chore(main): release 0.61.0 (#898)

🤖 I have created a release *beep* *boop*
---


##
[0.61.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.60.0...v0.61.0)
(2026-07-22)


### Features

* **deps:** Bump actions/setup-node from 6.4.0 to 7.0.0
([#897](https://github.com/agentclientprotocol/claude-agent-acp/issues/897))
([d9bd36d](d9bd36d8b0))
* **deps:** Update to @anthropic-ai/claude-agent-sdk 0.3.217
([#899](https://github.com/agentclientprotocol/claude-agent-acp/issues/899))
([edf3af0](edf3af043b))

---
This PR was generated with [Release
Please](https://github.com/googleapis/release-please). See
[documentation](https://github.com/googleapis/release-please#release-please).

Co-authored-by: acp-release-bot[bot] <246668977+acp-release-bot[bot]@users.noreply.github.com>

.github/workflows/ci.yml

@@ -21,10 +21,10 @@ jobs:
     timeout-minutes: 5
 
     steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0
 
       - name: Use Node.js
-        uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e
+        uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020
         with:
           node-version: "lts/*"
           cache: "npm"

.github/workflows/publish.yml

@@ -16,7 +16,7 @@ jobs:
       # Generating a GitHub token, so that PRs and tags created by
       # the release-please-action can trigger actions workflows.
       - name: Generate GitHub token
-        uses: actions/create-github-app-token@1b10c78c7865c340bc4f6099eb2f838309f1e8c3
+        uses: actions/create-github-app-token@bcd2ba49218906704ab6c1aa796996da409d3eb1
         id: generate-token
         with:
           # GitHub App ID secret name
@@ -44,14 +44,37 @@ jobs:
       contents: read
       id-token: write
     steps:
-      - uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0
       # Setup .npmrc file to publish to npm
-      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e
+      - uses: actions/setup-node@820762786026740c76f36085b0efc47a31fe5020
         with:
           node-version: "lts/*"
           registry-url: "https://registry.npmjs.org"
-      - name: Update npm
-        run: npm install -g npm@latest
       - run: npm ci
       - run: npm run build
       - run: npm publish
+
+  trigger-registry-update:
+    needs: publish-npm
+    runs-on: ubuntu-latest
+    environment: release
+    permissions: {}
+    steps:
+      - name: Generate token scoped to the registry repo
+        uses: actions/create-github-app-token@bcd2ba49218906704ab6c1aa796996da409d3eb1
+        id: registry-token
+        with:
+          app-id: ${{ secrets.REGISTRY_UPDATER_APP_ID }}
+          private-key: ${{ secrets.REGISTRY_UPDATER_APP_PRIVATE_KEY }}
+          owner: ${{ github.repository_owner }}
+          repositories: registry
+
+      - name: Dispatch registry version update
+        env:
+          GH_TOKEN: ${{ steps.registry-token.outputs.token }}
+        run: |
+          gh workflow run update-versions.yml \
+            --repo ${{ github.repository_owner }}/registry \
+            --ref main \
+            -f apply=true \
+            -f agents=claude-acp

.gitignore

@@ -3,3 +3,4 @@ dist
 .DS_Store
 .claude/*.local.*
 CLAUDE.local.md
+.idea/

.release-please-manifest.json

@@ -1,3 +1,3 @@
 {
-  ".": "0.31.4"
+  ".": "0.61.0"
 }

CHANGELOG.md

@@ -1,5 +1,418 @@
 # Changelog
 
+## [0.61.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.60.0...v0.61.0) (2026-07-22)
+
+
+### Features
+
+* **deps:** Bump actions/setup-node from 6.4.0 to 7.0.0 ([#897](https://github.com/agentclientprotocol/claude-agent-acp/issues/897)) ([d9bd36d](https://github.com/agentclientprotocol/claude-agent-acp/commit/d9bd36d8b06764d63656d0387ad9430ec9fcb27d))
+* **deps:** Update to @anthropic-ai/claude-agent-sdk 0.3.217 ([#899](https://github.com/agentclientprotocol/claude-agent-acp/issues/899)) ([edf3af0](https://github.com/agentclientprotocol/claude-agent-acp/commit/edf3af043b6d00e5caca2cc81a2285c477c8b2ab))
+
+## [0.60.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.59.0...v0.60.0) (2026-07-20)
+
+
+### Features
+
+* **deps:** Update to claude-agent-sdk 0.3.215 ([#890](https://github.com/agentclientprotocol/claude-agent-acp/issues/890)) ([92548f0](https://github.com/agentclientprotocol/claude-agent-acp/commit/92548f043547b0ddac95921ece020e69f7c12c5f))
+* implement configurable LLM providers ([#853](https://github.com/agentclientprotocol/claude-agent-acp/issues/853)) ([82cd692](https://github.com/agentclientprotocol/claude-agent-acp/commit/82cd692e500eedec182f817b18cebb005c4b98ce))
+
+
+### Bug Fixes
+
+* parse Agent/Task trailers without regex ([#879](https://github.com/agentclientprotocol/claude-agent-acp/issues/879)) ([06c3d7b](https://github.com/agentclientprotocol/claude-agent-acp/commit/06c3d7bdbd8cc9415c8cabac060a50e0951c758b))
+* remove ~15s stall on session/new and model switch by seeding the context window synchronously ([#894](https://github.com/agentclientprotocol/claude-agent-acp/issues/894)) ([ff9b96d](https://github.com/agentclientprotocol/claude-agent-acp/commit/ff9b96d462831b1c3b96722ea20215ff6e529cb1))
+* Silence missing PostToolUse callbacks ([#895](https://github.com/agentclientprotocol/claude-agent-acp/issues/895)) ([1887ada](https://github.com/agentclientprotocol/claude-agent-acp/commit/1887ada215b27bb1025d9b7696a46ae7a4ac0f7a)), closes [#889](https://github.com/agentclientprotocol/claude-agent-acp/issues/889)
+
+## [0.59.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.58.1...v0.59.0) (2026-07-13)
+
+
+### Features
+
+* **deps-dev:** bump nanoid from 3.3.15 to 3.3.16 ([#875](https://github.com/agentclientprotocol/claude-agent-acp/issues/875)) ([e67dacd](https://github.com/agentclientprotocol/claude-agent-acp/commit/e67dacdcac629b00e5bdcda94b3de57459696b1e))
+* **deps:** bump @anthropic-ai/claude-agent-sdk to 0.3.207 ([#874](https://github.com/agentclientprotocol/claude-agent-acp/issues/874)) ([c7f5b8f](https://github.com/agentclientprotocol/claude-agent-acp/commit/c7f5b8fe768507afa28cdffe5cc3c054c56306ad))
+
+
+### Bug Fixes
+
+* Add subagent parent tool use attribution ([#859](https://github.com/agentclientprotocol/claude-agent-acp/issues/859)) ([9cd48c5](https://github.com/agentclientprotocol/claude-agent-acp/commit/9cd48c597af68f630774ec00bc3adc85f6b0fd4b))
+* forward result text when the turn emitted no assistant message ([#858](https://github.com/agentclientprotocol/claude-agent-acp/issues/858)) ([61ae860](https://github.com/agentclientprotocol/claude-agent-acp/commit/61ae8609d4343da6758c2376c6cf684c7fb0956a))
+* hold a turn open while its background subagents are still live ([#870](https://github.com/agentclientprotocol/claude-agent-acp/issues/870)) ([7a70f82](https://github.com/agentclientprotocol/claude-agent-acp/commit/7a70f82739e085014cad878f08513cdef7b7fe16))
+* Refine tool calls from streamed input ([#867](https://github.com/agentclientprotocol/claude-agent-acp/issues/867)) ([2c19974](https://github.com/agentclientprotocol/claude-agent-acp/commit/2c19974d5b80d3178b35e80a4bb0a1050588aab9))
+* Seed context window from SDK usage report ([#868](https://github.com/agentclientprotocol/claude-agent-acp/issues/868)) ([3ba2d36](https://github.com/agentclientprotocol/claude-agent-acp/commit/3ba2d367a282fe3053a73804a416a435cb01ee08)), closes [#596](https://github.com/agentclientprotocol/claude-agent-acp/issues/596)
+* Skip synthetic login messages on replay ([#869](https://github.com/agentclientprotocol/claude-agent-acp/issues/869)) ([c7dff3c](https://github.com/agentclientprotocol/claude-agent-acp/commit/c7dff3cf7eeb03cb993493efafa96344f18de81b)), closes [#863](https://github.com/agentclientprotocol/claude-agent-acp/issues/863)
+
+## [0.58.1](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.58.0...v0.58.1) (2026-07-09)
+
+
+### Bug Fixes
+
+* Use valid npm version for publish ([#855](https://github.com/agentclientprotocol/claude-agent-acp/issues/855)) ([8b366d8](https://github.com/agentclientprotocol/claude-agent-acp/commit/8b366d8e9ad55e9237bf2d618cfc99b57fac570d))
+
+## [0.58.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.57.0...v0.58.0) (2026-07-09)
+
+
+### Features
+
+* **deps:** update to @anthropic-ai/claude-agent-sdk 0.3.205 ([#854](https://github.com/agentclientprotocol/claude-agent-acp/issues/854)) ([f664ced](https://github.com/agentclientprotocol/claude-agent-acp/commit/f664ced76df00674d88aedd018a4261da3bc3f3a))
+
+
+### Bug Fixes
+
+* Preserve live model on resumed sessions ([#848](https://github.com/agentclientprotocol/claude-agent-acp/issues/848)) ([f3d8ae3](https://github.com/agentclientprotocol/claude-agent-acp/commit/f3d8ae3eb389ee9367f48c8762562a55280ade0a)), closes [#845](https://github.com/agentclientprotocol/claude-agent-acp/issues/845)
+* Report usage for cancelled active turns ([#846](https://github.com/agentclientprotocol/claude-agent-acp/issues/846)) ([b03318f](https://github.com/agentclientprotocol/claude-agent-acp/commit/b03318f59c6165146c8d8e036227f35c3abd0ff4)), closes [#844](https://github.com/agentclientprotocol/claude-agent-acp/issues/844)
+* tolerate missing text in streamed thinking chunks ([#852](https://github.com/agentclientprotocol/claude-agent-acp/issues/852)) ([e944ced](https://github.com/agentclientprotocol/claude-agent-acp/commit/e944ceddea79950d6bfb6339266862507c4b964f))
+* Use SDK guards for elicitation validation ([#850](https://github.com/agentclientprotocol/claude-agent-acp/issues/850)) ([32b9350](https://github.com/agentclientprotocol/claude-agent-acp/commit/32b93501d7a35a3b405b5022b4caaad492eed11c))
+
+## [0.57.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.56.0...v0.57.0) (2026-07-07)
+
+
+### Features
+
+* Update @anthropic-ai/claude-agent-sdk to 0.3.202 ([#843](https://github.com/agentclientprotocol/claude-agent-acp/issues/843)) ([1612c07](https://github.com/agentclientprotocol/claude-agent-acp/commit/1612c078953c134e12368a2d968a9bcf14d31b55))
+
+
+### Bug Fixes
+
+* don't emit empty agent_message_chunk ([#841](https://github.com/agentclientprotocol/claude-agent-acp/issues/841)) ([d74dd1d](https://github.com/agentclientprotocol/claude-agent-acp/commit/d74dd1d9d356050fc87f9fdab8599f1546b12a2c))
+
+## [0.56.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.55.0...v0.56.0) (2026-07-06)
+
+
+### Features
+
+* **deps:** bump @anthropic-ai/claude-agent-sdk to 0.3.201 ([#837](https://github.com/agentclientprotocol/claude-agent-acp/issues/837)) ([ff2a8a9](https://github.com/agentclientprotocol/claude-agent-acp/commit/ff2a8a963536a4265700e4f2003a0ffff93c9087))
+
+## [0.55.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.54.1...v0.55.0) (2026-07-02)
+
+
+### Features
+
+* Add refusal fallback consent dialog support ([#834](https://github.com/agentclientprotocol/claude-agent-acp/issues/834)) ([15dfd83](https://github.com/agentclientprotocol/claude-agent-acp/commit/15dfd83daa84ced98ecbaa7c83fd5a53124030c1))
+* **deps-dev:** bump the minor group with 3 updates ([#831](https://github.com/agentclientprotocol/claude-agent-acp/issues/831)) ([8d5febf](https://github.com/agentclientprotocol/claude-agent-acp/commit/8d5febf85485c1ad19b81501fb5c6d4275448a7c))
+* Update to claude-agent-sdk 0.3.198 ([#836](https://github.com/agentclientprotocol/claude-agent-acp/issues/836)) ([307ab82](https://github.com/agentclientprotocol/claude-agent-acp/commit/307ab82712213413446466d808e161073e94fc7f))
+
+
+### Bug Fixes
+
+* Handle model refusal fallback updates ([#833](https://github.com/agentclientprotocol/claude-agent-acp/issues/833)) ([648e3f6](https://github.com/agentclientprotocol/claude-agent-acp/commit/648e3f60bd7e6eadb5f86b0dc475104ab12e7b02))
+* Handle SDK idle turns without results ([#835](https://github.com/agentclientprotocol/claude-agent-acp/issues/835)) ([57b00cc](https://github.com/agentclientprotocol/claude-agent-acp/commit/57b00ccac610e2eeaf0a753934a0c3b78d27ccb0))
+
+## [0.54.1](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.54.0...v0.54.1) (2026-06-30)
+
+
+### Bug Fixes
+
+* apply modelOverrides when resolving availableModels allowlist ([#827](https://github.com/agentclientprotocol/claude-agent-acp/issues/827)) ([98c284b](https://github.com/agentclientprotocol/claude-agent-acp/commit/98c284bb871a96710596dfdc2b9ea25d2400f6b2))
+
+## [0.54.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.53.0...v0.54.0) (2026-06-30)
+
+
+### Features
+
+* Add Fast mode session config support ([#828](https://github.com/agentclientprotocol/claude-agent-acp/issues/828)) ([fa949a2](https://github.com/agentclientprotocol/claude-agent-acp/commit/fa949a20db0e33f2d63aebeb37e9fea212d8ee95))
+* **deps-dev:** bump prettier from 3.9.1 to 3.9.3 in the minor group ([#821](https://github.com/agentclientprotocol/claude-agent-acp/issues/821)) ([b8df8e0](https://github.com/agentclientprotocol/claude-agent-acp/commit/b8df8e0e5460fd782214f4dde488f7476c80c454))
+* **deps:** bump @anthropic-ai/claude-agent-sdk to 0.3.197 for Sonnet 5 ([#826](https://github.com/agentclientprotocol/claude-agent-acp/issues/826)) ([ef42c46](https://github.com/agentclientprotocol/claude-agent-acp/commit/ef42c46e5aea3bb53f433b9f9dd36a62a9f2df6a))
+* **deps:** bump fast-uri from 3.1.2 to 3.1.3 ([#822](https://github.com/agentclientprotocol/claude-agent-acp/issues/822)) ([64aa130](https://github.com/agentclientprotocol/claude-agent-acp/commit/64aa130c66c299f9835193796ba372b3ed40e7c1))
+
+## [0.53.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.52.0...v0.53.0) (2026-06-29)
+
+
+### Features
+
+* Add ACP logout support ([#816](https://github.com/agentclientprotocol/claude-agent-acp/issues/816)) ([0a0468c](https://github.com/agentclientprotocol/claude-agent-acp/commit/0a0468c72092b7ceccc551b4353dd92f9c22cbb8))
+* **deps:** bump @anthropic-ai/claude-agent-sdk to 0.3.195 ([#818](https://github.com/agentclientprotocol/claude-agent-acp/issues/818)) ([5dd8746](https://github.com/agentclientprotocol/claude-agent-acp/commit/5dd87462376597eb7b695200f4c716af46e764d6))
+
+
+### Bug Fixes
+
+* Emit tool_call before permission request ([#820](https://github.com/agentclientprotocol/claude-agent-acp/issues/820)) ([c95fc88](https://github.com/agentclientprotocol/claude-agent-acp/commit/c95fc884e82e5a297d1b17b8965690260867835b))
+
+## [0.52.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.51.0...v0.52.0) (2026-06-25)
+
+
+### Features
+
+* Add version flag handling ([#813](https://github.com/agentclientprotocol/claude-agent-acp/issues/813)) ([9616bda](https://github.com/agentclientprotocol/claude-agent-acp/commit/9616bdac47505e4a14c36d667fcffc9ae97e1f2a)), closes [#809](https://github.com/agentclientprotocol/claude-agent-acp/issues/809)
+* **deps-dev:** bump expect-type from 1.3.0 to 1.4.0 in the minor group ([#814](https://github.com/agentclientprotocol/claude-agent-acp/issues/814)) ([61272ac](https://github.com/agentclientprotocol/claude-agent-acp/commit/61272acb30dcafaa2455d334b11ce2ad97339707))
+* **deps:** Update @anthropic-ai/claude-agent-sdk to 0.3.191 ([#810](https://github.com/agentclientprotocol/claude-agent-acp/issues/810)) ([228f02e](https://github.com/agentclientprotocol/claude-agent-acp/commit/228f02ecfb23be16e59e121c6b42c0f2b2f40a4e))
+* Push session title updates at turn end ([#812](https://github.com/agentclientprotocol/claude-agent-acp/issues/812)) ([1fe7ec0](https://github.com/agentclientprotocol/claude-agent-acp/commit/1fe7ec09a3a7bcb7501231dae4b0ffe6ef9b70a4))
+
+## [0.51.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.50.0...v0.51.0) (2026-06-24)
+
+
+### Features
+
+* **deps:** bump the minor group with 11 updates ([#807](https://github.com/agentclientprotocol/claude-agent-acp/issues/807)) ([8f6ebd1](https://github.com/agentclientprotocol/claude-agent-acp/commit/8f6ebd1d9198edf723f4c8c1aa2b49b906c46646))
+
+## [0.50.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.49.0...v0.50.0) (2026-06-23)
+
+
+### Features
+
+* **acp:** Handle ACP request cancellation signals ([#801](https://github.com/agentclientprotocol/claude-agent-acp/issues/801)) ([9013d1d](https://github.com/agentclientprotocol/claude-agent-acp/commit/9013d1d46883a7f3774a63a9dca16c4a0f634a97))
+* **deps:** bump actions/checkout from 6.0.3 to 7.0.0 ([#803](https://github.com/agentclientprotocol/claude-agent-acp/issues/803)) ([044c43e](https://github.com/agentclientprotocol/claude-agent-acp/commit/044c43e0c894082b9e747c01e9dbde6e21036823))
+* **deps:** upgrade to @anthropic-ai/claude-agent-sdk@0.3.186 ([#806](https://github.com/agentclientprotocol/claude-agent-acp/issues/806)) ([a7e6137](https://github.com/agentclientprotocol/claude-agent-acp/commit/a7e6137f6877b72b8daa39e74971c3559db8d28f))
+
+## [0.49.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.48.0...v0.49.0) (2026-06-22)
+
+
+### Features
+
+* Update to claude-agent-sdk 0.3.185 ([#798](https://github.com/agentclientprotocol/claude-agent-acp/issues/798)) ([8dc8c86](https://github.com/agentclientprotocol/claude-agent-acp/commit/8dc8c864263fa50d03bec7ebff7aee776604cb3d))
+
+
+### Bug Fixes
+
+* Deduplicate streamed assistant blocks by content ([#800](https://github.com/agentclientprotocol/claude-agent-acp/issues/800)) ([960f62d](https://github.com/agentclientprotocol/claude-agent-acp/commit/960f62d76582ae5c9c5575ab66974809049ce1d0))
+* Infer 1M context from model descriptions ([#799](https://github.com/agentclientprotocol/claude-agent-acp/issues/799)) ([508453c](https://github.com/agentclientprotocol/claude-agent-acp/commit/508453c288b4a12701abd507199e7fa0ab172171))
+
+## [0.48.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.47.0...v0.48.0) (2026-06-19)
+
+
+### Features
+
+* Agent selection dropdown in config options ([#794](https://github.com/agentclientprotocol/claude-agent-acp/issues/794)) ([5729c47](https://github.com/agentclientprotocol/claude-agent-acp/commit/5729c471f92e1d2a191eb2a6d18c2be47821e9ec))
+* **deps:** bump the minor group with 11 updates ([#787](https://github.com/agentclientprotocol/claude-agent-acp/issues/787)) ([ad3b5fe](https://github.com/agentclientprotocol/claude-agent-acp/commit/ad3b5fe74527f83964695a1e8056d010b981fc79))
+* Update to claude-agent-sdk 0.3.183 ([#791](https://github.com/agentclientprotocol/claude-agent-acp/issues/791)) ([744b2d4](https://github.com/agentclientprotocol/claude-agent-acp/commit/744b2d41128f67091d4136283594c0db11ea4db5))
+* Update to new ACP SDK patterns ([#790](https://github.com/agentclientprotocol/claude-agent-acp/issues/790)) ([2554c7b](https://github.com/agentclientprotocol/claude-agent-acp/commit/2554c7bf980472760a6e7810b826f030d2c3af25))
+
+
+### Bug Fixes
+
+* duplicate assistant text when turn activates mid-message ([#789](https://github.com/agentclientprotocol/claude-agent-acp/issues/789)) ([1c80bf8](https://github.com/agentclientprotocol/claude-agent-acp/commit/1c80bf8e56a9279dc799e7bbdcae87241e99c18b))
+* Skip empty thinking chunks ([#793](https://github.com/agentclientprotocol/claude-agent-acp/issues/793)) ([15fdf26](https://github.com/agentclientprotocol/claude-agent-acp/commit/15fdf26fc7d3a6c51e89d3e56fc91dd00cb3d7ae))
+* surface Bash tool image output instead of dropping it ([#617](https://github.com/agentclientprotocol/claude-agent-acp/issues/617)) ([a759e64](https://github.com/agentclientprotocol/claude-agent-acp/commit/a759e64ef6d9b7c5bfe9a6e6b182db835f8ed3b6))
+
+## [0.47.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.46.0...v0.47.0) (2026-06-17)
+
+
+### Features
+
+* Update to claude-agent-sdk 0.3.179 ([#783](https://github.com/agentclientprotocol/claude-agent-acp/issues/783)) ([59a098c](https://github.com/agentclientprotocol/claude-agent-acp/commit/59a098c2b530bbae034e9a2dfbd31f8b4ef2a4d0))
+
+
+### Bug Fixes
+
+* Duplicate assistant messages in feed ([#785](https://github.com/agentclientprotocol/claude-agent-acp/issues/785)) ([12d34e6](https://github.com/agentclientprotocol/claude-agent-acp/commit/12d34e64e53564602ac1c38a30127e234c5c25ff))
+
+## [0.46.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.45.1...v0.46.0) (2026-06-16)
+
+
+### Features
+
+* Update to claude-agent-sdk 0.3.178 ([#777](https://github.com/agentclientprotocol/claude-agent-acp/issues/777)) ([58549ff](https://github.com/agentclientprotocol/claude-agent-acp/commit/58549ffe6a8b02ce59894e567407bd4299c11428))
+
+
+### Bug Fixes
+
+* Better handle out of turn events ([#780](https://github.com/agentclientprotocol/claude-agent-acp/issues/780)) ([4f273a2](https://github.com/agentclientprotocol/claude-agent-acp/commit/4f273a20d870c9c69f71556b8e0519f1de30f285))
+* Forward option details in elicitation meta ([#779](https://github.com/agentclientprotocol/claude-agent-acp/issues/779)) ([b364059](https://github.com/agentclientprotocol/claude-agent-acp/commit/b3640599ae685beecacd93e012d5bbc9dac716f7)), closes [#764](https://github.com/agentclientprotocol/claude-agent-acp/issues/764)
+
+## [0.45.1](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.45.0...v0.45.1) (2026-06-16)
+
+
+### Bug Fixes
+
+* Fix terminal error printing as text instead of terminal output ([#776](https://github.com/agentclientprotocol/claude-agent-acp/issues/776)) ([db6eaaf](https://github.com/agentclientprotocol/claude-agent-acp/commit/db6eaaf71484a321e47093ad65bcf8994943cb31))
+* Scope custom answers per question ([#774](https://github.com/agentclientprotocol/claude-agent-acp/issues/774)) ([d58004a](https://github.com/agentclientprotocol/claude-agent-acp/commit/d58004a34880e0a76833697319eb2a9efa6a43c7))
+
+## [0.45.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.44.0...v0.45.0) (2026-06-15)
+
+
+### Features
+
+* **deps-dev:** bump the minor group with 3 updates ([#763](https://github.com/agentclientprotocol/claude-agent-acp/issues/763)) ([7de5e4b](https://github.com/agentclientprotocol/claude-agent-acp/commit/7de5e4bcca9bfea70593092060f82bc8abe33e0e))
+* **deps:** bump @anthropic-ai/claude-agent-sdk to 0.3.177 ([#771](https://github.com/agentclientprotocol/claude-agent-acp/issues/771)) ([1be5ca5](https://github.com/agentclientprotocol/claude-agent-acp/commit/1be5ca57ee772fe90e41126365dc4186a18ad257))
+
+
+### Bug Fixes
+
+* preserve ANTHROPIC_CUSTOM_MODEL_OPTION when availableModels is set ([#768](https://github.com/agentclientprotocol/claude-agent-acp/issues/768)) ([cc2885f](https://github.com/agentclientprotocol/claude-agent-acp/commit/cc2885f6a9993cf61e759c3c770015f94c218627))
+
+## [0.44.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.43.0...v0.44.0) (2026-06-09)
+
+
+### Features
+
+* **deps:** bump the minor group with 14 updates ([#758](https://github.com/agentclientprotocol/claude-agent-acp/issues/758)) ([7a70162](https://github.com/agentclientprotocol/claude-agent-acp/commit/7a701623df2e05aec8e552bdfd1cf573413c3471))
+* **deps:** update to @anthropic-ai/claude-agent-sdk 0.3.170 ([#761](https://github.com/agentclientprotocol/claude-agent-acp/issues/761)) ([d8af943](https://github.com/agentclientprotocol/claude-agent-acp/commit/d8af943a1efef9be27d771cec089d2f1cb56c523))
+
+## [0.43.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.42.0...v0.43.0) (2026-06-09)
+
+
+### Features
+
+* Add experimental elicitation support ([#756](https://github.com/agentclientprotocol/claude-agent-acp/issues/756)) ([12bd276](https://github.com/agentclientprotocol/claude-agent-acp/commit/12bd2762d9ba0ccb7497a3848a5928b8bb2ce820))
+* **deps:** update to @anthropic-ai/claude-agent-sdk 0.3.169 ([#754](https://github.com/agentclientprotocol/claude-agent-acp/issues/754)) ([bd0ae4d](https://github.com/agentclientprotocol/claude-agent-acp/commit/bd0ae4dde160f59559ab4ac1bf703fcf151eff17))
+* Update ACP SDK to 0.25.0 ([#753](https://github.com/agentclientprotocol/claude-agent-acp/issues/753)) ([0dbccf5](https://github.com/agentclientprotocol/claude-agent-acp/commit/0dbccf588ff702015494dd29969d47f0b7402feb))
+
+
+### Bug Fixes
+
+* Forward unstreamed assistant text blocks ([#757](https://github.com/agentclientprotocol/claude-agent-acp/issues/757)) ([7ff6b7f](https://github.com/agentclientprotocol/claude-agent-acp/commit/7ff6b7fd6b157aedbce0c636087f51ae7d1df3a7))
+* Validate cwd before creating sessions ([#751](https://github.com/agentclientprotocol/claude-agent-acp/issues/751)) ([9854b0c](https://github.com/agentclientprotocol/claude-agent-acp/commit/9854b0c6790ebafff514e3c4c8e03a131f065a4a)), closes [#749](https://github.com/agentclientprotocol/claude-agent-acp/issues/749)
+
+## [0.42.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.41.0...v0.42.0) (2026-06-05)
+
+
+### Features
+
+* **deps-dev:** bump obug from 2.1.1 to 2.1.2 in the minor group ([#743](https://github.com/agentclientprotocol/claude-agent-acp/issues/743)) ([7ffb78c](https://github.com/agentclientprotocol/claude-agent-acp/commit/7ffb78c57280fc32af77862896c0035ce0e690b7))
+* Send message ids for assistant and user messages ([#750](https://github.com/agentclientprotocol/claude-agent-acp/issues/750)) ([18516a3](https://github.com/agentclientprotocol/claude-agent-acp/commit/18516a3088f847a38f5eee0fe81ad5ca3e2a751c))
+* Update Claude Agent SDK to 0.3.165 ([#746](https://github.com/agentclientprotocol/claude-agent-acp/issues/746)) ([23cbe6f](https://github.com/agentclientprotocol/claude-agent-acp/commit/23cbe6f58fa9e10ad736f664c4cf94dbd1c6ad65))
+
+
+### Bug Fixes
+
+* Prevent cross-family model matching in resolveModelPreference ([#731](https://github.com/agentclientprotocol/claude-agent-acp/issues/731)) ([f4704c1](https://github.com/agentclientprotocol/claude-agent-acp/commit/f4704c168f917e6710f2b3cd49e1fef49613a469))
+* Prune tool cache per session after results ([#748](https://github.com/agentclientprotocol/claude-agent-acp/issues/748)) ([ec14211](https://github.com/agentclientprotocol/claude-agent-acp/commit/ec142110561f680330c6976783805685c5c069ec))
+* Update to better utilize existing SDK data ([#747](https://github.com/agentclientprotocol/claude-agent-acp/issues/747)) ([398f763](https://github.com/agentclientprotocol/claude-agent-acp/commit/398f763f71bd2fe2ca9e8269de120d26a744b7a8))
+
+## [0.41.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.40.0...v0.41.0) (2026-06-04)
+
+
+### Features
+
+* **deps:** upgrade to @anthropic-ai/claude-agent-sdk@0.3.162 ([#740](https://github.com/agentclientprotocol/claude-agent-acp/issues/740)) ([add7e31](https://github.com/agentclientprotocol/claude-agent-acp/commit/add7e31704581aa70febb6341ca0e8cee5d93b43))
+
+
+### Bug Fixes
+
+* Force cancellation when SDK query hangs ([#742](https://github.com/agentclientprotocol/claude-agent-acp/issues/742)) ([cffea4b](https://github.com/agentclientprotocol/claude-agent-acp/commit/cffea4be28e6def91027110433a99e80b1be8a0c)), closes [#680](https://github.com/agentclientprotocol/claude-agent-acp/issues/680)
+
+## [0.40.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.39.0...v0.40.0) (2026-06-02)
+
+
+### Features
+
+* **deps-dev:** bump the minor group with 2 updates ([#726](https://github.com/agentclientprotocol/claude-agent-acp/issues/726)) ([51a370e](https://github.com/agentclientprotocol/claude-agent-acp/commit/51a370e4feb8a74279698b156e35b96631a74d74))
+* **deps:** bump actions/checkout from 6.0.2 to 6.0.3 ([#736](https://github.com/agentclientprotocol/claude-agent-acp/issues/736)) ([9cda07b](https://github.com/agentclientprotocol/claude-agent-acp/commit/9cda07ba9df27b48bd94332e365eed982e3f5613))
+* **deps:** bump the minor group with 50 updates ([#737](https://github.com/agentclientprotocol/claude-agent-acp/issues/737)) ([32175b8](https://github.com/agentclientprotocol/claude-agent-acp/commit/32175b83a0227d0baa1f7d7eedabe743b4b870e8))
+
+
+### Bug Fixes
+
+* Optimize local command marker stripping ([#738](https://github.com/agentclientprotocol/claude-agent-acp/issues/738)) ([895422c](https://github.com/agentclientprotocol/claude-agent-acp/commit/895422c72d5f955903f227dc6887df823d2730e6)), closes [#727](https://github.com/agentclientprotocol/claude-agent-acp/issues/727)
+
+## [0.39.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.38.0...v0.39.0) (2026-05-29)
+
+
+### Features
+
+* **deps:** update Claude Agent SDK to 0.3.156 ([#722](https://github.com/agentclientprotocol/claude-agent-acp/issues/722)) ([3fb6db4](https://github.com/agentclientprotocol/claude-agent-acp/commit/3fb6db4e5ee5bd1f431cf5d2ce5331ce983f9597))
+
+## [0.38.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.37.0...v0.38.0) (2026-05-28)
+
+
+### Features
+
+* **deps:** bump to @anthropic-ai/claude-agent-sdk 0.3.154 (Opus 4.8) ([#716](https://github.com/agentclientprotocol/claude-agent-acp/issues/716)) ([a172885](https://github.com/agentclientprotocol/claude-agent-acp/commit/a172885c80fc32c108a8501268a543641088f543))
+* Support Opus 4.8 ([#718](https://github.com/agentclientprotocol/claude-agent-acp/issues/718)) ([98b54a0](https://github.com/agentclientprotocol/claude-agent-acp/commit/98b54a02c267cdef619becceccf6ef519fbf8f6f))
+
+
+### Bug Fixes
+
+* Remove hide Claude auth flag handling ([#707](https://github.com/agentclientprotocol/claude-agent-acp/issues/707)) ([7ed1daf](https://github.com/agentclientprotocol/claude-agent-acp/commit/7ed1daf7b42088ad149e9bf742f764752b82d093))
+
+## [0.37.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.36.1...v0.37.0) (2026-05-21)
+
+
+### Features
+
+* add 'Default' option to effort level config ([#701](https://github.com/agentclientprotocol/claude-agent-acp/issues/701)) ([9e259d1](https://github.com/agentclientprotocol/claude-agent-acp/commit/9e259d128aa58dbb8107e53362781ca4e8ee071e))
+* **deps:** bump the claude-agent-sdk to 0.3.146 ([#700](https://github.com/agentclientprotocol/claude-agent-acp/issues/700)) ([72875b2](https://github.com/agentclientprotocol/claude-agent-acp/commit/72875b248c7b90a89638de2f1709eec9e2153e7e))
+* **deps:** bump the minor group with 13 updates ([#689](https://github.com/agentclientprotocol/claude-agent-acp/issues/689)) ([f66bc31](https://github.com/agentclientprotocol/claude-agent-acp/commit/f66bc31ae4bea73b48eeeb7f1db90c7d69cc0dac))
+* Emit tool calls for memory recall events ([#703](https://github.com/agentclientprotocol/claude-agent-acp/issues/703)) ([a0bfb98](https://github.com/agentclientprotocol/claude-agent-acp/commit/a0bfb98ba73eaf37e9c071e886c054f90a5d5629)), closes [#650](https://github.com/agentclientprotocol/claude-agent-acp/issues/650)
+
+
+### Bug Fixes
+
+* Avoid cross-version model alias matches ([#702](https://github.com/agentclientprotocol/claude-agent-acp/issues/702)) ([e1e1c69](https://github.com/agentclientprotocol/claude-agent-acp/commit/e1e1c69029ef08579c841e6fbdffb32a5b94df06))
+* Avoid redundant initial model sync ([#704](https://github.com/agentclientprotocol/claude-agent-acp/issues/704)) ([b275f6f](https://github.com/agentclientprotocol/claude-agent-acp/commit/b275f6ff7f0f2e21adedbd8e63a6fc3d63cbbb8d)), closes [#646](https://github.com/agentclientprotocol/claude-agent-acp/issues/646)
+* Don't expose /clear in commands ([#705](https://github.com/agentclientprotocol/claude-agent-acp/issues/705)) ([cfce130](https://github.com/agentclientprotocol/claude-agent-acp/commit/cfce1307076e93f43a3ed8cc0134f9bb14a0f2d6))
+* emit  "cancelled" instead of "end_turn" when the session was interrupted. ([#694](https://github.com/agentclientprotocol/claude-agent-acp/issues/694)) ([2414a6f](https://github.com/agentclientprotocol/claude-agent-acp/commit/2414a6f98bec5bf50f3a13af528aa38d5a0fc974))
+* Recover prompt stream after a failed turn ([#706](https://github.com/agentclientprotocol/claude-agent-acp/issues/706)) ([2711f50](https://github.com/agentclientprotocol/claude-agent-acp/commit/2711f506d5799f0ae25160de311a4459ffb46c49)), closes [#654](https://github.com/agentclientprotocol/claude-agent-acp/issues/654)
+
+## [0.36.1](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.36.0...v0.36.1) (2026-05-18)
+
+
+### Bug Fixes
+
+* flaky authentication bypass for gateway ([#686](https://github.com/agentclientprotocol/claude-agent-acp/issues/686)) ([db852dc](https://github.com/agentclientprotocol/claude-agent-acp/commit/db852dcf7c8e3b461fabb1bdcf0c80a27d1da77d))
+
+## [0.36.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.35.0...v0.36.0) (2026-05-18)
+
+
+### Features
+
+* Add experimental session delete support ([#682](https://github.com/agentclientprotocol/claude-agent-acp/issues/682)) ([2162d5a](https://github.com/agentclientprotocol/claude-agent-acp/commit/2162d5af62d493a82381bd88bc8fc67d376e358b))
+* Support experimental additionalDirectories field ([#684](https://github.com/agentclientprotocol/claude-agent-acp/issues/684)) ([f37e9a0](https://github.com/agentclientprotocol/claude-agent-acp/commit/f37e9a0d47d8d201d98b2177083664a6781895cc))
+
+## [0.35.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.34.1...v0.35.0) (2026-05-16)
+
+
+### Features
+
+* **deps:** bump actions/create-github-app-token from 3.1.1 to 3.2.0 ([#669](https://github.com/agentclientprotocol/claude-agent-acp/issues/669)) ([47df1f2](https://github.com/agentclientprotocol/claude-agent-acp/commit/47df1f2a0be57ddb8873ae8e341f748471c7bc53))
+* **deps:** bump hono from 4.12.18 to 4.12.19 in the minor group ([#671](https://github.com/agentclientprotocol/claude-agent-acp/issues/671)) ([5d7165d](https://github.com/agentclientprotocol/claude-agent-acp/commit/5d7165d96158da2ddf3087676cdb0a19d1ff57aa))
+* **deps:** update to claude-agent-sdk 0.3.143 ([#664](https://github.com/agentclientprotocol/claude-agent-acp/issues/664)) ([27ca2e5](https://github.com/agentclientprotocol/claude-agent-acp/commit/27ca2e5d40917887671cc4c39a854fd8f92c01e6))
+* Use SDK settings resolution for defaults ([#677](https://github.com/agentclientprotocol/claude-agent-acp/issues/677)) ([eb1259c](https://github.com/agentclientprotocol/claude-agent-acp/commit/eb1259cf88375ca3c20a219ab99f12f5488fb360))
+
+
+### Bug Fixes
+
+* Add task hooks for plan state updates ([#676](https://github.com/agentclientprotocol/claude-agent-acp/issues/676)) ([5ff7d50](https://github.com/agentclientprotocol/claude-agent-acp/commit/5ff7d50f0ec7f34a7d9d901223aed73e2dbcce68))
+* render local-command-stdout messages instead of dropping them ([#649](https://github.com/agentclientprotocol/claude-agent-acp/issues/649)) ([3b9b7d5](https://github.com/agentclientprotocol/claude-agent-acp/commit/3b9b7d5a56defd925eb2038fa97b6484ad951587))
+
+## [0.34.1](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.34.0...v0.34.1) (2026-05-16)
+
+
+### Bug Fixes
+
+* load credential error for bedrock gateway ([#667](https://github.com/agentclientprotocol/claude-agent-acp/issues/667)) ([8d76be3](https://github.com/agentclientprotocol/claude-agent-acp/commit/8d76be356a3e74c58a7c93c3c201cca6a317d784))
+
+## [0.34.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.33.1...v0.34.0) (2026-05-15)
+
+
+### Features
+
+* add bedrock gateway authentication ([#665](https://github.com/agentclientprotocol/claude-agent-acp/issues/665)) ([002c63a](https://github.com/agentclientprotocol/claude-agent-acp/commit/002c63a78b844c168d3dfc744cf408362a6adf97))
+
+## [0.33.1](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.33.0...v0.33.1) (2026-05-07)
+
+
+### Bug Fixes
+
+* Honor availableModels settings allowlist ([#637](https://github.com/agentclientprotocol/claude-agent-acp/issues/637)) ([867a3a0](https://github.com/agentclientprotocol/claude-agent-acp/commit/867a3a0de2a050592d79aa76d7dd3dd5b478162d)), closes [#620](https://github.com/agentclientprotocol/claude-agent-acp/issues/620)
+
+## [0.33.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.32.0...v0.33.0) (2026-05-07)
+
+
+### Features
+
+* **deps:** bump the minor group with 14 updates ([#631](https://github.com/agentclientprotocol/claude-agent-acp/issues/631)) ([8b43ee8](https://github.com/agentclientprotocol/claude-agent-acp/commit/8b43ee813f05f5087a213a7f42154e39d6bc4a5a))
+* **deps:** Update to claude-agent-sdk 0.2.132 ([#636](https://github.com/agentclientprotocol/claude-agent-acp/issues/636)) ([0c8ff27](https://github.com/agentclientprotocol/claude-agent-acp/commit/0c8ff277f2af8d3085cc1b5cf33891630dede0b1))
+
+
+### Bug Fixes
+
+* Handle result origins in ACP agent ([#627](https://github.com/agentclientprotocol/claude-agent-acp/issues/627)) ([dba1998](https://github.com/agentclientprotocol/claude-agent-acp/commit/dba199839a25ba405f04b9d4409e647150fe0290))
+
+## [0.32.0](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.31.4...v0.32.0) (2026-05-03)
+
+
+### Features
+
+* **deps-dev:** Bump nanoid from 3.3.11 to 3.3.12 ([#622](https://github.com/agentclientprotocol/claude-agent-acp/issues/622)) ([c78ac62](https://github.com/agentclientprotocol/claude-agent-acp/commit/c78ac62d05283e5683b87b198923818f9f556a03))
+* **deps:** Bump the minor group with 2 updates ([#619](https://github.com/agentclientprotocol/claude-agent-acp/issues/619)) ([6ccd37c](https://github.com/agentclientprotocol/claude-agent-acp/commit/6ccd37ce947fbd55691f676344b15ef1266191ad))
+* **deps:** update to @anthropic-ai/claude-agent-sdk 0.2.126 ([#621](https://github.com/agentclientprotocol/claude-agent-acp/issues/621)) ([becc3b8](https://github.com/agentclientprotocol/claude-agent-acp/commit/becc3b86e60eb33e7823a0ef27d5aa99758750d7))
+* **deps:** Update to @anthropic-ai/claude-agent-sdk@0.2.123 ([#614](https://github.com/agentclientprotocol/claude-agent-acp/issues/614)) ([5b93119](https://github.com/agentclientprotocol/claude-agent-acp/commit/5b9311938d97f47400debd66f2b4792e50639c6b))
+
+
+### Bug Fixes
+
+* emit a real diff when Write overwrites an existing file ([#618](https://github.com/agentclientprotocol/claude-agent-acp/issues/618)) ([8d7e220](https://github.com/agentclientprotocol/claude-agent-acp/commit/8d7e22026e15d54208414bd9ad128eb86d41b451))
+
 ## [0.31.4](https://github.com/agentclientprotocol/claude-agent-acp/compare/v0.31.3...v0.31.4) (2026-04-28)
 
 

eslint.config.js

@@ -40,7 +40,7 @@ export default [
       "default-case": "error",
       "prefer-const": "error",
       "no-var": "error",
-      eqeqeq: ["error", "always"],
+      eqeqeq: ["error", "smart"],
       curly: ["error", "all"],
     },
   },

examples/steering.ts

@@ -0,0 +1,191 @@
+/**
+ * Example: mid-turn steering over ACP with the `_session/steering` extension.
+ *
+ * "Steering" lets a client deliver a follow-up message to a turn that is still
+ * running, instead of waiting for it to finish and sending a fresh
+ * `session/prompt`. This is what powers "the user typed something while the
+ * agent was still working" — the new message joins the in-flight turn so the
+ * agent can adapt immediately (it shines in multi-step / tool-using turns,
+ * where the message slots in between tool calls).
+ *
+ * The wire protocol (see ../steering_protocol.md) has three moving parts:
+ *
+ *   1. The agent advertises support in its `initialize` response, at the
+ *      top-level `_meta.steering.supported` (a sibling of `agentCapabilities`).
+ *   2. The client calls the `_session/steering` request with `{ sessionId,
+ *      prompt }` while a turn is running.
+ *   3. The agent replies with an `outcome`:
+ *        - "injected"       the message joined the running turn;
+ *        - "startedNewTurn" the turn had already finished (an unavoidable race),
+ *                           so the message began a fresh turn instead.
+ *      Both are success outcomes — the message is never dropped and the race is
+ *      never surfaced as an error.
+ *
+ * This example launches the agent as a subprocess, starts a deliberately
+ * long-running prompt, and — as soon as the agent begins streaming — injects a
+ * steering message and prints the outcome. All agent output is streamed to
+ * stdout so you can watch the turn change course.
+ *
+ * Run (build the agent first so `dist/index.js` exists):
+ *
+ *   npm run build
+ *   node examples/steering.ts
+ *
+ * (Node < 22.18 needs `node --experimental-strip-types examples/steering.ts`.)
+ *
+ * Override the prompts with the PROMPT / STEER env vars. Requires the agent to
+ * be authenticated, since it talks to the real model.
+ */
+import { spawn } from "node:child_process";
+import { Readable, Writable } from "node:stream";
+import { fileURLToPath } from "node:url";
+import path from "node:path";
+import { client as acpClient, methods, ndJsonStream } from "@agentclientprotocol/sdk";
+
+/** The steering extension method, per the ACP steering wire protocol. */
+const STEERING_METHOD = "_session/steering";
+
+/** Params for a `_session/steering` request — the same shape as the relevant
+ *  subset of a `session/prompt`. */
+type SteeringRequest = {
+  sessionId: string;
+  prompt: Array<{ type: "text"; text: string }>;
+};
+
+/** Result of a `_session/steering` request. Both values are successes: they
+ *  tell the client where the message landed, not whether it succeeded. */
+type SteeringResponse = {
+  outcome: "injected" | "startedNewTurn";
+};
+
+// The built agent entry. Run `npm run build` first so this exists.
+const repoRoot = path.resolve(path.dirname(fileURLToPath(import.meta.url)), "..");
+const AGENT_ENTRY = process.env.AGENT_ENTRY ?? path.join(repoRoot, "dist", "index.js");
+const CWD = process.env.CWD ?? process.cwd();
+
+// A deliberately long-running first prompt, and the follow-up injected while it
+// is still streaming. Override either via env vars to experiment.
+const PROMPT =
+  process.env.PROMPT ??
+  "Count slowly from 1 to 30, one number per line, with a short sentence of " +
+    "commentary after each. Do not stop early.";
+const STEER =
+  process.env.STEER ?? "Actually stop counting and instead reply with exactly one line: STEERED-OK";
+
+function log(msg: string) {
+  process.stderr.write(`\x1b[2m[client]\x1b[0m ${msg}\n`);
+}
+
+const delay = (ms: number) => new Promise((resolve) => setTimeout(resolve, ms));
+
+async function main() {
+  // An ACP client launches the agent as a subprocess and speaks JSON-RPC over
+  // its stdin/stdout. stderr is inherited so the agent's own logs stay visible.
+  const child = spawn(process.execPath, [AGENT_ENTRY], {
+    stdio: ["pipe", "pipe", "inherit"],
+    env: process.env,
+  });
+  child.on("error", (err) => {
+    log(`failed to spawn agent (${AGENT_ENTRY}): ${err}`);
+    process.exit(1);
+  });
+
+  const stream = ndJsonStream(
+    Writable.toWeb(child.stdin!) as WritableStream<Uint8Array>,
+    Readable.toWeb(child.stdout!) as unknown as ReadableStream<Uint8Array>,
+  );
+
+  // Resolves the first time the agent streams assistant text — our signal that
+  // the turn is genuinely underway and therefore steerable.
+  let signalFirstOutput = () => {};
+  const firstOutput = new Promise<void>((resolve) => (signalFirstOutput = resolve));
+
+  const connection = acpClient({ name: "steering-example" })
+    .onNotification(methods.client.session.update, (ctx) => {
+      const update = ctx.params.update;
+      if (update.sessionUpdate === "agent_message_chunk" && update.content.type === "text") {
+        process.stdout.write(update.content.text);
+        signalFirstOutput();
+      }
+    })
+    // Auto-approve permission prompts so the turn is never blocked on us.
+    .onRequest(methods.client.session.requestPermission, (ctx) => {
+      const options = ctx.params.options;
+      const option = options.find((o) => o.kind === "allow_once") ?? options[0];
+      return { outcome: { outcome: "selected", optionId: option.optionId } };
+    })
+    // Minimal file-system stubs; the example prompts don't touch files.
+    .onRequest(methods.client.fs.readTextFile, () => ({ content: "" }))
+    .onRequest(methods.client.fs.writeTextFile, () => ({}))
+    .connect(stream);
+
+  const agent = connection.agent;
+
+  // 1. Initialize and confirm the agent advertises steering. Per the wire
+  //    protocol the capability lives at the TOP-LEVEL `_meta` of the initialize
+  //    result — a sibling of `agentCapabilities`, not nested inside it.
+  const init = await agent.request(methods.agent.initialize, {
+    protocolVersion: 1,
+    clientCapabilities: { fs: { readTextFile: true, writeTextFile: true } },
+  });
+  const meta = init._meta as { steering?: { supported?: boolean } } | null | undefined;
+  const steeringSupported = meta?.steering?.supported === true;
+  log(`agent advertises steering: ${steeringSupported}`);
+  if (!steeringSupported) {
+    log("agent does not advertise steering; the steering request may be rejected.");
+  }
+
+  // 2. Open a session.
+  const { sessionId } = await agent.request(methods.agent.session.new, {
+    cwd: CWD,
+    mcpServers: [],
+  });
+  log(`session: ${sessionId}`);
+
+  // 3. Start a long turn, but DON'T await it yet — we need it in flight so we
+  //    can steer it. Its output streams through the notification handler above.
+  log(`prompt: ${PROMPT}`);
+  process.stdout.write("\n----- agent output -----\n");
+  const turn = agent.request(methods.agent.session.prompt, {
+    sessionId,
+    prompt: [{ type: "text", text: PROMPT }],
+  });
+
+  // 4. Once the turn is producing output, inject the follow-up. Wait for the
+  //    first streamed chunk (with a fallback) plus a beat, so the steer clearly
+  //    lands mid-turn.
+  await Promise.race([firstOutput, delay(5000)]);
+  await delay(1000);
+
+  process.stdout.write("\n");
+  log(`steer: ${STEER}`);
+  const steerRequest: SteeringRequest = {
+    sessionId,
+    prompt: [{ type: "text", text: STEER }],
+  };
+  try {
+    const result = await agent.request<SteeringResponse>(STEERING_METHOD, steerRequest);
+    log(`steer outcome: ${result.outcome}`);
+  } catch (err) {
+    log(`steer rejected: ${err}`);
+  }
+
+  // 5. Await the turn. With outcome "injected" the steer already reshaped the
+  //    output above; with "startedNewTurn" the follow-up runs as a fresh turn
+  //    that may still be streaming, so linger briefly to capture it.
+  const response = await turn.catch((err: unknown) => {
+    log(`turn error: ${err}`);
+    return undefined;
+  });
+  if (response) log(`turn stopReason: ${response.stopReason}`);
+  await delay(3000);
+  process.stdout.write("\n----- end of agent output -----\n");
+
+  connection.close();
+  child.kill();
+}
+
+main().catch((err) => {
+  log(`fatal: ${err?.stack ?? err}`);
+  process.exit(1);
+});

package.json

@@ -3,7 +3,7 @@
   "publishConfig": {
     "access": "public"
   },
-  "version": "0.31.4",
+  "version": "0.61.0",
   "description": "An ACP-compatible coding agent powered by the Claude Agent SDK (TypeScript)",
   "main": "dist/lib.js",
   "types": "dist/lib.d.ts",
@@ -59,23 +59,27 @@
   },
   "author": "Zed Industries",
   "license": "Apache-2.0",
+  "engines": {
+    "node": ">=22"
+  },
   "dependencies": {
-    "@agentclientprotocol/sdk": "0.21.0",
-    "@anthropic-ai/claude-agent-sdk": "0.2.121",
+    "@agentclientprotocol/sdk": "1.3.0",
+    "@anthropic-ai/claude-agent-sdk": "0.3.217",
     "zod": "^3.25.0 || ^4.0.0"
   },
   "devDependencies": {
-    "@anthropic-ai/sdk": "0.89.0",
+    "@anthropic-ai/sdk": "0.112.5",
     "@eslint/js": "10.0.1",
-    "@types/node": "25.6.0",
-    "@typescript-eslint/eslint-plugin": "8.59.1",
-    "@typescript-eslint/parser": "8.59.1",
-    "eslint": "10.2.1",
+    "@tsconfig/node22": "22.0.5",
+    "@types/node": "26.1.1",
+    "@typescript-eslint/eslint-plugin": "8.65.0",
+    "@typescript-eslint/parser": "8.65.0",
+    "eslint": "10.7.0",
     "eslint-config-prettier": "10.1.8",
-    "globals": "17.5.0",
-    "prettier": "3.8.3",
+    "globals": "17.7.0",
+    "prettier": "3.9.6",
     "ts-node": "10.9.2",
     "typescript": "6.0.3",
-    "vitest": "4.1.5"
+    "vitest": "4.1.10"
   }
 }

src/elicitation.ts

@@ -0,0 +1,404 @@
+import { randomUUID } from "node:crypto";
+import { CreateElicitationResponse } from "@agentclientprotocol/sdk";
+import type {
+  CreateElicitationRequest,
+  ElicitationAcceptAction,
+  ElicitationPropertySchema,
+  ElicitationSchema,
+  EnumOption,
+} from "@agentclientprotocol/sdk";
+import type { ElicitationRequest, ElicitationResult } from "@anthropic-ai/claude-agent-sdk";
+import type {
+  AskUserQuestionInput,
+  AskUserQuestionOutput,
+} from "@anthropic-ai/claude-agent-sdk/sdk-tools.js";
+
+/**
+ * Bridges between the Claude Agent SDK's elicitation/dialog callbacks and ACP's
+ * (unstable) elicitation protocol.
+ *
+ * Two distinct SDK surfaces flow through here:
+ *
+ *   1. `onElicitation` — fired when an MCP server requests user input. These map
+ *      directly onto ACP `session/create_elicitation` (form or url mode).
+ *   2. The built-in AskUserQuestion tool — when a `canUseTool` callback is
+ *      registered the SDK routes its permission check through `canUseTool`
+ *      (not the interactive `permission_ask_user_question` dialog). We render
+ *      its questions as an ACP form elicitation and feed the user's selections
+ *      back as the tool's `updatedInput`, which the tool's own `call()` reads.
+ */
+
+/** Modes the connected client advertised support for. */
+export type ElicitationSupport = {
+  form: boolean;
+  url: boolean;
+};
+
+/**
+ * Convert an MCP elicitation request (from the SDK's `onElicitation` callback)
+ * into an ACP `CreateElicitationRequest`. Returns `null` when the request can't
+ * be represented (e.g. a url-mode request with no url).
+ */
+export function mcpElicitationToCreateRequest(
+  request: ElicitationRequest,
+  sessionId: string,
+): CreateElicitationRequest | null {
+  if (request.mode === "url") {
+    if (!request.url) {
+      return null;
+    }
+    return {
+      mode: "url",
+      sessionId,
+      message: request.message,
+      url: request.url,
+      // URL elicitations need a stable id so the client can correlate the
+      // later `session/complete_elicitation` notification. MCP servers usually
+      // provide one; fall back to a generated id if not.
+      elicitationId: request.elicitationId ?? randomUUID(),
+    };
+  }
+
+  // Form mode (the default). The MCP `requestedSchema` is already a JSON Schema
+  // with primitive-typed properties, which is structurally what ACP expects.
+  return {
+    mode: "form",
+    sessionId,
+    message: request.message,
+    requestedSchema: normalizeElicitationSchema(request.requestedSchema),
+  };
+}
+
+/**
+ * Content of an accepted elicitation response.
+ *
+ * Uses the SDK's validating guard rather than an `action === "accept"` check:
+ * the guard both narrows past the union's custom/future variant and validates
+ * the payload, so a malformed accept (right tag, ill-typed content) yields
+ * empty content — the same classification the SDK's wire validators apply.
+ */
+function acceptedElicitationContent(
+  response: CreateElicitationResponse,
+): NonNullable<ElicitationAcceptAction["content"]> {
+  return CreateElicitationResponse.isAccept(response) ? (response.content ?? {}) : {};
+}
+
+/**
+ * Map an ACP elicitation response back to the MCP `ElicitResult` the SDK expects
+ * to hand back to the requesting server.
+ */
+export function createElicitationResponseToElicitResult(
+  response: CreateElicitationResponse,
+): ElicitationResult {
+  switch (response.action) {
+    case "accept":
+      return { action: "accept", content: acceptedElicitationContent(response) };
+    case "decline":
+      return { action: "decline" };
+    case "cancel":
+    default:
+      return { action: "cancel" };
+  }
+}
+
+/**
+ * A single question as supplied by the AskUserQuestion tool. Derived from the
+ * SDK's input type so the shape stays in sync; the SDK validates the model's
+ * tool call against this schema before it reaches us.
+ */
+export type AskUserQuestion = AskUserQuestionInput["questions"][number];
+
+/**
+ * Pull the well-formed questions out of an AskUserQuestion tool input. Returns
+ * `null` when there are no usable questions — including the case where every
+ * entry is malformed and filtering leaves an empty list — so callers can treat
+ * "nothing to ask" uniformly.
+ */
+export function extractAskUserQuestions(input: Record<string, unknown>): AskUserQuestion[] | null {
+  const questions = (input as { questions?: unknown }).questions;
+  if (!Array.isArray(questions)) {
+    return null;
+  }
+  const valid = questions.filter(
+    (q): q is AskUserQuestion =>
+      !!q && typeof q.question === "string" && Array.isArray(q.options) && q.options.length > 0,
+  );
+  return valid.length > 0 ? valid : null;
+}
+
+/** Stable form-field key for the question at the given index. */
+function questionFieldKey(index: number): string {
+  return `question_${index}`;
+}
+
+/**
+ * Form-field key for the per-question free-text "custom answer" field that sits
+ * alongside `question_<n>`. Mirrors the first-party clients, where every
+ * question carries its own "Other" box rather than one form-level field.
+ */
+function questionCustomFieldKey(index: number): string {
+  return `question_${index}_custom`;
+}
+
+/**
+ * `_meta` key under which a bridged enum option carries its `preview`, the one
+ * option field ACP's `EnumOption` still has no slot for (descriptions are
+ * first-class as of schema 1.19). Namespaced like the agent's other `_meta`
+ * extensions (`_claude/...`).
+ */
+const OPTION_META_KEY = "_claude/askUserQuestionOption";
+
+/**
+ * Render the AskUserQuestion tool's questions as an ACP form elicitation.
+ *
+ * Fields are keyed by a short stable id (`question_<n>`) rather than the full
+ * question text, so the question text appears in exactly one place per field.
+ * Single-select questions use a titled `oneOf` enum; multi-select questions use
+ * an array with a titled `anyOf` item enum. The enum `const` is always the
+ * option label, since that is what the tool records as the answer; an option's
+ * secondary text travels in the enum option's own `description` field.
+ *
+ * Each question is followed by its own optional free-text "custom answer" field
+ * (`question_<n>_custom`), mirroring the CLI's per-question "Other" box: the
+ * user can type their own answer instead of picking an option, scoped to that
+ * specific question. Nothing is marked required, so the user can also just skip
+ * — matching the built-in tool, which always offers Skip + a free-text box.
+ */
+export function askUserQuestionsToCreateRequest(
+  questions: AskUserQuestion[],
+  sessionId: string,
+  toolCallId: string | undefined,
+): CreateElicitationRequest {
+  const single = questions.length === 1;
+  const properties: Record<string, ElicitationPropertySchema> = {};
+
+  questions.forEach((question, index) => {
+    const options: EnumOption[] = question.options.map((option) => {
+      const enumOption: EnumOption = {
+        const: option.label,
+        title: option.label,
+      };
+      if (option.description) {
+        enumOption.description = option.description;
+      }
+      // The SDK option's `preview` (mockups, code snippets, comparisons shown
+      // on focus) still has no structural slot in `EnumOption`, so forward it
+      // under ACP's reserved `_meta` extension point for clients that render it.
+      if (option.preview) {
+        enumOption._meta = { [OPTION_META_KEY]: { preview: option.preview } };
+      }
+      return enumOption;
+    });
+
+    // For a single question the prompt is carried by `message`, so we don't
+    // repeat it in the field description. With multiple questions each field
+    // needs its own question text.
+    const description = single ? undefined : question.question;
+    const title = question.header || undefined;
+
+    properties[questionFieldKey(index)] = question.multiSelect
+      ? { type: "array", title, description, items: { anyOf: options } }
+      : { type: "string", title, description, oneOf: options };
+
+    properties[questionCustomFieldKey(index)] = {
+      type: "string",
+      title: "Other",
+      description: "Type your own answer instead of choosing an option above (optional).",
+    };
+  });
+
+  const requestedSchema: ElicitationSchema = {
+    type: "object",
+    properties,
+  };
+
+  const message = single ? questions[0].question : "Please answer the following questions.";
+
+  return {
+    mode: "form",
+    sessionId,
+    ...(toolCallId ? { toolCallId } : {}),
+    message,
+    requestedSchema,
+  };
+}
+
+/** Outcome of an AskUserQuestion elicitation, decoupled from any transport. */
+export type AskUserQuestionOutcome =
+  { action: "answered"; updatedInput: Record<string, unknown> } | { action: "cancel" };
+
+/**
+ * Fold an ACP elicitation response into the AskUserQuestion tool's input.
+ *
+ * Selected labels are read back from the indexed form fields and written into
+ * `answers` as a `{ [questionText]: label }` map (comma-joining multi-selects)
+ * — the key shape the tool's own `call()` reads. A non-empty per-question
+ * custom-answer field (`question_<n>_custom`) takes precedence over that
+ * question's selection, since the user typed their own answer instead of
+ * picking one. Decline yields empty answers (the model is told the user skipped
+ * rather than the turn aborting); cancel — and any custom/future action we
+ * don't understand — aborts the tool call.
+ */
+export function applyAskElicitationResponse(
+  response: CreateElicitationResponse,
+  toolInput: Record<string, unknown>,
+  questions: AskUserQuestion[],
+): AskUserQuestionOutcome {
+  if (response.action === "decline") {
+    return { action: "answered", updatedInput: { ...toolInput, answers: {} } };
+  }
+
+  if (response.action !== "accept") {
+    return { action: "cancel" };
+  }
+
+  const content = acceptedElicitationContent(response);
+  // Typed against the tool's own output schema so the answer/response shapes
+  // stay in sync with what the built-in tool's call() expects to read back.
+  const answers: AskUserQuestionOutput["answers"] = {};
+  questions.forEach((question, index) => {
+    // A typed custom answer wins over the selection: the user chose to write
+    // their own answer for this question instead of picking an option.
+    const custom = content[questionCustomFieldKey(index)];
+    if (typeof custom === "string" && custom.trim() !== "") {
+      answers[question.question] = custom.trim();
+      return;
+    }
+
+    const value = content[questionFieldKey(index)];
+    if (value === undefined || value === null) {
+      return;
+    }
+    const text = Array.isArray(value) ? value.join(", ") : String(value);
+    if (text === "") {
+      return;
+    }
+    answers[question.question] = text;
+  });
+
+  return { action: "answered", updatedInput: { ...toolInput, answers } };
+}
+
+/**
+ * Coerce an arbitrary MCP `requestedSchema` into an ACP `ElicitationSchema`.
+ * The two are structurally compatible JSON Schemas; we just guarantee the
+ * `type: "object"` discriminator is present.
+ */
+function normalizeElicitationSchema(
+  schema: Record<string, unknown> | undefined,
+): ElicitationSchema {
+  if (!schema || typeof schema !== "object") {
+    return { type: "object", properties: {} };
+  }
+  return { ...(schema as ElicitationSchema), type: "object" };
+}
+
+/**
+ * The `request_user_dialog` kind the CLI emits when a model refusal has a
+ * fallback available but needs user consent before retrying (e.g. Claude Fable
+ * declining a request with Opus available as the fallback). Declaring this
+ * kind in `supportedDialogKinds` is the opt-in: the CLI fails closed and never
+ * emits an undeclared kind — the flow degrades to the classic refusal error
+ * ending the turn.
+ */
+export const REFUSAL_FALLBACK_DIALOG_KIND = "refusal_fallback_prompt";
+
+/**
+ * Payload of the `refusal_fallback_prompt` dialog. The dialog protocol
+ * transports payloads opaquely, so this shape is recovered from the CLI's own
+ * schema (v2.1.177): `originalModel`/`fallbackModel` are required strings;
+ * `apiRefusalCategory` (nullable), `guidanceText`, and
+ * `retractedMessageUuids` are optional. We ignore `retractedMessageUuids` —
+ * ACP has no way to retract already-streamed chunks.
+ */
+export type RefusalFallbackPrompt = {
+  originalModel: string;
+  fallbackModel: string;
+  apiRefusalCategory: string | null;
+  guidanceText?: string;
+};
+
+/**
+ * Validate the opaque dialog payload into a {@link RefusalFallbackPrompt}.
+ * Returns `null` when the required fields are missing or mistyped (a newer CLI
+ * may reshape the payload), so the caller can cancel the dialog and let the
+ * CLI apply its default behavior instead of rendering something misleading.
+ */
+export function extractRefusalFallbackPrompt(
+  payload: Record<string, unknown>,
+): RefusalFallbackPrompt | null {
+  const { originalModel, fallbackModel, apiRefusalCategory, guidanceText } = payload;
+  if (typeof originalModel !== "string" || typeof fallbackModel !== "string") {
+    return null;
+  }
+  return {
+    originalModel,
+    fallbackModel,
+    apiRefusalCategory: typeof apiRefusalCategory === "string" ? apiRefusalCategory : null,
+    ...(typeof guidanceText === "string" && guidanceText ? { guidanceText } : {}),
+  };
+}
+
+/** Form-field key carrying the user's choice in the refusal-fallback form. */
+const REFUSAL_FALLBACK_CHOICE_KEY = "choice";
+
+/** Wire values of the dialog's result enum (CLI schema). `edit_prompt` is
+ *  deliberately not offered: in the CLI it prefills the composer with the
+ *  refused prompt for edit-and-retry, and ACP has no composer-prefill surface
+ *  — the user can simply edit and resend on their own. */
+const RETRY_FALLBACK_RESULT = "retry_fallback";
+const KEEP_REFUSAL_RESULT = "cancelled";
+
+/**
+ * Render the refusal-fallback consent prompt as an ACP form elicitation: a
+ * single-select between retrying on the fallback model and keeping the
+ * refusal. The enum `const`s are the dialog's wire result values, so the
+ * response maps back without a translation table.
+ */
+export function refusalFallbackToCreateRequest(
+  prompt: RefusalFallbackPrompt,
+  sessionId: string,
+): CreateElicitationRequest {
+  const category = prompt.apiRefusalCategory ? ` (${prompt.apiRefusalCategory})` : "";
+  const guidance = prompt.guidanceText ? `\n\n${prompt.guidanceText}` : "";
+  return {
+    mode: "form",
+    sessionId,
+    message:
+      `${prompt.originalModel} declined this request${category}. ` +
+      `Retry with ${prompt.fallbackModel}?` +
+      guidance,
+    requestedSchema: {
+      type: "object",
+      properties: {
+        [REFUSAL_FALLBACK_CHOICE_KEY]: {
+          type: "string",
+          oneOf: [
+            {
+              const: RETRY_FALLBACK_RESULT,
+              title: `Retry with ${prompt.fallbackModel}`,
+              description: `The session continues on ${prompt.fallbackModel}.`,
+            },
+            {
+              const: KEEP_REFUSAL_RESULT,
+              title: "Keep the refusal",
+              description: "You can send a new message.",
+            },
+          ],
+        },
+      },
+    },
+  };
+}
+
+/**
+ * Map the elicitation response back to the dialog's result enum. Only an
+ * explicit accept-with-retry resolves to `retry_fallback`; decline, cancel, a
+ * skipped field, or an unrecognized value all keep the refusal — the dialog's
+ * own default — so a dismissed or half-filled form can never trigger a model
+ * switch the user didn't ask for.
+ */
+export function refusalFallbackResultFromResponse(response: CreateElicitationResponse): string {
+  const choice = acceptedElicitationContent(response)[REFUSAL_FALLBACK_CHOICE_KEY];
+  return choice === RETRY_FALLBACK_RESULT ? RETRY_FALLBACK_RESULT : KEEP_REFUSAL_RESULT;
+}

src/index.ts

@@ -1,9 +1,12 @@
 #!/usr/bin/env node
 
-// Load managed settings and apply environment variables
-import { loadManagedSettings, applyEnvironmentSettings } from "./utils.js";
+import { resolveSettings } from "@anthropic-ai/claude-agent-sdk";
 import { claudeCliPath, runAcp } from "./acp-agent.js";
+import packageJson from "../package.json" with { type: "json" };
 
+// `--cli` is checked first so that `--version`/`-v` (and any other flags) are
+// forwarded to the wrapped native CLI rather than swallowed by our own version
+// handler below. Our version flag only applies when not delegating.
 if (process.argv.includes("--cli")) {
   const { spawn } = await import("node:child_process");
   const args = process.argv.slice(2).filter((arg) => arg !== "--cli");
@@ -34,10 +37,17 @@ if (process.argv.includes("--cli")) {
     console.error(err);
     process.exit(1);
   });
+} else if (process.argv.includes("--version") || process.argv.includes("-v")) {
+  console.log(packageJson.version);
+  process.exit(0);
 } else {
-  const managedSettings = loadManagedSettings();
-  if (managedSettings) {
-    applyEnvironmentSettings(managedSettings);
+  // Apply env vars from the managed-policy tier before any SDK call so the
+  // SDK subprocess inherits them. Going through resolveSettings (vs. a raw
+  // read of managed-settings.json) also picks up MDM sources on macOS and
+  // HKLM/HKCU on Windows.
+  const policy = await resolveSettings({ settingSources: [] });
+  for (const [key, value] of Object.entries(policy.effective.env ?? {})) {
+    process.env[key] = value;
   }
 
   // stdout is used to send messages to the client

src/lib.ts

@@ -10,25 +10,14 @@ export {
   type NewSessionMeta,
   type SDKMessageFilter,
 } from "./acp-agent.js";
-export {
-  loadManagedSettings,
-  applyEnvironmentSettings,
-  nodeToWebReadable,
-  nodeToWebWritable,
-  Pushable,
-  unreachable,
-} from "./utils.js";
+export { nodeToWebReadable, nodeToWebWritable, Pushable, unreachable } from "./utils.js";
 export {
   toolInfoFromToolUse,
   toDisplayPath,
   planEntries,
   toolUpdateFromToolResult,
 } from "./tools.js";
-export {
-  SettingsManager,
-  type ClaudeCodeSettings,
-  type SettingsManagerOptions,
-} from "./settings.js";
+export { SettingsManager, type SettingsManagerOptions } from "./settings.js";
 
 // Export types
 export type { ClaudePlanEntry } from "./tools.js";

src/settings.ts

@@ -1,5 +1,12 @@
 import * as fs from "node:fs";
 import * as path from "node:path";
+// resolveSettings and filterEscalatingDefaultMode are marked @alpha in the
+// SDK; API may shift in a future release.
+import {
+  filterEscalatingDefaultMode,
+  resolveSettings,
+  type Settings,
+} from "@anthropic-ai/claude-agent-sdk";
 import { CLAUDE_CONFIG_DIR } from "./acp-agent.js";
 
 /**
@@ -14,51 +21,10 @@ import { CLAUDE_CONFIG_DIR } from "./acp-agent.js";
  * Docs: https://code.claude.com/docs/en/iam#tool-specific-permission-rules
  */
 
-export interface PermissionSettings {
-  defaultMode?: string;
-}
-
-export interface ClaudeCodeSettings {
-  permissions?: PermissionSettings;
-  env?: Record<string, string>;
-  model?: string;
-  effortLevel?: string;
-}
-
-/**
- * Reads and parses a JSON settings file, returning an empty object if not found or invalid.
- * Silently ignores missing files (ENOENT) but logs warnings for other errors
- * (malformed JSON, permission errors, etc.) to aid debugging.
- */
-async function loadSettingsFile(
-  filePath: string | null,
-  logger?: { error: (...args: any[]) => void },
-): Promise<ClaudeCodeSettings> {
-  if (!filePath) {
-    return {};
-  }
-
-  try {
-    const content = await fs.promises.readFile(filePath, "utf-8");
-    return JSON.parse(content) as ClaudeCodeSettings;
-  } catch (error) {
-    if (error instanceof Error && "code" in error && error.code === "ENOENT") {
-      return {};
-    }
-    logger?.error(`Failed to load settings from ${filePath}:`, error);
-    return {};
-  }
-}
-
-/**
- * Gets the enterprise settings path based on the current platform
- */
-export function getManagedSettingsPath(): string {
+function getManagedSettingsPath(): string {
   switch (process.platform) {
     case "darwin":
       return "/Library/Application Support/ClaudeCode/managed-settings.json";
-    case "linux":
-      return "/etc/claude-code/managed-settings.json";
     case "win32":
       return "C:\\Program Files\\ClaudeCode\\managed-settings.json";
     default:
@@ -72,23 +38,17 @@ export interface SettingsManagerOptions {
 }
 
 /**
- * Manages Claude Code settings from multiple sources with proper precedence.
+ * Manages Claude Code settings using the SDK's `resolveSettings` merge engine
+ * so the values we see match what `query()` would observe.
  *
- * Settings are loaded from (in order of increasing precedence):
- * 1. User settings (~/.claude/settings.json)
- * 2. Project settings (<cwd>/.claude/settings.json)
- * 3. Local project settings (<cwd>/.claude/settings.local.json)
- * 4. Enterprise managed settings (platform-specific path)
- *
- * The manager watches all settings files for changes and automatically reloads.
+ * Watches the user/project/local/managed settings files for changes and
+ * re-resolves through the SDK on update. Escalating `permissions.defaultMode`
+ * values from repo-committed sources are filtered out via
+ * `filterEscalatingDefaultMode`, matching the CLI's trust policy.
  */
 export class SettingsManager {
   private cwd: string;
-  private userSettings: ClaudeCodeSettings = {};
-  private projectSettings: ClaudeCodeSettings = {};
-  private localSettings: ClaudeCodeSettings = {};
-  private enterpriseSettings: ClaudeCodeSettings = {};
-  private mergedSettings: ClaudeCodeSettings = {};
+  private effective: Settings = {};
   private watchers: fs.FSWatcher[] = [];
   private onChange?: () => void;
   private logger: { log: (...args: any[]) => void; error: (...args: any[]) => void };
@@ -126,98 +86,37 @@ export class SettingsManager {
   }
 
   /**
-   * Returns the path to the user settings file
+   * Paths the SDK reads when resolving settings for this cwd. Watching the
+   * containing directories means we pick up file creation as well as edits.
    */
-  private getUserSettingsPath(): string {
-    return path.join(CLAUDE_CONFIG_DIR, "settings.json");
+  private getWatchedPaths(): string[] {
+    return [
+      path.join(CLAUDE_CONFIG_DIR, "settings.json"),
+      path.join(this.cwd, ".claude", "settings.json"),
+      path.join(this.cwd, ".claude", "settings.local.json"),
+      getManagedSettingsPath(),
+    ];
   }
 
   /**
-   * Returns the path to the project settings file
-   */
-  private getProjectSettingsPath(): string {
-    return path.join(this.cwd, ".claude", "settings.json");
-  }
-
-  /**
-   * Returns the path to the local project settings file
-   */
-  private getLocalSettingsPath(): string {
-    return path.join(this.cwd, ".claude", "settings.local.json");
-  }
-
-  /**
-   * Loads settings from all sources
+   * Resolves the effective settings via the SDK and applies the CLI's trust
+   * filter for escalating `permissions.defaultMode` values.
    */
   private async loadAllSettings(): Promise<void> {
-    const [userSettings, projectSettings, localSettings, enterpriseSettings] = await Promise.all([
-      loadSettingsFile(this.getUserSettingsPath(), this.logger),
-      loadSettingsFile(this.getProjectSettingsPath(), this.logger),
-      loadSettingsFile(this.getLocalSettingsPath(), this.logger),
-      loadSettingsFile(getManagedSettingsPath(), this.logger),
-    ]);
-
-    this.userSettings = userSettings;
-    this.projectSettings = projectSettings;
-    this.localSettings = localSettings;
-    this.enterpriseSettings = enterpriseSettings;
-
-    this.mergeSettings();
-  }
-
-  /**
-   * Merges all settings sources with proper precedence.
-   * For permissions, rules from all sources are combined.
-   * Deny rules always take precedence during permission checks.
-   */
-  private mergeSettings(): void {
-    const allSettings = [
-      this.userSettings,
-      this.projectSettings,
-      this.localSettings,
-      this.enterpriseSettings,
-    ];
-
-    const merged: ClaudeCodeSettings = {};
-
-    for (const settings of allSettings) {
-      if (settings.env) {
-        merged.env = { ...merged.env, ...settings.env };
-      }
-
-      if (settings.model) {
-        merged.model = settings.model;
-      }
-
-      if (settings.effortLevel !== undefined) {
-        merged.effortLevel = settings.effortLevel;
-      }
-
-      if (settings.permissions?.defaultMode !== undefined) {
-        merged.permissions = {
-          ...merged.permissions,
-          defaultMode: settings.permissions.defaultMode,
-        };
-      }
+    try {
+      const resolved = await resolveSettings({ cwd: this.cwd });
+      this.effective = filterEscalatingDefaultMode(resolved);
+    } catch (error) {
+      this.logger.error("Failed to resolve settings:", error);
+      this.effective = {};
     }
-
-    this.mergedSettings = merged;
   }
 
   /**
    * Sets up file watchers for all settings files
    */
   private setupWatchers(): void {
-    const paths = [
-      this.getUserSettingsPath(),
-      this.getProjectSettingsPath(),
-      this.getLocalSettingsPath(),
-      getManagedSettingsPath(),
-    ];
-
-    for (const filePath of paths) {
-      if (!filePath) continue;
-
+    for (const filePath of this.getWatchedPaths()) {
       try {
         const dir = path.dirname(filePath);
         const filename = path.basename(filePath);
@@ -268,8 +167,8 @@ export class SettingsManager {
   /**
    * Returns the current merged settings
    */
-  getSettings(): ClaudeCodeSettings {
-    return this.mergedSettings;
+  getSettings(): Settings {
+    return this.effective;
   }
 
   /**

src/tests/.claude/commands/say-hello.md

@@ -3,4 +3,4 @@ description: Say hello
 argument-hint: name
 ---
 
-Respond with "Hello $1" and nothing else.
+Respond with "Hello $ARGUMENTS" and nothing else.

src/tests/acp-agent-settings.test.ts

@@ -2,8 +2,8 @@ import { describe, it, expect, beforeEach, afterEach, vi } from "vitest";
 import * as fs from "node:fs";
 import * as path from "node:path";
 import * as os from "node:os";
-import type { AgentSideConnection } from "@agentclientprotocol/sdk";
-import type { ClaudeAcpAgent as ClaudeAcpAgentType } from "../acp-agent.js";
+import type { AcpClient, ClaudeAcpAgent as ClaudeAcpAgentType } from "../acp-agent.js";
+import { makeMockQuery } from "./helpers.js";
 
 const { querySpy } = vi.hoisted(() => ({
   querySpy: vi.fn(),
@@ -21,13 +21,13 @@ describe("ClaudeAcpAgent settings", () => {
   let tempDir: string;
   let originalClaudeConfigDir: string | undefined;
 
-  function createMockClient(): AgentSideConnection {
+  function createMockClient(): AcpClient {
     return {
       sessionUpdate: async () => {},
       requestPermission: async () => ({ outcome: { outcome: "cancelled" } }),
       readTextFile: async () => ({ content: "" }),
       writeTextFile: async () => ({}),
-    } as unknown as AgentSideConnection;
+    } as unknown as AcpClient;
   }
 
   function mockQuery() {
@@ -35,19 +35,18 @@ describe("ClaudeAcpAgent settings", () => {
     const setModelSpy = vi.fn();
     querySpy.mockImplementation(({ options }: any) => {
       capturedOptions = options;
-      return {
+      return makeMockQuery({
         initializationResult: async () => ({
           models: [
             {
-              value: "claude-sonnet-4-5",
+              value: "claude-sonnet-4-6",
               displayName: "Claude Sonnet 4.5",
               description: "Default",
             },
           ],
         }),
         setModel: setModelSpy,
-        supportedCommands: async () => [],
-      } as any;
+      }) as any;
     });
     return { getCapturedOptions: () => capturedOptions, setModelSpy };
   }
@@ -126,6 +125,31 @@ describe("ClaudeAcpAgent settings", () => {
     expect(response.modes.currentModeId).toBe("acceptEdits");
   });
 
+  it("drops escalating defaultMode when it comes from project-tier settings", async () => {
+    // bypassPermissions in .claude/settings.json (a repo-committed tier) is
+    // filtered by the SDK's trust policy and clamps to 'default'.
+    const projectDir = path.join(tempDir, "project");
+    await fs.promises.mkdir(path.join(projectDir, ".claude"), { recursive: true });
+    await fs.promises.writeFile(
+      path.join(projectDir, ".claude", "settings.json"),
+      JSON.stringify({ permissions: { defaultMode: "bypassPermissions" } }),
+    );
+
+    const { getCapturedOptions } = mockQuery();
+
+    const { ClaudeAcpAgent } = await import("../acp-agent.js");
+    const agent: ClaudeAcpAgentType = new ClaudeAcpAgent(createMockClient());
+
+    const response = await (agent as any).createSession({
+      cwd: projectDir,
+      mcpServers: [],
+      _meta: { disableBuiltInTools: true },
+    });
+
+    expect(getCapturedOptions().permissionMode).toBe("default");
+    expect(response.modes.currentModeId).toBe("default");
+  });
+
   it("defaults to 'default' when no permissions.defaultMode is set", async () => {
     const projectDir = path.join(tempDir, "project");
     await fs.promises.mkdir(projectDir, { recursive: true });
@@ -197,8 +221,12 @@ describe("ClaudeAcpAgent settings", () => {
     });
 
     // Bad model is ignored at the usage site; falls back to the first SDK model.
-    expect(setModelSpy).toHaveBeenCalledWith("claude-sonnet-4-5");
-    expect(response.models.currentModelId).toBe("claude-sonnet-4-5");
+    // No setModel call is needed because no override was applied — the SDK is
+    // already on its own default.
+    expect(setModelSpy).not.toHaveBeenCalled();
+    expect(response.configOptions?.find((o: any) => o.id === "model")?.currentValue).toBe(
+      "claude-sonnet-4-6",
+    );
   });
 
   describe("auto mode availability per model", () => {
@@ -212,12 +240,11 @@ describe("ClaudeAcpAgent settings", () => {
       const setPermissionModeSpy = vi.fn();
       querySpy.mockImplementation(({ options }: any) => {
         capturedOptions = options;
-        return {
+        return makeMockQuery({
           initializationResult: async () => ({ models }),
           setModel: setModelSpy,
           setPermissionMode: setPermissionModeSpy,
-          supportedCommands: async () => [],
-        } as any;
+        }) as any;
       });
       return {
         getCapturedOptions: () => capturedOptions,
@@ -362,6 +389,477 @@ describe("ClaudeAcpAgent settings", () => {
     });
   });
 
+  describe("availableModels allowlist from settings", () => {
+    function mockQueryWithModels(models: any[]): {
+      setModelSpy: ReturnType<typeof vi.fn>;
+    } {
+      const setModelSpy = vi.fn();
+      querySpy.mockImplementation(() => {
+        return makeMockQuery({
+          initializationResult: async () => ({ models }),
+          setModel: setModelSpy,
+        }) as any;
+      });
+      return { setModelSpy };
+    }
+
+    it("restricts configOptions to the user's allowlist using their exact IDs", async () => {
+      // Reproduces the scenario from
+      // https://github.com/agentclientprotocol/claude-agent-acp/issues/620:
+      // user lists `claude-haiku-4-5` (no date pin) in availableModels, but
+      // the SDK still surfaces its `haiku` alias which resolves to a
+      // date-pinned variant the user doesn't have access to.
+      await fs.promises.writeFile(
+        path.join(tempDir, "settings.json"),
+        JSON.stringify({
+          availableModels: [
+            "claude-sonnet-4-6[1m]",
+            "claude-opus-4-6[1m]",
+            "claude-haiku-4-5",
+            "claude-opus-4-7[1m]",
+          ],
+        }),
+      );
+
+      const projectDir = path.join(tempDir, "project");
+      await fs.promises.mkdir(projectDir, { recursive: true });
+
+      mockQueryWithModels([
+        { value: "default", displayName: "Default", description: "Default model" },
+        {
+          value: "sonnet[1m]",
+          displayName: "Sonnet (1M context)",
+          description: "Sonnet 4.6 long context",
+        },
+        {
+          value: "opus[1m]",
+          displayName: "Opus (1M context)",
+          description: "Opus 1M context",
+        },
+        { value: "haiku", displayName: "Haiku", description: "Fast" },
+      ]);
+
+      const { ClaudeAcpAgent } = await import("../acp-agent.js");
+      const agent: ClaudeAcpAgentType = new ClaudeAcpAgent(createMockClient());
+
+      const response = await (agent as any).createSession({
+        cwd: projectDir,
+        mcpServers: [],
+        _meta: { disableBuiltInTools: true },
+      });
+
+      const modelOption = response.configOptions.find((o: any) => o.id === "model");
+      expect(modelOption.options.map((o: any) => o.value)).toEqual([
+        "default",
+        "claude-sonnet-4-6[1m]",
+        "claude-opus-4-6[1m]",
+        "claude-haiku-4-5",
+        "claude-opus-4-7[1m]",
+      ]);
+    });
+
+    it("unions availableModels across user and project settings", async () => {
+      // https://code.claude.com/docs/en/model-config#merge-behavior
+      await fs.promises.writeFile(
+        path.join(tempDir, "settings.json"),
+        JSON.stringify({ availableModels: ["claude-haiku-4-5"] }),
+      );
+
+      const projectDir = path.join(tempDir, "project");
+      await fs.promises.mkdir(path.join(projectDir, ".claude"), { recursive: true });
+      await fs.promises.writeFile(
+        path.join(projectDir, ".claude", "settings.json"),
+        JSON.stringify({
+          availableModels: ["claude-haiku-4-5", "claude-opus-4-7[1m]"],
+        }),
+      );
+
+      mockQueryWithModels([
+        { value: "default", displayName: "Default", description: "Default model" },
+        { value: "haiku", displayName: "Haiku", description: "Fast" },
+      ]);
+
+      const { ClaudeAcpAgent } = await import("../acp-agent.js");
+      const agent: ClaudeAcpAgentType = new ClaudeAcpAgent(createMockClient());
+
+      const response = await (agent as any).createSession({
+        cwd: projectDir,
+        mcpServers: [],
+        _meta: { disableBuiltInTools: true },
+      });
+
+      const modelOption = response.configOptions.find((o: any) => o.id === "model");
+      // User and project entries are unioned and deduplicated.
+      expect(modelOption.options.map((o: any) => o.value)).toEqual([
+        "default",
+        "claude-haiku-4-5",
+        "claude-opus-4-7[1m]",
+      ]);
+    });
+
+    it("returns only the default entry when availableModels is an empty array", async () => {
+      await fs.promises.writeFile(
+        path.join(tempDir, "settings.json"),
+        JSON.stringify({ availableModels: [] }),
+      );
+
+      const projectDir = path.join(tempDir, "project");
+      await fs.promises.mkdir(projectDir, { recursive: true });
+
+      mockQueryWithModels([
+        { value: "default", displayName: "Default", description: "Default model" },
+        { value: "haiku", displayName: "Haiku", description: "Fast" },
+      ]);
+
+      const { ClaudeAcpAgent } = await import("../acp-agent.js");
+      const agent: ClaudeAcpAgentType = new ClaudeAcpAgent(createMockClient());
+
+      const response = await (agent as any).createSession({
+        cwd: projectDir,
+        mcpServers: [],
+        _meta: { disableBuiltInTools: true },
+      });
+
+      const modelOption = response.configOptions.find((o: any) => o.id === "model");
+      expect(modelOption.options.map((o: any) => o.value)).toEqual(["default"]);
+    });
+
+    it("does not filter when availableModels is absent from settings", async () => {
+      const projectDir = path.join(tempDir, "project");
+      await fs.promises.mkdir(projectDir, { recursive: true });
+
+      mockQueryWithModels([
+        { value: "default", displayName: "Default", description: "Default model" },
+        { value: "haiku", displayName: "Haiku", description: "Fast" },
+      ]);
+
+      const { ClaudeAcpAgent } = await import("../acp-agent.js");
+      const agent: ClaudeAcpAgentType = new ClaudeAcpAgent(createMockClient());
+
+      const response = await (agent as any).createSession({
+        cwd: projectDir,
+        mcpServers: [],
+        _meta: { disableBuiltInTools: true },
+      });
+
+      const modelOption = response.configOptions.find((o: any) => o.id === "model");
+      expect(modelOption.options.map((o: any) => o.value)).toEqual(["default", "haiku"]);
+    });
+
+    it("passes the user's exact ID to setModel when it matches an SDK alias", async () => {
+      // Without the allowlist, the SDK would resolve `haiku` to a
+      // date-pinned variant. Forcing setModel to receive `claude-haiku-4-5`
+      // is exactly what the issue's workaround
+      // (`ANTHROPIC_DEFAULT_HAIKU_MODEL`) achieves manually.
+      await fs.promises.writeFile(
+        path.join(tempDir, "settings.json"),
+        JSON.stringify({
+          availableModels: ["claude-haiku-4-5"],
+          model: "claude-haiku-4-5",
+        }),
+      );
+
+      const projectDir = path.join(tempDir, "project");
+      await fs.promises.mkdir(projectDir, { recursive: true });
+
+      const { setModelSpy } = mockQueryWithModels([
+        { value: "default", displayName: "Default", description: "Default model" },
+        { value: "haiku", displayName: "Haiku", description: "Fast" },
+      ]);
+
+      const { ClaudeAcpAgent } = await import("../acp-agent.js");
+      const agent: ClaudeAcpAgentType = new ClaudeAcpAgent(createMockClient());
+
+      const response = await (agent as any).createSession({
+        cwd: projectDir,
+        mcpServers: [],
+        _meta: { disableBuiltInTools: true },
+      });
+
+      expect(setModelSpy).toHaveBeenCalledWith("claude-haiku-4-5");
+      expect(response.configOptions?.find((o: any) => o.id === "model")?.currentValue).toBe(
+        "claude-haiku-4-5",
+      );
+    });
+
+    it("does not inherit display info across mismatched model versions", async () => {
+      // https://github.com/agentclientprotocol/claude-agent-acp/issues/639:
+      // when the SDK's `opus` alias resolves to Opus 4.7, an allowlist entry
+      // of `claude-opus-4-6` (or `claude-opus-4-6[1m]`) used to substring-match
+      // `opus` and inherit the "Opus 4.7" display info. With version-aware
+      // matching, these entries fall back to showing their literal ID rather
+      // than a misleading newer name.
+      await fs.promises.writeFile(
+        path.join(tempDir, "settings.json"),
+        JSON.stringify({
+          availableModels: [
+            "claude-opus-4-6",
+            "claude-opus-4-6[1m]",
+            "claude-opus-4-7",
+            "claude-opus-4-7[1m]",
+          ],
+        }),
+      );
+
+      const projectDir = path.join(tempDir, "project");
+      await fs.promises.mkdir(projectDir, { recursive: true });
+
+      mockQueryWithModels([
+        { value: "default", displayName: "Default", description: "Default model" },
+        {
+          value: "opus",
+          displayName: "Opus 4.7",
+          description: "Claude Opus 4.7 — complex tasks, higher cost",
+        },
+        {
+          value: "opus[1m]",
+          displayName: "Opus 4.7 (1M context)",
+          description: "Opus 4.7 with 1M context",
+        },
+      ]);
+
+      const { ClaudeAcpAgent } = await import("../acp-agent.js");
+      const agent: ClaudeAcpAgentType = new ClaudeAcpAgent(createMockClient());
+
+      const response = await (agent as any).createSession({
+        cwd: projectDir,
+        mcpServers: [],
+        _meta: { disableBuiltInTools: true },
+      });
+
+      const modelOption = response.configOptions.find((o: any) => o.id === "model");
+      const byValue: Record<string, { name: string; description?: string }> = {};
+      for (const opt of modelOption.options) {
+        byValue[opt.value] = { name: opt.name, description: opt.description };
+      }
+
+      // 4-6 entries must NOT inherit the 4.7 SDK alias display info.
+      expect(byValue["claude-opus-4-6"].name).toBe("claude-opus-4-6");
+      expect(byValue["claude-opus-4-6"].description).toBe("");
+      expect(byValue["claude-opus-4-6[1m]"].name).toBe("claude-opus-4-6[1m]");
+      expect(byValue["claude-opus-4-6[1m]"].description).toBe("");
+
+      // 4-7 entries continue to inherit display info from a 4.7 SDK alias.
+      expect(byValue["claude-opus-4-7"].name).toBe("Opus 4.7");
+      expect(byValue["claude-opus-4-7[1m]"].name).toMatch(/Opus 4\.7/);
+    });
+
+    it("does not inherit display info across mismatched model families", async () => {
+      // https://github.com/agentclientprotocol/claude-agent-acp/issues/639:
+      // when the SDK's `opus` alias resolves to Opus 4.8, `claude-opus-4-6[1m]`
+      // fails the version check against `opus` but the tokenized matcher
+      // falls through to `sonnet[1m]` because the `1m` context hint alone
+      // is enough to score a match.
+      await fs.promises.writeFile(
+        path.join(tempDir, "settings.json"),
+        JSON.stringify({
+          availableModels: ["claude-opus-4-6", "claude-opus-4-6[1m]"],
+        }),
+      );
+
+      const projectDir = path.join(tempDir, "project");
+      await fs.promises.mkdir(projectDir, { recursive: true });
+
+      mockQueryWithModels([
+        { value: "default", displayName: "Default", description: "Default model" },
+        {
+          value: "sonnet",
+          displayName: "Sonnet",
+          description: "Sonnet 4.6 · Best for everyday tasks",
+        },
+        {
+          value: "sonnet[1m]",
+          displayName: "Sonnet (1M context)",
+          description: "Sonnet 4.6 for long sessions",
+        },
+        {
+          value: "opus",
+          displayName: "Opus 4.8",
+          description: "Claude Opus 4.8",
+        },
+      ]);
+
+      const { ClaudeAcpAgent } = await import("../acp-agent.js");
+      const agent: ClaudeAcpAgentType = new ClaudeAcpAgent(createMockClient());
+
+      const response = await (agent as any).createSession({
+        cwd: projectDir,
+        mcpServers: [],
+        _meta: { disableBuiltInTools: true },
+      });
+
+      const modelOption = response.configOptions.find((o: any) => o.id === "model");
+      const byValue: Record<string, { name: string; description?: string }> = {};
+      for (const opt of modelOption.options) {
+        byValue[opt.value] = { name: opt.name, description: opt.description };
+      }
+
+      expect(byValue["claude-opus-4-6[1m]"].name).toBe("claude-opus-4-6[1m]");
+      expect(byValue["claude-opus-4-6"].name).toBe("claude-opus-4-6");
+    });
+
+    it("preserves ANTHROPIC_CUSTOM_MODEL_OPTION even when absent from the allowlist", async () => {
+      // Per the model-config docs, ANTHROPIC_CUSTOM_MODEL_OPTION adds an entry
+      // "without replacing the built-in aliases" and "appears at the bottom of
+      // the /model picker", so it is exempt from the availableModels allowlist
+      // (the same way the Default option is "not affected by availableModels").
+      // ACP must match: a slim alias allowlist must not hide the custom model
+      // row, and it appears last, after the allowlisted entries.
+      // https://code.claude.com/docs/en/model-config#add-a-custom-model-option
+      const originalEnv = process.env.ANTHROPIC_CUSTOM_MODEL_OPTION;
+      process.env.ANTHROPIC_CUSTOM_MODEL_OPTION = "claude-opus-4-8[1m]";
+      try {
+        await fs.promises.writeFile(
+          path.join(tempDir, "settings.json"),
+          JSON.stringify({
+            availableModels: ["sonnet", "opus", "haiku"],
+          }),
+        );
+
+        const projectDir = path.join(tempDir, "project");
+        await fs.promises.mkdir(projectDir, { recursive: true });
+
+        mockQueryWithModels([
+          { value: "default", displayName: "Default", description: "Default model" },
+          { value: "sonnet", displayName: "Sonnet", description: "Claude Sonnet 4.6" },
+          { value: "opus", displayName: "Opus", description: "Claude Opus 4.6" },
+          { value: "haiku", displayName: "Haiku", description: "Claude Haiku 4.5" },
+          {
+            value: "claude-opus-4-8[1m]",
+            displayName: "Opus 4.8",
+            description: "Claude Opus 4.8",
+          },
+        ]);
+
+        const { ClaudeAcpAgent } = await import("../acp-agent.js");
+        const agent: ClaudeAcpAgentType = new ClaudeAcpAgent(createMockClient());
+
+        const response = await (agent as any).createSession({
+          cwd: projectDir,
+          mcpServers: [],
+          _meta: { disableBuiltInTools: true },
+        });
+
+        const modelOption = response.configOptions.find((o: any) => o.id === "model");
+        expect(modelOption.options.map((o: any) => o.value)).toEqual([
+          "default",
+          "sonnet",
+          "opus",
+          "haiku",
+          "claude-opus-4-8[1m]",
+        ]);
+        const custom = modelOption.options.find((o: any) => o.value === "claude-opus-4-8[1m]");
+        expect(custom.name).toBe("Opus 4.8");
+        expect(custom.description).toBe("Claude Opus 4.8");
+      } finally {
+        if (originalEnv === undefined) {
+          delete process.env.ANTHROPIC_CUSTOM_MODEL_OPTION;
+        } else {
+          process.env.ANTHROPIC_CUSTOM_MODEL_OPTION = originalEnv;
+        }
+      }
+    });
+
+    it("does not duplicate the custom model option when also in the allowlist", async () => {
+      // If the user lists the custom model's exact ID in availableModels AND it
+      // is set as ANTHROPIC_CUSTOM_MODEL_OPTION, it must appear exactly once.
+      const originalEnv = process.env.ANTHROPIC_CUSTOM_MODEL_OPTION;
+      process.env.ANTHROPIC_CUSTOM_MODEL_OPTION = "claude-opus-4-8[1m]";
+      try {
+        await fs.promises.writeFile(
+          path.join(tempDir, "settings.json"),
+          JSON.stringify({
+            availableModels: ["sonnet", "claude-opus-4-8[1m]"],
+          }),
+        );
+
+        const projectDir = path.join(tempDir, "project");
+        await fs.promises.mkdir(projectDir, { recursive: true });
+
+        mockQueryWithModels([
+          { value: "default", displayName: "Default", description: "Default model" },
+          { value: "sonnet", displayName: "Sonnet", description: "Claude Sonnet 4.6" },
+          {
+            value: "claude-opus-4-8[1m]",
+            displayName: "Opus 4.8",
+            description: "Claude Opus 4.8",
+          },
+        ]);
+
+        const { ClaudeAcpAgent } = await import("../acp-agent.js");
+        const agent: ClaudeAcpAgentType = new ClaudeAcpAgent(createMockClient());
+
+        const response = await (agent as any).createSession({
+          cwd: projectDir,
+          mcpServers: [],
+          _meta: { disableBuiltInTools: true },
+        });
+
+        const modelOption = response.configOptions.find((o: any) => o.id === "model");
+        expect(modelOption.options.map((o: any) => o.value)).toEqual([
+          "default",
+          "sonnet",
+          "claude-opus-4-8[1m]",
+        ]);
+      } finally {
+        if (originalEnv === undefined) {
+          delete process.env.ANTHROPIC_CUSTOM_MODEL_OPTION;
+        } else {
+          process.env.ANTHROPIC_CUSTOM_MODEL_OPTION = originalEnv;
+        }
+      }
+    });
+
+    it("maps allowlist aliases through modelOverrides to provider IDs", async () => {
+      // Bedrock/Vertex deployments enforce short aliases in availableModels
+      // but require provider-specific IDs at the API. Without applying
+      // modelOverrides during allowlist resolution, the alias is surfaced as
+      // the model value and passed verbatim to setModel, which the provider
+      // API rejects with a 400 "invalid model identifier".
+      await fs.promises.writeFile(
+        path.join(tempDir, "settings.json"),
+        JSON.stringify({
+          availableModels: ["claude-opus-4-6", "claude-sonnet-4-6"],
+          model: "claude-opus-4-6",
+          modelOverrides: {
+            "claude-opus-4-6": "global.anthropic.claude-opus-4-6-v1[1m]",
+            "claude-sonnet-4-6": "global.anthropic.claude-sonnet-4-6[1m]",
+          },
+        }),
+      );
+
+      const projectDir = path.join(tempDir, "project");
+      await fs.promises.mkdir(projectDir, { recursive: true });
+
+      const { setModelSpy } = mockQueryWithModels([
+        { value: "default", displayName: "Default", description: "Default model" },
+        { value: "opus", displayName: "Opus", description: "Claude Opus 4.6" },
+        { value: "sonnet", displayName: "Sonnet", description: "Claude Sonnet 4.6" },
+      ]);
+
+      const { ClaudeAcpAgent } = await import("../acp-agent.js");
+      const agent: ClaudeAcpAgentType = new ClaudeAcpAgent(createMockClient());
+
+      const response = await (agent as any).createSession({
+        cwd: projectDir,
+        mcpServers: [],
+        _meta: { disableBuiltInTools: true },
+      });
+
+      const modelOption = response.configOptions.find((o: any) => o.id === "model");
+      expect(modelOption.options.map((o: any) => o.value)).toEqual([
+        "default",
+        "global.anthropic.claude-opus-4-6-v1[1m]",
+        "global.anthropic.claude-sonnet-4-6[1m]",
+      ]);
+      // The allowlist synthesized the overridden ID (the SDK only surfaced the
+      // `opus` alias), so the adapter syncs the provider ID via setModel — never
+      // the raw alias that Bedrock would reject.
+      expect(setModelSpy).toHaveBeenCalledWith("global.anthropic.claude-opus-4-6-v1[1m]");
+      expect(modelOption.currentValue).toBe("global.anthropic.claude-opus-4-6-v1[1m]");
+    });
+  });
+
   it("resolves model aliases like opus[1m] to the correct model", async () => {
     await fs.promises.writeFile(
       path.join(tempDir, "settings.json"),
@@ -375,7 +873,7 @@ describe("ClaudeAcpAgent settings", () => {
 
     const setModelSpy = vi.fn();
     querySpy.mockImplementation(({ options: _options }: any) => {
-      return {
+      return makeMockQuery({
         initializationResult: async () => ({
           models: [
             {
@@ -391,8 +889,7 @@ describe("ClaudeAcpAgent settings", () => {
           ],
         }),
         setModel: setModelSpy,
-        supportedCommands: async () => [],
-      } as any;
+      }) as any;
     });
 
     const { ClaudeAcpAgent } = await import("../acp-agent.js");
@@ -405,6 +902,100 @@ describe("ClaudeAcpAgent settings", () => {
     });
 
     expect(setModelSpy).toHaveBeenCalledWith("claude-opus-4-6-1m");
-    expect(response.models.currentModelId).toBe("claude-opus-4-6-1m");
+    expect(response.configOptions?.find((o: any) => o.id === "model")?.currentValue).toBe(
+      "claude-opus-4-6-1m",
+    );
+  });
+
+  it("skips the initial setModel when the resolved value matches the SDK's model list verbatim", async () => {
+    // Covers the launcher case from PR #646: the launcher bakes the model into
+    // ANTHROPIC_MODEL, the SDK already starts on that model, and a second
+    // setModel call would be a redundant round-trip (and on some launcher
+    // setups, more fragile than launch-time selection).
+    const originalEnv = process.env.ANTHROPIC_MODEL;
+    process.env.ANTHROPIC_MODEL = "claude-opus-4-6";
+
+    const projectDir = path.join(tempDir, "project");
+    await fs.promises.mkdir(projectDir, { recursive: true });
+
+    const setModelSpy = vi.fn();
+    querySpy.mockImplementation(() => {
+      return makeMockQuery({
+        initializationResult: async () => ({
+          models: [
+            { value: "default", displayName: "Default", description: "" },
+            { value: "claude-sonnet-4-6", displayName: "Claude Sonnet 4.6", description: "" },
+            { value: "claude-opus-4-6", displayName: "Claude Opus 4.6", description: "" },
+          ],
+        }),
+        setModel: setModelSpy,
+      }) as any;
+    });
+
+    try {
+      const { ClaudeAcpAgent } = await import("../acp-agent.js");
+      const agent: ClaudeAcpAgentType = new ClaudeAcpAgent(createMockClient());
+
+      const response = await (agent as any).createSession({
+        cwd: projectDir,
+        mcpServers: [],
+        _meta: { disableBuiltInTools: true },
+      });
+
+      expect(setModelSpy).not.toHaveBeenCalled();
+      expect(response.configOptions?.find((o: any) => o.id === "model")?.currentValue).toBe(
+        "claude-opus-4-6",
+      );
+    } finally {
+      if (originalEnv === undefined) {
+        delete process.env.ANTHROPIC_MODEL;
+      } else {
+        process.env.ANTHROPIC_MODEL = originalEnv;
+      }
+    }
+  });
+
+  it("still calls setModel when the allowlist synthesizes a value the SDK has not surfaced", async () => {
+    // The allowlist may rewrite a model's `value` to the user's literal ID
+    // (e.g., `claude-haiku-4-5`) even when the SDK only exposed an alias
+    // (`haiku`). In that case the SDK has not independently arrived at the
+    // user's preferred ID, so we must sync via setModel.
+    await fs.promises.writeFile(
+      path.join(tempDir, "settings.json"),
+      JSON.stringify({
+        availableModels: ["claude-haiku-4-5"],
+        model: "claude-haiku-4-5",
+      }),
+    );
+
+    const projectDir = path.join(tempDir, "project");
+    await fs.promises.mkdir(projectDir, { recursive: true });
+
+    const setModelSpy = vi.fn();
+    querySpy.mockImplementation(() => {
+      return makeMockQuery({
+        initializationResult: async () => ({
+          models: [
+            { value: "default", displayName: "Default", description: "" },
+            { value: "haiku", displayName: "Haiku", description: "Fast" },
+          ],
+        }),
+        setModel: setModelSpy,
+      }) as any;
+    });
+
+    const { ClaudeAcpAgent } = await import("../acp-agent.js");
+    const agent: ClaudeAcpAgentType = new ClaudeAcpAgent(createMockClient());
+
+    const response = await (agent as any).createSession({
+      cwd: projectDir,
+      mcpServers: [],
+      _meta: { disableBuiltInTools: true },
+    });
+
+    expect(setModelSpy).toHaveBeenCalledWith("claude-haiku-4-5");
+    expect(response.configOptions?.find((o: any) => o.id === "model")?.currentValue).toBe(
+      "claude-haiku-4-5",
+    );
   });
 });

src/tests/additional-roots.test.ts

@@ -1,28 +1,36 @@
 import { mkdtemp, rm } from "node:fs/promises";
 import * as os from "node:os";
 import * as path from "node:path";
-import { AgentSideConnection, SessionNotification } from "@agentclientprotocol/sdk";
+import { SessionNotification } from "@agentclientprotocol/sdk";
 import type { Options } from "@anthropic-ai/claude-agent-sdk";
-import type { ClaudeAcpAgent as ClaudeAcpAgentType } from "../acp-agent.js";
+import type { AcpClient, ClaudeAcpAgent as ClaudeAcpAgentType } from "../acp-agent.js";
 import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
 
 let capturedOptions: Options | undefined;
-vi.mock("@anthropic-ai/claude-agent-sdk", async () => ({
-  ...(await vi.importActual<typeof import("@anthropic-ai/claude-agent-sdk")>(
+vi.mock("@anthropic-ai/claude-agent-sdk", async () => {
+  const actual = await vi.importActual<typeof import("@anthropic-ai/claude-agent-sdk")>(
     "@anthropic-ai/claude-agent-sdk",
-  )),
-  query: ({ options }: { options: Options }) => {
-    capturedOptions = options;
-    return {
-      initializationResult: async () => ({
-        models: [{ value: "claude-sonnet-4-5", displayName: "Claude Sonnet", description: "Fast" }],
-      }),
-      setModel: async () => {},
-      supportedCommands: async () => [],
-      [Symbol.asyncIterator]: async function* () {},
-    };
-  },
-}));
+  );
+  const { makeMockQuery } = await import("./helpers.js");
+  return {
+    ...actual,
+    query: ({ options }: { options: Options }) => {
+      capturedOptions = options;
+      return makeMockQuery({
+        initializationResult: async () => ({
+          models: [
+            {
+              value: "claude-sonnet-4-6",
+              displayName: "Claude Sonnet",
+              description: "Fast",
+              supportsAutoMode: true,
+            },
+          ],
+        }),
+      });
+    },
+  };
+});
 vi.mock("../tools.js", async () => ({
   ...(await vi.importActual<typeof import("../tools.js")>("../tools.js")),
   registerHookCallback: vi.fn(),
@@ -31,7 +39,7 @@ vi.mock("../tools.js", async () => ({
 describe("additionalRoots", () => {
   let agent: ClaudeAcpAgentType;
   const tempDirs: string[] = [];
-  const newSession = (meta: Record<string, unknown>, cwd = "/test") =>
+  const newSession = (meta: Record<string, unknown>, cwd = process.cwd()) =>
     agent.newSession({ cwd, mcpServers: [], _meta: meta });
 
   beforeEach(async () => {
@@ -44,7 +52,7 @@ describe("additionalRoots", () => {
       requestPermission: async () => ({ outcome: { outcome: "cancelled" } }),
       readTextFile: async () => ({ content: "" }),
       writeTextFile: async () => ({}),
-    } as unknown as AgentSideConnection);
+    } as unknown as AcpClient);
   });
 
   afterEach(
@@ -68,4 +76,33 @@ describe("additionalRoots", () => {
     });
     expect(capturedOptions!.additionalDirectories).toEqual(["/workspace/shared", "", root]);
   });
+
+  it("prefers the official ACP additionalDirectories field over _meta.additionalRoots", async () => {
+    await agent.newSession({
+      cwd: process.cwd(),
+      mcpServers: [],
+      additionalDirectories: ["/from/official"],
+      _meta: { additionalRoots: ["/from/meta"] },
+    });
+    expect(capturedOptions!.additionalDirectories).toEqual(["/from/official"]);
+  });
+
+  it("merges official ACP additionalDirectories with claudeCode SDK additionalDirectories", async () => {
+    await agent.newSession({
+      cwd: process.cwd(),
+      mcpServers: [],
+      additionalDirectories: ["/from/official"],
+      _meta: { claudeCode: { options: { additionalDirectories: ["/from/sdk"] } } },
+    });
+    expect(capturedOptions!.additionalDirectories).toEqual(["/from/sdk", "/from/official"]);
+  });
+
+  it("falls back to _meta.additionalRoots when the official field is omitted", async () => {
+    await agent.newSession({
+      cwd: process.cwd(),
+      mcpServers: [],
+      _meta: { additionalRoots: ["/from/meta"] },
+    });
+    expect(capturedOptions!.additionalDirectories).toEqual(["/from/meta"]);
+  });
 });

src/tests/agent-discovery.test.ts

@@ -0,0 +1,42 @@
+import { describe, it, expect } from "vitest";
+import { randomUUID } from "crypto";
+import { query } from "@anthropic-ai/claude-agent-sdk";
+import type { SDKUserMessage } from "@anthropic-ai/claude-agent-sdk";
+import { Pushable } from "../utils.js";
+import { discoverCustomAgents, BUILTIN_AGENT_NAMES } from "../acp-agent.js";
+
+// `discoverCustomAgents` distinguishes user/plugin-configured agents from
+// Claude Code's built-in subagents by name (BUILTIN_AGENT_NAMES), since the
+// SDK's `AgentInfo` carries no built-in/source flag. That hardcoded set drifts
+// silently if the SDK changes its default roster: a newly added built-in would
+// leak into the ACP "Agent" picker as if the user had configured it. This
+// integration test runs the real SDK so the set is flagged the moment it does.
+describe.skipIf(!process.env.RUN_INTEGRATION_TESTS)("agent discovery (SDK)", () => {
+  it("BUILTIN_AGENT_NAMES exactly covers the SDK's default agent roster", async () => {
+    const input = new Pushable<SDKUserMessage>();
+    const q = query({
+      prompt: input,
+      options: {
+        systemPrompt: { type: "preset", preset: "claude_code" },
+        sessionId: randomUUID(),
+        // No setting sources → no user/project/local custom agents are loaded,
+        // so `supportedAgents()` reports exactly the intrinsic built-ins
+        // regardless of what the developer has in ~/.claude/agents.
+        settingSources: [],
+        includePartialMessages: true,
+      },
+    });
+
+    try {
+      const reported = (await q.supportedAgents()).map((a) => a.name).sort();
+      expect(reported).toEqual([...BUILTIN_AGENT_NAMES].sort());
+
+      // With the built-ins being the only agents present, discovery surfaces
+      // nothing — the ACP "Agent" picker is correctly omitted.
+      expect(await discoverCustomAgents(q)).toEqual([]);
+    } finally {
+      input.end();
+      q.return(undefined);
+    }
+  }, 30000);
+});

src/tests/authorization.test.ts

@@ -1,24 +1,35 @@
 import { describe, expect, it, Mock, vi, afterEach, beforeEach } from "vitest";
-import { ClaudeAcpAgent } from "../acp-agent.js";
-import { AgentSideConnection } from "@agentclientprotocol/sdk";
+import { AcpClient, ClaudeAcpAgent } from "../acp-agent.js";
+import { makeMockQuery } from "./helpers.js";
 
-const mockQuery = vi.hoisted(() =>
-  vi.fn(() => ({
-    initializationResult: vi.fn().mockResolvedValue({
-      models: [{ value: "id", displayName: "name", description: "description" }],
-    }),
-    setModel: vi.fn(),
-    supportedCommands: vi.fn().mockResolvedValue([]),
-  })),
-);
+const mockQuery = vi.hoisted(() => vi.fn());
 
-vi.mock("@anthropic-ai/claude-agent-sdk", () => ({
+vi.mock("@anthropic-ai/claude-agent-sdk", async () => ({
+  ...(await vi.importActual<typeof import("@anthropic-ai/claude-agent-sdk")>(
+    "@anthropic-ai/claude-agent-sdk",
+  )),
   query: mockQuery,
 }));
 
 describe("authorization", () => {
   beforeEach(() => {
     vi.useFakeTimers();
+    // Set here (not in vi.fn(impl) at hoist time) so the helper import is
+    // available; afterEach's resetAllMocks clears it, beforeEach re-sets it.
+    mockQuery.mockImplementation(() =>
+      makeMockQuery({
+        initializationResult: async () => ({
+          models: [
+            {
+              value: "id",
+              displayName: "name",
+              description: "description",
+              supportsAutoMode: true,
+            },
+          ],
+        }),
+      }),
+    );
   });
 
   afterEach(() => {
@@ -33,7 +44,7 @@ describe("authorization", () => {
   async function createAgentMock(): Promise<[ClaudeAcpAgent, Mock]> {
     const connectionMock = {
       sessionUpdate: async (_: any) => {},
-    } as AgentSideConnection;
+    } as AcpClient;
 
     const agent = new ClaudeAcpAgent(connectionMock);
 
@@ -50,6 +61,9 @@ describe("authorization", () => {
     expect(initializeResponse.authMethods).not.toContainEqual(
       expect.objectContaining({ id: "gateway" }),
     );
+    expect(initializeResponse.authMethods).not.toContainEqual(
+      expect.objectContaining({ id: "gateway-bedrock" }),
+    );
   });
 
   it("gateway auth offered when client advertises auth._meta.gateway capability", async () => {
@@ -62,11 +76,20 @@ describe("authorization", () => {
       } as any,
     });
     expect(initializeResponse.authMethods).toContainEqual(
-      expect.objectContaining({ id: "gateway" }),
+      expect.objectContaining({
+        id: "gateway",
+        _meta: { gateway: { protocol: "anthropic" } },
+      }),
+    );
+    expect(initializeResponse.authMethods).toContainEqual(
+      expect.objectContaining({
+        id: "gateway-bedrock",
+        _meta: { gateway: { protocol: "bedrock" } },
+      }),
     );
   });
 
-  it("uses gateway env after gateway auth", async () => {
+  it("uses gateway env after anthropic gateway auth", async () => {
     const [agent, mockQuery] = await createAgentMock();
 
     const initializeResponse = await agent.initialize({
@@ -85,7 +108,7 @@ describe("authorization", () => {
     });
 
     await agent.newSession({
-      cwd: "testRoot",
+      cwd: process.cwd(),
       mcpServers: [],
       _meta: {
         claudeCode: {
@@ -102,7 +125,7 @@ describe("authorization", () => {
       expect.objectContaining({
         options: expect.objectContaining({
           env: expect.objectContaining({
-            ANTHROPIC_AUTH_TOKEN: "",
+            ANTHROPIC_AUTH_TOKEN: " ",
             ANTHROPIC_BASE_URL: "https://gateway.example",
             ANTHROPIC_CUSTOM_HEADERS: "x-api-key: test",
             userEnv: "userEnv",
@@ -112,6 +135,42 @@ describe("authorization", () => {
     );
   });
 
+  it("uses gateway env after gateway auth", async () => {
+    const [agent, mockQuery] = await createAgentMock();
+
+    await agent.initialize({
+      protocolVersion: 1,
+      clientCapabilities: {
+        auth: { terminal: true, _meta: { gateway: true } },
+      } as any,
+    });
+
+    await agent.authenticate({
+      methodId: "gateway-bedrock",
+      _meta: {
+        gateway: { baseUrl: "https://gateway.example", headers: { "custom-header": "test" } },
+      },
+    });
+
+    await agent.newSession({
+      cwd: process.cwd(),
+      mcpServers: [],
+    });
+
+    expect(mockQuery).toHaveBeenCalledWith(
+      expect.objectContaining({
+        options: expect.objectContaining({
+          env: expect.objectContaining({
+            CLAUDE_CODE_USE_BEDROCK: "1",
+            AWS_BEARER_TOKEN_BEDROCK: " ",
+            ANTHROPIC_BEDROCK_BASE_URL: "https://gateway.example",
+            ANTHROPIC_CUSTOM_HEADERS: "custom-header: test",
+          }),
+        }),
+      }),
+    );
+  });
+
   it("hide claude authentication without terminal-auth", async () => {
     const [agent] = await createAgentMock();
     vi.stubGlobal("process", { ...process, argv: ["--hide-claude-auth"] });

src/tests/create-session-options.test.ts

@@ -1,27 +1,36 @@
 import { describe, it, expect, beforeEach, afterEach, vi } from "vitest";
-import { AgentSideConnection, SessionNotification } from "@agentclientprotocol/sdk";
+import { RequestError, SessionNotification } from "@agentclientprotocol/sdk";
 import type { Options } from "@anthropic-ai/claude-agent-sdk";
-import type { ClaudeAcpAgent as ClaudeAcpAgentType } from "../acp-agent.js";
+import type { AcpClient, ClaudeAcpAgent as ClaudeAcpAgentType } from "../acp-agent.js";
+import * as fs from "node:fs";
+import * as os from "node:os";
+import * as path from "node:path";
 
 let capturedOptions: Options | undefined;
+let contextUsageResult: (() => Promise<{ rawMaxTokens: number; model?: string }>) | undefined;
 vi.mock("@anthropic-ai/claude-agent-sdk", async () => {
   const actual = await vi.importActual<typeof import("@anthropic-ai/claude-agent-sdk")>(
     "@anthropic-ai/claude-agent-sdk",
   );
+  const { makeMockQuery, DEFAULT_CONTEXT_USAGE } = await import("./helpers.js");
   return {
     ...actual,
     query: (args: { prompt: unknown; options: Options }) => {
       capturedOptions = args.options;
-      return {
+      return makeMockQuery({
         initializationResult: async () => ({
           models: [
-            { value: "claude-sonnet-4-5", displayName: "Claude Sonnet", description: "Fast" },
+            {
+              value: "claude-sonnet-4-6",
+              displayName: "Claude Sonnet",
+              description: "Fast",
+              supportsAutoMode: true,
+            },
           ],
         }),
-        setModel: async () => {},
-        supportedCommands: async () => [],
-        [Symbol.asyncIterator]: async function* () {},
-      };
+        getContextUsage: () =>
+          contextUsageResult ? contextUsageResult() : Promise.resolve(DEFAULT_CONTEXT_USAGE),
+      });
     },
   };
 });
@@ -38,17 +47,18 @@ describe("createSession options merging", () => {
   let agent: ClaudeAcpAgentType;
   let ClaudeAcpAgent: typeof ClaudeAcpAgentType;
 
-  function createMockClient(): AgentSideConnection {
+  function createMockClient(): AcpClient {
     return {
       sessionUpdate: async (_notification: SessionNotification) => {},
       requestPermission: async () => ({ outcome: { outcome: "cancelled" } }),
       readTextFile: async () => ({ content: "" }),
       writeTextFile: async () => ({}),
-    } as unknown as AgentSideConnection;
+    } as unknown as AcpClient;
   }
 
   beforeEach(async () => {
     capturedOptions = undefined;
+    contextUsageResult = undefined;
 
     vi.resetModules();
     const acpAgent = await import("../acp-agent.js");
@@ -59,7 +69,7 @@ describe("createSession options merging", () => {
 
   it("merges user-provided disallowedTools with ACP internal list", async () => {
     await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
       mcpServers: [],
       _meta: {
         claudeCode: {
@@ -79,7 +89,7 @@ describe("createSession options merging", () => {
 
   it("works when user provides no disallowedTools", async () => {
     await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
       mcpServers: [],
     });
 
@@ -88,7 +98,7 @@ describe("createSession options merging", () => {
 
   it("works when user provides empty disallowedTools", async () => {
     await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
       mcpServers: [],
       _meta: {
         claudeCode: {
@@ -104,7 +114,7 @@ describe("createSession options merging", () => {
 
   it("sets tools to empty array when disableBuiltInTools is true", async () => {
     await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
       mcpServers: [],
       _meta: {
         disableBuiltInTools: true,
@@ -127,7 +137,7 @@ describe("createSession options merging", () => {
     const userPreToolUseHook = { hooks: [{ command: "echo pre" }] };
 
     await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
       mcpServers: [],
       _meta: {
         claudeCode: {
@@ -149,7 +159,7 @@ describe("createSession options merging", () => {
 
   it("inherits HOME and PATH from process.env when no env is provided", async () => {
     await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
       mcpServers: [],
     });
 
@@ -159,7 +169,7 @@ describe("createSession options merging", () => {
 
   it("merges user-provided env vars on top of process.env", async () => {
     await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
       mcpServers: [],
       _meta: {
         claudeCode: {
@@ -179,7 +189,7 @@ describe("createSession options merging", () => {
 
   it("allows user-provided env vars to override process.env entries", async () => {
     await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
       mcpServers: [],
       _meta: {
         claudeCode: {
@@ -197,7 +207,7 @@ describe("createSession options merging", () => {
 
   it("defaults tools to claude_code preset when not provided", async () => {
     await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
       mcpServers: [],
     });
 
@@ -206,7 +216,7 @@ describe("createSession options merging", () => {
 
   it("passes through user-provided tools string array", async () => {
     await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
       mcpServers: [],
       _meta: {
         claudeCode: {
@@ -222,7 +232,7 @@ describe("createSession options merging", () => {
 
   it("explicit tools array takes precedence over disableBuiltInTools", async () => {
     await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
       mcpServers: [],
       _meta: {
         disableBuiltInTools: true,
@@ -239,7 +249,7 @@ describe("createSession options merging", () => {
 
   it("passes through empty tools array to disable all built-in tools", async () => {
     await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
       mcpServers: [],
       _meta: {
         claudeCode: {
@@ -255,7 +265,7 @@ describe("createSession options merging", () => {
 
   describe("systemPrompt via _meta", () => {
     it("defaults to the claude_code preset when not provided", async () => {
-      await agent.newSession({ cwd: "/test", mcpServers: [] });
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
 
       expect(capturedOptions!.systemPrompt).toEqual({
         type: "preset",
@@ -265,7 +275,7 @@ describe("createSession options merging", () => {
 
     it("replaces the preset when a string is provided", async () => {
       await agent.newSession({
-        cwd: "/test",
+        cwd: process.cwd(),
         mcpServers: [],
         _meta: { systemPrompt: "custom prompt" },
       });
@@ -275,7 +285,7 @@ describe("createSession options merging", () => {
 
     it("forwards append", async () => {
       await agent.newSession({
-        cwd: "/test",
+        cwd: process.cwd(),
         mcpServers: [],
         _meta: { systemPrompt: { append: "extra instructions" } },
       });
@@ -289,7 +299,7 @@ describe("createSession options merging", () => {
 
     it("forwards excludeDynamicSections", async () => {
       await agent.newSession({
-        cwd: "/test",
+        cwd: process.cwd(),
         mcpServers: [],
         _meta: { systemPrompt: { excludeDynamicSections: true } },
       });
@@ -303,7 +313,7 @@ describe("createSession options merging", () => {
 
     it("forwards append and excludeDynamicSections together", async () => {
       await agent.newSession({
-        cwd: "/test",
+        cwd: process.cwd(),
         mcpServers: [],
         _meta: {
           systemPrompt: {
@@ -323,7 +333,7 @@ describe("createSession options merging", () => {
 
     it("ignores caller-provided type/preset overrides", async () => {
       await agent.newSession({
-        cwd: "/test",
+        cwd: process.cwd(),
         mcpServers: [],
         _meta: {
           systemPrompt: {
@@ -363,7 +373,7 @@ describe("createSession options merging", () => {
         modelOverrides: { "claude-opus-4-6": "us.anthropic.claude-opus-4-6-v1" },
       });
 
-      await agent.newSession({ cwd: "/test", mcpServers: [] });
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
 
       expect(capturedOptions!.settings).toEqual({
         modelOverrides: { "claude-opus-4-6": "us.anthropic.claude-opus-4-6-v1" },
@@ -375,7 +385,7 @@ describe("createSession options merging", () => {
         availableModels: ["opus", "sonnet"],
       });
 
-      await agent.newSession({ cwd: "/test", mcpServers: [] });
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
 
       expect(capturedOptions!.settings).toEqual({
         availableModels: ["opus", "sonnet"],
@@ -388,7 +398,7 @@ describe("createSession options merging", () => {
         availableModels: ["opus"],
       });
 
-      await agent.newSession({ cwd: "/test", mcpServers: [] });
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
 
       expect(capturedOptions!.settings).toEqual({
         modelOverrides: { "claude-opus-4-6": "us.anthropic.claude-opus-4-6-v1" },
@@ -397,7 +407,7 @@ describe("createSession options merging", () => {
     });
 
     it("does not add settings when env var is not set", async () => {
-      await agent.newSession({ cwd: "/test", mcpServers: [] });
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
 
       expect(capturedOptions!.settings).toBeUndefined();
     });
@@ -408,13 +418,13 @@ describe("createSession options merging", () => {
       });
 
       await agent.newSession({
-        cwd: "/test",
+        cwd: process.cwd(),
         mcpServers: [],
         _meta: {
           claudeCode: {
             options: {
               settings: {
-                model: "claude-sonnet-4-5",
+                model: "claude-sonnet-4-6",
                 modelOverrides: { "claude-opus-4-6": "meta-value" },
               },
             },
@@ -424,7 +434,7 @@ describe("createSession options merging", () => {
 
       // _meta settings take precedence; env var is ignored entirely
       expect(capturedOptions!.settings).toEqual({
-        model: "claude-sonnet-4-5",
+        model: "claude-sonnet-4-6",
         modelOverrides: { "claude-opus-4-6": "meta-value" },
       });
     });
@@ -432,13 +442,13 @@ describe("createSession options merging", () => {
     it("throws on invalid JSON", async () => {
       process.env.CLAUDE_MODEL_CONFIG = "not-json";
 
-      await expect(agent.newSession({ cwd: "/test", mcpServers: [] })).rejects.toThrow();
+      await expect(agent.newSession({ cwd: process.cwd(), mcpServers: [] })).rejects.toThrow();
     });
   });
 
   it("merges user-provided mcpServers with ACP mcpServers", async () => {
     await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
       mcpServers: [
         {
           name: "acp-server",
@@ -467,4 +477,382 @@ describe("createSession options merging", () => {
     // ACP-provided MCP server should also be present
     expect(capturedOptions!.mcpServers).toHaveProperty("acp-server");
   });
+
+  describe("thinking config from MAX_THINKING_TOKENS", () => {
+    let originalMaxThinking: string | undefined;
+
+    beforeEach(() => {
+      originalMaxThinking = process.env.MAX_THINKING_TOKENS;
+      delete process.env.MAX_THINKING_TOKENS;
+    });
+
+    afterEach(() => {
+      if (originalMaxThinking !== undefined) {
+        process.env.MAX_THINKING_TOKENS = originalMaxThinking;
+      } else {
+        delete process.env.MAX_THINKING_TOKENS;
+      }
+    });
+
+    it("leaves thinking unset (SDK default) when env var is absent", async () => {
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+      expect(capturedOptions!.thinking).toBeUndefined();
+      // The deprecated option must not be set either.
+      expect(capturedOptions!.maxThinkingTokens).toBeUndefined();
+    });
+
+    it("maps 0 to disabled thinking", async () => {
+      process.env.MAX_THINKING_TOKENS = "0";
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+      expect(capturedOptions!.thinking).toEqual({ type: "disabled" });
+    });
+
+    it("maps a positive value to a fixed thinking budget", async () => {
+      process.env.MAX_THINKING_TOKENS = "12000";
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+      expect(capturedOptions!.thinking).toEqual({ type: "enabled", budgetTokens: 12000 });
+    });
+
+    it("ignores a non-numeric value", async () => {
+      process.env.MAX_THINKING_TOKENS = "lots";
+      // The bad value is deliberate — capture the agent's warning instead of
+      // letting it hit the console.
+      const errorSpy = vi.fn();
+      (agent as any).logger = { log: () => {}, error: errorSpy };
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+      expect(capturedOptions!.thinking).toBeUndefined();
+      expect(errorSpy).toHaveBeenCalledWith(expect.stringContaining("MAX_THINKING_TOKENS"));
+    });
+
+    it("lets a user-provided thinking option override the env default", async () => {
+      process.env.MAX_THINKING_TOKENS = "12000";
+      await agent.newSession({
+        cwd: process.cwd(),
+        mcpServers: [],
+        _meta: {
+          claudeCode: {
+            options: {
+              thinking: { type: "adaptive" },
+            },
+          },
+        },
+      });
+      expect(capturedOptions!.thinking).toEqual({ type: "adaptive" });
+    });
+  });
+
+  describe("cwd validation", () => {
+    it("rejects a relative cwd with invalidParams", async () => {
+      await expect(
+        agent.newSession({ cwd: "relative/path", mcpServers: [] }),
+      ).rejects.toMatchObject({ code: RequestError.invalidParams().code });
+    });
+
+    it("rejects a non-existent cwd with invalidParams", async () => {
+      const missing = path.join(os.tmpdir(), "claude-acp-does-not-exist-xyz");
+      await expect(agent.newSession({ cwd: missing, mcpServers: [] })).rejects.toMatchObject({
+        code: RequestError.invalidParams().code,
+      });
+    });
+
+    it("rejects a cwd that points at a file with invalidParams", async () => {
+      const dir = fs.mkdtempSync(path.join(os.tmpdir(), "claude-acp-cwd-is-a-file-"));
+      const file = path.join(dir, "not-a-directory.txt");
+      fs.writeFileSync(file, "not a directory");
+      try {
+        await expect(agent.newSession({ cwd: file, mcpServers: [] })).rejects.toMatchObject({
+          code: RequestError.invalidParams().code,
+        });
+      } finally {
+        fs.rmSync(dir, { recursive: true, force: true });
+      }
+    });
+
+    it("accepts an existing absolute directory", async () => {
+      await expect(agent.newSession({ cwd: process.cwd(), mcpServers: [] })).resolves.toBeDefined();
+    });
+  });
+
+  describe("elicitation", () => {
+    it("keeps AskUserQuestion disabled and omits callbacks without elicitation capability", async () => {
+      await agent.initialize({ protocolVersion: 1, clientCapabilities: {} });
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+
+      expect(capturedOptions!.disallowedTools).toContain("AskUserQuestion");
+      expect(capturedOptions!.onElicitation).toBeUndefined();
+    });
+
+    it("enables AskUserQuestion and wires the elicitation callback when form is supported", async () => {
+      await agent.initialize({
+        protocolVersion: 1,
+        clientCapabilities: { elicitation: { form: {} } },
+      });
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+
+      expect(capturedOptions!.disallowedTools).not.toContain("AskUserQuestion");
+      expect(typeof capturedOptions!.onElicitation).toBe("function");
+    });
+
+    it("wires callbacks for url-only elicitation but keeps AskUserQuestion disabled", async () => {
+      await agent.initialize({
+        protocolVersion: 1,
+        clientCapabilities: { elicitation: { url: {} } },
+      });
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+
+      expect(capturedOptions!.disallowedTools).toContain("AskUserQuestion");
+      expect(typeof capturedOptions!.onElicitation).toBe("function");
+    });
+
+    it("still merges user-provided disallowedTools when AskUserQuestion is enabled", async () => {
+      await agent.initialize({
+        protocolVersion: 1,
+        clientCapabilities: { elicitation: { form: {} } },
+      });
+      await agent.newSession({
+        cwd: process.cwd(),
+        mcpServers: [],
+        _meta: { claudeCode: { options: { disallowedTools: ["WebSearch"] } } },
+      });
+
+      expect(capturedOptions!.disallowedTools).toContain("WebSearch");
+      expect(capturedOptions!.disallowedTools).not.toContain("AskUserQuestion");
+    });
+  });
+
+  describe("context window seeding", () => {
+    function sessionFor(sessionId: string) {
+      return (agent as unknown as { sessions: Record<string, any> }).sessions[sessionId];
+    }
+
+    it("does not call getContextUsage during session creation", async () => {
+      // getContextUsage stalls until the session's first prompt turn has run
+      // (it is not serviced pre-turn), so session/new must never call it —
+      // awaiting it inline is what regressed session/new latency in 0.59.0.
+      const ctxSpy = vi.fn(async () => ({ rawMaxTokens: 967000 }));
+      contextUsageResult = ctxSpy;
+
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+
+      expect(ctxSpy).not.toHaveBeenCalled();
+    });
+
+    it("seeds contextWindowSize from text inference, falling back to the default when it misses", async () => {
+      // The mock model ("claude-sonnet-4-6" / "Claude Sonnet" / "Fast") carries
+      // no "1m" token anywhere, so inference misses and the window falls back to
+      // the default; the authoritative value arrives later via result.modelUsage.
+      contextUsageResult = async () => ({ rawMaxTokens: 967000 });
+
+      const response = await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+
+      expect(sessionFor(response.sessionId).contextWindowSize).toBe(200000);
+      expect(sessionFor(response.sessionId).contextWindowAuthoritative).toBe(false);
+    });
+
+    it("session/load seeds the window from the resumed session's getContextUsage report", async () => {
+      // Resumed sessions get getContextUsage serviced pre-turn (issue #845 uses
+      // it to restore the live model), and the same response carries the
+      // authoritative window (`rawMaxTokens`). After a process restart the
+      // module cache is empty and text inference misses natively-1M aliases, so
+      // discarding this in-hand value would replay the issue-#596 flicker on
+      // every reload — the flagship scenario. 888_000 can only come from the
+      // report: inference on the mock model yields null → 200_000 default.
+      contextUsageResult = async () => ({ rawMaxTokens: 888_000, model: "claude-sonnet-4-6" });
+
+      await (
+        agent as unknown as {
+          createSession: (params: object, opts: { resume?: string }) => Promise<unknown>;
+        }
+      ).createSession({ cwd: process.cwd(), mcpServers: [] }, { resume: "resumed-window-probe" });
+
+      const session = sessionFor("resumed-window-probe");
+      expect(session.contextWindowSize).toBe(888_000);
+      expect(session.contextWindowAuthoritative).toBe(true);
+    });
+
+    it("scopes providerCacheKey by per-session env routing", async () => {
+      // The context-window cache key must distinguish backends exactly as the
+      // CLI will see them: a session routed to a proxy via _meta env shares a
+      // model id spelling with default-routed sessions but not a context lane,
+      // so it must land in its own cache bucket (and two default-routed
+      // sessions must share one).
+      const r1 = await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+      const r2 = await agent.newSession({
+        cwd: process.cwd(),
+        mcpServers: [],
+        _meta: {
+          claudeCode: {
+            options: { env: { ANTHROPIC_BASE_URL: "https://window-probe-proxy.example" } },
+          },
+        },
+      });
+      const r3 = await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+
+      expect(sessionFor(r2.sessionId).providerCacheKey).not.toBe(
+        sessionFor(r1.sessionId).providerCacheKey,
+      );
+      expect(sessionFor(r3.sessionId).providerCacheKey).toBe(
+        sessionFor(r1.sessionId).providerCacheKey,
+      );
+    });
+
+    it("scopes providerCacheKey by provider headers: same endpoint, different headers → different buckets", async () => {
+      // Two providers/set configs sharing apiType+baseUrl but differing in
+      // headers (e.g. an `anthropic-beta: context-1m-…` routing header) can
+      // serve different context lanes for the same model id, so they must not
+      // share a window-cache bucket.
+      await agent.unstable_setProvider({
+        providerId: "main",
+        apiType: "anthropic",
+        baseUrl: "https://gw.example",
+        headers: {},
+      });
+      const plain = await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+
+      await agent.unstable_setProvider({
+        providerId: "main",
+        apiType: "anthropic",
+        baseUrl: "https://gw.example",
+        headers: { "anthropic-beta": "context-1m-2025-08-07" },
+      });
+      const beta = await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+
+      expect(sessionFor(beta.sessionId).providerCacheKey).not.toBe(
+        sessionFor(plain.sessionId).providerCacheKey,
+      );
+    });
+  });
+
+  describe("refusal fallback dialog", () => {
+    it("declares the dialog kind and callback when the client supports form elicitation", async () => {
+      await agent.initialize({
+        protocolVersion: 1,
+        clientCapabilities: { elicitation: { form: {} } },
+      });
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+
+      expect(capturedOptions!.supportedDialogKinds).toEqual(["refusal_fallback_prompt"]);
+      expect(typeof capturedOptions!.onUserDialog).toBe("function");
+    });
+
+    it("omits the dialog wiring without form elicitation support (url-only included)", async () => {
+      await agent.initialize({
+        protocolVersion: 1,
+        clientCapabilities: { elicitation: { url: {} } },
+      });
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+
+      // The CLI fails closed on the undeclared kind, so the flow degrades to
+      // the classic refusal error instead of parking a dialog we can't render.
+      expect(capturedOptions!.supportedDialogKinds).toBeUndefined();
+      expect(capturedOptions!.onUserDialog).toBeUndefined();
+    });
+
+    /** Wire the session up with form support and return the captured dialog
+     *  callback plus a mock the test can point the elicitation response at. */
+    async function setupDialog() {
+      await agent.initialize({
+        protocolVersion: 1,
+        clientCapabilities: { elicitation: { form: {} } },
+      });
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+      const createElicitation = vi.fn();
+      (
+        agent as unknown as { client: { unstable_createElicitation: unknown } }
+      ).client.unstable_createElicitation = createElicitation;
+      return { onUserDialog: capturedOptions!.onUserDialog!, createElicitation };
+    }
+
+    const signal = () => ({ signal: new AbortController().signal });
+
+    it("renders the prompt as a form elicitation and maps the retry choice", async () => {
+      const { onUserDialog, createElicitation } = await setupDialog();
+      createElicitation.mockResolvedValue({
+        action: "accept",
+        content: { choice: "retry_fallback" },
+      });
+
+      const result = await onUserDialog(
+        {
+          dialogKind: "refusal_fallback_prompt",
+          payload: {
+            originalModel: "claude-fable-5",
+            fallbackModel: "claude-opus-4-8",
+            apiRefusalCategory: "cyber",
+          },
+        },
+        signal(),
+      );
+
+      expect(result).toEqual({ behavior: "completed", result: "retry_fallback" });
+      const request = createElicitation.mock.calls[0][0];
+      expect(request.mode).toBe("form");
+      expect(request.message).toContain("claude-fable-5");
+      expect(request.message).toContain("claude-opus-4-8");
+    });
+
+    it("completes with cancelled when the user keeps the refusal", async () => {
+      const { onUserDialog, createElicitation } = await setupDialog();
+      createElicitation.mockResolvedValue({ action: "cancel" });
+
+      const result = await onUserDialog(
+        {
+          dialogKind: "refusal_fallback_prompt",
+          payload: { originalModel: "a", fallbackModel: "b" },
+        },
+        signal(),
+      );
+
+      expect(result).toEqual({ behavior: "completed", result: "cancelled" });
+    });
+
+    it("cancels unrecognized dialog kinds without presenting anything", async () => {
+      const { onUserDialog, createElicitation } = await setupDialog();
+
+      const result = await onUserDialog(
+        { dialogKind: "some_future_dialog", payload: {} },
+        signal(),
+      );
+
+      expect(result).toEqual({ behavior: "cancelled" });
+      expect(createElicitation).not.toHaveBeenCalled();
+    });
+
+    it("cancels on a malformed payload without presenting anything", async () => {
+      const { onUserDialog, createElicitation } = await setupDialog();
+      // The malformed payload is deliberate — capture the agent's warning
+      // instead of letting it hit the console.
+      const errorSpy = vi.fn();
+      (agent as any).logger = { log: () => {}, error: errorSpy };
+
+      const result = await onUserDialog(
+        { dialogKind: "refusal_fallback_prompt", payload: { fallbackModel: 42 } },
+        signal(),
+      );
+
+      expect(result).toEqual({ behavior: "cancelled" });
+      expect(createElicitation).not.toHaveBeenCalled();
+      expect(errorSpy).toHaveBeenCalledWith(expect.stringContaining("unexpected shape"));
+    });
+
+    it("cancels when the elicitation request fails", async () => {
+      const { onUserDialog, createElicitation } = await setupDialog();
+      createElicitation.mockRejectedValue(new Error("client exploded"));
+      // The client failure is deliberate — capture the agent's warning
+      // instead of letting it hit the console.
+      const errorSpy = vi.fn();
+      (agent as any).logger = { log: () => {}, error: errorSpy };
+
+      const result = await onUserDialog(
+        {
+          dialogKind: "refusal_fallback_prompt",
+          payload: { originalModel: "a", fallbackModel: "b" },
+        },
+        signal(),
+      );
+
+      expect(result).toEqual({ behavior: "cancelled" });
+      expect(errorSpy).toHaveBeenCalledWith(expect.stringContaining("client exploded"));
+    });
+  });
 });

src/tests/elicitation.test.ts

@@ -0,0 +1,504 @@
+import { describe, it, expect } from "vitest";
+import type {
+  CreateElicitationRequest,
+  CreateElicitationResponse,
+  EnumOption,
+} from "@agentclientprotocol/sdk";
+import type { ElicitationRequest } from "@anthropic-ai/claude-agent-sdk";
+import {
+  applyAskElicitationResponse,
+  askUserQuestionsToCreateRequest,
+  AskUserQuestion,
+  createElicitationResponseToElicitResult,
+  extractAskUserQuestions,
+  extractRefusalFallbackPrompt,
+  mcpElicitationToCreateRequest,
+  refusalFallbackResultFromResponse,
+  refusalFallbackToCreateRequest,
+} from "../elicitation.js";
+
+const SESSION_ID = "session-1";
+
+/**
+ * Build a question matching the SDK's (strict) AskUserQuestion schema while
+ * keeping test sites terse. Option descriptions default to "" (the tool always
+ * provides one); pass a description to exercise the enum option's description.
+ */
+function mkQuestion(
+  question: string,
+  options: Array<{ label: string; description?: string; preview?: string }>,
+  opts: { header?: string; multiSelect?: boolean } = {},
+): AskUserQuestion {
+  return {
+    question,
+    header: opts.header ?? "",
+    multiSelect: opts.multiSelect ?? false,
+    options: options.map((o) => ({
+      label: o.label,
+      description: o.description ?? "",
+      ...(o.preview ? { preview: o.preview } : {}),
+    })),
+  } as AskUserQuestion;
+}
+
+describe("mcpElicitationToCreateRequest", () => {
+  it("maps a form-mode request, defaulting the schema type to object", () => {
+    const request: ElicitationRequest = {
+      serverName: "server",
+      message: "Need your name",
+      mode: "form",
+      requestedSchema: { properties: { name: { type: "string" } } },
+    };
+
+    const result = mcpElicitationToCreateRequest(request, SESSION_ID);
+
+    expect(result).toEqual({
+      mode: "form",
+      sessionId: SESSION_ID,
+      message: "Need your name",
+      requestedSchema: { type: "object", properties: { name: { type: "string" } } },
+    });
+  });
+
+  it("maps a url-mode request and preserves the elicitation id", () => {
+    const request: ElicitationRequest = {
+      serverName: "server",
+      message: "Authorize",
+      mode: "url",
+      url: "https://example.com/auth",
+      elicitationId: "el-123",
+    };
+
+    expect(mcpElicitationToCreateRequest(request, SESSION_ID)).toEqual({
+      mode: "url",
+      sessionId: SESSION_ID,
+      message: "Authorize",
+      url: "https://example.com/auth",
+      elicitationId: "el-123",
+    });
+  });
+
+  it("generates an elicitation id when a url request omits one", () => {
+    const request: ElicitationRequest = {
+      serverName: "server",
+      message: "Authorize",
+      mode: "url",
+      url: "https://example.com/auth",
+    };
+
+    const result = mcpElicitationToCreateRequest(request, SESSION_ID);
+    expect(result?.mode).toBe("url");
+    expect((result as { elicitationId: string }).elicitationId).toMatch(/.+/);
+  });
+
+  it("returns null for a url request with no url", () => {
+    const request: ElicitationRequest = {
+      serverName: "server",
+      message: "Authorize",
+      mode: "url",
+    };
+
+    expect(mcpElicitationToCreateRequest(request, SESSION_ID)).toBeNull();
+  });
+});
+
+describe("createElicitationResponseToElicitResult", () => {
+  it("maps accept with content", () => {
+    const response = { action: "accept", content: { name: "Ada" } } as CreateElicitationResponse;
+    expect(createElicitationResponseToElicitResult(response)).toEqual({
+      action: "accept",
+      content: { name: "Ada" },
+    });
+  });
+
+  it("defaults accept content to an empty object", () => {
+    const response = { action: "accept" } as CreateElicitationResponse;
+    expect(createElicitationResponseToElicitResult(response)).toEqual({
+      action: "accept",
+      content: {},
+    });
+  });
+
+  it("maps decline and cancel", () => {
+    expect(
+      createElicitationResponseToElicitResult({ action: "decline" } as CreateElicitationResponse),
+    ).toEqual({ action: "decline" });
+    expect(
+      createElicitationResponseToElicitResult({ action: "cancel" } as CreateElicitationResponse),
+    ).toEqual({ action: "cancel" });
+  });
+
+  it("maps a custom/future action to cancel", () => {
+    expect(
+      createElicitationResponseToElicitResult({
+        action: "_zed/snooze",
+      } as CreateElicitationResponse),
+    ).toEqual({ action: "cancel" });
+  });
+});
+
+describe("extractAskUserQuestions", () => {
+  const question = mkQuestion("Which?", [{ label: "A" }, { label: "B" }]);
+
+  it("reads questions off the tool input", () => {
+    expect(extractAskUserQuestions({ questions: [question] })).toEqual([question]);
+  });
+
+  it("returns null when there are no questions", () => {
+    expect(extractAskUserQuestions({})).toBeNull();
+    expect(extractAskUserQuestions({ questions: [] })).toBeNull();
+  });
+
+  it("filters out malformed questions (missing options, empty options, bad question)", () => {
+    const input = {
+      questions: [
+        question,
+        { question: "missing options" },
+        { options: [] },
+        { options: [], question: "empty" },
+      ],
+    } as unknown as Record<string, unknown>;
+    expect(extractAskUserQuestions(input)).toEqual([question]);
+  });
+
+  it("returns null when every question is filtered out", () => {
+    const input = {
+      questions: [{ question: "no options" }, { options: [{ label: "A" }] }],
+    } as unknown as Record<string, unknown>;
+    expect(extractAskUserQuestions(input)).toBeNull();
+  });
+});
+
+describe("askUserQuestionsToCreateRequest", () => {
+  it("keys a single-select question by a stable id and carries the prompt in message only", () => {
+    const questions = [
+      mkQuestion(
+        "Which library?",
+        [{ label: "date-fns", description: "Lightweight" }, { label: "luxon" }],
+        {
+          header: "Library",
+        },
+      ),
+    ];
+
+    const result = askUserQuestionsToCreateRequest(questions, SESSION_ID, "tool-1");
+
+    expect(result).toMatchObject({
+      mode: "form",
+      sessionId: SESSION_ID,
+      toolCallId: "tool-1",
+      message: "Which library?",
+    });
+    const schema = (result as Extract<CreateElicitationRequest, { mode: "form" }>).requestedSchema;
+    // Nothing is required; the user can pick, type a custom answer, or skip.
+    expect(schema.required).toBeUndefined();
+    // Single question → no duplicated description (prompt is in `message`).
+    expect(schema.properties?.["question_0"]).toEqual({
+      type: "string",
+      title: "Library",
+      description: undefined,
+      oneOf: [
+        // The description travels in the enum option's first-class field; the
+        // title stays the clean label (no "label — description" flattening).
+        { const: "date-fns", title: "date-fns", description: "Lightweight" },
+        // No description → the field is omitted entirely.
+        { const: "luxon", title: "luxon" },
+      ],
+    });
+    // The full question text is not used as a property key.
+    expect(schema.properties?.["Which library?"]).toBeUndefined();
+  });
+
+  it("carries descriptions structurally and forwards previews under _meta", () => {
+    const questions = [
+      mkQuestion("Which layout?", [
+        {
+          label: "Grid",
+          description: "Cards in a responsive grid",
+          preview: "```\n[ ] [ ] [ ]\n[ ] [ ] [ ]\n```",
+        },
+        { label: "List", description: "Stacked rows" },
+        { label: "Plain" },
+      ]),
+    ];
+
+    const schema = (
+      askUserQuestionsToCreateRequest(questions, SESSION_ID, undefined) as Extract<
+        CreateElicitationRequest,
+        { mode: "form" }
+      >
+    ).requestedSchema;
+    const oneOf = (schema.properties?.["question_0"] as { oneOf: EnumOption[] }).oneOf;
+
+    // The description uses the enum option's own field; only the preview —
+    // which still has no structural slot — rides under `_meta`.
+    expect(oneOf[0]).toEqual({
+      const: "Grid",
+      title: "Grid",
+      description: "Cards in a responsive grid",
+      _meta: {
+        "_claude/askUserQuestionOption": {
+          preview: "```\n[ ] [ ] [ ]\n[ ] [ ] [ ]\n```",
+        },
+      },
+    });
+    // Description only → no _meta needed.
+    expect(oneOf[1]).toEqual({ const: "List", title: "List", description: "Stacked rows" });
+    // Neither → just the label.
+    expect(oneOf[2]).toEqual({ const: "Plain", title: "Plain" });
+  });
+
+  it("includes a per-question optional free-text custom-answer field", () => {
+    const questions = [
+      mkQuestion("Which?", [{ label: "A" }, { label: "B" }]),
+      mkQuestion("What else?", [{ label: "C" }, { label: "D" }]),
+    ];
+    const schema = (
+      askUserQuestionsToCreateRequest(questions, SESSION_ID, undefined) as Extract<
+        CreateElicitationRequest,
+        { mode: "form" }
+      >
+    ).requestedSchema;
+
+    // Each question gets its own "Other" box, not one shared form-level field.
+    expect(schema.properties?.["question_0_custom"]).toMatchObject({
+      type: "string",
+      title: "Other",
+    });
+    expect(schema.properties?.["question_1_custom"]).toMatchObject({
+      type: "string",
+      title: "Other",
+    });
+  });
+
+  it("builds an array property for multi-select questions and includes per-field question text", () => {
+    const questions = [
+      mkQuestion("Pick one?", [{ label: "A" }, { label: "B" }]),
+      mkQuestion("Which features?", [{ label: "auth" }, { label: "logging" }], {
+        multiSelect: true,
+      }),
+    ];
+
+    const result = askUserQuestionsToCreateRequest(questions, SESSION_ID, undefined);
+    const schema = (result as Extract<CreateElicitationRequest, { mode: "form" }>).requestedSchema;
+
+    expect(result).not.toHaveProperty("toolCallId");
+    // Multiple questions → generic message, each field carries its own question.
+    expect(result.message).toBe("Please answer the following questions.");
+    expect(schema.properties?.["question_1"]).toEqual({
+      type: "array",
+      title: undefined,
+      description: "Which features?",
+      items: {
+        anyOf: [
+          { const: "auth", title: "auth" },
+          { const: "logging", title: "logging" },
+        ],
+      },
+    });
+  });
+});
+
+describe("applyAskElicitationResponse", () => {
+  const questions = [
+    mkQuestion("Single?", [{ label: "A" }, { label: "B" }]),
+    mkQuestion("Multi?", [{ label: "X" }, { label: "Y" }], { multiSelect: true }),
+  ];
+  const toolInput = { questions, metadata: { source: "test" } };
+
+  it("writes selected labels (keyed by indexed fields) into updatedInput on accept", () => {
+    const response = {
+      action: "accept",
+      content: { question_0: "A", question_1: ["X", "Y"] },
+    } as CreateElicitationResponse;
+
+    expect(applyAskElicitationResponse(response, toolInput, questions)).toEqual({
+      action: "answered",
+      updatedInput: {
+        questions,
+        metadata: { source: "test" },
+        answers: { "Single?": "A", "Multi?": "X, Y" },
+      },
+    });
+  });
+
+  it("folds a per-question custom answer into that question's answer", () => {
+    const response = {
+      action: "accept",
+      content: { question_0: "A", question_1_custom: "something else entirely" },
+    } as CreateElicitationResponse;
+
+    expect(applyAskElicitationResponse(response, toolInput, questions)).toEqual({
+      action: "answered",
+      updatedInput: {
+        questions,
+        metadata: { source: "test" },
+        answers: { "Single?": "A", "Multi?": "something else entirely" },
+      },
+    });
+  });
+
+  it("prefers a question's custom answer over its selection", () => {
+    const response = {
+      action: "accept",
+      content: { question_0: "A", question_0_custom: "  my own take  " },
+    } as CreateElicitationResponse;
+
+    expect(applyAskElicitationResponse(response, toolInput, questions)).toEqual({
+      action: "answered",
+      updatedInput: {
+        questions,
+        metadata: { source: "test" },
+        answers: { "Single?": "my own take" },
+      },
+    });
+  });
+
+  it("ignores empty selections and blank custom answers", () => {
+    const response = {
+      action: "accept",
+      content: { question_1: [], question_0_custom: "   " },
+    } as CreateElicitationResponse;
+
+    expect(applyAskElicitationResponse(response, toolInput, questions)).toEqual({
+      action: "answered",
+      updatedInput: { questions, metadata: { source: "test" }, answers: {} },
+    });
+  });
+
+  it("treats decline as answered with no answers", () => {
+    const response = { action: "decline" } as CreateElicitationResponse;
+    expect(applyAskElicitationResponse(response, toolInput, questions)).toEqual({
+      action: "answered",
+      updatedInput: { questions, metadata: { source: "test" }, answers: {} },
+    });
+  });
+
+  it("returns cancel on cancel", () => {
+    const response = { action: "cancel" } as CreateElicitationResponse;
+    expect(applyAskElicitationResponse(response, toolInput, questions)).toEqual({
+      action: "cancel",
+    });
+  });
+
+  it("returns cancel for a custom/future action it does not understand", () => {
+    const response = { action: "_zed/snooze" } as CreateElicitationResponse;
+    expect(applyAskElicitationResponse(response, toolInput, questions)).toEqual({
+      action: "cancel",
+    });
+  });
+
+  it("omits answers for questions the user left unanswered", () => {
+    const response = {
+      action: "accept",
+      content: { question_0: "B" },
+    } as CreateElicitationResponse;
+
+    const result = applyAskElicitationResponse(response, toolInput, questions);
+    expect(result.action).toBe("answered");
+    expect((result as Extract<typeof result, { action: "answered" }>).updatedInput.answers).toEqual(
+      { "Single?": "B" },
+    );
+  });
+});
+
+describe("extractRefusalFallbackPrompt", () => {
+  it("extracts required and optional fields", () => {
+    expect(
+      extractRefusalFallbackPrompt({
+        originalModel: "claude-fable-5",
+        fallbackModel: "claude-opus-4-8",
+        apiRefusalCategory: "cyber",
+        guidanceText: "You can retry with the fallback model.",
+        retractedMessageUuids: ["u1"],
+      }),
+    ).toEqual({
+      originalModel: "claude-fable-5",
+      fallbackModel: "claude-opus-4-8",
+      apiRefusalCategory: "cyber",
+      guidanceText: "You can retry with the fallback model.",
+    });
+  });
+
+  it("returns null when a required model field is missing or mistyped", () => {
+    expect(extractRefusalFallbackPrompt({ fallbackModel: "claude-opus-4-8" })).toBeNull();
+    expect(
+      extractRefusalFallbackPrompt({ originalModel: 5, fallbackModel: "claude-opus-4-8" }),
+    ).toBeNull();
+  });
+
+  it("normalizes a non-string category to null and drops empty guidance", () => {
+    expect(
+      extractRefusalFallbackPrompt({
+        originalModel: "a",
+        fallbackModel: "b",
+        apiRefusalCategory: null,
+        guidanceText: "",
+      }),
+    ).toEqual({ originalModel: "a", fallbackModel: "b", apiRefusalCategory: null });
+  });
+});
+
+describe("refusalFallbackToCreateRequest", () => {
+  const prompt = {
+    originalModel: "claude-fable-5",
+    fallbackModel: "claude-opus-4-8",
+    apiRefusalCategory: "cyber" as string | null,
+  };
+
+  it("builds a single-choice form whose enum consts are the dialog's wire results", () => {
+    const request = refusalFallbackToCreateRequest(prompt, SESSION_ID) as Extract<
+      CreateElicitationRequest,
+      { mode: "form" }
+    >;
+    expect(request.mode).toBe("form");
+    expect(request).toMatchObject({ sessionId: SESSION_ID });
+    expect(request.message).toContain("claude-fable-5");
+    expect(request.message).toContain("(cyber)");
+    expect(request.message).toContain("claude-opus-4-8");
+
+    const choice = request.requestedSchema.properties?.choice as {
+      type: string;
+      oneOf: Array<{ const: string; title: string; description?: string }>;
+    };
+    expect(choice.type).toBe("string");
+    expect(choice.oneOf.map((o) => o.const)).toEqual(["retry_fallback", "cancelled"]);
+    expect(choice.oneOf[0].title).toContain("claude-opus-4-8");
+    // Each option explains its consequence in the description field.
+    expect(choice.oneOf[0].description).toContain("claude-opus-4-8");
+    expect(choice.oneOf[1].description).toBeTruthy();
+  });
+
+  it("omits the category marker and appends guidance when provided", () => {
+    const request = refusalFallbackToCreateRequest(
+      { ...prompt, apiRefusalCategory: null, guidanceText: "Consider rephrasing." },
+      SESSION_ID,
+    );
+    expect(request.message).not.toContain("(");
+    expect(request.message).toContain("Consider rephrasing.");
+  });
+});
+
+describe("refusalFallbackResultFromResponse", () => {
+  it("maps an accepted retry choice to retry_fallback", () => {
+    const response = {
+      action: "accept",
+      content: { choice: "retry_fallback" },
+    } as CreateElicitationResponse;
+    expect(refusalFallbackResultFromResponse(response)).toBe("retry_fallback");
+  });
+
+  it("keeps the refusal for every other outcome", () => {
+    const cases = [
+      { action: "accept", content: { choice: "cancelled" } },
+      { action: "accept", content: {} },
+      { action: "accept", content: { choice: "something-new" } },
+      { action: "accept" },
+      { action: "decline" },
+      { action: "cancel" },
+    ] as CreateElicitationResponse[];
+    for (const response of cases) {
+      expect(refusalFallbackResultFromResponse(response)).toBe("cancelled");
+    }
+  });
+});

src/tests/fast-mode-config.test.ts

@@ -0,0 +1,341 @@
+import { describe, it, expect, vi } from "vitest";
+import type { ClientCapabilities, SessionNotification } from "@agentclientprotocol/sdk";
+import type { ModelInfo } from "@anthropic-ai/claude-agent-sdk";
+import {
+  buildConfigOptions,
+  clientSupportsBooleanConfigOptions,
+  createFastModeConfigOption,
+  fastModeStateEnabled,
+  resolveFastModeEnabled,
+  FAST_MODE_CONFIG_ID,
+  FAST_MODE_ON,
+  FAST_MODE_OFF,
+  ClaudeAcpAgent,
+  type AcpClient,
+} from "../acp-agent.js";
+
+const MODES = {
+  currentModeId: "default",
+  availableModes: [{ id: "default", name: "Default", description: "Standard behavior" }],
+};
+
+const MODELS = {
+  currentModelId: "claude-opus-4-8",
+  availableModels: [
+    { modelId: "claude-opus-4-8", name: "Claude Opus", description: "Most capable" },
+  ],
+};
+
+const MODEL_INFOS: ModelInfo[] = [
+  { value: "claude-opus-4-8", displayName: "Claude Opus", description: "Most capable" },
+];
+
+describe("createFastModeConfigOption", () => {
+  it("produces a native boolean toggle when the client opted in", () => {
+    expect(createFastModeConfigOption(true, true)).toEqual({
+      id: FAST_MODE_CONFIG_ID,
+      name: "Fast mode",
+      description: expect.any(String),
+      category: "model_config",
+      type: "boolean",
+      currentValue: true,
+    });
+  });
+
+  it("falls back to an on/off select when the client did not opt in", () => {
+    const option = createFastModeConfigOption(false, false);
+    expect(option).toMatchObject({
+      id: FAST_MODE_CONFIG_ID,
+      type: "select",
+      category: "model_config",
+      currentValue: FAST_MODE_OFF,
+      options: [
+        { value: FAST_MODE_ON, name: "On" },
+        { value: FAST_MODE_OFF, name: "Off" },
+      ],
+    });
+    expect(option).not.toHaveProperty("currentValue", true);
+  });
+});
+
+describe("clientSupportsBooleanConfigOptions", () => {
+  it("is true only when session.configOptions.boolean is present", () => {
+    expect(
+      clientSupportsBooleanConfigOptions({ session: { configOptions: { boolean: {} } } }),
+    ).toBe(true);
+  });
+
+  it("is false when the capability is omitted or null at any level", () => {
+    expect(clientSupportsBooleanConfigOptions(undefined)).toBe(false);
+    expect(clientSupportsBooleanConfigOptions(null)).toBe(false);
+    expect(clientSupportsBooleanConfigOptions({})).toBe(false);
+    expect(clientSupportsBooleanConfigOptions({ session: { configOptions: {} } })).toBe(false);
+    expect(
+      clientSupportsBooleanConfigOptions({ session: { configOptions: { boolean: null } } }),
+    ).toBe(false);
+  });
+});
+
+describe("resolveFastModeEnabled", () => {
+  const base = { sessionId: "s", configId: FAST_MODE_CONFIG_ID };
+
+  it("accepts native boolean values", () => {
+    expect(resolveFastModeEnabled({ ...base, type: "boolean", value: true })).toBe(true);
+    expect(resolveFastModeEnabled({ ...base, type: "boolean", value: false })).toBe(false);
+  });
+
+  it("accepts the on/off select fallback", () => {
+    expect(resolveFastModeEnabled({ ...base, value: FAST_MODE_ON })).toBe(true);
+    expect(resolveFastModeEnabled({ ...base, value: FAST_MODE_OFF })).toBe(false);
+  });
+
+  it("rejects any other value", () => {
+    expect(() => resolveFastModeEnabled({ ...base, value: "maybe" })).toThrow(
+      /Invalid value for config option fast/,
+    );
+  });
+});
+
+describe("fastModeStateEnabled", () => {
+  it("treats cooldown as on (the user's intent persists through rate-limit cooldown)", () => {
+    expect(fastModeStateEnabled("on")).toBe(true);
+    expect(fastModeStateEnabled("cooldown")).toBe(true);
+    expect(fastModeStateEnabled("off")).toBe(false);
+  });
+});
+
+describe("buildConfigOptions Fast mode", () => {
+  it("omits the Fast mode option when the model does not support it", () => {
+    const options = buildConfigOptions(MODES, MODELS, MODEL_INFOS, undefined, [], "default", {
+      supported: false,
+      enabled: false,
+      useBooleanOption: true,
+    });
+    expect(options.find((o) => o.id === FAST_MODE_CONFIG_ID)).toBeUndefined();
+  });
+
+  it("omits the Fast mode option when no fast mode state is provided", () => {
+    const options = buildConfigOptions(MODES, MODELS, MODEL_INFOS, undefined, [], "default");
+    expect(options.find((o) => o.id === FAST_MODE_CONFIG_ID)).toBeUndefined();
+  });
+
+  it("surfaces a boolean toggle when supported and the client opted in", () => {
+    const options = buildConfigOptions(MODES, MODELS, MODEL_INFOS, undefined, [], "default", {
+      supported: true,
+      enabled: true,
+      useBooleanOption: true,
+    });
+    expect(options).toContainEqual(createFastModeConfigOption(true, true));
+  });
+
+  it("surfaces a select fallback when supported but the client did not opt in", () => {
+    const options = buildConfigOptions(MODES, MODELS, MODEL_INFOS, undefined, [], "default", {
+      supported: true,
+      enabled: false,
+      useBooleanOption: false,
+    });
+    expect(options).toContainEqual(createFastModeConfigOption(false, false));
+  });
+});
+
+describe("setSessionConfigOption Fast mode toggle", () => {
+  const SESSION_ID = "fast-session";
+
+  function setup(opts: { useBooleanOption: boolean }) {
+    const sessionUpdates: SessionNotification[] = [];
+    const client = {
+      sessionUpdate: async (n: SessionNotification) => {
+        sessionUpdates.push(n);
+      },
+      requestPermission: async () => ({ outcome: { outcome: "cancelled" } }),
+      readTextFile: async () => ({ content: "" }),
+      writeTextFile: async () => ({}),
+    } as unknown as AcpClient;
+
+    const agent = new ClaudeAcpAgent(client);
+    const clientCapabilities: ClientCapabilities = opts.useBooleanOption
+      ? { session: { configOptions: { boolean: {} } } }
+      : {};
+    (agent as unknown as { clientCapabilities: ClientCapabilities }).clientCapabilities =
+      clientCapabilities;
+
+    const applyFlagSettings = vi.fn();
+    (agent as unknown as { sessions: Record<string, unknown> }).sessions[SESSION_ID] = {
+      query: { applyFlagSettings },
+      fastModeEnabled: false,
+      configOptions: [createFastModeConfigOption(false, opts.useBooleanOption)],
+    };
+
+    return { agent, applyFlagSettings, sessionUpdates };
+  }
+
+  it("toggles Fast mode on/off through a boolean value", async () => {
+    const { agent, applyFlagSettings } = setup({ useBooleanOption: true });
+
+    const onResponse = await agent.setSessionConfigOption({
+      sessionId: SESSION_ID,
+      configId: FAST_MODE_CONFIG_ID,
+      type: "boolean",
+      value: true,
+    });
+    expect(applyFlagSettings).toHaveBeenCalledWith({ fastMode: true });
+    expect(onResponse.configOptions).toContainEqual(createFastModeConfigOption(true, true));
+    expect(
+      (agent as unknown as { sessions: Record<string, { fastModeEnabled: boolean }> }).sessions[
+        SESSION_ID
+      ].fastModeEnabled,
+    ).toBe(true);
+
+    const offResponse = await agent.setSessionConfigOption({
+      sessionId: SESSION_ID,
+      configId: FAST_MODE_CONFIG_ID,
+      type: "boolean",
+      value: false,
+    });
+    expect(applyFlagSettings).toHaveBeenLastCalledWith({ fastMode: false });
+    expect(offResponse.configOptions).toContainEqual(createFastModeConfigOption(false, true));
+  });
+
+  it("toggles Fast mode through the on/off select fallback", async () => {
+    const { agent, applyFlagSettings } = setup({ useBooleanOption: false });
+
+    const response = await agent.setSessionConfigOption({
+      sessionId: SESSION_ID,
+      configId: FAST_MODE_CONFIG_ID,
+      value: FAST_MODE_ON,
+    });
+    expect(applyFlagSettings).toHaveBeenCalledWith({ fastMode: true });
+    expect(response.configOptions).toContainEqual(createFastModeConfigOption(true, false));
+  });
+
+  it("does not change session state when the SDK rejects the flag", async () => {
+    const { agent, applyFlagSettings } = setup({ useBooleanOption: true });
+    applyFlagSettings.mockRejectedValueOnce(new Error("nope"));
+
+    await expect(
+      agent.setSessionConfigOption({
+        sessionId: SESSION_ID,
+        configId: FAST_MODE_CONFIG_ID,
+        type: "boolean",
+        value: true,
+      }),
+    ).rejects.toThrow("nope");
+
+    const session = (agent as unknown as { sessions: Record<string, { fastModeEnabled: boolean }> })
+      .sessions[SESSION_ID];
+    expect(session.fastModeEnabled).toBe(false);
+  });
+});
+
+describe("syncFastModeState (SDK-driven state changes)", () => {
+  const SESSION_ID = "fast-session";
+
+  function setup(opts: { fastModeEnabled: boolean; withOption: boolean }) {
+    const sessionUpdates: SessionNotification[] = [];
+    const client = {
+      sessionUpdate: async (n: SessionNotification) => {
+        sessionUpdates.push(n);
+      },
+      requestPermission: async () => ({ outcome: { outcome: "cancelled" } }),
+      readTextFile: async () => ({ content: "" }),
+      writeTextFile: async () => ({}),
+    } as unknown as AcpClient;
+
+    const agent = new ClaudeAcpAgent(client);
+    (agent as unknown as { clientCapabilities: ClientCapabilities }).clientCapabilities = {
+      session: { configOptions: { boolean: {} } },
+    };
+
+    const session = {
+      query: {},
+      fastModeEnabled: opts.fastModeEnabled,
+      configOptions: opts.withOption
+        ? [createFastModeConfigOption(opts.fastModeEnabled, true)]
+        : [],
+    };
+    (agent as unknown as { sessions: Record<string, unknown> }).sessions[SESSION_ID] = session;
+
+    const sync = (
+      agent as unknown as {
+        syncFastModeState: (
+          sessionId: string,
+          session: unknown,
+          state: string | undefined,
+        ) => Promise<void>;
+      }
+    ).syncFastModeState.bind(agent);
+
+    return { sync, session, sessionUpdates };
+  }
+
+  it("emits a config_option_update when the SDK reports a new state", async () => {
+    const { sync, session, sessionUpdates } = setup({ fastModeEnabled: false, withOption: true });
+
+    await sync(SESSION_ID, session, "on");
+
+    expect(session.fastModeEnabled).toBe(true);
+    expect(session.configOptions).toContainEqual(createFastModeConfigOption(true, true));
+    expect(sessionUpdates).toHaveLength(1);
+    expect(sessionUpdates[0].update).toMatchObject({
+      sessionUpdate: "config_option_update",
+    });
+    const updated = (
+      sessionUpdates[0].update as { configOptions: ReturnType<typeof createFastModeConfigOption>[] }
+    ).configOptions;
+    expect(updated).toContainEqual(createFastModeConfigOption(true, true));
+  });
+
+  it("leaves the toggle on and quiet during a rate-limit cooldown", async () => {
+    const { sync, session, sessionUpdates } = setup({ fastModeEnabled: true, withOption: true });
+
+    // cooldown is a transient suspension of an already-enabled fast mode.
+    await sync(SESSION_ID, session, "cooldown");
+
+    expect(session.fastModeEnabled).toBe(true);
+    expect(sessionUpdates).toHaveLength(0);
+  });
+
+  it("never lets a stray cooldown spuriously enable a toggle the user has off", async () => {
+    const { sync, session, sessionUpdates } = setup({ fastModeEnabled: false, withOption: true });
+
+    await sync(SESSION_ID, session, "cooldown");
+
+    expect(session.fastModeEnabled).toBe(false);
+    expect(sessionUpdates).toHaveLength(0);
+  });
+
+  it("clears the toggle when the SDK reports off", async () => {
+    const { sync, session, sessionUpdates } = setup({ fastModeEnabled: true, withOption: true });
+
+    await sync(SESSION_ID, session, "off");
+
+    expect(session.fastModeEnabled).toBe(false);
+    expect(session.configOptions).toContainEqual(createFastModeConfigOption(false, true));
+    expect(sessionUpdates).toHaveLength(1);
+  });
+
+  it("is a no-op when the reported state is undefined or unchanged", async () => {
+    const { sync, session, sessionUpdates } = setup({ fastModeEnabled: false, withOption: true });
+
+    await sync(SESSION_ID, session, undefined);
+    await sync(SESSION_ID, session, "off");
+
+    expect(sessionUpdates).toHaveLength(0);
+  });
+
+  it("preserves the retained setting (no clobber) when the model has no Fast mode option", async () => {
+    // Model without fast support: the SDK reports a capability-driven state, not
+    // the user's intent. We must leave session.fastModeEnabled untouched so it's
+    // correct when a supporting model is reselected — reconciling here was the
+    // original intent-clobber bug.
+    const enabledCase = setup({ fastModeEnabled: true, withOption: false });
+    await enabledCase.sync(SESSION_ID, enabledCase.session, "off");
+    expect(enabledCase.session.fastModeEnabled).toBe(true);
+    expect(enabledCase.sessionUpdates).toHaveLength(0);
+
+    const disabledCase = setup({ fastModeEnabled: false, withOption: false });
+    await disabledCase.sync(SESSION_ID, disabledCase.session, "on");
+    expect(disabledCase.session.fastModeEnabled).toBe(false);
+    expect(disabledCase.sessionUpdates).toHaveLength(0);
+  });
+});

src/tests/helpers.ts

@@ -0,0 +1,34 @@
+/**
+ * Shared test doubles. Deliberately vitest-free so `vi.mock` async factories
+ * can `await import("./helpers.js")` without ordering hazards; tests supply
+ * their own vi.fn spies via `overrides`.
+ */
+
+/** The context-usage report the base mock query returns. `rawMaxTokens`
+ *  matches the agent's DEFAULT_CONTEXT_WINDOW so window-related assertions
+ *  don't shift in tests that don't care about context usage. */
+export const DEFAULT_CONTEXT_USAGE = { totalTokens: 0, rawMaxTokens: 200000 };
+
+/**
+ * Base stub for the SDK `query()` return object, covering the surface
+ * ClaudeAcpAgent touches unconditionally at session creation. Tests pass
+ * `overrides` for the parts they assert on (spies, custom models, rejecting
+ * getContextUsage, …).
+ *
+ * When the agent starts calling a new SDK method on every session, add it
+ * here once — the getContextUsage adoption required hand-editing ~10 inline
+ * mocks across five files, and any missed copy didn't fail: it silently
+ * rerouted that test through the error-fallback branch and re-polluted test
+ * output.
+ */
+export function makeMockQuery(overrides: Record<string, unknown> = {}) {
+  return {
+    initializationResult: async () => ({ models: [] }),
+    setModel: async () => {},
+    setPermissionMode: async () => {},
+    supportedCommands: async () => [],
+    getContextUsage: async () => DEFAULT_CONTEXT_USAGE,
+    [Symbol.asyncIterator]: async function* () {},
+    ...overrides,
+  };
+}

src/tests/model-resolution.test.ts

@@ -0,0 +1,272 @@
+import { describe, it, expect } from "vitest";
+import { randomUUID } from "crypto";
+import { query } from "@anthropic-ai/claude-agent-sdk";
+import type { ModelInfo } from "@anthropic-ai/claude-agent-sdk";
+import {
+  resolveModelPreference,
+  applyAvailableModelsAllowlist,
+  matchResumedModel,
+} from "../acp-agent.js";
+
+// Mirrors a real `supportedModels()` response: alias rows carry
+// `resolvedModel`, and "Sonnet 5" has no `major.minor` dot unlike older
+// "claude-opus-4-6"-style ids.
+const LIVE_SHAPED_MODELS: ModelInfo[] = [
+  {
+    value: "default",
+    resolvedModel: "claude-opus-4-8[1m]",
+    displayName: "Default (recommended)",
+    description: "Use the default model (currently Opus 4.8 (1M context))",
+  },
+  {
+    value: "opus[1m]",
+    resolvedModel: "claude-opus-4-8[1m]",
+    displayName: "Opus",
+    description: "Opus 4.8 with 1M context · Best for everyday, complex tasks",
+  },
+  {
+    value: "sonnet",
+    resolvedModel: "claude-sonnet-5",
+    displayName: "Sonnet",
+    description: "Sonnet 5 · Efficient for routine tasks",
+  },
+  {
+    value: "haiku",
+    resolvedModel: "claude-haiku-4-5-20251001",
+    displayName: "Haiku",
+    description: "Haiku 4.5 · Fastest for quick answers",
+  },
+];
+
+describe("resolveModelPreference - resolvedModel matching", () => {
+  it("matches a preference that equals an alias's resolvedModel exactly", () => {
+    expect(resolveModelPreference(LIVE_SHAPED_MODELS, "claude-sonnet-5")?.value).toBe("sonnet");
+    expect(resolveModelPreference(LIVE_SHAPED_MODELS, "claude-opus-4-8[1m]")?.value).toBe(
+      "opus[1m]",
+    );
+    expect(resolveModelPreference(LIVE_SHAPED_MODELS, "claude-haiku-4-5-20251001")?.value).toBe(
+      "haiku",
+    );
+  });
+
+  it("is case-insensitive", () => {
+    expect(resolveModelPreference(LIVE_SHAPED_MODELS, "CLAUDE-SONNET-5")?.value).toBe("sonnet");
+  });
+
+  it("still resolves plain aliases via the direct-match tier", () => {
+    expect(resolveModelPreference(LIVE_SHAPED_MODELS, "sonnet")?.value).toBe("sonnet");
+    expect(resolveModelPreference(LIVE_SHAPED_MODELS, "opus[1m]")?.value).toBe("opus[1m]");
+  });
+
+  it("does not let a retired dated id drift onto a same-family alias of a different generation", () => {
+    expect(resolveModelPreference(LIVE_SHAPED_MODELS, "claude-sonnet-4-6")).toBeNull();
+  });
+
+  // "opus[1m]"'s `value` has no dotted version, just the context-hint digit
+  // in "[1m]" — unstripped, that reads as version "1" and blocks the match
+  // before the matcher reaches the description's real "4.8".
+  it("does not mistake an alias's context-hint suffix for its version", () => {
+    expect(resolveModelPreference(LIVE_SHAPED_MODELS, "claude-opus-4-8")?.value).toBe("opus[1m]");
+  });
+
+  it("returns null for a preference that matches nothing", () => {
+    expect(resolveModelPreference(LIVE_SHAPED_MODELS, "claude-gpt-99")).toBeNull();
+  });
+
+  it("matches a '-1m' id-suffix preference against a '[1m]'-spelled resolvedModel", () => {
+    // Without hint canonicalization the exact tier misses and the substring
+    // tier lands on the bare 200k sibling, silently downgrading a 1M pin.
+    const models: ModelInfo[] = [
+      {
+        value: "sonnet",
+        resolvedModel: "claude-sonnet-5",
+        displayName: "Sonnet",
+        description: "Sonnet 5 · Efficient for routine tasks",
+      },
+      {
+        value: "sonnet[1m]",
+        resolvedModel: "claude-sonnet-5[1m]",
+        displayName: "Sonnet",
+        description: "Sonnet 5 with 1M context",
+      },
+    ];
+    expect(resolveModelPreference(models, "claude-sonnet-5-1m")?.value).toBe("sonnet[1m]");
+  });
+
+  it("matches a '-1m'-suffix alias directly against a '[1m]'-spelled value", () => {
+    expect(resolveModelPreference(LIVE_SHAPED_MODELS, "opus-1m")?.value).toBe("opus[1m]");
+  });
+
+  // Rows without resolvedModel (older SDKs, allowlist-synthesized entries,
+  // option-derived lists) can't use the exact resolved tier, so the hinted
+  // preference must not be absorbed by the bare sibling in the substring
+  // tier — the tokenized tier scores the hinted row correctly instead.
+  it("does not let a bare sibling absorb a hinted preference when resolvedModel is absent", () => {
+    const models: ModelInfo[] = [
+      { value: "sonnet", displayName: "Sonnet", description: "" },
+      { value: "sonnet[1m]", displayName: "Sonnet", description: "" },
+    ];
+    expect(resolveModelPreference(models, "claude-sonnet-5[1m]")?.value).toBe("sonnet[1m]");
+    expect(resolveModelPreference(models, "claude-sonnet-5-1m")?.value).toBe("sonnet[1m]");
+    // Bare preferences still land on the bare row via the substring tier.
+    expect(resolveModelPreference(models, "claude-sonnet-5")?.value).toBe("sonnet");
+  });
+});
+
+describe("matchResumedModel", () => {
+  // Environment shape from issue #845: the default resolves to Sonnet, the
+  // session was explicitly switched to Opus before the process restart.
+  const SONNET_DEFAULT_MODELS: ModelInfo[] = [
+    {
+      value: "default",
+      resolvedModel: "claude-sonnet-5",
+      displayName: "Default (recommended)",
+      description: "Use the default model (currently Sonnet 5)",
+    },
+    {
+      value: "opus",
+      resolvedModel: "claude-opus-4-6[1m]",
+      displayName: "Opus",
+      description: "Opus 4.6 · Most capable for complex work",
+    },
+    {
+      value: "sonnet",
+      resolvedModel: "claude-sonnet-5",
+      displayName: "Sonnet",
+      description: "Sonnet 5 · Efficient for routine tasks",
+    },
+    {
+      value: "sonnet[1m]",
+      resolvedModel: "claude-sonnet-5[1m]",
+      displayName: "Sonnet",
+      description: "Sonnet 5 with 1M context",
+    },
+  ];
+
+  it("maps an explicitly-switched session's live model to its picker alias", () => {
+    expect(matchResumedModel(SONNET_DEFAULT_MODELS, "claude-opus-4-6").value).toBe("opus");
+    expect(matchResumedModel(SONNET_DEFAULT_MODELS, "claude-opus-4-6[1m]").value).toBe("opus");
+  });
+
+  it("keeps Default selected when the live model is what Default resolves to", () => {
+    expect(matchResumedModel(SONNET_DEFAULT_MODELS, "claude-sonnet-5").value).toBe("default");
+  });
+
+  it("prefers an exact named-row resolvedModel match over the hint-stripped Default tier", () => {
+    // "claude-sonnet-5[1m]" differs from Default's "claude-sonnet-5" only by
+    // the context hint; stripping both would misreport the explicit 1M row as
+    // Default (and infer the wrong context window downstream). Both hint
+    // spellings must land on the row regardless of which one it uses.
+    expect(matchResumedModel(SONNET_DEFAULT_MODELS, "claude-sonnet-5[1m]").value).toBe(
+      "sonnet[1m]",
+    );
+    expect(matchResumedModel(SONNET_DEFAULT_MODELS, "claude-sonnet-5-1m").value).toBe("sonnet[1m]");
+  });
+
+  it("normalizes the '-1m' id-suffix hint form and casing against Default's resolution", () => {
+    expect(matchResumedModel(LIVE_SHAPED_MODELS, "claude-opus-4-8-1m").value).toBe("default");
+    expect(matchResumedModel(LIVE_SHAPED_MODELS, "Claude-Opus-4-8[1M]").value).toBe("default");
+  });
+
+  it("prefers Default over a named row sharing the identical resolvedModel", () => {
+    // "opus[1m]" carries the same resolvedModel as Default; the live id can't
+    // tell the two apart, so a never-customized session stays on Default.
+    expect(matchResumedModel(LIVE_SHAPED_MODELS, "claude-opus-4-8[1m]").value).toBe("default");
+  });
+
+  it("ignores '[1m]' context hints when comparing against Default's resolution", () => {
+    // The transcript-restored model drops the "[1m]" context hint, so a
+    // session that never left the 1M-context default resumes as the bare id.
+    expect(matchResumedModel(LIVE_SHAPED_MODELS, "claude-opus-4-8").value).toBe("default");
+  });
+
+  it("still resolves non-default live models against the picker entries", () => {
+    expect(matchResumedModel(LIVE_SHAPED_MODELS, "claude-sonnet-5").value).toBe("sonnet");
+    expect(matchResumedModel(LIVE_SHAPED_MODELS, "claude-haiku-4-5-20251001").value).toBe("haiku");
+  });
+
+  it("tracks a live model with no picker counterpart verbatim", () => {
+    const result = matchResumedModel(LIVE_SHAPED_MODELS, "claude-gpt-99");
+    expect(result.value).toBe("claude-gpt-99");
+    expect(result.displayName).toBe("claude-gpt-99");
+  });
+});
+
+describe("applyAvailableModelsAllowlist - resolvedModel matching", () => {
+  it("resolves an allowlist entry pinned to a canonical resolvedModel id", () => {
+    const result = applyAvailableModelsAllowlist(LIVE_SHAPED_MODELS, ["claude-sonnet-5"]);
+    const entry = result.find((m) => m.value === "claude-sonnet-5");
+    expect(entry).toBeDefined();
+    expect(entry?.displayName).toBe("Sonnet");
+  });
+});
+
+describe("applyAvailableModelsAllowlist - modelOverrides", () => {
+  const SDK_MODELS: ModelInfo[] = [
+    { value: "default", displayName: "Default", description: "Default model" },
+    {
+      value: "opus",
+      displayName: "Opus",
+      description: "Claude Opus 4.6",
+      supportsEffort: true,
+      supportedEffortLevels: ["low", "high"],
+    },
+  ];
+
+  it("surfaces the override target as the value while keeping the alias's display metadata", () => {
+    // The override target is an opaque Bedrock ARN with no "opus" substring,
+    // so matching against it (instead of the "claude-opus-4-6" alias) would
+    // find nothing and silently drop displayName/description/supportsEffort.
+    const overrides = {
+      "claude-opus-4-6": "arn:aws:bedrock:us-west-2:111122223333:inference-profile/custom-7f3a",
+    };
+    const result = applyAvailableModelsAllowlist(SDK_MODELS, ["claude-opus-4-6"], overrides);
+
+    const entry = result.find((m) => m.value === overrides["claude-opus-4-6"]);
+    expect(entry).toEqual({
+      value: overrides["claude-opus-4-6"],
+      displayName: "Opus",
+      description: "Claude Opus 4.6",
+      supportsEffort: true,
+      supportedEffortLevels: ["low", "high"],
+    });
+  });
+});
+
+// Runs against the real SDK (no mocks) so the fixture above is flagged the
+// moment the SDK's actual model list shape drifts from it. Requires a usable
+// ANTHROPIC_API_KEY, hence opt-in via RUN_INTEGRATION_TESTS.
+describe.skipIf(!process.env.RUN_INTEGRATION_TESTS)("model resolution (real SDK)", () => {
+  it("every model's resolvedModel (when present) resolves back to that same row", async () => {
+    const q = query({ prompt: "hi", options: { sessionId: randomUUID() } });
+    try {
+      const models = await q.supportedModels();
+      expect(models.length).toBeGreaterThan(0);
+
+      for (const model of models) {
+        // "default" shares a resolvedModel with the CLI's recommended named
+        // alias, and resolution intentionally prefers that alias on a tie.
+        if (!model.resolvedModel || model.value === "default") continue;
+        const resolved = resolveModelPreference(models, model.resolvedModel);
+        expect(resolved?.value).toBe(model.value);
+      }
+    } finally {
+      q.return(undefined);
+    }
+  }, 30000);
+
+  it("resolves the well-known 'sonnet' and 'opus' aliases to a model bearing their family name", async () => {
+    const q = query({ prompt: "hi", options: { sessionId: randomUUID() } });
+    try {
+      const models = await q.supportedModels();
+
+      const sonnet = resolveModelPreference(models, "sonnet");
+      expect(sonnet?.resolvedModel?.toLowerCase()).toContain("sonnet");
+
+      const opus = resolveModelPreference(models, "opus");
+      expect(opus?.resolvedModel?.toLowerCase()).toContain("opus");
+    } finally {
+      q.return(undefined);
+    }
+  }, 30000);
+});

src/tests/providers.test.ts

@@ -0,0 +1,303 @@
+import { describe, expect, it, Mock, vi, afterEach, beforeEach } from "vitest";
+import { AcpClient, ClaudeAcpAgent } from "../acp-agent.js";
+
+const mockQuery = vi.hoisted(() =>
+  vi.fn(() => ({
+    initializationResult: vi.fn().mockResolvedValue({
+      models: [
+        { value: "id", displayName: "name", description: "description", supportsAutoMode: true },
+      ],
+    }),
+    setModel: vi.fn(),
+    setPermissionMode: vi.fn(),
+    supportedCommands: vi.fn().mockResolvedValue([]),
+  })),
+);
+
+vi.mock("@anthropic-ai/claude-agent-sdk", async () => ({
+  ...(await vi.importActual<typeof import("@anthropic-ai/claude-agent-sdk")>(
+    "@anthropic-ai/claude-agent-sdk",
+  )),
+  query: mockQuery,
+}));
+
+// `logout` shells out to `claude auth logout`; make execFile succeed so the
+// real logout body runs (and clears provider config) without a live CLI.
+vi.mock("node:child_process", async () => {
+  const actual = await vi.importActual<typeof import("node:child_process")>("node:child_process");
+  return {
+    ...actual,
+    execFile: vi.fn((_file: string, _args: string[], cb: (...a: unknown[]) => void) =>
+      cb(null, { stdout: "", stderr: "" }),
+    ),
+  };
+});
+
+describe("providers", () => {
+  beforeEach(() => {
+    vi.useFakeTimers();
+  });
+
+  afterEach(() => {
+    vi.runAllTimers();
+    vi.useRealTimers();
+    vi.unstubAllGlobals();
+    vi.resetAllMocks();
+  });
+
+  async function createAgentMock(): Promise<[ClaudeAcpAgent, Mock]> {
+    const connectionMock = {
+      sessionUpdate: async (_: any) => {},
+    } as AcpClient;
+    const agent = new ClaudeAcpAgent(connectionMock);
+    return [agent, mockQuery];
+  }
+
+  it("advertises the providers capability in initialize", async () => {
+    const [agent] = await createAgentMock();
+    const response = await agent.initialize({ protocolVersion: 1, clientCapabilities: {} });
+    expect(response.agentCapabilities?.providers).toEqual({});
+  });
+
+  it("lists a single optional 'main' provider, unconfigured by default", async () => {
+    const [agent] = await createAgentMock();
+    const response = await agent.unstable_listProviders({});
+    expect(response.providers).toEqual([
+      {
+        providerId: "main",
+        supported: ["anthropic", "bedrock", "vertex"],
+        required: false,
+        current: null,
+      },
+    ]);
+  });
+
+  it("reflects apiType/baseUrl after set, and never echoes headers", async () => {
+    const [agent] = await createAgentMock();
+    await agent.unstable_setProvider({
+      providerId: "main",
+      apiType: "anthropic",
+      baseUrl: "https://gateway.example/v1",
+      headers: { authorization: "Bearer secret" },
+    });
+
+    const response = await agent.unstable_listProviders({});
+    expect(response.providers[0].current).toEqual({
+      apiType: "anthropic",
+      baseUrl: "https://gateway.example/v1",
+    });
+    // headers must not leak through list
+    expect(JSON.stringify(response.providers)).not.toContain("secret");
+  });
+
+  it("rejects set for an unknown providerId", async () => {
+    const [agent] = await createAgentMock();
+    await expect(
+      agent.unstable_setProvider({
+        providerId: "openai",
+        apiType: "anthropic",
+        baseUrl: "https://gateway.example",
+      }),
+    ).rejects.toMatchObject({ code: -32602 });
+  });
+
+  it("rejects set for an unsupported apiType", async () => {
+    const [agent] = await createAgentMock();
+    await expect(
+      agent.unstable_setProvider({
+        providerId: "main",
+        apiType: "openai",
+        baseUrl: "https://gateway.example",
+      }),
+    ).rejects.toMatchObject({ code: -32602 });
+  });
+
+  it("rejects set for an empty or non-http baseUrl", async () => {
+    const [agent] = await createAgentMock();
+    await expect(
+      agent.unstable_setProvider({ providerId: "main", apiType: "anthropic", baseUrl: "" }),
+    ).rejects.toMatchObject({ code: -32602 });
+    await expect(
+      agent.unstable_setProvider({
+        providerId: "main",
+        apiType: "anthropic",
+        baseUrl: "ftp://nope.example",
+      }),
+    ).rejects.toMatchObject({ code: -32602 });
+  });
+
+  it("disables the 'main' provider by clearing config and reporting current: null", async () => {
+    const [agent] = await createAgentMock();
+    await agent.unstable_setProvider({
+      providerId: "main",
+      apiType: "anthropic",
+      baseUrl: "https://gateway.example/v1",
+    });
+    expect((await agent.unstable_listProviders({})).providers[0].current).not.toBeNull();
+
+    await expect(agent.unstable_disableProvider({ providerId: "main" })).resolves.toEqual({});
+
+    expect((await agent.unstable_listProviders({})).providers[0].current).toBeNull();
+  });
+
+  it("treats disabling an unknown provider as an idempotent no-op", async () => {
+    const [agent] = await createAgentMock();
+    await expect(agent.unstable_disableProvider({ providerId: "openai" })).resolves.toEqual({});
+  });
+
+  it("routes anthropic provider config into session env", async () => {
+    const [agent, mockQuery] = await createAgentMock();
+    await agent.initialize({ protocolVersion: 1, clientCapabilities: {} });
+    await agent.unstable_setProvider({
+      providerId: "main",
+      apiType: "anthropic",
+      baseUrl: "https://gateway.example",
+      headers: { "x-api-key": "test" },
+    });
+
+    await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+
+    expect(mockQuery).toHaveBeenCalledWith(
+      expect.objectContaining({
+        options: expect.objectContaining({
+          env: expect.objectContaining({
+            ANTHROPIC_AUTH_TOKEN: " ",
+            ANTHROPIC_BASE_URL: "https://gateway.example",
+            ANTHROPIC_CUSTOM_HEADERS: "x-api-key: test",
+          }),
+        }),
+      }),
+    );
+  });
+
+  it("routes bedrock provider config into session env", async () => {
+    const [agent, mockQuery] = await createAgentMock();
+    await agent.initialize({ protocolVersion: 1, clientCapabilities: {} });
+    await agent.unstable_setProvider({
+      providerId: "main",
+      apiType: "bedrock",
+      baseUrl: "https://gateway.example",
+      headers: { "custom-header": "test" },
+    });
+
+    await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+
+    expect(mockQuery).toHaveBeenCalledWith(
+      expect.objectContaining({
+        options: expect.objectContaining({
+          env: expect.objectContaining({
+            CLAUDE_CODE_USE_BEDROCK: "1",
+            AWS_BEARER_TOKEN_BEDROCK: " ",
+            ANTHROPIC_BEDROCK_BASE_URL: "https://gateway.example",
+            ANTHROPIC_CUSTOM_HEADERS: "custom-header: test",
+          }),
+        }),
+      }),
+    );
+  });
+
+  it("accepts vertex config via _meta and routes it into session env", async () => {
+    const [agent, mockQuery] = await createAgentMock();
+    await agent.initialize({ protocolVersion: 1, clientCapabilities: {} });
+    await agent.unstable_setProvider({
+      providerId: "main",
+      apiType: "vertex",
+      baseUrl: "https://vertex.example",
+      headers: { "custom-header": "test" },
+      _meta: { claudeCode: { vertex: { projectId: "my-project", region: "us-east5" } } },
+    });
+
+    // list surfaces apiType/baseUrl but not the _meta extras
+    const listed = await agent.unstable_listProviders({});
+    expect(listed.providers[0].current).toEqual({
+      apiType: "vertex",
+      baseUrl: "https://vertex.example",
+    });
+
+    await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+
+    expect(mockQuery).toHaveBeenCalledWith(
+      expect.objectContaining({
+        options: expect.objectContaining({
+          env: expect.objectContaining({
+            CLAUDE_CODE_USE_VERTEX: "1",
+            ANTHROPIC_VERTEX_BASE_URL: "https://vertex.example",
+            ANTHROPIC_VERTEX_PROJECT_ID: "my-project",
+            CLOUD_ML_REGION: "us-east5",
+            ANTHROPIC_CUSTOM_HEADERS: "custom-header: test",
+          }),
+        }),
+      }),
+    );
+  });
+
+  it("rejects vertex set without _meta project/region", async () => {
+    const [agent] = await createAgentMock();
+    await expect(
+      agent.unstable_setProvider({
+        providerId: "main",
+        apiType: "vertex",
+        baseUrl: "https://vertex.example",
+      }),
+    ).rejects.toMatchObject({ code: -32602 });
+
+    // partial _meta (missing region) is also rejected
+    await expect(
+      agent.unstable_setProvider({
+        providerId: "main",
+        apiType: "vertex",
+        baseUrl: "https://vertex.example",
+        _meta: { claudeCode: { vertex: { projectId: "my-project" } } } as any,
+      }),
+    ).rejects.toMatchObject({ code: -32602 });
+  });
+
+  it("providers/set takes precedence over gateway auth", async () => {
+    const [agent, mockQuery] = await createAgentMock();
+    await agent.initialize({
+      protocolVersion: 1,
+      clientCapabilities: { auth: { terminal: true, _meta: { gateway: true } } } as any,
+    });
+
+    // Gateway auth first...
+    await agent.authenticate({
+      methodId: "gateway",
+      _meta: { gateway: { baseUrl: "https://gateway.example", headers: {} } },
+    });
+    // ...then providers/set wins.
+    await agent.unstable_setProvider({
+      providerId: "main",
+      apiType: "anthropic",
+      baseUrl: "https://provider.example",
+    });
+
+    await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
+
+    expect(mockQuery).toHaveBeenCalledWith(
+      expect.objectContaining({
+        options: expect.objectContaining({
+          env: expect.objectContaining({
+            ANTHROPIC_BASE_URL: "https://provider.example",
+          }),
+        }),
+      }),
+    );
+  });
+
+  it("clears provider config on logout", async () => {
+    const [agent] = await createAgentMock();
+    // Avoids resolving the native CLI binary in claudeCliPath().
+    vi.stubEnv("CLAUDE_CODE_EXECUTABLE", "/bin/true");
+
+    await agent.unstable_setProvider({
+      providerId: "main",
+      apiType: "anthropic",
+      baseUrl: "https://provider.example",
+    });
+    expect((await agent.unstable_listProviders({})).providers[0].current).not.toBeNull();
+
+    await agent.logout({});
+
+    expect((await agent.unstable_listProviders({})).providers[0].current).toBeNull();
+  });
+});

src/tests/resolve-permission-mode.test.ts

@@ -29,6 +29,13 @@ describe("resolvePermissionMode", () => {
     expect(resolvePermissionMode("bypass")).toBe("bypassPermissions");
   });
 
+  it("resolves 'manual' as an alias for 'default'", () => {
+    const { logger, error } = mockLogger();
+    expect(resolvePermissionMode("manual", logger)).toBe("default");
+    expect(resolvePermissionMode("Manual", logger)).toBe("default");
+    expect(error).not.toHaveBeenCalled();
+  });
+
   it("trims whitespace", () => {
     expect(resolvePermissionMode("  dontAsk  ")).toBe("dontAsk");
   });

src/tests/session-config-options.test.ts

@@ -1,7 +1,8 @@
 import { describe, it, expect, beforeEach, vi } from "vitest";
-import { AgentSideConnection, SessionNotification } from "@agentclientprotocol/sdk";
+import { SessionNotification } from "@agentclientprotocol/sdk";
 import type { ModelInfo } from "@anthropic-ai/claude-agent-sdk";
-import type { ClaudeAcpAgent as ClaudeAcpAgentType } from "../acp-agent.js";
+import type { AcpClient, ClaudeAcpAgent as ClaudeAcpAgentType } from "../acp-agent.js";
+import { makeMockQuery } from "./helpers.js";
 
 const { registerHookCallbackSpy } = vi.hoisted(() => ({
   registerHookCallbackSpy: vi.fn(),
@@ -30,7 +31,7 @@ const MOCK_MODELS = {
   currentModelId: "claude-opus-4-5",
   availableModels: [
     { modelId: "claude-opus-4-5", name: "Claude Opus", description: "Most capable" },
-    { modelId: "claude-sonnet-4-5", name: "Claude Sonnet", description: "Balanced" },
+    { modelId: "claude-sonnet-4-6", name: "Claude Sonnet", description: "Balanced" },
   ],
 };
 
@@ -65,8 +66,9 @@ const MOCK_CONFIG_OPTIONS = [
     description: "Available effort levels for this model",
     type: "select",
     category: "effort",
-    currentValue: "high",
+    currentValue: "default",
     options: [
+      { value: "default", name: "Default" },
       { value: "low", name: "Low" },
       { value: "medium", name: "Medium" },
       { value: "high", name: "High" },
@@ -83,7 +85,7 @@ describe("session config options", () => {
   let setModelSpy: ReturnType<typeof vi.fn>;
   let applyFlagSettingsSpy: ReturnType<typeof vi.fn>;
 
-  function createMockClient(): AgentSideConnection {
+  function createMockClient(): AcpClient {
     return {
       sessionUpdate: async (notification: SessionNotification) => {
         sessionUpdates.push(notification);
@@ -91,7 +93,7 @@ describe("session config options", () => {
       requestPermission: async () => ({ outcome: { outcome: "cancelled" } }),
       readTextFile: async () => ({ content: "" }),
       writeTextFile: async () => ({}),
-    } as unknown as AgentSideConnection;
+    } as unknown as AcpClient;
   }
 
   function populateSession() {
@@ -100,29 +102,28 @@ describe("session config options", () => {
     applyFlagSettingsSpy = vi.fn();
 
     (agent as unknown as { sessions: Record<string, unknown> }).sessions[SESSION_ID] = {
-      query: {
+      query: makeMockQuery({
         setPermissionMode: setPermissionModeSpy,
         setModel: setModelSpy,
         applyFlagSettings: applyFlagSettingsSpy,
-        supportedCommands: async () => [],
-      },
+      }),
       input: null,
       cancelled: false,
       permissionMode: "default",
       settingsManager: {},
       modes: structuredClone(MOCK_MODES),
       models: structuredClone(MOCK_MODELS),
-      modelInfos: MOCK_MODELS.availableModels.map(
-        (m): ModelInfo => ({
-          value: m.modelId,
-          displayName: m.name,
-          description: m.description,
-          supportsEffort: true,
-          supportedEffortLevels: ["low", "medium", "high"],
-        }),
-      ),
+      modelInfos: MOCK_MODELS.availableModels.map((m): ModelInfo => ({
+        value: m.modelId,
+        displayName: m.name,
+        description: m.description,
+        supportsEffort: true,
+        supportedEffortLevels: ["low", "medium", "high"],
+      })),
       configOptions: structuredClone(MOCK_CONFIG_OPTIONS),
       contextWindowSize: 200000,
+      toolUseCache: {},
+      emittedToolCalls: new Set(),
     };
   }
 
@@ -147,13 +148,13 @@ describe("session config options", () => {
 
   describe("newSession returns configOptions", () => {
     it("includes configOptions in the response", async () => {
-      const response = await agent.newSession({ cwd: "/test", mcpServers: [] });
+      const response = await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
       expect(response.configOptions).toBeDefined();
       expect(response.configOptions).toEqual(MOCK_CONFIG_OPTIONS);
     });
 
     it("includes mode and model config options", async () => {
-      const response = await agent.newSession({ cwd: "/test", mcpServers: [] });
+      const response = await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
       const modeOption = response.configOptions?.find((o) => o.id === "mode");
       const modelOption = response.configOptions?.find((o) => o.id === "model");
       expect(modeOption).toBeDefined();
@@ -172,7 +173,7 @@ describe("session config options", () => {
       (agent as unknown as { loadSession: typeof loadSessionSpy }).loadSession = loadSessionSpy;
 
       const response = await agent.loadSession({
-        cwd: "/test",
+        cwd: process.cwd(),
         sessionId: SESSION_ID,
         mcpServers: [],
       });
@@ -215,6 +216,31 @@ describe("session config options", () => {
       ).rejects.toThrow("Invalid value for config option mode: invalid-mode");
     });
 
+    it("rejects mode and config changes once the query stream has closed (husk session)", async () => {
+      // After an unexpected stream death the session lingers as a husk
+      // (queryClosed=true) so prompt() can answer with a clear error. The
+      // config/mode handlers must do the same rather than calling setModel/
+      // setPermissionMode on the closed query.
+      const session = (agent as unknown as { sessions: Record<string, { queryClosed?: boolean }> })
+        .sessions[SESSION_ID];
+      session.queryClosed = true;
+
+      await expect(
+        agent.setSessionConfigOption({
+          sessionId: SESSION_ID,
+          configId: "model",
+          value: "claude-sonnet-4-6",
+        }),
+      ).rejects.toThrow(/start a new session/);
+      await expect(agent.setSessionMode({ sessionId: SESSION_ID, modeId: "plan" })).rejects.toThrow(
+        /start a new session/,
+      );
+
+      // Short-circuited before touching the (closed) query.
+      expect(setModelSpy).not.toHaveBeenCalled();
+      expect(setPermissionModeSpy).not.toHaveBeenCalled();
+    });
+
     it("changes mode, sends current_mode_update but not config_option_update", async () => {
       await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
@@ -242,10 +268,10 @@ describe("session config options", () => {
       await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
         configId: "model",
-        value: "claude-sonnet-4-5",
+        value: "claude-sonnet-4-6",
       });
 
-      expect(setModelSpy).toHaveBeenCalledWith("claude-sonnet-4-5");
+      expect(setModelSpy).toHaveBeenCalledWith("claude-sonnet-4-6");
 
       const configUpdate = sessionUpdates.find(
         (n) => n.update.sessionUpdate === "config_option_update",
@@ -273,7 +299,7 @@ describe("session config options", () => {
         value: "sonnet",
       });
 
-      expect(setModelSpy).toHaveBeenCalledWith("claude-sonnet-4-5");
+      expect(setModelSpy).toHaveBeenCalledWith("claude-sonnet-4-6");
     });
 
     it("resolves display name to model ID", async () => {
@@ -283,19 +309,19 @@ describe("session config options", () => {
         value: "Claude Sonnet",
       });
 
-      expect(setModelSpy).toHaveBeenCalledWith("claude-sonnet-4-5");
+      expect(setModelSpy).toHaveBeenCalledWith("claude-sonnet-4-6");
     });
 
     it("still works with exact model ID", async () => {
       const response = await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
         configId: "model",
-        value: "claude-sonnet-4-5",
+        value: "claude-sonnet-4-6",
       });
 
-      expect(setModelSpy).toHaveBeenCalledWith("claude-sonnet-4-5");
+      expect(setModelSpy).toHaveBeenCalledWith("claude-sonnet-4-6");
       const modelOption = response.configOptions.find((o) => o.id === "model");
-      expect(modelOption?.currentValue).toBe("claude-sonnet-4-5");
+      expect(modelOption?.currentValue).toBe("claude-sonnet-4-6");
     });
 
     it("throws for completely invalid model value", async () => {
@@ -376,33 +402,30 @@ describe("session config options", () => {
     });
   });
 
-  describe("unstable_setSessionModel sends config_option_update", () => {
+  describe("setSessionConfigOption(model) returns updated configOptions", () => {
     beforeEach(() => {
       populateSession();
     });
 
-    it("sends config_option_update when model is changed via setSessionModel", async () => {
-      await agent.unstable_setSessionModel({
+    it("returns configOptions with the new model when changed", async () => {
+      const response = await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
-        modelId: "claude-sonnet-4-5",
+        configId: "model",
+        value: "claude-sonnet-4-6",
       });
 
-      const configUpdate = sessionUpdates.find(
-        (n) => n.update.sessionUpdate === "config_option_update",
-      );
-      expect(configUpdate).toBeDefined();
-      expect(configUpdate?.update).toMatchObject({
-        sessionUpdate: "config_option_update",
-        configOptions: expect.arrayContaining([
-          expect.objectContaining({ id: "model", currentValue: "claude-sonnet-4-5" }),
+      expect(response.configOptions).toEqual(
+        expect.arrayContaining([
+          expect.objectContaining({ id: "model", currentValue: "claude-sonnet-4-6" }),
         ]),
-      });
+      );
     });
 
     it("updates stored configOptions currentValue when model changes", async () => {
-      await agent.unstable_setSessionModel({
+      await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
-        modelId: "claude-sonnet-4-5",
+        configId: "model",
+        value: "claude-sonnet-4-6",
       });
 
       const session = (
@@ -411,10 +434,10 @@ describe("session config options", () => {
         }
       ).sessions[SESSION_ID];
       const modelOption = session.configOptions.find((o) => o.id === "model");
-      expect(modelOption?.currentValue).toBe("claude-sonnet-4-5");
+      expect(modelOption?.currentValue).toBe("claude-sonnet-4-6");
     });
 
-    it("includes updated effort in config_option_update when model drops effort support", async () => {
+    it("drops effort from returned configOptions when model drops effort support", async () => {
       const session = (agent as unknown as { sessions: Record<string, any> }).sessions[SESSION_ID];
       session.modelInfos = [
         {
@@ -425,30 +448,25 @@ describe("session config options", () => {
           supportedEffortLevels: ["low", "medium", "high"],
         },
         {
-          value: "claude-sonnet-4-5",
+          value: "claude-sonnet-4-6",
           displayName: "Claude Sonnet",
           description: "Balanced",
           supportsEffort: false,
         },
       ];
 
-      await agent.unstable_setSessionModel({
+      const response = await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
-        modelId: "claude-sonnet-4-5",
+        configId: "model",
+        value: "claude-sonnet-4-6",
       });
 
-      const configUpdate = sessionUpdates.find(
-        (n) => n.update.sessionUpdate === "config_option_update",
-      );
-      expect(configUpdate).toBeDefined();
-      const effortOption = (configUpdate?.update as any).configOptions.find(
-        (o: any) => o.id === "effort",
-      );
+      const effortOption = response.configOptions.find((o) => o.id === "effort");
       expect(effortOption).toBeUndefined();
-      expect(applyFlagSettingsSpy).toHaveBeenCalledWith({ effortLevel: undefined });
+      expect(applyFlagSettingsSpy).toHaveBeenCalledWith({ effortLevel: null });
     });
 
-    it("clamps effort in config_option_update when new model has different supported levels", async () => {
+    it("clamps effort in returned configOptions when new model has different supported levels", async () => {
       // Set current effort to "max" which the new model won't support
       const session = (agent as unknown as { sessions: Record<string, any> }).sessions[SESSION_ID];
       const effortOpt = session.configOptions.find((o: any) => o.id === "effort");
@@ -463,7 +481,7 @@ describe("session config options", () => {
           supportedEffortLevels: ["low", "medium", "high", "max"],
         },
         {
-          value: "claude-sonnet-4-5",
+          value: "claude-sonnet-4-6",
           displayName: "Claude Sonnet",
           description: "Balanced",
           supportsEffort: true,
@@ -471,39 +489,31 @@ describe("session config options", () => {
         },
       ];
 
-      await agent.unstable_setSessionModel({
+      const response = await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
-        modelId: "claude-sonnet-4-5",
+        configId: "model",
+        value: "claude-sonnet-4-6",
       });
 
-      const configUpdate = sessionUpdates.find(
-        (n) => n.update.sessionUpdate === "config_option_update",
-      );
-      const effortOption = (configUpdate?.update as any).configOptions.find(
-        (o: any) => o.id === "effort",
-      );
+      const effortOption = response.configOptions.find((o) => o.id === "effort");
       expect(effortOption).toBeDefined();
-      expect(effortOption.currentValue).toBe("high");
-      expect(applyFlagSettingsSpy).toHaveBeenCalledWith({ effortLevel: "high" });
+      expect(effortOption?.currentValue).toBe("default");
+      expect(applyFlagSettingsSpy).toHaveBeenCalledWith({ effortLevel: null });
     });
 
-    it("preserves effort in config_option_update when new model supports same level", async () => {
+    it("preserves effort in returned configOptions when new model supports same level", async () => {
       // Set effort to "low" first
       const session = (agent as unknown as { sessions: Record<string, any> }).sessions[SESSION_ID];
       const effortOpt = session.configOptions.find((o: any) => o.id === "effort");
       if (effortOpt) effortOpt.currentValue = "low";
 
-      await agent.unstable_setSessionModel({
+      const response = await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
-        modelId: "claude-sonnet-4-5",
+        configId: "model",
+        value: "claude-sonnet-4-6",
       });
 
-      const configUpdate = sessionUpdates.find(
-        (n) => n.update.sessionUpdate === "config_option_update",
-      );
-      const effortOption = (configUpdate?.update as any).configOptions.find(
-        (o: any) => o.id === "effort",
-      );
+      const effortOption = response.configOptions.find((o) => o.id === "effort");
       expect(effortOption?.currentValue).toBe("low");
       // Effort didn't change, so applyFlagSettings should NOT be called
       expect(applyFlagSettingsSpy).not.toHaveBeenCalled();
@@ -532,7 +542,7 @@ describe("session config options", () => {
       await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
         configId: "model",
-        value: "claude-sonnet-4-5",
+        value: "claude-sonnet-4-6",
       });
 
       const configUpdates = sessionUpdates.filter(
@@ -557,6 +567,31 @@ describe("session config options", () => {
       expect(applyFlagSettingsSpy).toHaveBeenCalledWith({ effortLevel: "low" });
     });
 
+    it("calls applyFlagSettings with null effortLevel for 'default'", async () => {
+      // Set effort to a non-default value first
+      const session = (agent as unknown as { sessions: Record<string, any> }).sessions[SESSION_ID];
+      const effortOpt = session.configOptions.find((o: any) => o.id === "effort");
+      if (effortOpt) effortOpt.currentValue = "high";
+
+      await agent.setSessionConfigOption({
+        sessionId: SESSION_ID,
+        configId: "effort",
+        value: "default",
+      });
+
+      expect(applyFlagSettingsSpy).toHaveBeenCalledWith({ effortLevel: null });
+
+      // The SDK's applyFlagSettings travels over a JSON pipe and only clears a
+      // flag-layer key when an explicit `null` is sent — `undefined` is
+      // dropped during JSON.stringify, which would leave the previous effort
+      // override in place. Round-trip the call args through JSON to make sure
+      // the key actually reaches the SDK.
+      const calls = applyFlagSettingsSpy.mock.calls;
+      const lastCallArgs = calls[calls.length - 1]?.[0];
+      const serialized = JSON.parse(JSON.stringify(lastCallArgs));
+      expect(serialized).toHaveProperty("effortLevel", null);
+    });
+
     it("updates effort currentValue in returned configOptions", async () => {
       const response = await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
@@ -622,7 +657,7 @@ describe("session config options", () => {
           supportedEffortLevels: ["low", "medium", "high"],
         },
         {
-          value: "claude-sonnet-4-5",
+          value: "claude-sonnet-4-6",
           displayName: "Claude Sonnet",
           description: "Balanced",
           supportsEffort: false,
@@ -632,7 +667,7 @@ describe("session config options", () => {
       const response = await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
         configId: "model",
-        value: "claude-sonnet-4-5",
+        value: "claude-sonnet-4-6",
       });
 
       const effortOption = response.configOptions.find((o) => o.id === "effort");
@@ -650,7 +685,7 @@ describe("session config options", () => {
           supportedEffortLevels: ["low", "medium", "high"],
         },
         {
-          value: "claude-sonnet-4-5",
+          value: "claude-sonnet-4-6",
           displayName: "Claude Sonnet",
           description: "Balanced",
           supportsEffort: false,
@@ -660,16 +695,16 @@ describe("session config options", () => {
       await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
         configId: "model",
-        value: "claude-sonnet-4-5",
+        value: "claude-sonnet-4-6",
       });
 
-      expect(applyFlagSettingsSpy).toHaveBeenCalledWith({ effortLevel: undefined });
+      expect(applyFlagSettingsSpy).toHaveBeenCalledWith({ effortLevel: null });
     });
 
     it("adds effort option when switching to a model that supports effort", async () => {
       const session = (agent as unknown as { sessions: Record<string, any> }).sessions[SESSION_ID];
       // Start with sonnet (no effort) as current
-      session.models = { ...session.models, currentModelId: "claude-sonnet-4-5" };
+      session.models = { ...session.models, currentModelId: "claude-sonnet-4-6" };
       session.modelInfos = [
         {
           value: "claude-opus-4-5",
@@ -679,7 +714,7 @@ describe("session config options", () => {
           supportedEffortLevels: ["low", "medium", "high"],
         },
         {
-          value: "claude-sonnet-4-5",
+          value: "claude-sonnet-4-6",
           displayName: "Claude Sonnet",
           description: "Balanced",
           supportsEffort: false,
@@ -696,8 +731,8 @@ describe("session config options", () => {
 
       const effortOption = response.configOptions.find((o) => o.id === "effort");
       expect(effortOption).toBeDefined();
-      // No previous effort, so defaults to "high" (the API default)
-      expect(effortOption?.currentValue).toBe("high");
+      // No previous effort, so defaults to "default" (no effort override)
+      expect(effortOption?.currentValue).toBe("default");
     });
 
     it("clamps effort to valid value when new model has different supported levels", async () => {
@@ -715,7 +750,7 @@ describe("session config options", () => {
           supportedEffortLevels: ["low", "medium", "high", "max"],
         },
         {
-          value: "claude-sonnet-4-5",
+          value: "claude-sonnet-4-6",
           displayName: "Claude Sonnet",
           description: "Balanced",
           supportsEffort: true,
@@ -726,15 +761,15 @@ describe("session config options", () => {
       const response = await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
         configId: "model",
-        value: "claude-sonnet-4-5",
+        value: "claude-sonnet-4-6",
       });
 
       const effortOption = response.configOptions.find((o) => o.id === "effort");
       expect(effortOption).toBeDefined();
-      // "max" is not in sonnet's levels, so should fall back to "high" (the API default)
-      expect(effortOption?.currentValue).toBe("high");
-      // SDK should be told about the clamped value
-      expect(applyFlagSettingsSpy).toHaveBeenCalledWith({ effortLevel: "high" });
+      // "max" is not in sonnet's levels, so should fall back to "default" (no effort override)
+      expect(effortOption?.currentValue).toBe("default");
+      // SDK should be told to clear the effort override
+      expect(applyFlagSettingsSpy).toHaveBeenCalledWith({ effortLevel: null });
     });
 
     it("preserves effort value when new model supports the same level", async () => {
@@ -749,7 +784,7 @@ describe("session config options", () => {
       const response = await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
         configId: "model",
-        value: "claude-sonnet-4-5",
+        value: "claude-sonnet-4-6",
       });
 
       const effortOption = response.configOptions.find((o) => o.id === "effort");
@@ -778,10 +813,89 @@ describe("session config options", () => {
       await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
         configId: "model",
-        value: "claude-sonnet-4-5",
+        value: "claude-sonnet-4-6",
       });
 
-      expect(setModelSpy).toHaveBeenCalledWith("claude-sonnet-4-5");
+      expect(setModelSpy).toHaveBeenCalledWith("claude-sonnet-4-6");
+    });
+
+    // Option entries carry no resolvedModel, so alias resolution must consult
+    // session.modelInfos — otherwise a full model id in either hint spelling
+    // ("[1m]"/"-1m") falls to the substring tier and lands on the bare 200k
+    // sibling, silently downgrading the session's context lane.
+    it("resolves a full model id onto its hinted row via session.modelInfos", async () => {
+      const session = (agent as unknown as { sessions: Record<string, any> }).sessions[SESSION_ID];
+      session.models = {
+        currentModelId: "sonnet",
+        availableModels: [
+          { modelId: "sonnet", name: "Sonnet", description: "" },
+          { modelId: "sonnet[1m]", name: "Sonnet", description: "" },
+        ],
+      };
+      session.modelInfos = [
+        {
+          value: "sonnet",
+          resolvedModel: "claude-sonnet-5",
+          displayName: "Sonnet",
+          description: "",
+        },
+        {
+          value: "sonnet[1m]",
+          resolvedModel: "claude-sonnet-5[1m]",
+          displayName: "Sonnet",
+          description: "",
+        },
+      ];
+      session.configOptions = session.configOptions.map((o: { id: string }) =>
+        o.id === "model"
+          ? {
+              ...o,
+              currentValue: "sonnet",
+              options: [
+                { value: "sonnet", name: "Sonnet" },
+                { value: "sonnet[1m]", name: "Sonnet" },
+              ],
+            }
+          : o,
+      );
+
+      await agent.setSessionConfigOption({
+        sessionId: SESSION_ID,
+        configId: "model",
+        value: "claude-sonnet-5[1m]",
+      });
+      expect(setModelSpy).toHaveBeenCalledWith("sonnet[1m]");
+
+      setModelSpy.mockClear();
+      await agent.setSessionConfigOption({
+        sessionId: SESSION_ID,
+        configId: "model",
+        value: "claude-sonnet-5-1m",
+      });
+      expect(setModelSpy).toHaveBeenCalledWith("sonnet[1m]");
+    });
+
+    // A session can be running a model with no picker entry (resumed onto a
+    // model excluded by the availableModels allowlist, or a refusal
+    // fallback); its verbatim id is then the option's currentValue. A client
+    // round-tripping that reported value must not get "Invalid value".
+    it("accepts the reported currentValue even when it has no options entry", async () => {
+      const session = (agent as unknown as { sessions: Record<string, any> }).sessions[SESSION_ID];
+      session.models = { ...session.models, currentModelId: "claude-offlist-9" };
+      session.configOptions = session.configOptions.map((o: { id: string }) =>
+        o.id === "model" ? { ...o, currentValue: "claude-offlist-9" } : o,
+      );
+
+      const response = await agent.setSessionConfigOption({
+        sessionId: SESSION_ID,
+        configId: "model",
+        value: "claude-offlist-9",
+      });
+
+      expect(setModelSpy).toHaveBeenCalledWith("claude-offlist-9");
+      expect(response.configOptions?.find((o) => o.id === "model")?.currentValue).toBe(
+        "claude-offlist-9",
+      );
     });
 
     it("setSessionMode also syncs configOptions", async () => {
@@ -795,10 +909,11 @@ describe("session config options", () => {
       expect(session.configOptions.find((o) => o.id === "mode")?.currentValue).toBe("plan");
     });
 
-    it("setSessionModel also syncs configOptions", async () => {
-      await agent.unstable_setSessionModel({
+    it("setSessionConfigOption(model) also syncs configOptions", async () => {
+      await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
-        modelId: "claude-sonnet-4-5",
+        configId: "model",
+        value: "claude-sonnet-4-6",
       });
 
       const session = (
@@ -807,11 +922,89 @@ describe("session config options", () => {
         }
       ).sessions[SESSION_ID];
       expect(session.configOptions.find((o) => o.id === "model")?.currentValue).toBe(
-        "claude-sonnet-4-5",
+        "claude-sonnet-4-6",
       );
     });
   });
 
+  describe("context window on model change", () => {
+    beforeEach(() => {
+      populateSession();
+    });
+
+    function getSession() {
+      return (agent as unknown as { sessions: Record<string, any> }).sessions[SESSION_ID];
+    }
+
+    it("sets the window from text inference on model switch, without any getContextUsage IPC", async () => {
+      // getContextUsage stalls until the session's first prompt turn, so the
+      // switch path must never call it; the window is seeded from the text
+      // heuristic (here via the new model's resolvedModel) and later confirmed
+      // by result.modelUsage.
+      const session = getSession();
+      session.query.getContextUsage = vi.fn(async () => ({ rawMaxTokens: 967000 }));
+      session.modelInfos = session.modelInfos.map((m: ModelInfo) =>
+        m.value === "claude-sonnet-4-6" ? { ...m, resolvedModel: "claude-sonnet-5[1m]" } : m,
+      );
+
+      await agent.setSessionConfigOption({
+        sessionId: SESSION_ID,
+        configId: "model",
+        value: "claude-sonnet-4-6",
+      });
+
+      expect(session.query.getContextUsage).not.toHaveBeenCalled();
+      expect(session.contextWindowSize).toBe(1_000_000);
+    });
+
+    it("falls back to the default window when inference misses, without any getContextUsage IPC", async () => {
+      const session = getSession();
+      session.contextWindowSize = 1_000_000;
+      // Present but must NOT be called; the switch never consults it.
+      session.query.getContextUsage = vi.fn(async () => ({ rawMaxTokens: 967000 }));
+      // claude-sonnet-4-6 carries no "1m" token in its id, resolvedModel,
+      // displayName, or description, so inference misses → default window.
+
+      await agent.setSessionConfigOption({
+        sessionId: SESSION_ID,
+        configId: "model",
+        value: "claude-sonnet-4-6",
+      });
+
+      expect(session.query.getContextUsage).not.toHaveBeenCalled();
+      expect(session.contextWindowSize).toBe(200000);
+    });
+
+    it("does not call getContextUsage even when switching to a fresh model", async () => {
+      const session = getSession();
+      const spy = vi.fn(async () => ({ rawMaxTokens: 967000 }));
+      session.query.getContextUsage = spy;
+
+      await agent.setSessionConfigOption({
+        sessionId: SESSION_ID,
+        configId: "model",
+        value: "claude-sonnet-4-6",
+      });
+
+      expect(spy).not.toHaveBeenCalled();
+    });
+
+    it("keeps the learned window when re-asserting the current model", async () => {
+      const session = getSession();
+      session.contextWindowSize = 1_000_000;
+      session.query.getContextUsage = vi.fn(async () => ({ rawMaxTokens: 200000 }));
+
+      await agent.setSessionConfigOption({
+        sessionId: SESSION_ID,
+        configId: "model",
+        value: "claude-opus-4-5",
+      });
+
+      expect(session.query.getContextUsage).not.toHaveBeenCalled();
+      expect(session.contextWindowSize).toBe(1_000_000);
+    });
+  });
+
   describe("auto mode availability per model", () => {
     /**
      * Augment the session populated by `populateSession()` with a Haiku entry
@@ -832,7 +1025,7 @@ describe("session config options", () => {
           supportsAutoMode: true,
         },
         {
-          value: "claude-sonnet-4-5",
+          value: "claude-sonnet-4-6",
           displayName: "Claude Sonnet",
           description: "Balanced",
           supportsEffort: true,
@@ -852,7 +1045,7 @@ describe("session config options", () => {
         currentModelId: "claude-opus-4-5",
         availableModels: [
           { modelId: "claude-opus-4-5", name: "Claude Opus", description: "Most capable" },
-          { modelId: "claude-sonnet-4-5", name: "Claude Sonnet", description: "Balanced" },
+          { modelId: "claude-sonnet-4-6", name: "Claude Sonnet", description: "Balanced" },
           { modelId: "claude-haiku-4-5", name: "Claude Haiku", description: "Fast" },
         ],
       };
@@ -921,58 +1114,21 @@ describe("session config options", () => {
     it("drops `auto` from available modes when switching to Haiku", async () => {
       setupHaikuOpusSession("default");
 
-      await agent.unstable_setSessionModel({
+      const response = await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
-        modelId: "claude-haiku-4-5",
+        configId: "model",
+        value: "claude-haiku-4-5",
       });
 
-      const configUpdate = sessionUpdates.find(
-        (n) => n.update.sessionUpdate === "config_option_update",
-      );
-      expect(configUpdate).toBeDefined();
-      const modeOption = (configUpdate?.update as any).configOptions.find(
-        (o: any) => o.id === "mode",
-      );
+      const modeOption = response.configOptions.find((o) => o.id === "mode");
       expect(modeOption).toBeDefined();
-      const modeValues = modeOption.options.map((o: any) => o.value);
+      const modeValues = (modeOption as any).options.map((o: any) => o.value);
       expect(modeValues).not.toContain("auto");
       expect(modeValues).toEqual(
         expect.arrayContaining(["default", "acceptEdits", "plan", "dontAsk"]),
       );
     });
 
-    it("clamps to `default` and emits current_mode_update when Opus(auto) → Haiku", async () => {
-      setupHaikuOpusSession("auto");
-
-      await agent.unstable_setSessionModel({
-        sessionId: SESSION_ID,
-        modelId: "claude-haiku-4-5",
-      });
-
-      // SDK was synced to "default".
-      expect(setPermissionModeSpy).toHaveBeenCalledWith("default");
-
-      // current_mode_update was emitted before config_option_update so a
-      // client applying notifications in order observes the mode change
-      // before re-rendering the config-option list.
-      const modeUpdateIdx = sessionUpdates.findIndex(
-        (n) => n.update.sessionUpdate === "current_mode_update",
-      );
-      const configUpdateIdx = sessionUpdates.findIndex(
-        (n) => n.update.sessionUpdate === "config_option_update",
-      );
-      expect(modeUpdateIdx).toBeGreaterThanOrEqual(0);
-      expect(configUpdateIdx).toBeGreaterThanOrEqual(0);
-      expect(modeUpdateIdx).toBeLessThan(configUpdateIdx);
-      expect((sessionUpdates[modeUpdateIdx].update as any).currentModeId).toBe("default");
-
-      // configOptions reflect the clamped mode.
-      const modeOption = (sessionUpdates[configUpdateIdx].update as any).configOptions.find(
-        (o: any) => o.id === "mode",
-      );
-      expect(modeOption.currentValue).toBe("default");
-    });
-
     it("re-adds `auto` when switching from Haiku back to Opus", async () => {
       const session = setupHaikuOpusSession("default");
       // Pretend Haiku is the current model with no `auto`.
@@ -987,19 +1143,15 @@ describe("session config options", () => {
         description: m.description,
       }));
 
-      await agent.unstable_setSessionModel({
+      const response = await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
-        modelId: "claude-opus-4-5",
+        configId: "model",
+        value: "claude-opus-4-5",
       });
 
-      const configUpdate = sessionUpdates.find(
-        (n) => n.update.sessionUpdate === "config_option_update",
-      );
-      expect(configUpdate).toBeDefined();
-      const modeOption = (configUpdate?.update as any).configOptions.find(
-        (o: any) => o.id === "mode",
-      );
-      const modeValues = modeOption.options.map((o: any) => o.value);
+      const modeOption = response.configOptions.find((o) => o.id === "mode");
+      expect(modeOption).toBeDefined();
+      const modeValues = (modeOption as any).options.map((o: any) => o.value);
       expect(modeValues).toContain("auto");
 
       // The current mode ("default") is still valid on Opus, so no
@@ -1013,9 +1165,10 @@ describe("session config options", () => {
     it("preserves the current mode when it remains valid after a model switch", async () => {
       setupHaikuOpusSession("plan");
 
-      await agent.unstable_setSessionModel({
+      const response = await agent.setSessionConfigOption({
         sessionId: SESSION_ID,
-        modelId: "claude-haiku-4-5",
+        configId: "model",
+        value: "claude-haiku-4-5",
       });
 
       // `plan` is in availableModes for both Opus and Haiku, so no clamp.
@@ -1026,20 +1179,15 @@ describe("session config options", () => {
       );
       expect(modeUpdates).toHaveLength(0);
 
-      const configUpdate = sessionUpdates.find(
-        (n) => n.update.sessionUpdate === "config_option_update",
-      );
-      const modeOption = (configUpdate?.update as any).configOptions.find(
-        (o: any) => o.id === "mode",
-      );
-      expect(modeOption.currentValue).toBe("plan");
+      const modeOption = response.configOptions.find((o) => o.id === "mode");
+      expect((modeOption as any).currentValue).toBe("plan");
     });
 
     it("clamps mode and emits current_mode_update via setSessionConfigOption(model)", async () => {
-      // Mirrors the unstable_setSessionModel(auto → Haiku) test, but goes
-      // through the request/response API. The `current_mode_update` side
-      // effect must still fire so clients learn about the clamp regardless of
-      // which entry point triggered the model switch.
+      // Switching Opus(auto) → Haiku clamps the mode to "default". The
+      // `current_mode_update` side effect must fire so clients learn about the
+      // clamp even though the request/response API returns the new
+      // configOptions rather than emitting a config_option_update.
       setupHaikuOpusSession("auto");
 
       const response = await agent.setSessionConfigOption({
@@ -1122,6 +1270,10 @@ describe("session config options", () => {
         ],
       };
 
+      // The tool_call was already surfaced (by the streamed tool_use chunk), so
+      // the permission request won't re-emit one — keep this focused on options.
+      session.emittedToolCalls.add("toolu_1");
+
       const canUseTool = (agent as any).canUseTool(SESSION_ID);
       const signal = new AbortController().signal;
       try {
@@ -1153,6 +1305,10 @@ describe("session config options", () => {
         ],
       };
       permissionResponse = { outcome: { outcome: "selected", optionId: "auto" } };
+      // The tool_call was already surfaced (by the streamed tool_use chunk), so
+      // the permission request won't re-emit one — the deny path below should
+      // produce no session updates at all.
+      session.emittedToolCalls.add("toolu_2");
 
       const canUseTool = (agent as any).canUseTool(SESSION_ID);
       const result = await canUseTool(
@@ -1181,6 +1337,10 @@ describe("session config options", () => {
         ],
       };
 
+      // The tool_call was already surfaced (by the streamed tool_use chunk), so
+      // the permission request won't re-emit one — keep this focused on options.
+      session.emittedToolCalls.add("toolu_3");
+
       const canUseTool = (agent as any).canUseTool(SESSION_ID);
       const signal = new AbortController().signal;
       try {

src/tests/session-load.test.ts

@@ -1,18 +1,20 @@
 import { describe, it, expect } from "vitest";
 import { randomUUID } from "crypto";
-import { AgentSideConnection, RequestError, SessionNotification } from "@agentclientprotocol/sdk";
+import { RequestError, SessionNotification } from "@agentclientprotocol/sdk";
 import { query, getSessionMessages } from "@anthropic-ai/claude-agent-sdk";
-import { ClaudeAcpAgent } from "../acp-agent.js";
+import { AcpClient, ClaudeAcpAgent } from "../acp-agent.js";
 import { Pushable } from "../utils.js";
 import type { SDKUserMessage } from "@anthropic-ai/claude-agent-sdk";
 
-function createMockClient(): AgentSideConnection {
+function createMockClient(
+  sessionUpdate: (notification: SessionNotification) => Promise<void> = async () => {},
+): AcpClient {
   return {
-    sessionUpdate: async (_notification: SessionNotification) => {},
+    sessionUpdate,
     requestPermission: async () => ({ outcome: { outcome: "cancelled" } }),
     readTextFile: async () => ({ content: "" }),
     writeTextFile: async () => ({}),
-  } as unknown as AgentSideConnection;
+  } as unknown as AcpClient;
 }
 
 describe.skipIf(!process.env.RUN_INTEGRATION_TESTS)("session load/resume lifecycle", () => {
@@ -165,7 +167,7 @@ describe.skipIf(!process.env.RUN_INTEGRATION_TESTS)("session load/resume lifecyc
       requestPermission: async () => ({ outcome: { outcome: "cancelled" } }),
       readTextFile: async () => ({ content: "" }),
       writeTextFile: async () => ({}),
-    } as unknown as AgentSideConnection;
+    } as unknown as AcpClient;
 
     const commandName =
       "<command-name>/model</command-name>\n            <command-message>model</command-message>\n            <command-args>opus</command-args>";
@@ -274,7 +276,7 @@ describe.skipIf(!process.env.RUN_INTEGRATION_TESTS)("session load/resume lifecyc
       requestPermission: async () => ({ outcome: { outcome: "cancelled" } }),
       readTextFile: async () => ({ content: "" }),
       writeTextFile: async () => ({}),
-    } as unknown as AgentSideConnection;
+    } as unknown as AcpClient;
 
     // Exact shape from https://github.com/zed-industries/claude-code-acp/issues/579.
     const concatenated =
@@ -330,4 +332,52 @@ describe.skipIf(!process.env.RUN_INTEGRATION_TESTS)("session load/resume lifecyc
     }
     expect(recordedUserChunks.some((c) => c.includes("hi"))).toBe(true);
   }, 60000);
+
+  // Regression test for https://github.com/agentclientprotocol/claude-agent-acp/issues/845
+  // On resume the CLI restores the model the session's transcript was running,
+  // so loadSession must report that live model instead of recomputing the
+  // env/settings/default choice from scratch. Haiku is used because no account
+  // resolves its default there, so the restored model provably differs from
+  // the freshly-computed default. Requires no ANTHROPIC_MODEL env var and no
+  // "model" key in settings.json (either would pin the model on resume and
+  // mask the reconciliation this test covers).
+  it("ACP: loadSession reports the model the resumed session is actually running", async () => {
+    const agentChunks: string[] = [];
+    const client = createMockClient(async (notification) => {
+      if (notification.update.sessionUpdate === "agent_message_chunk") {
+        const content = notification.update.content;
+        if (content.type === "text") agentChunks.push(content.text);
+      }
+    });
+
+    const modelOf = (opts: { id: string; currentValue?: unknown }[] | null | undefined) =>
+      opts?.find((o) => o.id === "model")?.currentValue;
+
+    const agent = new ClaudeAcpAgent(client);
+    try {
+      const cwd = process.cwd();
+      const { sessionId, configOptions } = await agent.newSession({ cwd, mcpServers: [] });
+      expect(modelOf(configOptions)).not.toBe("haiku");
+
+      await agent.setSessionConfigOption({ sessionId, configId: "model", value: "haiku" });
+      // A real turn on Haiku, so the transcript's last assistant message —
+      // which is what the CLI restores the model from — records it.
+      await agent.prompt({
+        sessionId,
+        prompt: [{ type: "text", text: "Reply with only the word OK" }],
+      });
+      await agent.closeSession({ sessionId });
+
+      const loadResp = await agent.loadSession({ sessionId, cwd, mcpServers: [] });
+      expect(modelOf(loadResp.configOptions)).toBe("haiku");
+
+      // And the reported option agrees with the live session (`/context`).
+      agentChunks.length = 0;
+      await agent.prompt({ sessionId, prompt: [{ type: "text", text: "/context" }] });
+      const liveModel = agentChunks.join("\n").match(/\*\*Model:\*\*\s*(\S+)/)?.[1];
+      expect(liveModel).toContain("haiku");
+    } finally {
+      await agent.dispose();
+    }
+  }, 120000);
 });

src/tests/settings.test.ts

@@ -7,13 +7,26 @@ import * as os from "node:os";
 describe("SettingsManager", () => {
   let tempDir: string;
   let settingsManager: SettingsManager;
+  let originalConfigDir: string | undefined;
 
   beforeEach(async () => {
     tempDir = await fs.promises.mkdtemp(path.join(os.tmpdir(), "settings-test-"));
+    // Isolate the user tier. resolveSettings reads the user settings.json from
+    // CLAUDE_CONFIG_DIR at call time; without this it falls back to the real
+    // ~/.claude and the developer's own settings (e.g. availableModels) leak
+    // into the merge, making union-based assertions fail on their machine.
+    originalConfigDir = process.env.CLAUDE_CONFIG_DIR;
+    process.env.CLAUDE_CONFIG_DIR = path.join(tempDir, "user-config");
+    await fs.promises.mkdir(process.env.CLAUDE_CONFIG_DIR, { recursive: true });
   });
 
   afterEach(async () => {
     settingsManager?.dispose();
+    if (originalConfigDir === undefined) {
+      delete process.env.CLAUDE_CONFIG_DIR;
+    } else {
+      process.env.CLAUDE_CONFIG_DIR = originalConfigDir;
+    }
     await fs.promises.rm(tempDir, { recursive: true, force: true });
   });
 
@@ -53,6 +66,57 @@ describe("SettingsManager", () => {
       expect(settings.model).toBe("claude-3-5-haiku");
     });
 
+    it("should expose availableModels from settings", async () => {
+      const claudeDir = path.join(tempDir, ".claude");
+      await fs.promises.mkdir(claudeDir, { recursive: true });
+
+      await fs.promises.writeFile(
+        path.join(claudeDir, "settings.json"),
+        JSON.stringify({
+          availableModels: ["claude-haiku-4-5", "claude-opus-4-7[1m]"],
+        }),
+      );
+
+      settingsManager = new SettingsManager(tempDir);
+      await settingsManager.initialize();
+
+      const settings = settingsManager.getSettings();
+      expect(settings.availableModels).toEqual(["claude-haiku-4-5", "claude-opus-4-7[1m]"]);
+    });
+
+    it("should union and dedupe availableModels across sources", async () => {
+      // Per Claude Code docs: "When `availableModels` is set at multiple
+      // levels, such as user settings and project settings, arrays are
+      // merged and deduplicated."
+      // https://code.claude.com/docs/en/model-config#merge-behavior
+      const claudeDir = path.join(tempDir, ".claude");
+      await fs.promises.mkdir(claudeDir, { recursive: true });
+
+      await fs.promises.writeFile(
+        path.join(claudeDir, "settings.json"),
+        JSON.stringify({
+          availableModels: ["claude-haiku-4-5", "claude-opus-4-7[1m]"],
+        }),
+      );
+      await fs.promises.writeFile(
+        path.join(claudeDir, "settings.local.json"),
+        JSON.stringify({
+          // claude-opus-4-7[1m] overlaps with project; should be deduped.
+          availableModels: ["claude-opus-4-7[1m]", "claude-sonnet-4-6[1m]"],
+        }),
+      );
+
+      settingsManager = new SettingsManager(tempDir);
+      await settingsManager.initialize();
+
+      const settings = settingsManager.getSettings();
+      expect(settings.availableModels).toEqual([
+        "claude-haiku-4-5",
+        "claude-opus-4-7[1m]",
+        "claude-sonnet-4-6[1m]",
+      ]);
+    });
+
     it("should merge permissions.defaultMode with later sources taking precedence", async () => {
       const claudeDir = path.join(tempDir, ".claude");
       await fs.promises.mkdir(claudeDir, { recursive: true });
@@ -90,4 +154,66 @@ describe("SettingsManager", () => {
       expect(settings.permissions?.defaultMode).toBe("plan");
     });
   });
+
+  describe("escalating defaultMode trust filter", () => {
+    // The SDK's filterEscalatingDefaultMode drops escalating values
+    // (bypassPermissions / auto / acceptEdits) that came from a repo-committed
+    // tier (.claude/settings.json), preventing a checked-in file from
+    // silently escalating permissions. Local (.claude/settings.local.json)
+    // and user-tier sources are not committed by convention, so escalating
+    // values from those tiers are preserved.
+
+    it.each(["bypassPermissions", "auto", "acceptEdits"] as const)(
+      "drops %s when set in project-tier settings",
+      async (mode) => {
+        const claudeDir = path.join(tempDir, ".claude");
+        await fs.promises.mkdir(claudeDir, { recursive: true });
+
+        await fs.promises.writeFile(
+          path.join(claudeDir, "settings.json"),
+          JSON.stringify({ permissions: { defaultMode: mode } }),
+        );
+
+        settingsManager = new SettingsManager(tempDir);
+        await settingsManager.initialize();
+
+        expect(settingsManager.getSettings().permissions?.defaultMode).toBeUndefined();
+      },
+    );
+
+    it.each(["plan", "dontAsk"] as const)(
+      "preserves non-escalating %s from project-tier settings",
+      async (mode) => {
+        const claudeDir = path.join(tempDir, ".claude");
+        await fs.promises.mkdir(claudeDir, { recursive: true });
+
+        await fs.promises.writeFile(
+          path.join(claudeDir, "settings.json"),
+          JSON.stringify({ permissions: { defaultMode: mode } }),
+        );
+
+        settingsManager = new SettingsManager(tempDir);
+        await settingsManager.initialize();
+
+        expect(settingsManager.getSettings().permissions?.defaultMode).toBe(mode);
+      },
+    );
+
+    it("preserves escalating defaultMode when it comes from local-tier settings", async () => {
+      // settings.local.json is git-ignored by convention, so the trust
+      // filter does not apply.
+      const claudeDir = path.join(tempDir, ".claude");
+      await fs.promises.mkdir(claudeDir, { recursive: true });
+
+      await fs.promises.writeFile(
+        path.join(claudeDir, "settings.local.json"),
+        JSON.stringify({ permissions: { defaultMode: "acceptEdits" } }),
+      );
+
+      settingsManager = new SettingsManager(tempDir);
+      await settingsManager.initialize();
+
+      expect(settingsManager.getSettings().permissions?.defaultMode).toBe("acceptEdits");
+    });
+  });
 });

src/tools.ts

@@ -8,15 +8,24 @@ import {
 import { HookCallback } from "@anthropic-ai/claude-agent-sdk";
 import {
   AgentInput,
+  AgentOutput,
+  AskUserQuestionInput,
   BashInput,
+  BashOutput,
   FileEditInput,
   FileReadInput,
+  FileReadOutput,
   FileWriteInput,
   GlobInput,
   GrepInput,
+  ReportFindingsInput,
+  TaskCreateInput,
+  TaskCreateOutput,
+  TaskUpdateInput,
   TodoWriteInput,
   WebFetchInput,
   WebSearchInput,
+  WebSearchOutput,
 } from "@anthropic-ai/claude-agent-sdk/sdk-tools.js";
 import {
   ImageBlockParam,
@@ -51,7 +60,6 @@ import {
   BetaWebSearchToolResultBlockParam,
 } from "@anthropic-ai/sdk/resources/beta.mjs";
 import path from "node:path";
-import { Logger } from "./acp-agent.js";
 
 /**
  * Union of all possible content types that can appear in tool results from the Anthropic SDK.
@@ -367,6 +375,59 @@ export function toolInfoFromToolUse(
       };
     }
 
+    case "ReportFindings": {
+      const input = toolUse.input as ReportFindingsInput | undefined;
+      const findings = input?.findings ?? [];
+      return {
+        title:
+          findings.length === 0
+            ? "Report findings: none found"
+            : `Report ${findings.length} finding${findings.length === 1 ? "" : "s"}`,
+        kind: "think",
+        content: findings.map((finding) => ({
+          type: "content" as const,
+          content: {
+            type: "text" as const,
+            text: `**${finding.file}${finding.line ? `:${finding.line}` : ""}** — ${finding.summary}`,
+          },
+        })),
+      };
+    }
+
+    case "TaskCreate": {
+      const input = toolUse.input as TaskCreateInput | undefined;
+      return {
+        title: input?.subject ? `Create task: ${input.subject}` : "Create task",
+        kind: "think",
+        content: [],
+      };
+    }
+
+    case "TaskUpdate": {
+      const input = toolUse.input as TaskUpdateInput | undefined;
+      return {
+        title: input?.subject ? `Update task: ${input.subject}` : "Update task",
+        kind: "think",
+        content: [],
+      };
+    }
+
+    case "TaskList": {
+      return {
+        title: "List tasks",
+        kind: "think",
+        content: [],
+      };
+    }
+
+    case "TaskGet": {
+      return {
+        title: "Get task",
+        kind: "think",
+        content: [],
+      };
+    }
+
     case "ExitPlanMode": {
       const planInput = toolUse.input as { plan?: string } | undefined;
       return {
@@ -378,6 +439,24 @@ export function toolInfoFromToolUse(
       };
     }
 
+    case "AskUserQuestion": {
+      const input = toolUse.input as Partial<AskUserQuestionInput> | undefined;
+      const questions = Array.isArray(input?.questions) ? input.questions : [];
+      return {
+        title:
+          questions.length === 1 && questions[0]?.question
+            ? questions[0].question
+            : "Asking for your input",
+        kind: "other",
+        content: questions
+          .filter((q) => typeof q?.question === "string")
+          .map((q) => ({
+            type: "content" as const,
+            content: { type: "text" as const, text: q.question },
+          })),
+      };
+    }
+
     case "Other": {
       const input = toolUse.input;
       let output;
@@ -410,6 +489,88 @@ export function toolInfoFromToolUse(
   }
 }
 
+/**
+ * Narrow the untyped message-level `tool_use_result` toward a per-tool Output
+ * shape: rejects everything but a plain non-null object (arrays pass a bare
+ * `typeof === "object"` check, so they're excluded here). The returned value
+ * is only *nominally* typed — it arrives over the wire from arbitrary CLI
+ * versions, so each caller must still guard the specific fields it reads
+ * before trusting them.
+ */
+function structuredResult<T extends object>(toolUseResult: unknown): T | undefined {
+  return toolUseResult !== null &&
+    typeof toolUseResult === "object" &&
+    !Array.isArray(toolUseResult)
+    ? (toolUseResult as T)
+    : undefined;
+}
+
+/**
+ * Strip the model-directed trailer from a raw Agent/Task tool_result text:
+ * a `<usage>…</usage>` totals block and/or an
+ * `agentId: <id> (use SendMessage …)` continuation line at the end of the
+ * text. Both patterns are tail-anchored and independent (older CLIs emit
+ * variants with only one of them), so a format change makes them stop
+ * matching rather than mangle the report.
+ */
+function stripAgentTrailer(text: string): string {
+  return stripAgentIdLine(stripUsageBlock(text));
+}
+
+const USAGE_OPEN = "<usage>";
+const USAGE_CLOSE = "</usage>";
+
+/** Remove a trailing `<usage>…</usage>` block, plus trailing whitespace and
+ *  one preceding newline. Matches from the *last* `<usage>` so a report that
+ *  merely mentions the marker earlier isn't truncated at the mention. */
+function stripUsageBlock(text: string): string {
+  const body = text.trimEnd();
+  if (!body.endsWith(USAGE_CLOSE)) {
+    return text;
+  }
+  const open = body.lastIndexOf(USAGE_OPEN, body.length - USAGE_CLOSE.length - USAGE_OPEN.length);
+  if (open === -1) {
+    return text;
+  }
+  return body.slice(0, open > 0 && body[open - 1] === "\n" ? open - 1 : open);
+}
+
+/** The continuation line, anchored to a whole line so the regex has a single
+ *  start position and no ambiguous repetition (`[\w-]+` can't consume the
+ *  following space, `[^)]*` can't consume the closing paren) — it runs in
+ *  linear time on any input. */
+const AGENT_ID_LINE = /^agentId: [\w-]+ \([^)]*\)$/;
+
+/** Remove a final `agentId: <id> (…)` line, plus trailing whitespace and the
+ *  newline that preceded the line. */
+function stripAgentIdLine(text: string): string {
+  const body = text.trimEnd();
+  const lineStart = body.lastIndexOf("\n") + 1;
+  if (!AGENT_ID_LINE.test(body.slice(lineStart))) {
+    return text;
+  }
+  return body.slice(0, Math.max(lineStart - 1, 0));
+}
+
+/** Apply {@link stripAgentTrailer} across a raw tool_result `content` (plain
+ *  string or block array), leaving non-text blocks untouched. */
+function stripAgentTrailerFromContent(content: unknown): unknown {
+  if (typeof content === "string") {
+    return stripAgentTrailer(content);
+  }
+  if (Array.isArray(content)) {
+    return content.map((block) =>
+      block !== null &&
+      typeof block === "object" &&
+      block.type === "text" &&
+      typeof block.text === "string"
+        ? { ...block, text: stripAgentTrailer(block.text) }
+        : block,
+    );
+  }
+  return content;
+}
+
 export function toolUpdateFromToolResult(
   toolResult:
     | ToolResultBlockParam
@@ -424,19 +585,76 @@ export function toolUpdateFromToolResult(
     | BetaToolSearchToolResultBlockParam,
   toolUse: any | undefined,
   supportsTerminalOutput: boolean = false,
+  toolUseResult?: unknown,
 ): ToolUpdate {
   if (
     "is_error" in toolResult &&
     toolResult.is_error &&
     toolResult.content &&
-    toolResult.content.length > 0
+    toolResult.content.length > 0 &&
+    !(toolUse?.name === "Bash" && supportsTerminalOutput)
   ) {
     // Only return errors
     return toAcpContentUpdate(toolResult.content, true);
   }
 
+  // Shared raw-text fallback: renders the tool_result content the model saw.
+  // The structured cases below fall back to this when `tool_use_result` is
+  // absent or fails its shape guard (older CLIs, replayed sessions).
+  const rawContentUpdate = () =>
+    toAcpContentUpdate(toolResult.content, "is_error" in toolResult ? toolResult.is_error : false);
+
   switch (toolUse?.name) {
-    case "Read":
+    case "Read": {
+      // The raw tool_result text is the model-facing view: line-numbered
+      // content plus any appended <system-reminder> blocks (malicious-code
+      // checks, memory staleness notes, …) that clients shouldn't see. The
+      // structured FileReadOutput carries the clean content — rebuild the
+      // line-numbered view from it. Non-text variants (image/notebook/pdf)
+      // fall back to the raw content blocks, which already render fine.
+      const structuredRead = structuredResult<FileReadOutput>(toolUseResult);
+      if (
+        structuredRead?.type === "text" &&
+        typeof structuredRead.file?.content === "string" &&
+        // An empty file has nothing to line-number; keep the raw view (the
+        // model-facing "file is empty" note) rather than a phantom blank line.
+        structuredRead.file.content.length > 0
+      ) {
+        // startLine is typed non-optional but defended anyway; a Read's
+        // `offset` input is the same 1-based starting line, so it beats a
+        // blind 1 when an emitter omits the field.
+        const startLine =
+          structuredRead.file.startLine ??
+          (toolUse.input as FileReadInput | undefined)?.offset ??
+          1;
+        // A trailing newline is a line terminator, not an extra line — don't
+        // number a phantom empty line after it.
+        let numbered = structuredRead.file.content
+          .replace(/\n$/, "")
+          .split("\n")
+          .map((line, i) => `${startLine + i}\t${line}`)
+          .join("\n");
+        // The model-facing truncation banner doesn't survive reconstruction
+        // from file.content (the SDK flag exists for exactly this case) —
+        // re-establish it so a partial first page doesn't read as the whole
+        // file.
+        if (structuredRead.file.truncatedByTokenCap) {
+          const { numLines, totalLines } = structuredRead.file;
+          const detail =
+            typeof numLines === "number" && typeof totalLines === "number"
+              ? `: showing ${numLines} of ${totalLines} lines`
+              : "";
+          numbered += `\n[File truncated${detail}]`;
+        }
+        return {
+          content: [
+            {
+              type: "content",
+              content: { type: "text", text: markdownEscape(numbered) },
+            },
+          ],
+        };
+      }
       if (Array.isArray(toolResult.content) && toolResult.content.length > 0) {
         return {
           content: toolResult.content.map((content: any) => ({
@@ -464,6 +682,7 @@ export function toolUpdateFromToolResult(
         };
       }
       return {};
+    }
 
     case "Bash": {
       const result = toolResult.content;
@@ -471,13 +690,57 @@ export function toolUpdateFromToolResult(
       const isError = "is_error" in toolResult && toolResult.is_error;
 
       // Extract output and exit code from either format:
-      // 1. BetaBashCodeExecutionResultBlock: { type: "bash_code_execution_result", stdout, stderr, return_code }
-      // 2. Plain string content from a regular tool_result
-      // 3. Array content (e.g. [{ type: "text", text: "..." }])
+      // 1. The structured BashOutput (message-level tool_use_result): its
+      //    stdout/stderr exclude the model-directed suffixes the raw text
+      //    carries (stale-read hints, gh rate-limit hints, the
+      //    persisted-output wrapper for too-large outputs — the interruption
+      //    and truncation facts those carried are re-established from the
+      //    structured flags below). Skipped for image output (the raw content
+      //    array carries the actual image blocks) and backgrounded commands
+      //    (the raw text carries the background-task notice; structured
+      //    stdout may be empty).
+      // 2. BetaBashCodeExecutionResultBlock: { type: "bash_code_execution_result", stdout, stderr, return_code }
+      // 3. Plain string content from a regular tool_result
+      // 4. Array content (e.g. [{ type: "text", text: "..." }] for stdout,
+      //    or [{ type: "image", source: {...} }] when the local Bash tool
+      //    produces an image, e.g. piping a base64 data URI)
       let output = "";
       let exitCode = isError ? 1 : 0;
 
+      const structuredBash = structuredResult<BashOutput>(toolUseResult);
       if (
+        structuredBash &&
+        typeof structuredBash.stdout === "string" &&
+        typeof structuredBash.stderr === "string" &&
+        !structuredBash.isImage &&
+        structuredBash.backgroundTaskId === undefined
+      ) {
+        output = [structuredBash.stdout, structuredBash.stderr].filter(Boolean).join("\n");
+        // Two raw-text notices don't survive the structured stdout/stderr —
+        // re-establish them so the client isn't shown a clean-looking result:
+        // the CLI appends its abort marker only to the model-facing text, and
+        // an aborted command isn't a success, so synthesize a failing exit
+        // code when the result wasn't already an error.
+        if (structuredBash.interrupted) {
+          output = [output, "[Command was aborted before completion]"].filter(Boolean).join("\n");
+          exitCode = 1;
+        }
+        // Structured stdout is clipped (~30k chars) when the full output was
+        // persisted to disk; without this note the clip is silent and the
+        // path to the full output is lost.
+        if (typeof structuredBash.persistedOutputPath === "string") {
+          const size =
+            typeof structuredBash.persistedOutputSize === "number"
+              ? ` (${structuredBash.persistedOutputSize} bytes total)`
+              : "";
+          output = [
+            output,
+            `[Output truncated${size}: full output saved to ${structuredBash.persistedOutputPath}]`,
+          ]
+            .filter(Boolean)
+            .join("\n");
+        }
+      } else if (
         result &&
         typeof result === "object" &&
         "type" in result &&
@@ -488,13 +751,20 @@ export function toolUpdateFromToolResult(
         exitCode = bashResult.return_code;
       } else if (typeof result === "string") {
         output = result;
-      } else if (
-        Array.isArray(result) &&
-        result.length > 0 &&
-        "text" in result[0] &&
-        typeof result[0].text === "string"
-      ) {
-        output = result.map((c: any) => c.text).join("\n");
+      } else if (Array.isArray(result) && result.length > 0) {
+        const textOnly = result.every(
+          (c: any) => c && typeof c === "object" && typeof c.text === "string",
+        );
+        if (textOnly) {
+          output = result.map((c: any) => c.text).join("\n");
+        } else {
+          // Image (or mixed non-text) content. Binary payloads can't be
+          // streamed through the terminal-output _meta channel, so bypass
+          // it and surface the blocks as ACP content. This handles the
+          // local Bash tool's image output, which previously failed the
+          // text-only guard and was silently dropped.
+          return toAcpContentUpdate(result, isError);
+        }
       }
 
       if (supportsTerminalOutput) {
@@ -533,6 +803,45 @@ export function toolUpdateFromToolResult(
       return {};
     }
 
+    case "Agent":
+    case "Task": {
+      // The raw tool_result text ends with a model-directed trailer (an
+      // `agentId: … (use SendMessage …)` line plus a `<usage>` totals block)
+      // that ACP clients shouldn't see. The message-level `tool_use_result`
+      // carries the structured AgentOutput whose `content` is the subagent's
+      // report without the trailer — render from it when present (per the SDK
+      // 0.3.207 guidance) and fall back to the raw text otherwise (older CLIs,
+      // replayed sessions).
+      // Narrowed to the full union, not the completed variant — the status
+      // check below is what discriminates it, and pre-narrowing would let
+      // future field reads typecheck against a variant the runtime value may
+      // not be.
+      const structured = structuredResult<AgentOutput>(toolUseResult);
+      if (
+        structured?.status === "completed" &&
+        Array.isArray(structured.content) &&
+        // A completed subagent can end with zero text blocks; an empty
+        // structured render would beat the raw fallback for no benefit.
+        structured.content.length > 0
+      ) {
+        return toAcpContentUpdate(
+          structured.content,
+          "is_error" in toolResult ? toolResult.is_error : false,
+        );
+      }
+      // No structured report to render from (replayed sessions —
+      // getSessionMessages doesn't expose the transcript's toolUseResult —
+      // and older CLIs). The SDK advises rendering from tool_use_result
+      // instead of parsing the text, but with no structured value the
+      // tail-anchored strip is the only cleanup available; if the trailer
+      // format changes it simply stops matching and the full raw text
+      // renders, no worse than before.
+      return toAcpContentUpdate(
+        stripAgentTrailerFromContent(toolResult.content),
+        "is_error" in toolResult ? toolResult.is_error : false,
+      );
+    }
+
     case "Edit": // Edit is handled in hooks
     case "Write": {
       return {};
@@ -542,15 +851,54 @@ export function toolUpdateFromToolResult(
       return { title: "Exited Plan Mode" };
     }
 
+    case "WebSearch": {
+      // The raw tool_result text is a model-directed dump ("Web search
+      // results for query: …\n\nLinks: [{…json…}]"). The structured
+      // WebSearchOutput carries the hits — render them the way server-side
+      // web_search_result blocks render ("Title (url)").
+      const structuredSearch = structuredResult<WebSearchOutput>(toolUseResult);
+      if (structuredSearch && Array.isArray(structuredSearch.results)) {
+        const lines = structuredSearch.results.flatMap((entry) =>
+          typeof entry === "string"
+            ? [entry]
+            : Array.isArray(entry?.content)
+              ? // tool_use_result arrives untyped across CLI version skew —
+                // skip off-spec hits rather than rendering
+                // "undefined (undefined)" lines.
+                entry.content.flatMap((hit) =>
+                  typeof hit?.title === "string" && typeof hit?.url === "string"
+                    ? [formatWebSearchHit(hit)]
+                    : [],
+                )
+              : [],
+        );
+        if (lines.length > 0) {
+          return {
+            content: [
+              {
+                type: "content",
+                content: { type: "text", text: lines.join("\n") },
+              },
+            ],
+          };
+        }
+      }
+      return rawContentUpdate();
+    }
+
     default: {
-      return toAcpContentUpdate(
-        toolResult.content,
-        "is_error" in toolResult ? toolResult.is_error : false,
-      );
+      return rawContentUpdate();
     }
   }
 }
 
+/** One display format for a web-search hit, shared by the structured
+ *  WebSearchOutput render and the server-side `web_search_result` block so
+ *  the two paths can't drift. */
+function formatWebSearchHit(hit: { title: string; url: string }): string {
+  return `${hit.title} (${hit.url})`;
+}
+
 function toAcpContentUpdate(
   content: any,
   isError: boolean = false,
@@ -625,7 +973,7 @@ function toAcpContentBlock(content: ToolResultContent, isError: boolean): Conten
         `Error: ${content.error_code}${content.error_message ? ` - ${content.error_message}` : ""}`,
       );
     case "web_search_result":
-      return wrapText(`${content.title} (${content.url})`);
+      return wrapText(formatWebSearchHit(content));
     case "web_search_tool_result_error":
       return wrapText(`Error: ${content.error_code}`);
     case "web_fetch_result":
@@ -669,6 +1017,103 @@ export function planEntries(input: { todos: ClaudePlanEntry[] } | undefined): Pl
   }));
 }
 
+/**
+ * Per-session task list accumulated from Task* tool calls (TaskCreate /
+ * TaskUpdate). The headless/SDK session emits these as incremental tool
+ * calls keyed by task ID, replacing the snapshot-style TodoWrite tool.
+ * Iteration order is insertion order (Map semantics), matching the order
+ * tasks are created.
+ */
+export type TaskEntry = {
+  subject: string;
+  status: "pending" | "in_progress" | "completed";
+  activeForm?: string;
+  description?: string;
+};
+export type TaskState = Map<string, TaskEntry>;
+
+/**
+ * Best-effort parse of a TaskCreate tool_result content into the structured
+ * TaskCreateOutput. The SDK delivers tool outputs either as a string or as
+ * an array of TextBlockParam-like blocks containing JSON text; try both.
+ */
+export function parseTaskCreateOutput(content: unknown): TaskCreateOutput | undefined {
+  const tryParse = (text: string): TaskCreateOutput | undefined => {
+    try {
+      const parsed = JSON.parse(text);
+      if (
+        parsed &&
+        typeof parsed === "object" &&
+        parsed.task &&
+        typeof parsed.task.id === "string"
+      ) {
+        return parsed as TaskCreateOutput;
+      }
+    } catch {
+      // ignore
+    }
+    return undefined;
+  };
+
+  if (typeof content === "string") {
+    return tryParse(content);
+  }
+  if (Array.isArray(content)) {
+    for (const block of content) {
+      if (block && typeof block === "object" && "type" in block && block.type === "text") {
+        const text = (block as { text?: unknown }).text;
+        if (typeof text === "string") {
+          const parsed = tryParse(text);
+          if (parsed) return parsed;
+        }
+      }
+    }
+  }
+  return undefined;
+}
+
+export function applyTaskCreate(
+  state: TaskState,
+  input: TaskCreateInput | undefined,
+  output: TaskCreateOutput | undefined,
+): void {
+  const taskId = output?.task?.id;
+  if (!taskId || !input) return;
+  state.set(taskId, {
+    subject: input.subject,
+    status: "pending",
+    activeForm: input.activeForm,
+    description: input.description,
+  });
+}
+
+export function applyTaskUpdate(state: TaskState, input: TaskUpdateInput | undefined): void {
+  if (!input?.taskId) return;
+  if (input.status === "deleted") {
+    state.delete(input.taskId);
+    return;
+  }
+  const existing = state.get(input.taskId);
+  // Without a subject from either the existing entry or the update payload,
+  // we'd produce a plan entry with empty `content` — drop the update.
+  const subject = input.subject ?? existing?.subject;
+  if (!subject) return;
+  state.set(input.taskId, {
+    subject,
+    status: input.status ?? existing?.status ?? "pending",
+    activeForm: input.activeForm ?? existing?.activeForm,
+    description: input.description ?? existing?.description,
+  });
+}
+
+export function taskStateToPlanEntries(state: TaskState): PlanEntry[] {
+  return Array.from(state.values()).map((task) => ({
+    content: task.subject,
+    status: task.status,
+    priority: "medium",
+  }));
+}
+
 export function markdownEscape(text: string): string {
   let escape = "```";
   for (const [m] of text.matchAll(/^```+/gm)) {
@@ -679,7 +1124,7 @@ export function markdownEscape(text: string): string {
   return escape + "\n" + text + (text.endsWith("\n") ? "" : "\n") + escape;
 }
 
-interface EditToolResponseHunk {
+interface DiffToolResponseHunk {
   oldStart: number;
   oldLines: number;
   newStart: number;
@@ -687,23 +1132,24 @@ interface EditToolResponseHunk {
   lines: string[];
 }
 
-interface EditToolResponse {
+interface DiffToolResponse {
   filePath?: string;
-  structuredPatch?: EditToolResponseHunk[];
+  structuredPatch?: DiffToolResponseHunk[];
 }
 
 /**
- * Builds diff ToolUpdate content from the structured Edit toolResponse provided
- * by the PostToolUse hook. Unlike parsing the plain unified diff string, this uses
- * the pre-parsed structuredPatch which supports multiple replacement sites (replaceAll)
- * and always includes context lines for better readability.
+ * Builds diff ToolUpdate content from the structured toolResponse provided by
+ * the PostToolUse hook for diff-producing tools (Edit, Write). Unlike parsing
+ * the plain unified diff string, this uses the pre-parsed structuredPatch
+ * which supports multiple replacement sites (replaceAll) and always includes
+ * context lines for better readability.
  */
-export function toolUpdateFromEditToolResponse(toolResponse: unknown): {
+export function toolUpdateFromDiffToolResponse(toolResponse: unknown): {
   content?: ToolCallContent[];
   locations?: ToolCallLocation[];
 } {
   if (!toolResponse || typeof toolResponse !== "object") return {};
-  const response = toolResponse as EditToolResponse;
+  const response = toolResponse as DiffToolResponse;
   if (!response.filePath || !Array.isArray(response.structuredPatch)) return {};
 
   const content: ToolCallContent[] = [];
@@ -770,12 +1216,7 @@ export const registerHookCallback = (
 
 /* A callback for Claude Code that is called when receiving a PostToolUse hook */
 export const createPostToolUseHook =
-  (
-    logger: Logger = console,
-    options?: {
-      onEnterPlanMode?: () => Promise<void>;
-    },
-  ): HookCallback =>
+  (options?: { onEnterPlanMode?: () => Promise<void> }): HookCallback =>
   async (input: any, toolUseID: string | undefined): Promise<{ continue: boolean }> => {
     if (input.hook_event_name === "PostToolUse") {
       // Handle EnterPlanMode tool - notify client of mode change after successful execution
@@ -787,12 +1228,44 @@ export const createPostToolUseHook =
         const onPostToolUseHook = toolUseCallbacks[toolUseID]?.onPostToolUseHook;
         if (onPostToolUseHook) {
           await onPostToolUseHook(toolUseID, input.tool_input, input.tool_response);
-          delete toolUseCallbacks[toolUseID]; // Cleanup after execution
-        } else {
-          logger.error(`No onPostToolUseHook found for tool use ID: ${toolUseID}`);
-          delete toolUseCallbacks[toolUseID];
         }
+        delete toolUseCallbacks[toolUseID]; // Cleanup after execution
       }
     }
     return { continue: true };
   };
+
+/**
+ * Hook callback for `TaskCreated` / `TaskCompleted` events. The SDK fires
+ * these for both user-facing TaskCreate tool calls and subagent task
+ * creation, giving us `task_id` + `task_subject` without having to parse
+ * tool_result payloads.
+ *
+ * Populating `taskState` from the hook means a later `TaskUpdate` (which
+ * typically only carries `taskId` + `status`) finds an existing entry with
+ * a real subject, instead of synthesizing a placeholder with empty content.
+ */
+export const createTaskHook =
+  (options: { taskState: TaskState; onChange?: () => Promise<void> }): HookCallback =>
+  async (input): Promise<{ continue: boolean }> => {
+    const taskId =
+      "task_id" in input && typeof input.task_id === "string" ? input.task_id : undefined;
+    if (!taskId) return { continue: true };
+
+    if (input.hook_event_name === "TaskCreated") {
+      if (!input.task_subject) return { continue: true };
+      if (options.taskState.has(taskId)) return { continue: true };
+      options.taskState.set(taskId, {
+        subject: input.task_subject,
+        status: "pending",
+        description: input.task_description,
+      });
+      if (options.onChange) await options.onChange();
+    } else if (input.hook_event_name === "TaskCompleted") {
+      const existing = options.taskState.get(taskId);
+      if (!existing || existing.status === "completed") return { continue: true };
+      options.taskState.set(taskId, { ...existing, status: "completed" });
+      if (options.onChange) await options.onChange();
+    }
+    return { continue: true };
+  };

src/utils.ts

@@ -2,9 +2,7 @@
 
 import { Readable, Writable } from "node:stream";
 import { WritableStream, ReadableStream } from "node:stream/web";
-import { readFileSync } from "node:fs";
 import { Logger } from "./acp-agent.js";
-import { ClaudeCodeSettings, getManagedSettingsPath } from "./settings.js";
 
 // Useful for bridging push-based and async-iterator-based code.
 export class Pushable<T> implements AsyncIterable<T> {
@@ -89,19 +87,3 @@ export function unreachable(value: never, logger: Logger = console) {
 export function sleep(time: number): Promise<void> {
   return new Promise((resolve) => setTimeout(resolve, time));
 }
-
-export function loadManagedSettings(): ClaudeCodeSettings | null {
-  try {
-    return JSON.parse(readFileSync(getManagedSettingsPath(), "utf8")) as ClaudeCodeSettings;
-  } catch {
-    return null;
-  }
-}
-
-export function applyEnvironmentSettings(settings: ClaudeCodeSettings): void {
-  if (settings.env) {
-    for (const [key, value] of Object.entries(settings.env)) {
-      process.env[key] = value;
-    }
-  }
-}

tsconfig.json

@@ -1,13 +1,7 @@
 {
+  "extends": "@tsconfig/node22/tsconfig.json",
   "compilerOptions": {
-    "target": "ES2021",
-    "module": "NodeNext",
-    "moduleResolution": "NodeNext",
-    "lib": ["ES2021"],
-    "strict": true,
-    "esModuleInterop": true,
     "forceConsistentCasingInFileNames": true,
-    "skipLibCheck": true,
     "outDir": "dist",
     "rootDir": "src",
     "declaration": true,