chore(main): release 0.52.0 (#811 )

🤖 I have created a release *beep* *boop* --- ## [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](9616bdac47)), 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](61272acb30)) * **deps:** Update @anthropic-ai/claude-agent-sdk to 0.3.191 ([#810](https://github.com/agentclientprotocol/claude-agent-acp/issues/810)) ([228f02e](228f02ecfb)) * Push session title updates at turn end ([#812](https://github.com/agentclientprotocol/claude-agent-acp/issues/812)) ([1fe7ec0](1fe7ec09a3)) --- 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>
feat(deps-dev): bump expect-type from 1.3.0 to 1.4.0 in the minor group (#814 )
2026-10-27 12:40:32 +00:00 · 2026-06-25 17:11:07 +02:00 · 2026-06-25 17:08:42 +02:00 · 2026-06-25 12:37:02 +00:00 · 2026-06-25 11:45:50 +00:00 · 2026-06-25 13:42:57 +02:00
23 changed files with 4874 additions and 985 deletions
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -21,7 +21,7 @@ jobs:
    timeout-minutes: 5

    steps:
-      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0

      - name: Use Node.js
        uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e
--- a/.github/workflows/publish.yml
+++ b/.github/workflows/publish.yml
@@ -44,7 +44,7 @@ jobs:
      contents: read
      id-token: write
    steps:
-      - uses: actions/checkout@df4cb1c069e1874edd31b4311f1884172cec0e10
+      - uses: actions/checkout@9c091bb21b7c1c1d1991bb908d89e4e9dddfe3e0
      # Setup .npmrc file to publish to npm
      - uses: actions/setup-node@48b55a011bda9f5d6aeb4c2d9c7362e8dae4041e
        with:
--- a/.release-please-manifest.json
+++ b/.release-please-manifest.json
@@ -1,3 +1,3 @@
 {
-  ".": "0.42.0"
+  ".": "0.52.0"
 }
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -1,5 +1,130 @@
 # Changelog

+## [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)


--- a/eslint.config.js
+++ b/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"],
    },
  },
--- a/package-lock.json
+++ b/package-lock.json
--- a/package.json
+++ b/package.json
@@ -3,7 +3,7 @@
  "publishConfig": {
    "access": "public"
  },
-  "version": "0.42.0",
+  "version": "0.52.0",
  "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": "0.24.0",
-    "@anthropic-ai/claude-agent-sdk": "0.3.165",
+    "@agentclientprotocol/sdk": "1.0.0",
+    "@anthropic-ai/claude-agent-sdk": "0.3.191",
    "zod": "^3.25.0 || ^4.0.0"
  },
  "devDependencies": {
-    "@anthropic-ai/sdk": "0.100.1",
+    "@anthropic-ai/sdk": "0.106.0",
    "@eslint/js": "10.0.1",
-    "@types/node": "25.9.1",
-    "@typescript-eslint/eslint-plugin": "8.60.1",
-    "@typescript-eslint/parser": "8.60.1",
-    "eslint": "10.4.1",
+    "@tsconfig/node22": "22.0.5",
+    "@types/node": "26.0.1",
+    "@typescript-eslint/eslint-plugin": "8.62.0",
+    "@typescript-eslint/parser": "8.62.0",
+    "eslint": "10.5.0",
    "eslint-config-prettier": "10.1.8",
-    "globals": "17.6.0",
-    "prettier": "3.8.3",
+    "globals": "17.7.0",
+    "prettier": "3.8.4",
    "ts-node": "10.9.2",
    "typescript": "6.0.3",
-    "vitest": "4.1.8"
+    "vitest": "4.1.9"
  }
 }
--- a/src/acp-agent.ts
+++ b/src/acp-agent.ts
--- a/src/elicitation.ts
+++ b/src/elicitation.ts
@@ -0,0 +1,301 @@
+import { randomUUID } from "node:crypto";
+import type {
+  CreateElicitationRequest,
+  CreateElicitationResponse,
+  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),
+  };
+}
+
+/**
+ * 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: response.content ?? {} };
+    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 structured
+ * `description`/`preview`, since ACP's `EnumOption` has no field for either.
+ * Namespaced like the agent's other `_meta` extensions (`_claude/...`).
+ */
+const OPTION_META_KEY = "_claude/askUserQuestionOption";
+
+/** A single option as supplied by the AskUserQuestion tool. */
+type AskUserQuestionOption = AskUserQuestion["options"][number];
+
+/**
+ * Build the `_meta` payload that carries an option's structured `description`
+ * and optional `preview` for clients that can render them as distinct fields.
+ * Returns `undefined` when neither is present, so we don't emit an empty `_meta`.
+ */
+function askUserQuestionOptionMeta(
+  option: AskUserQuestionOption,
+): NonNullable<EnumOption["_meta"]> | undefined {
+  const detail: { description?: string; preview?: string } = {};
+  if (option.description) {
+    detail.description = option.description;
+  }
+  if (option.preview) {
+    detail.preview = option.preview;
+  }
+  return Object.keys(detail).length > 0 ? { [OPTION_META_KEY]: detail } : undefined;
+}
+
+/**
+ * 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.
+ *
+ * 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` stays a flattened "label — description" string so clients that
+        // only read `const`/`title` still surface the description. Clients that
+        // understand the `_meta` below should prefer its structured fields (and
+        // treat `const` as the clean label) rather than parsing this title.
+        title: option.description ? `${option.label} — ${option.description}` : option.label,
+      };
+      // The SDK option carries a `description` (secondary text) and an optional
+      // `preview` (mockups, code snippets, comparisons shown on focus). Neither
+      // has a structural slot in `EnumOption`, so forward them under ACP's
+      // reserved `_meta` extension point for clients that can render them.
+      const meta = askUserQuestionOptionMeta(option);
+      if (meta) {
+        enumOption._meta = meta;
+      }
+      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 aborts the tool call.
+ */
+export function applyAskElicitationResponse(
+  response: CreateElicitationResponse,
+  toolInput: Record<string, unknown>,
+  questions: AskUserQuestion[],
+): AskUserQuestionOutcome {
+  if (response.action === "cancel") {
+    return { action: "cancel" };
+  }
+
+  if (response.action === "decline") {
+    return { action: "answered", updatedInput: { ...toolInput, answers: {} } };
+  }
+
+  const content = response.content ?? {};
+  // 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" };
+}
--- a/src/index.ts
+++ b/src/index.ts
@@ -2,8 +2,12 @@

 import { resolveSettings } from "@anthropic-ai/claude-agent-sdk";
 import { claudeCliPath, runAcp } from "./acp-agent.js";
+import packageJson from "../package.json" with { type: "json" };

-if (process.argv.includes("--cli")) {
+if (process.argv.includes("--version") || process.argv.includes("-v")) {
+  console.log(packageJson.version);
+  process.exit(0);
+} else if (process.argv.includes("--cli")) {
  const { spawn } = await import("node:child_process");
  const args = process.argv.slice(2).filter((arg) => arg !== "--cli");
  const child = spawn(await claudeCliPath(), args, { stdio: "inherit" });
--- a/src/tests/acp-agent-settings.test.ts
+++ b/src/tests/acp-agent-settings.test.ts
@@ -2,8 +2,7 @@ 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";

 const { querySpy } = vi.hoisted(() => ({
  querySpy: vi.fn(),
@@ -21,13 +20,13 @@ 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() {
@@ -700,6 +699,118 @@ describe("ClaudeAcpAgent settings", () => {
      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("resolves model aliases like opus[1m] to the correct model", async () => {
--- a/src/tests/acp-agent.test.ts
+++ b/src/tests/acp-agent.test.ts
--- a/src/tests/additional-roots.test.ts
+++ b/src/tests/additional-roots.test.ts
@@ -1,9 +1,9 @@
 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;
@@ -39,7 +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 () => {
@@ -52,7 +52,7 @@ describe("additionalRoots", () => {
      requestPermission: async () => ({ outcome: { outcome: "cancelled" } }),
      readTextFile: async () => ({ content: "" }),
      writeTextFile: async () => ({}),
-    } as unknown as AgentSideConnection);
+    } as unknown as AcpClient);
  });

  afterEach(
@@ -79,7 +79,7 @@ describe("additionalRoots", () => {

  it("prefers the official ACP additionalDirectories field over _meta.additionalRoots", async () => {
    await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
      mcpServers: [],
      additionalDirectories: ["/from/official"],
      _meta: { additionalRoots: ["/from/meta"] },
@@ -89,7 +89,7 @@ describe("additionalRoots", () => {

  it("merges official ACP additionalDirectories with claudeCode SDK additionalDirectories", async () => {
    await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
      mcpServers: [],
      additionalDirectories: ["/from/official"],
      _meta: { claudeCode: { options: { additionalDirectories: ["/from/sdk"] } } },
@@ -99,7 +99,7 @@ describe("additionalRoots", () => {

  it("falls back to _meta.additionalRoots when the official field is omitted", async () => {
    await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
      mcpServers: [],
      _meta: { additionalRoots: ["/from/meta"] },
    });
--- a/src/tests/agent-discovery.test.ts
+++ b/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);
+});
--- a/src/tests/authorization.test.ts
+++ b/src/tests/authorization.test.ts
@@ -1,6 +1,5 @@
 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";

 const mockQuery = vi.hoisted(() =>
  vi.fn(() => ({
@@ -39,7 +38,7 @@ describe("authorization", () => {
  async function createAgentMock(): Promise<[ClaudeAcpAgent, Mock]> {
    const connectionMock = {
      sessionUpdate: async (_: any) => {},
-    } as AgentSideConnection;
+    } as AcpClient;

    const agent = new ClaudeAcpAgent(connectionMock);

@@ -103,7 +102,7 @@ describe("authorization", () => {
    });

    await agent.newSession({
-      cwd: "testRoot",
+      cwd: process.cwd(),
      mcpServers: [],
      _meta: {
        claudeCode: {
@@ -148,7 +147,7 @@ describe("authorization", () => {
    });

    await agent.newSession({
-      cwd: "testRoot",
+      cwd: process.cwd(),
      mcpServers: [],
    });

--- a/src/tests/create-session-options.test.ts
+++ b/src/tests/create-session-options.test.ts
@@ -1,7 +1,10 @@
 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;
 vi.mock("@anthropic-ai/claude-agent-sdk", async () => {
@@ -44,13 +47,13 @@ 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 () => {
@@ -65,7 +68,7 @@ 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: {
@@ -85,7 +88,7 @@ describe("createSession options merging", () => {

  it("works when user provides no disallowedTools", async () => {
    await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
      mcpServers: [],
    });

@@ -94,7 +97,7 @@ describe("createSession options merging", () => {

  it("works when user provides empty disallowedTools", async () => {
    await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
      mcpServers: [],
      _meta: {
        claudeCode: {
@@ -110,7 +113,7 @@ 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,
@@ -133,7 +136,7 @@ describe("createSession options merging", () => {
    const userPreToolUseHook = { hooks: [{ command: "echo pre" }] };

    await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
      mcpServers: [],
      _meta: {
        claudeCode: {
@@ -155,7 +158,7 @@ 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: [],
    });

@@ -165,7 +168,7 @@ 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: {
@@ -185,7 +188,7 @@ 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: {
@@ -203,7 +206,7 @@ describe("createSession options merging", () => {

  it("defaults tools to claude_code preset when not provided", async () => {
    await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
      mcpServers: [],
    });

@@ -212,7 +215,7 @@ describe("createSession options merging", () => {

  it("passes through user-provided tools string array", async () => {
    await agent.newSession({
-      cwd: "/test",
+      cwd: process.cwd(),
      mcpServers: [],
      _meta: {
        claudeCode: {
@@ -228,7 +231,7 @@ 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,
@@ -245,7 +248,7 @@ 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: {
@@ -261,7 +264,7 @@ 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",
@@ -271,7 +274,7 @@ 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" },
      });
@@ -281,7 +284,7 @@ describe("createSession options merging", () => {

    it("forwards append", async () => {
      await agent.newSession({
-        cwd: "/test",
+        cwd: process.cwd(),
        mcpServers: [],
        _meta: { systemPrompt: { append: "extra instructions" } },
      });
@@ -295,7 +298,7 @@ describe("createSession options merging", () => {

    it("forwards excludeDynamicSections", async () => {
      await agent.newSession({
-        cwd: "/test",
+        cwd: process.cwd(),
        mcpServers: [],
        _meta: { systemPrompt: { excludeDynamicSections: true } },
      });
@@ -309,7 +312,7 @@ describe("createSession options merging", () => {

    it("forwards append and excludeDynamicSections together", async () => {
      await agent.newSession({
-        cwd: "/test",
+        cwd: process.cwd(),
        mcpServers: [],
        _meta: {
          systemPrompt: {
@@ -329,7 +332,7 @@ describe("createSession options merging", () => {

    it("ignores caller-provided type/preset overrides", async () => {
      await agent.newSession({
-        cwd: "/test",
+        cwd: process.cwd(),
        mcpServers: [],
        _meta: {
          systemPrompt: {
@@ -369,7 +372,7 @@ 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" },
@@ -381,7 +384,7 @@ 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"],
@@ -394,7 +397,7 @@ 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" },
@@ -403,7 +406,7 @@ 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();
    });
@@ -414,7 +417,7 @@ describe("createSession options merging", () => {
      });

      await agent.newSession({
-        cwd: "/test",
+        cwd: process.cwd(),
        mcpServers: [],
        _meta: {
          claudeCode: {
@@ -438,13 +441,13 @@ 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",
@@ -491,7 +494,7 @@ describe("createSession options merging", () => {
    });

    it("leaves thinking unset (SDK default) when env var is absent", async () => {
-      await agent.newSession({ cwd: "/test", mcpServers: [] });
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
      expect(capturedOptions!.thinking).toBeUndefined();
      // The deprecated option must not be set either.
      expect(capturedOptions!.maxThinkingTokens).toBeUndefined();
@@ -499,26 +502,26 @@ describe("createSession options merging", () => {

    it("maps 0 to disabled thinking", async () => {
      process.env.MAX_THINKING_TOKENS = "0";
-      await agent.newSession({ cwd: "/test", mcpServers: [] });
+      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: "/test", mcpServers: [] });
+      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";
-      await agent.newSession({ cwd: "/test", mcpServers: [] });
+      await agent.newSession({ cwd: process.cwd(), mcpServers: [] });
      expect(capturedOptions!.thinking).toBeUndefined();
    });

    it("lets a user-provided thinking option override the env default", async () => {
      process.env.MAX_THINKING_TOKENS = "12000";
      await agent.newSession({
-        cwd: "/test",
+        cwd: process.cwd(),
        mcpServers: [],
        _meta: {
          claudeCode: {
@@ -531,4 +534,83 @@ describe("createSession options merging", () => {
      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");
+    });
+  });
 });
--- a/src/tests/elicitation.test.ts
+++ b/src/tests/elicitation.test.ts
@@ -0,0 +1,391 @@
+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,
+  mcpElicitationToCreateRequest,
+} 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 "label — description" title.
+ */
+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" });
+  });
+});
+
+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: [
+        {
+          const: "date-fns",
+          title: "date-fns — Lightweight",
+          // Structured description forwarded under `_meta` so clients can render
+          // it as secondary text instead of parsing it out of the title.
+          _meta: { "_claude/askUserQuestionOption": { description: "Lightweight" } },
+        },
+        // No description → no `_meta` emitted.
+        { const: "luxon", title: "luxon" },
+      ],
+    });
+    // The full question text is not used as a property key.
+    expect(schema.properties?.["Which library?"]).toBeUndefined();
+  });
+
+  it("forwards option description and preview under _meta for rich clients", () => {
+    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;
+
+    // Both description and preview travel structurally; the title still flattens
+    // the description for clients that only read const/title.
+    expect(oneOf[0]).toEqual({
+      const: "Grid",
+      title: "Grid — Cards in a responsive grid",
+      _meta: {
+        "_claude/askUserQuestionOption": {
+          description: "Cards in a responsive grid",
+          preview: "```\n[ ] [ ] [ ]\n[ ] [ ] [ ]\n```",
+        },
+      },
+    });
+    // Description only → preview omitted from _meta.
+    expect(oneOf[1]._meta).toEqual({
+      "_claude/askUserQuestionOption": { description: "Stacked rows" },
+    });
+    // Neither → no _meta at all.
+    expect(oneOf[2]._meta).toBeUndefined();
+  });
+
+  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("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" },
+    );
+  });
+});
--- a/src/tests/session-config-options.test.ts
+++ b/src/tests/session-config-options.test.ts
@@ -1,7 +1,7 @@
 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";

 const { registerHookCallbackSpy } = vi.hoisted(() => ({
  registerHookCallbackSpy: vi.fn(),
@@ -84,7 +84,7 @@ 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);
@@ -92,7 +92,7 @@ describe("session config options", () => {
      requestPermission: async () => ({ outcome: { outcome: "cancelled" } }),
      readTextFile: async () => ({ content: "" }),
      writeTextFile: async () => ({}),
-    } as unknown as AgentSideConnection;
+    } as unknown as AcpClient;
  }

  function populateSession() {
@@ -148,13 +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();
@@ -173,7 +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: [],
      });
@@ -216,6 +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,
--- a/src/tests/session-load.test.ts
+++ b/src/tests/session-load.test.ts
@@ -1,18 +1,18 @@
 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(): AcpClient {
  return {
    sessionUpdate: async (_notification: SessionNotification) => {},
    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 +165,7 @@ 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 +274,7 @@ 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 =
--- a/src/tests/settings.test.ts
+++ b/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 });
  });

--- a/src/tests/tools.test.ts
+++ b/src/tests/tools.test.ts
@@ -1,5 +1,5 @@
 import { describe, it, expect, beforeEach } from "vitest";
-import { AgentSideConnection, ClientCapabilities } from "@agentclientprotocol/sdk";
+import { ClientCapabilities } from "@agentclientprotocol/sdk";
 import { ImageBlockParam, ToolResultBlockParam } from "@anthropic-ai/sdk/resources";
 import {
  BetaMCPToolResultBlock,
@@ -10,7 +10,7 @@ import {
  BetaBashCodeExecutionResultBlock,
  BetaBashCodeExecutionToolResultBlockParam,
 } from "@anthropic-ai/sdk/resources/beta.mjs";
-import { toAcpNotifications, ToolUseCache, Logger } from "../acp-agent.js";
+import { AcpClient, toAcpNotifications, ToolUseCache, Logger } from "../acp-agent.js";
 import {
  toolUpdateFromToolResult,
  createPostToolUseHook,
@@ -25,7 +25,7 @@ import {
 } from "../tools.js";

 describe("rawOutput in tool call updates", () => {
-  const mockClient = {} as AgentSideConnection;
+  const mockClient = {} as AcpClient;
  const mockLogger: Logger = { log: () => {}, error: () => {} };

  it("should include rawOutput with string content for tool_result", () => {
@@ -425,7 +425,7 @@ describe("rawOutput in tool call updates", () => {
 });

 describe("Bash terminal output", () => {
-  const mockClient = {} as AgentSideConnection;
+  const mockClient = {} as AcpClient;
  const mockLogger: Logger = { log: () => {}, error: () => {} };

  const bashToolUse = {
@@ -502,6 +502,23 @@ describe("Bash terminal output", () => {
      });
    });

+    it("should route failed commands through the terminal when supportsTerminalOutput is true", () => {
+      const toolResult: ToolResultBlockParam = {
+        type: "tool_result",
+        tool_use_id: "toolu_bash",
+        content: "some error output",
+        is_error: true,
+      };
+      const update = toolUpdateFromToolResult(toolResult, bashToolUse, true);
+
+      expect(update.content).toEqual([{ type: "terminal", terminalId: "toolu_bash" }]);
+      expect(update._meta).toEqual({
+        terminal_info: { terminal_id: "toolu_bash" },
+        terminal_output: { terminal_id: "toolu_bash", data: "some error output" },
+        terminal_exit: { terminal_id: "toolu_bash", exit_code: 1, signal: null },
+      });
+    });
+
    it("should fall back to stderr when stdout is empty", () => {
      const toolResult = makeBashResult("", "some error output", 1);
      const update = toolUpdateFromToolResult(toolResult, bashToolUse, false);
@@ -610,10 +627,25 @@ describe("Bash terminal output", () => {
        });
      });

-      it("should use error handler when is_error is true (early return before Bash case)", () => {
+      it("should route is_error through the terminal when supportsTerminalOutput is true", () => {
        const toolResult = makeStringBashResult("command not found: bad_cmd", true);
        const update = toolUpdateFromToolResult(toolResult, bashToolUse, true);

+        // Failed Bash commands skip the early error return and reach the Bash
+        // case so the client receives terminal output with a non-zero exit code
+        // instead of plain markdown details.
+        expect(update.content).toEqual([{ type: "terminal", terminalId: "toolu_bash" }]);
+        expect(update._meta).toEqual({
+          terminal_info: { terminal_id: "toolu_bash" },
+          terminal_output: { terminal_id: "toolu_bash", data: "command not found: bad_cmd" },
+          terminal_exit: { terminal_id: "toolu_bash", exit_code: 1, signal: null },
+        });
+      });
+
+      it("should use error handler when is_error is true and terminal output is unsupported", () => {
+        const toolResult = makeStringBashResult("command not found: bad_cmd", true);
+        const update = toolUpdateFromToolResult(toolResult, bashToolUse, false);
+
        // is_error with content hits the early error return at the top of
        // toolUpdateFromToolResult, before reaching the Bash switch case.
        // So there's no terminal _meta, just error-formatted content.
@@ -670,6 +702,77 @@ describe("Bash terminal output", () => {
        });
      });
    });
+
+    describe("with image array tool_result (local Bash image output path)", () => {
+      // The local Bash tool emits image content as
+      // `[{ type: "image", source: { type: "base64", ... } }]` when a
+      // command produces an image (e.g. piping a base64 data URI).
+      const makeImageBashResult = (
+        data: string,
+        media_type: "image/png" | "image/jpeg" | "image/gif" | "image/webp" = "image/png",
+      ): ToolResultBlockParam => ({
+        type: "tool_result",
+        tool_use_id: "toolu_bash",
+        content: [
+          {
+            type: "image",
+            source: { type: "base64", media_type, data },
+          } as ImageBlockParam,
+        ],
+      });
+
+      it("should surface image content as ACP image content (terminal disabled)", () => {
+        const toolResult = makeImageBashResult("iVBORw0KGgo=");
+        const update = toolUpdateFromToolResult(toolResult, bashToolUse, false);
+
+        expect(update.content).toEqual([
+          {
+            type: "content",
+            content: { type: "image", data: "iVBORw0KGgo=", mimeType: "image/png" },
+          },
+        ]);
+        expect(update._meta).toBeUndefined();
+      });
+
+      it("should bypass terminal _meta even when supportsTerminalOutput is true", () => {
+        // Binary content cannot be streamed through the terminal-output
+        // _meta channel, so the fix returns ACP content directly and skips
+        // the terminal info/output/exit triple.
+        const toolResult = makeImageBashResult("iVBORw0KGgo=", "image/jpeg");
+        const update = toolUpdateFromToolResult(toolResult, bashToolUse, true);
+
+        expect(update.content).toEqual([
+          {
+            type: "content",
+            content: { type: "image", data: "iVBORw0KGgo=", mimeType: "image/jpeg" },
+          },
+        ]);
+        expect(update._meta).toBeUndefined();
+      });
+
+      it("should still surface multi-block content with text + image mixed", () => {
+        const toolResult: ToolResultBlockParam = {
+          type: "tool_result",
+          tool_use_id: "toolu_bash",
+          content: [
+            { type: "text", text: "generated:" },
+            {
+              type: "image",
+              source: { type: "base64", media_type: "image/png", data: "AAAA" },
+            } as ImageBlockParam,
+          ],
+        };
+        const update = toolUpdateFromToolResult(toolResult, bashToolUse, false);
+
+        expect(update.content).toEqual([
+          { type: "content", content: { type: "text", text: "generated:" } },
+          {
+            type: "content",
+            content: { type: "image", data: "AAAA", mimeType: "image/png" },
+          },
+        ]);
+      });
+    });
  });

  describe("toAcpNotifications with clientCapabilities", () => {
@@ -897,7 +1000,7 @@ describe("Bash terminal output", () => {
        sessionUpdate: async (notification: any) => {
          hookUpdates.push(notification);
        },
-      } as unknown as AgentSideConnection;
+      } as unknown as AcpClient;

      // Register hook callback by processing tool_use
      toAcpNotifications(
@@ -976,7 +1079,7 @@ describe("Bash terminal output", () => {
        sessionUpdate: async (notification: any) => {
          hookUpdates.push(notification);
        },
-      } as unknown as AgentSideConnection;
+      } as unknown as AcpClient;

      toAcpNotifications(
        [
@@ -1061,7 +1164,7 @@ describe("Bash terminal output", () => {
        sessionUpdate: async (notification: any) => {
          hookUpdates.push(notification);
        },
-      } as unknown as AgentSideConnection;
+      } as unknown as AcpClient;

      toAcpNotifications(
        [
@@ -1115,7 +1218,7 @@ describe("Bash terminal output", () => {
        sessionUpdate: async (notification: any) => {
          hookUpdates.push(notification);
        },
-      } as unknown as AgentSideConnection;
+      } as unknown as AcpClient;

      toAcpNotifications(
        [
@@ -1191,7 +1294,7 @@ describe("Bash terminal output", () => {
        sessionUpdate: async (notification: any) => {
          hookUpdates.push(notification);
        },
-      } as unknown as AgentSideConnection;
+      } as unknown as AcpClient;

      toAcpNotifications(
        [
@@ -1272,7 +1375,7 @@ describe("Bash terminal output", () => {
        sessionUpdate: async (notification: any) => {
          hookUpdates.push(notification);
        },
-      } as unknown as AgentSideConnection;
+      } as unknown as AcpClient;

      // Step 1: Process tool_use chunk — registers the PostToolUse hook callback
      const toolUseChunk = {
@@ -1372,7 +1475,7 @@ describe("Bash terminal output", () => {
        sessionUpdate: async (notification: any) => {
          hookUpdates.push(notification);
        },
-      } as unknown as AgentSideConnection;
+      } as unknown as AcpClient;

      // Process tool_use (registers hook)
      toAcpNotifications(
@@ -1537,7 +1640,7 @@ describe("planEntries - undefined input regression", () => {
 });

 describe("toAcpNotifications - TodoWrite with undefined input regression", () => {
-  const mockClient = {} as AgentSideConnection;
+  const mockClient = {} as AcpClient;
  const mockLogger: Logger = { log: () => {}, error: () => {} };

  it("should not throw when TodoWrite tool_use has undefined input", () => {
@@ -1672,7 +1775,7 @@ describe("applyTaskCreate / applyTaskUpdate", () => {
 });

 describe("toAcpNotifications - Task* tools", () => {
-  const mockClient = {} as AgentSideConnection;
+  const mockClient = {} as AcpClient;
  const mockLogger: Logger = { log: () => {}, error: () => {} };

  it("suppresses tool_call for TaskCreate/TaskUpdate/TaskList/TaskGet on tool_use", () => {
--- a/src/tools.ts
+++ b/src/tools.ts
@@ -8,6 +8,7 @@ import {
 import { HookCallback } from "@anthropic-ai/claude-agent-sdk";
 import {
  AgentInput,
+  AskUserQuestionInput,
  BashInput,
  FileEditInput,
  FileReadInput,
@@ -415,6 +416,24 @@ 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;
@@ -466,7 +485,8 @@ export function toolUpdateFromToolResult(
    "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);
@@ -510,7 +530,9 @@ export function toolUpdateFromToolResult(
      // 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: "..." }])
+      // 3. 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;

@@ -525,13 +547,20 @@ 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) {
--- a/tsconfig.json
+++ b/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,