: {} })
+    await wait(() => reads.length === 2)
+    reads[1]?.resolve(Response.json({ data: { ses_new: { type: "running" } } }))
+    await wait(() => setup.data.session.status("ses_new") === "running")
+    reads[0]?.resolve(Response.json({ data: { ses_old: { type: "running" } } }))
+    await Bun.sleep(20)
+    expect(setup.data.session.status("ses_new")).toBe("running")
+    expect(setup.data.session.status("ses_old")).toBe("idle")
+  } finally {
+    reads.forEach((read) => read.resolve(Response.json({ data: {} })))
+    setup.dispose()
+  }
+})
+
+function activityFixture(read: () => Response | Promise<Response>) {
+  const listeners = new Set<Parameters<CreateDataInput["event"]["listen"]>[0]>()
+  const api = OpenCode.make({
+    baseUrl: "http://opencode.local",
+    fetch: async (input, init) => {
+      const request = input instanceof Request ? input : new Request(input, init)
+      const path = new URL(request.url).pathname
+      if (path === "/api/session/active") return read()
+      if (path === "/api/project") return Response.json([])
+      if (path === "/api/location") return Response.json({ directory: "/project" })
+      return Response.json({ location: { directory: "/project" }, data: { branch: "main" } })
+    },
+  })
+  return createRoot((dispose) => ({
+    data: createData({
+      api: () => api,
+      directory: "/project",
+      event: {
+        on: () => () => {},
+        listen(handler) {
+          listeners.add(handler)
+          return () => listeners.delete(handler)
+        },
+      },
+    }),
+    emit: (details: OpenCodeEvent) => listeners.forEach((listener) => listener({ name: details.type, details })),
+    dispose,
+  }))
+}
+
 async function wait(check: () => boolean) {
  const started = Date.now()
  while (!check()) {
@@ -19,7 +19,7 @@

 - If a captured user-visible output channel returns (an earlier `output.text`/`output.file`/`output.image` API was removed from v1), keep `output` as its name, distinct from the program return value: `return` stays the structured result for the model, while `output.*` describes artifacts the host may render into a conversation or UI after execution. Keep this host-neutral and let applications decide how captured output is delivered. In v1, hosts collect media host-side (outside CodeMode) instead.
 - Improve the failure taxonomy. Distinguish parse/compile mistakes, unsupported syntax, user-thrown errors, invalid returned data, tool refusal, tool internal failure, timeout, and genuine runtime defects so agents can recover accurately instead of treating everything as a generic execution failure.
- Preserve the public/private error split. Tool authors should be able to return a safe model-visible message while retaining a private cause for host diagnostics. Unknown host failures must remain sanitized by default.
+- Report host failure messages and underlying causes rather than replacing them with generic diagnostics. Preserve interruption behavior.
 - Think deliberately about richer binary boundaries before allowing `Blob`, `File`, `ArrayBuffer`, streams, or typed arrays beyond today's JSON-like values. If CodeMode supports binary tool args/results, use explicit tagged data shapes and clear size limits rather than relying on ambient runtime serialization.
 - Keep host capabilities explicit. Globals such as `fetch`, `crypto`, filesystem handles, extra modules, or network clients should be opt-in runtime capabilities with obvious policy defaults, not ambient authority. Default to unavailable unless a host deliberately provides the capability.
 - If `fetch` is added, model it as a host-provided outbound capability with policy controls: allowed origins, methods, headers, response size, timeout, and whether response bodies may be returned, emitted, or only summarized through a tool.
@@ -139,8 +139,8 @@ Diagnostic kinds:
 | `ExecutionFailure`      | The program threw or another execution error occurred.                                         |
 | `Truncated`             | Warning only: additional warnings were omitted by `maxOutputBytes`.                            |

-Unknown host failures, defects, and invalid outputs are sanitized. `toolError("safe message")` explicitly exposes a
-safe refusal to the model; its optional cause remains private.
+Host failures and defects report their messages and underlying causes. Invalid outputs include the validation or
+copying error. Interruption propagates without becoming an error diagnostic.

 ## Discovery

@@ -365,7 +365,7 @@ ultimate source of truth.
      `catch`.
 - [x] Source locations on unsupported-syntax diagnostics for JavaScript-shaped input; TypeScript transpilation may
      shift them.
- [x] Sanitized model-visible diagnostics and explicit safe `ToolError` messages.
- [ ] Distinguish user-thrown failures from interpreter defects and explicit tool refusals from sanitized internal tool
+- [x] Model-visible host failure messages and underlying causes, including output-validation errors.
+- [ ] Distinguish user-thrown failures from interpreter defects and explicit tool refusals from internal tool
      failures; preserve those categories in caught errors, promise rejection handlers, and `Promise.allSettled`
      reasons.
@@ -1,11 +1,11 @@
 import { Schema } from "effect"

-/** Safe operational refusal from a standard tool pack, reported as `ToolFailure`. */
+/** Tool failure reported as `ToolFailure`. */
 export class ToolError extends Schema.TaggedError<ToolError>()("ToolError", {
  message: Schema.String,
  cause: Schema.optionalKey(Schema.Defect()),
 }) {}

-/** Creates a tool refusal whose message is safe to include in an execution diagnostic. */
+/** Creates a tool failure with an optional underlying cause. */
 export const toolError = (message: string, cause?: unknown): ToolError =>
  new ToolError({ message, ...(cause === undefined ? {} : { cause }) })
@@ -1,5 +1,5 @@
-import { Cause, Effect, Exit, Schema } from "effect"
-import { ToolError, toolError } from "./tool-error.js"
+import { Cause, Effect, Exit, Formatter, Schema } from "effect"
+import { toolError } from "./tool-error.js"
 import {
  decodeInput as decodeToolInput,
  decodeOutput as decodeToolOutput,
@@ -117,15 +117,6 @@ export class ToolRuntimeError extends Error {
  }
 }

-const runHost = <A, E, R>(effect: Effect.Effect<A, E, R>): Effect.Effect<A, ToolError, R> =>
-  effect.pipe(
-    Effect.catchCause((cause) => {
-      if (Cause.hasInterruptsOnly(cause)) return Effect.interrupt
-      const error = Cause.squash(cause)
-      return Effect.fail(error instanceof ToolError ? error : toolError("Tool execution failed", error))
-    }),
-  )
-
 const blockedMemberNames = new Set(["__proto__", "constructor", "prototype"])

 export const isBlockedMember = (name: string): boolean => blockedMemberNames.has(name)
@@ -507,19 +498,12 @@ export const make = <R>(
        if (Exit.isSuccess(exit)) return onEnd({ ...call, durationMs, outcome: "success" })
        if (Cause.hasInterruptsOnly(exit.cause)) return onEnd({ ...call, durationMs, outcome: "interrupted" })
        const error = Cause.squash(exit.cause)
-        const message =
-          error instanceof ToolError || error instanceof ToolRuntimeError ? error.message : "Tool execution failed"
+        const message = error instanceof Error ? error.message : Cause.pretty(exit.cause)
        return onEnd({ ...call, durationMs, outcome: "failure", message })
      }),
    )
  }

-  const decodeOutput = (value: unknown, name: string) =>
-    Effect.try({
-      try: () => copyIn(value, `Result from tool '${name}'`),
-      catch: () => new ToolRuntimeError("InvalidToolOutput", `Invalid output from tool '${name}'.`),
-    })
-
  const recordCall = (call: ToolCall): void => {
    if (maxToolCalls !== undefined && calls.length >= maxToolCalls) {
      throw new ToolRuntimeError("ToolCallLimitExceeded", `Execution exceeded its tool-call limit of ${maxToolCalls}.`)
@@ -548,12 +532,22 @@ export const make = <R>(
      return yield* observeEnd(
        Effect.gen(function* () {
          if (hooks?.onToolCallStart !== undefined) yield* hooks.onToolCallStart(call)
-          const raw = yield* runHost(Effect.suspend(() => tool.execute(input)))
-          const result = yield* Effect.try({
-            try: () => decodeToolOutput(tool, raw),
-            catch: () => new ToolRuntimeError("InvalidToolOutput", `Invalid output from tool '${name}'.`),
+          const raw = yield* Effect.suspend(() => tool.execute(input)).pipe(
+            Effect.catchCause((cause) => {
+              if (Cause.hasInterruptsOnly(cause)) return Effect.interrupt
+              return Effect.fail(
+                toolError(
+                  Cause.prettyErrors(cause)
+                    .map((error) => (error.cause ? Formatter.format(error) : error.message || error.name))
+                    .join("\n"),
+                ),
+              )
+            }),
+          )
+          return yield* Effect.try({
+            try: () => copyIn(decodeToolOutput(tool, raw), `Result from tool '${name}'`),
+            catch: (cause) => new ToolRuntimeError("InvalidToolOutput", `Invalid output from tool '${name}': ${cause}`),
          })
-          return yield* decodeOutput(result, name)
        }),
        call,
      )
@@ -5,15 +5,15 @@ import { CodeMode, Tool, toolError } from "../src/index.js"
 const run = (tool: Tool.Tool<never>) =>
  Effect.runPromise(CodeMode.make({ tools: { host: { call: tool } } }).execute("return await tools.host.call({})"))

-class UnsafeHostError extends Schema.TaggedError<UnsafeHostError>()("UnsafeHostError", {
-  reason: Schema.String,
+class HostError extends Schema.TaggedError<HostError>()("HostError", {
+  message: Schema.String,
 }) {}

 describe("CodeMode host failure boundary", () => {
-  test("preserves explicit safe tool failures", async () => {
+  test("preserves explicit tool failures", async () => {
    const result = await run(
      Tool.make({
-        description: "Fail safely",
+        description: "Fail",
        input: Schema.Struct({}),
        output: Schema.String,
        execute: () => Effect.fail(toolError("Authorized request was refused")),
@@ -26,10 +26,10 @@ describe("CodeMode host failure boundary", () => {
    })
  })

-  test("does not rewrite explicit safe tool failures", async () => {
+  test("does not rewrite explicit tool failures", async () => {
    const result = await run(
      Tool.make({
-        description: "Fail safely",
+        description: "Fail",
        input: Schema.Struct({}),
        output: Schema.String,
        execute: () => Effect.fail(toolError("File not found: /tmp/report.json")),
@@ -42,10 +42,15 @@ describe("CodeMode host failure boundary", () => {
    })
  })

-  test("sanitizes unknown host failures and defects", async () => {
+  test("reports failures, defects, rejected Promises, and nested causes", async () => {
    for (const failure of [
-      Effect.fail(new UnsafeHostError({ reason: "Authorization: Bearer typed-secret" })),
-      Effect.die(new Error("postgres://user:defect-secret@example.invalid")),
+      Effect.fail(new HostError({ message: "Connection refused" })),
+      Effect.die(new Error("Connection refused")),
+      Effect.promise(async () => {
+        throw new Error("Connection refused")
+      }),
+      Effect.fail(toolError("Request failed", new Error("Connection refused"))),
+      Effect.failCause(Cause.combine(Cause.fail("Request failed"), Cause.die("Connection refused"))),
    ]) {
      const result = await run(
        Tool.make({
@@ -58,31 +63,28 @@ describe("CodeMode host failure boundary", () => {

      expect(result.ok ? undefined : result.error).toStrictEqual({
        kind: "ToolFailure",
-        message: "Tool execution failed",
+        message: expect.stringContaining("Connection refused"),
      })
-      expect(JSON.stringify(result)).not.toMatch(/typed-secret|defect-secret|Authorization: Bearer/)
    }
  })

-  test("sanitizes invalid host output", async () => {
-    const secret = "invalid-output-secret"
+  test("reports invalid host output", async () => {
    const result = await run(
      Tool.make({
        description: "Return invalid output",
        input: Schema.Struct({}),
-        output: Schema.Struct({ safe: Schema.String }),
-        execute: () => Effect.succeed({ safe: 1, secret } as unknown as { readonly safe: string }),
+        output: Schema.Struct({ value: Schema.String }),
+        execute: () => Effect.succeed({ value: 1 } as unknown as { readonly value: string }),
      }),
    )

    expect(result.ok ? undefined : result.error).toStrictEqual({
      kind: "InvalidToolOutput",
-      message: "Invalid output from tool 'host.call'.",
+      message: "Invalid output from tool 'host.call': SchemaError(Expected string\n  at [\"value\"])",
    })
-    expect(JSON.stringify(result)).not.toMatch(/invalid-output-secret/)
  })

-  test("sanitizes host output that throws while being copied", async () => {
+  test("reports host output copying errors", async () => {
    const result = await run(
      Tool.make({
        description: "Return hostile output",
@@ -94,7 +96,7 @@ describe("CodeMode host failure boundary", () => {
              {},
              {
                ownKeys: () => {
-                  throw new Error("host-output-secret")
+                  throw new Error("Cannot enumerate output")
                },
              },
            ),
@@ -104,9 +106,8 @@ describe("CodeMode host failure boundary", () => {

    expect(result.ok ? undefined : result.error).toStrictEqual({
      kind: "InvalidToolOutput",
-      message: "Invalid output from tool 'host.call'.",
+      message: "Invalid output from tool 'host.call': Error: Cannot enumerate output",
    })
-    expect(JSON.stringify(result)).not.toMatch(/host-output-secret/)
  })

  test("caught tool failures are Error values in-program", async () => {
@@ -229,7 +230,7 @@ describe("CodeMode tool-call observation", () => {
      { phase: "start", index: 0, name: "context.lookup" },
      { phase: "end", index: 0, name: "context.lookup", outcome: "failure", message: "Lookup refused" },
      { phase: "start", index: 0, name: "context.lookup" },
-      { phase: "end", index: 0, name: "context.lookup", outcome: "failure", message: "Tool execution failed" },
+      { phase: "end", index: 0, name: "context.lookup", outcome: "failure", message: "broken" },
    ])
  })

@@ -84,10 +84,7 @@ export function map(input: MapInput): Mapping | undefined {
          ...mapAPIKey(input.settings),
          ...(typeof input.settings.location === "string" ? { location: input.settings.location } : {}),
          ...(typeof input.settings.project === "string" ? { project: input.settings.project } : {}),
-          ...mapGoogleOptions(
-            input.settings,
-            isStringRecord(input.settings.labels) ? { labels: input.settings.labels } : {},
-          ),
+          ...mapGoogleOptions(input.settings),
        },
        ...(isStringRecord(input.settings.headers) ? { headers: input.settings.headers } : {}),
      }
@@ -296,7 +293,7 @@ function mapAPIKey(settings: Readonly<Record<string, unknown>>) {
  return typeof settings.apiKey === "string" ? { apiKey: settings.apiKey } : {}
 }

-function mapGoogleOptions(settings: Readonly<Record<string, unknown>>, extra: Readonly<Record<string, unknown>> = {}) {
+function mapGoogleOptions(settings: Readonly<Record<string, unknown>>) {
  const input = settings.thinkingConfig
  const thinkingConfig = {
    ...(isRecord(input) && typeof input.thinkingBudget === "number" ? { thinkingBudget: input.thinkingBudget } : {}),
@@ -311,7 +308,6 @@ function mapGoogleOptions(settings: Readonly<Record<string, unknown>>, extra: Re
    ...(Array.isArray(settings.safetySettings) ? { safetySettings: settings.safetySettings } : {}),
    ...(typeof settings.serviceTier === "string" ? { serviceTier: settings.serviceTier } : {}),
    ...(Object.keys(thinkingConfig).length > 0 ? { thinkingConfig } : {}),
-    ...extra,
  }
  if (Object.keys(options).length === 0) return {}
  return { providerOptions: options }
@@ -345,28 +341,21 @@ function mapOpenRouter(
 }

 function mapOpenRouterOptions(settings: Readonly<Record<string, unknown>>) {
-  const options = Object.fromEntries(
-    Object.entries(settings).filter(
-      ([key]) =>
-        ![
-          "apiKey",
-          "api_keys",
-          "appName",
-          "appUrl",
-          "authToken",
-          "baseURL",
-          "chunkTimeout",
-          "compatibility",
-          "extraBody",
-          "fetch",
-          "headers",
-          "promptCacheKey",
-          "timeout",
-        ].includes(key),
-    ),
-  )
-  if (Object.keys(options).length === 0) return {}
-  return { providerOptions: options }
+  return mapProviderOptions(settings, [
+    "apiKey",
+    "api_keys",
+    "appName",
+    "appUrl",
+    "authToken",
+    "baseURL",
+    "chunkTimeout",
+    "compatibility",
+    "extraBody",
+    "fetch",
+    "headers",
+    "promptCacheKey",
+    "timeout",
+  ])
 }

 function isStringRecord(value: unknown): value is Readonly<Record<string, string>> {
@@ -514,7 +514,7 @@ function userPart(part: ContentPart): UserContent {
 function assistantPart(part: ContentPart): AssistantContent {
  switch (part.type) {
    case "text":
-      return [{ type: "text", text: part.text }]
+      return [{ type: "text", text: part.text, providerOptions: metadataProviderOptions(part.providerMetadata) }]
    case "media":
      return [{ type: "file", mediaType: part.mediaType, data: part.data, filename: part.filename }]
    case "reasoning":
@@ -612,12 +612,12 @@ function streamLanguage(language: LanguageModelV3, options: LanguageModelV3CallO
    Stream.unwrap(
      Effect.tryPromise({
        try: () => language.doStream(options),
-        catch: llmError,
+        catch: (error) => llmError(error, "request"),
      }).pipe(
        Effect.map((result) =>
          Stream.fromReadableStream({
            evaluate: () => result.stream,
-            onError: llmError,
+            onError: (error) => llmError(error, "read"),
          }).pipe(
            Stream.mapEffect((event) => streamPartEvents(state, event)),
            Stream.flatMap((events) => Stream.fromIterable(events)),
@@ -744,7 +744,7 @@ function streamPartEvents(
        }),
      ])
    case "error":
-      return Effect.fail(llmError(event.error))
+      return Effect.fail(llmError(event.error, "read"))
  }
 }

@@ -794,9 +794,20 @@ function messageValue(input: unknown) {
  }
 }

-function llmError(error: unknown) {
+function llmError(error: unknown, operation: "request" | "read") {
  if (error instanceof AIError) return error
  if (APICallError.isInstance(error)) return apiCallError(error)
+  const network = networkFailure(error)
+  if (network)
+    return new AIError({
+      reason: new TransportError({
+        message: network.message.trim() === "" ? unknownErrorMessage(error) : network.message,
+        cause: error,
+        transport: "http",
+        operation,
+        code: network.code,
+      }),
+    })
  return new AIError({
    reason: new UnknownProviderError({
      message: unknownErrorMessage(error),
@@ -806,6 +817,50 @@ function llmError(error: unknown) {
  })
 }

+// Runtime-generated network failure shapes. The codes mirror the AI SDK's own
+// Bun network error list in handleFetchError; the messages are undici's fetch
+// TypeError and stream termination strings plus our SSE chunk timeout error.
+// Unrecognized shapes still retry via the UnknownProvider default; this match
+// only adds transport semantics (continuation eligibility, display).
+const NETWORK_ERROR_CODES = new Set([
+  "ECONNRESET",
+  "ECONNREFUSED",
+  "ETIMEDOUT",
+  "EPIPE",
+  "ConnectionRefused",
+  "ConnectionClosed",
+  "FailedToOpenSocket",
+])
+const NETWORK_ERROR_MESSAGES = new Set([
+  "fetch failed",
+  "failed to fetch",
+  "terminated",
+  "other side closed",
+  "sse read timed out",
+])
+
+const NativeErrorShape = Schema.Struct({
+  message: Schema.String,
+  code: Schema.optionalKey(Schema.String),
+  cause: Schema.optionalKey(Schema.Unknown),
+})
+const decodeNativeErrorShape = Schema.decodeUnknownOption(NativeErrorShape)
+
+function networkFailure(error: unknown, depth = 0): { message: string; code?: string } | undefined {
+  if (depth > 4) return undefined
+  const shape = Option.getOrUndefined(decodeNativeErrorShape(error))
+  if (!shape) return undefined
+  // Prefer the deepest match: wrappers like undici's "fetch failed" TypeError
+  // carry the specific network code on their cause.
+  const cause = networkFailure(shape.cause, depth + 1)
+  if (cause) return cause
+  if (shape.code !== undefined && (NETWORK_ERROR_CODES.has(shape.code) || shape.code.startsWith("UND_ERR")))
+    return { message: shape.message, code: shape.code }
+  if (NETWORK_ERROR_MESSAGES.has(shape.message.trim().toLowerCase()))
+    return { message: shape.message, code: shape.code }
+  return undefined
+}
+
 function apiCallError(error: APICallError) {
  const failure = RequestExecutor.httpFailure({
    message: providerErrorMessage(error),
@@ -65,9 +65,7 @@ export const create = (
          (name, tool, input) =>
            Effect.gen(function* () {
              const index = yield* Ref.getAndUpdate(callIndex, (index) => index + 1)
-              const executed = yield* executeTool(name, tool, input, context).pipe(
-                Effect.mapError((failure) => toolError(failure.message, failure)),
-              )
+              const executed = yield* executeTool(name, tool, input, context)
              const content =
                typeof executed.content === "string"
                  ? [{ type: "text" as const, text: executed.content }]
@@ -4,9 +4,7 @@ import { makeLocationNode } from "@opencode-ai/util/effect/app-node"
 import path from "path"
 import { isDeepStrictEqual } from "node:util"
 import { type ParseError, parse } from "jsonc-parser"
-import { applyEdits, modify } from "jsonc-parser"
 import { Context, Effect, Layer, Option, PubSub, Ref, Schema, Semaphore, Stream } from "effect"
-import { produce, type Draft } from "immer"
 import {
  AgentsDirectory,
  ClaudeDirectory,
@@ -16,7 +14,6 @@ import {
  type Entry,
  Event,
 } from "@opencode-ai/schema/config"
-import { isRecord } from "@opencode-ai/ai/utils/record"
 import { Credential } from "./credential.js"
 import { Bus } from "./bus.js"
 import { Watcher } from "./filesystem/watcher.js"
@@ -36,8 +33,6 @@ export function latest<K extends keyof Info>(entries: readonly Entry[], key: K):
 export interface Interface {
  /** Returns location config documents and discovery sources from lowest to highest priority. */
  readonly entries: () => Effect.Effect<Entry[]>
-  /** Updates the first file-backed configuration document. */
-  readonly update: (update: (draft: Draft<Info>) => void) => Effect.Effect<Info, UpdateError>
  /**
   * Streams raw filesystem updates under config roots. Config owns root
   * topology and watch reconciliation; domain owners filter this feed for the
@@ -46,13 +41,11 @@ export interface Interface {
  readonly changes: () => Stream.Stream<Watcher.Update>
 }

-export class UpdateError extends Schema.TaggedError<UpdateError>()("Config.UpdateError", {
-  message: Schema.String,
-  cause: Schema.optional(Schema.Defect()),
-}) {}
-
 export const Options = Schema.Struct({
  project: Schema.optional(Schema.Boolean),
+  // false skips the global config dir, ~/.claude, and ~/.agents; wellknown,
+  // file, and content entries still load.
+  global: Schema.optional(Schema.Boolean),
  file: Schema.optional(Schema.String),
  content: Schema.optional(Schema.String),
 })
@@ -77,20 +70,6 @@ export const testLayer = (initial: Entry[] = []) =>
      const updates = yield* PubSub.unbounded<Watcher.Update>()
      const service = Test.of({
        entries: () => Ref.get(entries),
-        update: (update) =>
-          Effect.gen(function* () {
-            const current = yield* Ref.get(entries)
-            const index = current.findIndex((entry) => entry.type === "document" && entry.path !== undefined)
-            const entry = current[index]
-            if (!entry || entry.type !== "document")
-              return yield* Effect.fail(new UpdateError({ message: "No editable config document found" }))
-            const info = yield* Effect.try({
-              try: () => produce(entry.info, update),
-              catch: (cause) => new UpdateError({ message: "Config update failed", cause }),
-            })
-            yield* Ref.set(entries, current.with(index, new Document({ type: "document", path: entry.path, info })))
-            return info
-          }),
        changes: () => Stream.fromPubSub(updates),
        setEntries: (next) => Ref.set(entries, next),
        emitChange: (update) => PubSub.publish(updates, update).pipe(Effect.asVoid),
@@ -220,28 +199,39 @@ export const layer = (options?: Options) =>
                })
                .pipe(Effect.orDie)

+        const globalEnabled = options?.global !== false
+        // A walked path that resolves into a global root is global config
+        // however the walk reached it (home above the project, or a location
+        // beneath the global config dir), so global: false excludes it
+        // uniformly — classified once here, not per consumer below.
+        const globalRoots = [globalDirectory, globalClaudeDirectory, globalAgentsDirectory].map((item) =>
+          path.resolve(item),
+        )
+        const visible = globalEnabled
+          ? discovered
+          : discovered.filter((item) => {
+              const resolved = path.resolve(item)
+              return !globalRoots.some((root) => resolved === root || resolved.startsWith(root + path.sep))
+            })
        // We load certain files from a few other folders in the ecosystem
        const claude = [
          ...new Set([
-            ...((yield* fs.isDir(globalClaudeDirectory)) ? [globalClaudeDirectory] : []),
-            ...discovered.filter((item) => path.basename(item) === ".claude").toReversed(),
+            ...(globalEnabled && (yield* fs.isDir(globalClaudeDirectory)) ? [globalClaudeDirectory] : []),
+            ...visible.filter((item) => path.basename(item) === ".claude").toReversed(),
          ]),
        ].map((directory) => new ClaudeDirectory({ type: "claude", path: AbsolutePath.make(directory) }))
        const agents = [
          ...new Set([
-            ...((yield* fs.isDir(globalAgentsDirectory)) ? [globalAgentsDirectory] : []),
-            ...discovered.filter((item) => path.basename(item) === ".agents").toReversed(),
+            ...(globalEnabled && (yield* fs.isDir(globalAgentsDirectory)) ? [globalAgentsDirectory] : []),
+            ...visible.filter((item) => path.basename(item) === ".agents").toReversed(),
          ]),
        ].map((directory) => new AgentsDirectory({ type: "agents", path: AbsolutePath.make(directory) }))

-        const directories = [
-          globalDirectory,
-          ...discovered
-            .filter((item) => path.basename(item) === ".opencode")
-            .toReversed()
-            .map((directory) => AbsolutePath.make(directory)),
-        ]
-        const directPaths = discovered
+        const projectDirectories = visible
+          .filter((item) => path.basename(item) === ".opencode")
+          .toReversed()
+          .map((directory) => AbsolutePath.make(directory))
+        const directPaths = visible
          .filter((item) => ![".agents", ".claude", ".opencode"].includes(path.basename(item)))
          .toReversed()
        fileTargets.clear()
@@ -273,15 +263,21 @@ export const layer = (options?: Options) =>
              )
            : []

-        const supplementary = yield* Effect.forEach(directories, loadDirectory).pipe(Effect.orDie)
+        // Global entries sit below explicit and direct files; project
+        // directories rank above them.
+        const globalSupplementary = globalEnabled ? yield* loadDirectory(globalDirectory).pipe(Effect.orDie) : []
+        const projectSupplementary = yield* Effect.forEach(projectDirectories, loadDirectory).pipe(
+          Effect.orDie,
+          Effect.map((entries) => entries.flat()),
+        )
        return [
          ...(yield* loadWellknown().pipe(Effect.orDie)),
          ...claude,
          ...agents,
-          ...(supplementary[0] ?? []),
+          ...globalSupplementary,
          ...explicit,
          ...direct,
-          ...supplementary.slice(1).flat(),
+          ...projectSupplementary,
          ...content,
        ]
      })
@@ -380,54 +376,10 @@ export const layer = (options?: Options) =>
      )
      yield* reconcile(initial)

-      const update = Effect.fn("Config.update")((mutate: (draft: Draft<Info>) => void) =>
-        reloadLock.withPermit(
-          Effect.gen(function* () {
-            // TODO: Replace entry-order selection with an explicit config scope/target model.
-            const document = configs.find((entry) => entry.type === "document" && entry.path !== undefined)
-            if (!document || document.type !== "document" || !document.path)
-              return yield* Effect.fail(new UpdateError({ message: "No editable config document found" }))
-            const next = yield* Effect.try({
-              try: () => produce(document.info, mutate),
-              catch: (cause) => new UpdateError({ message: "Config update failed", cause }),
-            })
-            const edits = changes(document.info, next)
-            if (!edits.length) return document.info
-            const text = yield* fs
-              .readFileString(document.path)
-              .pipe(
-                Effect.mapError(
-                  (cause) => new UpdateError({ message: `Failed to read config: ${document.path}`, cause }),
-                ),
-              )
-            const updated = edits.reduce(
-              (text, edit) =>
-                applyEdits(
-                  text,
-                  modify(text, edit.path, edit.value, { formattingOptions: { tabSize: 2, insertSpaces: true } }),
-                ),
-              text,
-            )
-            const info = yield* parseInfo(updated, document.path)
-            if (!info)
-              return yield* Effect.fail(new UpdateError({ message: `Invalid config update: ${document.path}` }))
-            const temporary = document.path + ".tmp"
-            yield* fs.writeFileString(temporary, updated.endsWith("\n") ? updated : updated + "\n").pipe(
-              Effect.andThen(fs.rename(temporary, document.path)),
-              Effect.mapError(
-                (cause) => new UpdateError({ message: `Failed to write config: ${document.path}`, cause }),
-              ),
-            )
-            return info
-          }),
-        ),
-      )
-
      return Service.of({
        entries: Effect.fnUntraced(function* () {
          return configs
        }),
-        update,
        changes: () => Stream.fromPubSub(updates),
      })
    }),
@@ -442,17 +394,3 @@ export function configured(options?: Options) {
 }

 export const node = configured()
-
-type Edit = { readonly path: (string | number)[]; readonly value: unknown }
-
-function changes(before: unknown, after: unknown, path: (string | number)[] = []): Edit[] {
-  if (Object.is(before, after)) return []
-  if (isRecord(before) && isRecord(after)) {
-    return [...new Set([...Object.keys(before), ...Object.keys(after)])].flatMap((key) => {
-      if (!(key in after)) return [{ path: [...path, key], value: undefined }]
-      if (!(key in before)) return [{ path: [...path, key], value: after[key] }]
-      return changes(before[key], after[key], [...path, key])
-    })
-  }
-  return [{ path, value: after }]
-}
@@ -0,0 +1,150 @@
+export * as ConfigFile from "./file.js"
+
+import { isDeepStrictEqual } from "node:util"
+import { isRecord } from "@opencode-ai/ai/utils/record"
+import { FSUtil } from "@opencode-ai/util/fs-util"
+import { Effect, Schema, Semaphore } from "effect"
+import {
+  applyEdits,
+  createScanner,
+  findNodeAtLocation,
+  modify,
+  parseTree,
+  type Node,
+  type ParseError,
+} from "jsonc-parser"
+
+export class UpdateError extends Schema.TaggedError<UpdateError>()("ConfigFile.UpdateError", {
+  message: Schema.String,
+  cause: Schema.optional(Schema.Defect()),
+}) {}
+
+const isJson = Schema.is(Schema.MutableJson)
+const isDocument = (value: unknown): value is Schema.MutableJsonObject => isRecord(value) && isJson(value)
+const lock = Semaphore.makeUnsafe(1)
+
+/**
+ * Edits an existing JSON(C) file using raw source values, not resolved Config.Info.
+ * The synchronous callback mutates a source clone; its return value is ignored.
+ * Validates JSON only; normalization and substitution remain the reader's job.
+ * Does not discover files, start watchers, or refresh Config state.
+ * Read-modify-write calls are serialized within this process.
+ */
+export const update = Effect.fn("ConfigFile.update")(
+  function* (
+    filepath: string,
+    mutate: (draft: Schema.MutableJsonObject) => void,
+  ): Effect.fn.Return<Schema.JsonObject, UpdateError, FSUtil.Service> {
+    const fs = yield* FSUtil.Service
+    const text = yield* fs
+      .readFileString(filepath)
+      .pipe(Effect.mapError((cause) => new UpdateError({ message: `Failed to read config: ${filepath}`, cause })))
+    const errors: ParseError[] = []
+    const current = parseSource(text, errors)
+    if (errors.length || !isDocument(current))
+      return yield* Effect.fail(new UpdateError({ message: `Invalid config file: ${filepath}` }))
+
+    const next = yield* Effect.try({
+      try: () => {
+        const draft = structuredClone(current)
+        mutate(draft)
+        return draft
+      },
+      catch: (cause) => new UpdateError({ message: "Config update failed", cause }),
+    })
+    if (!isDocument(next))
+      return yield* Effect.fail(new UpdateError({ message: `Config update must produce a JSON object: ${filepath}` }))
+
+    const edits = changes(current, next)
+    if (!edits.length) return next
+    const updated = yield* Effect.try({
+      try: () => edits.reduce(patch, text),
+      catch: (cause) => new UpdateError({ message: `Failed to patch config: ${filepath}`, cause }),
+    })
+    // Duplicate keys can make parse choose the last value while modify edits the first.
+    const written = parseSource(updated, errors)
+    if (errors.length || !isDeepStrictEqual(written, next))
+      return yield* Effect.fail(
+        new UpdateError({ message: `Config patch does not match the requested update: ${filepath}` }),
+      )
+    const temporary = filepath + ".tmp"
+    yield* fs.writeFileString(temporary, updated.endsWith("\n") ? updated : updated + "\n").pipe(
+      Effect.andThen(fs.rename(temporary, filepath)),
+      Effect.mapError((cause) => new UpdateError({ message: `Failed to write config: ${filepath}`, cause })),
+    )
+    return next
+  },
+  (effect) => lock.withPermit(effect),
+)
+
+type Edit = { readonly path: (string | number)[]; readonly value: unknown }
+
+function parseSource(text: string, errors: ParseError[]) {
+  const root = parseTree(text, errors, { allowTrailingComma: true })
+  if (!root || errors.length) return undefined
+  // parse() assigns onto {}, invoking the __proto__ setter instead of retaining
+  // an own JSON key. Construct object entries from the AST without those setters.
+  const value = (node: Node): unknown => {
+    if (node.type === "array") return (node.children ?? []).map(value)
+    if (node.type === "object")
+      return Object.fromEntries(
+        (node.children ?? []).map((property) => {
+          const child = property.children?.[1]
+          return [property.children?.[0]?.value, child && value(child)]
+        }),
+      )
+    return node.value
+  }
+  return value(root)
+}
+
+function patch(text: string, edit: Edit) {
+  if (edit.value !== undefined)
+    return applyEdits(
+      text,
+      modify(text, edit.path, edit.value, { formattingOptions: { tabSize: 2, insertSpaces: true } }),
+    )
+
+  const tree = parseTree(text)
+  const node = tree && findNodeAtLocation(tree, edit.path)
+  if (!node) return text
+  // jsonc-parser removes adjacent comments along with the separator. Remove only
+  // the property/element itself and one comma, leaving surrounding comments intact.
+  const target = node.parent?.type === "property" ? node.parent : node
+  const siblings = target.parent?.children ?? []
+  const previous = siblings[siblings.indexOf(target) - 1]
+  const scanner = createScanner(text, true)
+  scanner.setPosition(target.offset + target.length)
+  scanner.scan()
+  const following = text[scanner.getTokenOffset()] === ","
+  if (!following && previous) {
+    scanner.setPosition(previous.offset + previous.length)
+    scanner.scan()
+  }
+  return applyEdits(text, [
+    { offset: target.offset, length: target.length, content: "" },
+    ...(following || previous ? [{ offset: scanner.getTokenOffset(), length: 1, content: "" }] : []),
+  ])
+}
+
+function changes(before: unknown, after: unknown, path: (string | number)[] = []): Edit[] {
+  if (isDeepStrictEqual(before, after)) return []
+  if (Array.isArray(before) && Array.isArray(after)) {
+    return [
+      ...after.flatMap((value, index) => changes(before[index], value, [...path, index])),
+      // Remove from the end so earlier deletions cannot shift later paths.
+      ...before
+        .slice(after.length)
+        .map((_, index) => ({ path: [...path, after.length + index], value: undefined }))
+        .toReversed(),
+    ]
+  }
+  if (isRecord(before) && isRecord(after)) {
+    return [...new Set([...Object.keys(before), ...Object.keys(after)])].flatMap((key) => {
+      if (!Object.hasOwn(after, key)) return [{ path: [...path, key], value: undefined }]
+      if (!Object.hasOwn(before, key)) return [{ path: [...path, key], value: after[key] }]
+      return changes(before[key], after[key], [...path, key])
+    })
+  }
+  return [{ path, value: after }]
+}
@@ -87,7 +87,6 @@ export const Plugin = define({
                  ...input.prompt,
                  sessionID: input.sessionID,
                  text: yield* evaluateTemplate(command.template, input.prompt.text, {
-                    config,
                    location,
                    processes,
                    shell,
@@ -152,7 +151,6 @@ function evaluateTemplate(
  template: string,
  input: string,
  services: {
-    readonly config: Config.Interface
    readonly location: Location.Info
    readonly processes: AppProcess.Interface
    readonly shell: ShellSelect.Interface
@@ -19,6 +19,9 @@ export const Plugin = define({
  id: "opencode.config.instruction",
  effect: Effect.fn(function* () {
    const discovery = yield* InstructionDiscovery.Service
+    // Nothing this plugin watches or loads can contribute when both scopes
+    // are disabled; skip the resolves, the watcher fiber, and the transform.
+    if (!discovery.project && !discovery.global) return
    yield* Effect.gen(function* () {
      const fs = yield* FSUtil.Service
      const global = yield* Global.Service
@@ -35,9 +38,15 @@ export const Plugin = define({
      const loaded: { current: Loaded } = { current: { type: "available", files: [] } }

      const publish = (update: Watcher.Update) => PubSub.publish(changes, update.path).pipe(Effect.asVoid)
+      // The ancestor walk can reach the global file when the location sits
+      // beneath the global config dir; global: false excludes it there too.
      const candidates = [
-        globalFile,
-        ...(project ? ancestorDirectories(start, stop).map((directory) => join(directory, "AGENTS.md")) : []),
+        ...(discovery.global ? [globalFile] : []),
+        ...(project
+          ? ancestorDirectories(start, stop)
+              .map((directory) => join(directory, "AGENTS.md"))
+              .filter((file) => discovery.global || file !== globalFile)
+          : []),
      ]
      for (const path of new Set(candidates)) {
        const updates = yield* watcher.subscribe({ path, type: "file" })
@@ -51,15 +60,15 @@ export const Plugin = define({
      })

      const globalSource = Effect.fn("ConfigInstructionPlugin.globalSource")(function* () {
+        if (!discovery.global) return []
        const file = yield* read(globalFile)
        return file ? [file] : []
      })

      const projectSource = Effect.fn("ConfigInstructionPlugin.projectSource")(function* () {
        if (!project) return []
-        const discovered = new Set(
-          yield* Effect.forEach(yield* fs.up({ targets: ["AGENTS.md"], start, stop }), fs.resolve),
-        )
+        const walked = yield* Effect.forEach(yield* fs.up({ targets: ["AGENTS.md"], start, stop }), fs.resolve)
+        const discovered = new Set(walked.filter((file) => discovery.global || file !== globalFile))
        const files = yield* Effect.forEach(discovered, read, { concurrency: "unbounded" })
        if (files.some((file) => file === undefined)) return Instructions.unavailable
        return files.filter((file): file is InstructionDiscovery.File => file !== undefined)
@@ -135,7 +135,7 @@ export const fffLayer = Layer.effect(
        find: () => Effect.succeed([]),
      })
    }
-    yield* Effect.addFinalizer(() => Effect.sync(() => result.value.destroy()).pipe(Effect.ignore))
+    yield* Effect.addFinalizer(() => Effect.sync(() => result.value.destroy()))
    return Service.of({
      find: (input) =>
        Effect.sync(() => {
@@ -311,7 +311,7 @@ function validateField(field: InputField, value: Form.Value): string | undefined
    }
    if (field.format === "email" && !/^[^\s@]+@[^\s@]+\.[^\s@]+$/.test(value))
      return `Expected email for form field: ${field.key}`
-    if (field.format === "uri" && !isUri(value)) return `Expected URI for form field: ${field.key}`
+    if (field.format === "uri" && !URL.canParse(value)) return `Expected URI for form field: ${field.key}`
    if (field.format === "date" && !isDate(value)) return `Expected date for form field: ${field.key}`
    if (field.format === "date-time" && !isDateTime(value)) return `Expected date-time for form field: ${field.key}`
    if (field.options && !field.custom && !field.options.some((option) => option.value === value)) {
@@ -347,15 +347,6 @@ function isStringArray(value: Form.Value): value is ReadonlyArray<string> {
  return Array.isArray(value) && value.every((item): item is string => typeof item === "string")
 }

-function isUri(value: string) {
-  try {
-    new URL(value)
-    return true
-  } catch {
-    return false
-  }
-}
-
 function isDate(value: string) {
  if (!/^\d{4}-\d{2}-\d{2}$/.test(value)) return false
  const date = new Date(`${value}T00:00:00.000Z`)
@@ -37,20 +37,19 @@ export const layer = Layer.effect(

    const runText = Effect.fn("Generate.text")(function* (input: TextInput) {
      const resolved = yield* resolver.resolve(input.model).pipe(
-        Effect.catchTags({
-          "SessionRunnerModel.VariantUnavailableError": (error) =>
-            input.model
+        Effect.catchTag(
+          [
+            "SessionRunnerModel.VariantUnavailableError",
+            "SessionRunnerModel.UnsupportedPackageError",
+            "SessionRunnerModel.UnresolvedProviderVariablesError",
+          ],
+          (error) => {
+            const mapped: Error = input.model
              ? new ModelSelectionError({ message: error.message })
-              : new UnavailableError({ message: error.message, service: error.providerID }),
-          "SessionRunnerModel.UnsupportedPackageError": (error) =>
-            input.model
-              ? new ModelSelectionError({ message: error.message })
-              : new UnavailableError({ message: error.message, service: error.providerID }),
-          "SessionRunnerModel.UnresolvedProviderVariablesError": (error) =>
-            input.model
-              ? new ModelSelectionError({ message: error.message })
-              : new UnavailableError({ message: error.message, service: error.providerID }),
-        }),
+              : new UnavailableError({ message: error.message, service: error.providerID })
+            return Effect.fail(mapped)
+          },
+        ),
      )
      if (!resolved)
        return yield* new ModelSelectionError({
@@ -47,7 +47,7 @@ export function prepareResponsesTools({
          name: tool.name,
          description: tool.description,
          parameters: tool.inputSchema,
-          strict: strictJsonSchema,
+          strict: tool.strict ?? strictJsonSchema,
        })
        break
      case "provider": {
@@ -114,17 +114,42 @@ export type Error = LayerNode.Error<typeof graph>
 export interface Options {
  // Plugins this instance is born with; empty and absent are equivalent.
  readonly plugins?: InstancePlugins.List
+  // Filesystem config discovery; true (default) is today's behavior. When
+  // false the instance boots vanilla: no upward config scan, no global config
+  // dir, no plugin-dir loading or ambient plugin module imports, no boot-time
+  // AGENTS.md discovery. Wellknown integration config and host-injected values
+  // still apply, and a config value that explicitly names something
+  // file-backed (a skill path, an MCP command) is a request, not discovery.
+  // Use-triggered behavior also stays: reading a file still injects nested
+  // AGENTS.md instructions — that is session-time context, not boot-time
+  // population.
+  readonly discovery?: boolean
  readonly replacements?: LayerNode.Replacements
 }

+// Vanilla neutralizes the discovery inputs; the plugin list stays untouched.
+// These are defaults ahead of caller replacements: a host that replaces a
+// gated node itself owns that node's discovery flags. Plugin-directory
+// discovery needs no swap here — implicit scanning only triggers on Directory
+// config entries, which a no-scan Config never produces, and the default
+// source still honors explicit plugin operations from wellknown and
+// host-injected config.
+const vanillaReplacements: LayerNode.Replacements = [
+  [Config.node, Config.configured({ project: false, global: false })],
+  [InstructionDiscovery.node, InstructionDiscovery.configured({ project: false, global: false })],
+]
+
 // One instance is one compiled, fresh copy of the graph standing on a directory.
 export function layer(ref: Location.Ref, options: Options = {}) {
  const startedAt = performance.now()
-  // Bound pairs come last, so they win over caller replacements of the same nodes.
-  const allReplacements = (options.replacements ?? []).concat([
-    [Location.node, Location.boundNode(ref)],
+  // Ordered: vanilla defaults, then caller replacements (which win over the
+  // defaults), then bound pairs (which win over everything).
+  const allReplacements: LayerNode.Replacements = [
+    ...(options.discovery === false ? vanillaReplacements : []),
+    ...(options.replacements ?? []),
+    [Location.node, Location.boundNode(ref, { discovery: options.discovery })],
    [InstancePlugins.node, InstancePlugins.bound(options.plugins ?? [])],
-  ])
+  ]
  // Apply replacements during hoist, not afterward: replacements can
  // introduce new tagged dependencies (Location.boundNode depends on
  // Project), and the hoist walk is the only pass that can still slice
--- a/Show More
+++ b/Show More
Author	SHA1	Message	Date
opencode-agent[bot]andrekram1-node	f6992059be	fix(core): isolate shell jobs by shell ID (#45985 ) Co-authored-by: rekram1-node <rekram1-node@users.noreply.github.com>	2026-08-28 13:13:16 -05:00
Kit Langton	732f949a65	fix(tui): admit compaction requests optimistically (#45973 ) Render queued compaction feedback before model setup, coalesce repeated gestures, and reconcile canonical admissions without restoring consumed rows. Serialize prompt and compaction preparation through the existing per-session admission chain.	2026-08-28 14:00:09 -04:00
James Long	cd3b12c579	feat(tui): resize terminal panes (#45979 )	2026-08-28 13:47:50 -04:00
Kit Langton	0d6232ffef	test(core): share git test fixtures (#45640 )	2026-08-28 13:37:42 -04:00
Kit Langton	6da20f0efe	chore(protocol): refresh committed OpenAPI baseline (#45969 )	2026-08-28 13:37:11 -04:00
Aiden Cline	8e25e83e5a	fix(core): continue retryable failures after durable output (#45861 )	2026-08-28 12:35:00 -05:00
Kit Langton	fe188f8722	fix(core): normalize directory page headings (#45641 )	2026-08-28 13:29:41 -04:00
Kit Langton	e4bc8b765b	refactor(core): reuse MCP prompt helpers (#45660 )	2026-08-28 13:26:51 -04:00
Kit Langton	d28b6e9ac2	test(core): construct valid tool contexts (#45657 )	2026-08-28 13:26:41 -04:00
Kit Langton	ac3cd1b183	test(core): simplify provider smoke tests (#45656 )	2026-08-28 13:26:27 -04:00
Kit Langton	c601d3b021	refactor(core): share provider factory loading (#45654 )	2026-08-28 13:26:15 -04:00
Kit Langton	9bc2165e5c	refactor(core): reuse lazy shell parser initialization (#45651 )	2026-08-28 13:26:04 -04:00
Kit Langton	374d317412	test(core): simplify provider test setup (#45635 )	2026-08-28 13:25:53 -04:00
Kit Langton	e50c89834e	refactor(core): reuse plugin response wrapper (#45648 )	2026-08-28 13:25:40 -04:00
Kit Langton	f367c202d9	refactor(core): remove unused command dependency (#45623 )	2026-08-28 13:25:27 -04:00
Kit Langton	9d33d83bb4	refactor(core): reuse native option projection (#45647 )	2026-08-28 13:23:33 -04:00
Kit Langton	8f1eff50aa	refactor(core): deepen snapshot comparison helper (#45653 )	2026-08-28 13:22:33 -04:00
Kit Langton	e7918e25fd	refactor(core): remove ineffective typed recovery (#45631 )	2026-08-28 13:16:13 -04:00
Aiden Cline	1eff84615c	fix(ai): keep chat reasoning in one lifecycle (#45864 )	2026-08-28 12:14:58 -05:00
Kit Langton	f1fd6ba3c0	test(core): await registry readiness barriers (#45630 )	2026-08-28 13:14:41 -04:00
Kit Langton	39daec9bec	refactor(core): narrow session internals (#45638 )	2026-08-28 13:05:16 -04:00
Kit Langton	484f5faf8d	docs(plugin): update current API examples (#45627 )	2026-08-28 13:04:59 -04:00
Kit Langton	9a227d186f	fix(core): remove unreachable MCP authorization path (#45633 )	2026-08-28 13:03:17 -04:00
Kit Langton	fc27061838	refactor(core): remove tool context aliases (#45639 )	2026-08-28 13:03:01 -04:00
Kit Langton	d71cc3be77	test(core): guarantee lock fixture cleanup (#45636 )	2026-08-28 13:02:50 -04:00
Kit Langton	503e680672	test(core): preserve session layer inference (#45629 )	2026-08-28 13:02:21 -04:00
Kit Langton	1f77f4a4ed	refactor(core): inline auxiliary model operations (#45974 )	2026-08-28 12:34:01 -04:00
Kit Langton	e8fa7daed5	fix(client): preserve live activity during hydration (#45975 ) Merge session status changes received during activity hydration and ignore superseded reconnect responses. Cover terminal events, starts, deletion, and overlapping connections so idle sessions do not regain stale TUI spinners.	2026-08-28 12:30:21 -04:00
Kit Langton	936c73b54d	refactor(core): simplify tool settlement (#45967 )	2026-08-28 12:04:27 -04:00
Kit Langton	b9cb4fc36a	chore(core): retire prompt-cache diagnostics (#45965 )	2026-08-28 11:55:55 -04:00
Kit Langton	a808a02f05	refactor(core): separate config file editing (#45818 ) Separate source-file editing from Config with explicit targets, fresh raw JSON reads, owned synchronous mutation, and comment-preserving verified writeback. Preserve own JSON source keys and leave configuration discovery, precedence, and watching unchanged.	2026-08-28 11:23:48 -04:00
James Long	e142a783f7	refactor(tui): extract reusable pane resize logic (#45939 )	2026-08-28 11:20:23 -04:00
Kit Langton	201536f265	test: repair stale V2 baseline coverage (#45826 )	2026-08-28 10:40:04 -04:00
Kit Langton	0d42e76006	feat(ai): add first-class TestLLM controls (#45828 )	2026-08-28 09:49:47 -04:00
James Long	ca47949475	feat(plugin): expose experimental terminal reads (#45792 )	2026-08-28 08:03:35 -04:00
Filip	1c8e557eb4	Revert "fix(cli): prevent repeated updates and npm cache growth" (#45865 )	2026-08-28 07:23:43 +00:00
Aiden Cline	11ca603490	fix(ai): finish chat streams at done sentinel (#45850 )	2026-08-28 00:53:53 -05:00
Aiden Cline	85d8b07f09	fix(ai): ignore late converse tool deltas (#45847 )	2026-08-28 00:38:01 -05:00
Aiden Cline	d2ee536c16	fix(core): classify AISDK network failures as transport errors (#45840 )	2026-08-28 00:21:17 -05:00
Aiden Cline	e12e04f482	fix(core): strengthen background shell guidance (#45843 )	2026-08-28 00:19:15 -05:00
Kit Langton	5743537945	test(core): use effectful temp fixtures (#45637 )	2026-08-28 01:01:04 -04:00
Aiden Cline	18e22cd82e	refactor(ai): simplify response stream state (#45835 )	2026-08-27 23:56:42 -05:00
Aiden Cline	bdf019a9ac	fix(ai): detect DashScope input length overflow errors (#45834 )	2026-08-27 23:43:14 -05:00
Aiden Cline	5cbafc57c0	refactor(ai): default unrecognized provider failures to retry (#45825 )	2026-08-27 23:40:52 -05:00
Aiden Cline	55674b858b	fix(ai): make final snapshots authoritative for text and reasoning (#45831 )	2026-08-27 23:38:19 -05:00
Luke Parker	bb390f435c	fix(session-ui): enable word diffs in unified view (#45833 )	2026-08-28 04:31:18 +00:00
Kit Langton	92b9eebab2	feat(tui): streamline diff review workflow (#45817 )	2026-08-28 00:11:29 -04:00
Aiden Cline	074413a96d	fix(ai): normalize response item boundaries (#45789 )	2026-08-27 23:08:50 -05:00
opencode-agent[bot]	4685ba8d3e	chore: update nix node_modules hashes	2026-08-28 04:03:16 +00:00
Kit Langton	d2c307ab30	test(core): remove unused test imports (#45674 )	2026-08-27 23:58:41 -04:00
Kit Langton	a5193634d0	test(core): use direct test imports (#45683 )	2026-08-27 23:58:34 -04:00
Kit Langton	95725d121a	chore: remove unused changesets (#45827 )	2026-08-27 23:55:57 -04:00
Aiden Cline	56f2559798	fix(codemode): report original tool errors (#45820 )	2026-08-27 22:54:36 -05:00
Kit Langton	38bffc9db1	chore(deps): upgrade Effect to rc.112 (#45819 )	2026-08-27 23:43:37 -04:00
Kit Langton	1d3308d904	test(core): refresh test vocabulary (#45714 )	2026-08-27 23:26:45 -04:00
opencode-agent[bot]	3c9e5adefa	chore: update nix node_modules hashes	2026-08-28 03:23:31 +00:00
Kit Langton	a065ad4ba7	feat(core): vanilla instance discovery option (#45752 )	2026-08-28 03:16:29 +00:00
Kit Langton	048aec32e7	refactor(core): share local provider settings fold (#45645 )	2026-08-27 23:13:13 -04:00
Kit Langton	a258b17fac	test(core): use partial service mocks (#45703 )	2026-08-27 23:13:04 -04:00
Kit Langton	07facd2086	refactor(util): share process stop policy (#45725 )	2026-08-27 23:12:57 -04:00
Kit Langton	6bf342a861	refactor(core): consolidate HTML inline handling (#45649 )	2026-08-27 23:12:50 -04:00
Kit Langton	2cca1ae860	refactor(core): use nonthrowing URL parsing (#45726 )	2026-08-27 23:12:43 -04:00
Kit Langton	becf5bf482	refactor(core): retain changed worktrees directly (#45709 )	2026-08-27 23:12:36 -04:00
Kit Langton	afe9e579d6	refactor(core): simplify instruction preview blobs (#45702 )	2026-08-27 23:12:28 -04:00
Kit Langton	0b82fe60ea	refactor(core): consolidate model error mapping (#45722 )	2026-08-27 23:12:19 -04:00
Luke Parker	6fad330efc	chore(deps): upgrade Solid and Router (#45808 )	2026-08-28 03:03:08 +00:00
Kit Langton	87f21a301d	fix(core): forward plugin interrupt options (#45632 )	2026-08-27 23:01:57 -04:00
Kit Langton	0123fed65d	fix(core): preserve Responses tool strictness (#45663 )	2026-08-27 23:01:43 -04:00
Kit Langton	64ef85159d	fix(core): preserve assistant text metadata (#45650 )	2026-08-27 23:01:33 -04:00
Kit Langton	e63779d9e1	fix(core): keep vertex settings serializable (#45704 )	2026-08-27 23:01:20 -04:00
Kit Langton	ac55c05c30	fix(core): correct compaction history direction (#45698 )	2026-08-27 23:01:07 -04:00
Dax Raad	8fa672ed71	docs: document diagnostic profile signals	2026-08-27 22:53:28 -04:00