Recover malformed agent interview responses (#2146)

## Summary
- preserve valid onboarding JSON returned in Pi thinking-only assistant
blocks
- retry one bounded JSON-only reformat turn when the model returns prose
or malformed output
- keep streamed output as a final extraction fallback instead of
overwriting it with an empty content array

## Verification
- `pnpm --filter @fusion/dashboard exec vitest run
src/__tests__/agent-onboarding.test.ts` — 20 passed
- `pnpm --filter @fusion/dashboard typecheck`
- `pnpm lint`
- `pnpm check:changesets --strict`
- live local-runtime AI Interview produced a structured
Hermes/computer-use onboarding question after restart

Follow-up to #2142, which fixed the missing planning-model fallback and
runtime-hint prompt.

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **Bug Fixes**
* Improved agent onboarding recovery when assistant replies include
thinking-only content or malformed JSON.
* Added a single automatic retry that re-formats invalid output into
valid onboarding JSON.
* Preserved structured “thinking” content as part of valid onboarding
responses.
* Normalized optional onboarding fields so null/empty/whitespace-only
values are treated as missing.
* Tightened Hermes automation so the runtime hint is set exactly to
`hermes`.
* **Tests**
* Added onboarding event synchronization and expanded coverage for
recovery and field normalization.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

.changeset/calm-interview-recovery.md

@@ -0,0 +1,7 @@
+---
+"@runfusion/fusion": patch
+---
+
+summary: Recover agent interviews when models return thinking-only or malformed JSON responses.
+category: fix
+dev: Preserves structured thinking output and retries one JSON-only reformat turn before surfacing an error.

packages/core/src/prompt-overrides.ts

@@ -301,6 +301,10 @@ For questions:
 For completion:
 {\n  "type": "complete",\n  "data": {\n    "title": "Task title",\n    "description": "Detailed description",\n    "suggestedSize": "S|M|L",\n    "suggestedDependencies": [],\n    "keyDeliverables": ["Item 1", "Item 2"]\n  }\n}`,
   },
+  /**
+   * FNXC:AgentOnboardingRuntime 2026-07-15-15:25:
+   * Agent onboarding must emit the exact `hermes` runtime hint for computer-use, desktop-automation, and UI-testing agents so runtime resolution selects the Hermes tool surface instead of treating a descriptive label as an unknown runtime.
+   */
   "agent-onboarding-system": {
     key: "agent-onboarding-system",
     name: "Agent Onboarding System",
@@ -325,6 +329,7 @@ Rules:
 - Prefer structuring instructionsText with these markdown sections when drafting: ## Description, ## Expertise, ## Priorities, ## Boundaries, ## Communication, ## Collaboration & Escalation
 - Freeform instructionsText is still acceptable for compatibility; sectioned structure is preferred for new agents
 - modelHint and runtimeHint are optional draft suggestions only (not final runtime selection)
+- When the user requests Hermes, computer use, desktop automation, or UI testing, use the exact runtimeHint "hermes"; never invent a descriptive runtime name
 - heartbeatProcedurePath, heartbeatIntervalMs, and heartbeatEnabled are optional draft hints only.`,
   },
   "subtask-breakdown-system": {

packages/dashboard/src/__tests__/agent-onboarding.test.ts

@@ -28,6 +28,7 @@ vi.mock("@fusion/engine", () => ({
 
 import {
   __resetAgentOnboardingState,
+  agentOnboardingStreamManager,
   cancelAgentOnboardingSession,
   createAgentOnboardingSessionPrompt,
   getAgentOnboardingSession,
@@ -38,6 +39,7 @@ import {
   retryAgentOnboardingSession,
   SessionNotFoundError,
   startAgentOnboardingSession,
+  stopAgentOnboardingGeneration,
 } from "../agent-onboarding.js";
 
 function createMockAgent(responses: string[]) {
@@ -65,6 +67,26 @@ async function waitFor(check: () => boolean, timeoutMs = 2000): Promise<void> {
   }
 }
 
+/*
+FNXC:AgentOnboarding 2026-07-15-14:32:
+Recovery regressions must await the onboarding event seam rather than poll wall-clock session state. Buffered-event replay covers generations that complete before the test subscribes.
+
+FNXC:AgentOnboarding 2026-07-15-16:42:
+Live subscribers receive AgentOnboardingStreamEvent objects keyed by `type`, while SessionEventBuffer replay records are keyed by `event` with JSON-serialized `data`. The test seam must honor both shapes.
+*/
+function waitForOnboardingEvent(sessionId: string, eventTypes: string[]): Promise<void> {
+  if (agentOnboardingStreamManager.getBufferedEvents(sessionId, 0).some((event) => eventTypes.includes(event.event))) {
+    return Promise.resolve();
+  }
+  return new Promise((resolve) => {
+    const unsubscribe = agentOnboardingStreamManager.subscribe(sessionId, (event) => {
+      if (!eventTypes.includes(event.type)) return;
+      unsubscribe();
+      resolve();
+    });
+  });
+}
+
 function createSkillPluginRunner(skills: Array<{ name: string; enabled?: boolean }>) {
   return {
     getPluginSkills: () => skills.map((skill) => ({ pluginId: "fusion-plugin-compound-engineering", skill })),
@@ -80,6 +102,7 @@ describe("agent-onboarding", () => {
 
   afterEach(() => {
     __resetAgentOnboardingState();
+    vi.useRealTimers();
   });
 
   it("parses question responses", () => {
@@ -173,6 +196,31 @@ describe("agent-onboarding", () => {
     ).toThrow(/Invalid summary/);
   });
 
+  it("normalizes absent-like optional string hints", () => {
+    const parsed = parseAgentOnboardingResponse(
+      JSON.stringify({
+        type: "complete",
+        data: {
+          name: "Hermes Desktop Tester",
+          role: "executor",
+          instructionsText: "Exercise desktop workflows with Hermes computer-use tools.",
+          thinkingLevel: "medium",
+          maxTurns: 25,
+          heartbeatProcedurePath: "  ",
+          modelHint: "",
+          runtimeHint: null,
+        },
+      }),
+    );
+
+    expect(parsed.type).toBe("complete");
+    if (parsed.type === "complete") {
+      expect(parsed.data.heartbeatProcedurePath).toBeUndefined();
+      expect(parsed.data.modelHint).toBeUndefined();
+      expect(parsed.data.runtimeHint).toBeUndefined();
+    }
+  });
+
   it("rejects malformed rich draft fields", () => {
     expect(() =>
       parseAgentOnboardingResponse(
@@ -184,7 +232,7 @@ describe("agent-onboarding", () => {
             instructionsText: "Valid instructions",
             thinkingLevel: "medium",
             maxTurns: 20,
-            heartbeatProcedurePath: "",
+            heartbeatProcedurePath: 42,
           },
         }),
       ),
@@ -290,6 +338,7 @@ describe("agent-onboarding", () => {
     const options = mockCreateFnAgent.mock.calls.at(-1)?.[0] as { systemPrompt?: string };
     expect(options.systemPrompt).toContain('"runtimeHint"');
     expect(options.systemPrompt).toContain("optional draft suggestions");
+    expect(options.systemPrompt).toContain('use the exact runtimeHint "hermes"');
     expect(options.systemPrompt).not.toContain("Do not include runtimeMode/model/runtimeHint");
   });
 
@@ -406,6 +455,224 @@ describe("agent-onboarding", () => {
     expect(options.skillSelection?.requestedSkillNames).toEqual(["fusion"]);
   });
 
+  it("preserves valid JSON from thinking-only assistant responses", async () => {
+    const response = JSON.stringify({
+      type: "question",
+      data: { id: "goal", type: "text", question: "What is the primary goal?" },
+    });
+    const messages: Array<{
+      role: string;
+      content: Array<{ type: "thinking"; thinking: string }>;
+    }> = [];
+    mockCreateFnAgent.mockImplementationOnce(async (options: unknown) => {
+      const callbacks = options as { onThinking?: (delta: string) => void };
+      return {
+        session: {
+          state: { messages },
+          prompt: vi.fn(async () => {
+            callbacks.onThinking?.(response);
+            messages.push({ role: "assistant", content: [{ type: "thinking", thinking: response }] });
+          }),
+          dispose: vi.fn(),
+        },
+      };
+    });
+
+    const sessionId = await startAgentOnboardingSession(
+      "127.0.0.1",
+      { intent: "thinking-only response", existingAgents: [], templates: [] },
+      process.cwd(),
+    );
+
+    await waitForOnboardingEvent(sessionId, ["question", "error"]);
+
+    const session = getAgentOnboardingSession(sessionId);
+    expect(session?.error).toBeUndefined();
+    expect(session?.currentQuestion?.id).toBe("goal");
+  });
+
+  it("prefers parseable thinking JSON over non-JSON text in the same assistant response", async () => {
+    const response = JSON.stringify({
+      type: "question",
+      data: { id: "goal", type: "text", question: "What is the primary goal?" },
+    });
+    const messages: Array<{
+      role: string;
+      content: Array<{ type: "text"; text: string } | { type: "thinking"; thinking: string }>;
+    }> = [];
+    const prompt = vi.fn(async () => {
+      messages.push({
+        role: "assistant",
+        content: [
+          { type: "thinking", thinking: response },
+          { type: "text", text: "I worked through the onboarding request." },
+        ],
+      });
+    });
+    mockCreateFnAgent.mockResolvedValueOnce({
+      session: { state: { messages }, prompt, dispose: vi.fn() },
+    });
+
+    const sessionId = await startAgentOnboardingSession(
+      "127.0.0.1",
+      { intent: "mixed thinking and text response", existingAgents: [], templates: [] },
+      process.cwd(),
+    );
+
+    await waitForOnboardingEvent(sessionId, ["question", "error"]);
+
+    const session = getAgentOnboardingSession(sessionId);
+    expect(session?.error).toBeUndefined();
+    expect(session?.currentQuestion?.id).toBe("goal");
+    expect(prompt).toHaveBeenCalledTimes(1);
+  });
+
+  it("uses streamed output when assistant string content is blank", async () => {
+    const response = JSON.stringify({
+      type: "question",
+      data: { id: "goal", type: "text", question: "What is the primary goal?" },
+    });
+    const messages: Array<{ role: string; content: string }> = [];
+    let prompt: ReturnType<typeof vi.fn>;
+    mockCreateFnAgent.mockImplementationOnce(async (options: unknown) => {
+      const callbacks = options as { onText?: (delta: string) => void };
+      prompt = vi.fn(async () => {
+        callbacks.onText?.(response);
+        messages.push({ role: "assistant", content: "   " });
+      });
+      return {
+        session: { state: { messages }, prompt, dispose: vi.fn() },
+      };
+    });
+
+    const sessionId = await startAgentOnboardingSession(
+      "127.0.0.1",
+      { intent: "blank assistant content", existingAgents: [], templates: [] },
+      process.cwd(),
+    );
+
+    await waitForOnboardingEvent(sessionId, ["question", "error"]);
+
+    const session = getAgentOnboardingSession(sessionId);
+    expect(session?.error).toBeUndefined();
+    expect(session?.currentQuestion?.id).toBe("goal");
+    expect(prompt!).toHaveBeenCalledTimes(1);
+  });
+
+  it("retries once when the model response is not valid JSON", async () => {
+    const agent = createMockAgent([
+      "I can help design that agent.",
+      JSON.stringify({
+        type: "question",
+        data: { id: "goal", type: "text", question: "What is the primary goal?" },
+      }),
+    ]);
+    mockCreateFnAgent.mockResolvedValueOnce(agent);
+
+    const sessionId = await startAgentOnboardingSession(
+      "127.0.0.1",
+      { intent: "recover malformed output", existingAgents: [], templates: [] },
+      process.cwd(),
+    );
+
+    await waitForOnboardingEvent(sessionId, ["question", "error"]);
+
+    const session = getAgentOnboardingSession(sessionId);
+    expect(session?.error).toBeUndefined();
+    expect(session?.currentQuestion?.id).toBe("goal");
+    expect(agent.session.prompt).toHaveBeenCalledTimes(2);
+  });
+
+  it("settles a stalled reformat turn when its timeout expires", async () => {
+    vi.useFakeTimers();
+    const messages: Array<{ role: string; content: string }> = [];
+    let callCount = 0;
+    const prompt = vi.fn(() => {
+      callCount += 1;
+      if (callCount === 1) {
+        messages.push({ role: "assistant", content: "not valid JSON" });
+        return Promise.resolve();
+      }
+      return new Promise<void>(() => {});
+    });
+    const dispose = vi.fn();
+    let expiredOnText: ((delta: string) => void) | undefined;
+    mockCreateFnAgent.mockImplementationOnce(async (options: unknown) => {
+      expiredOnText = (options as { onText?: (delta: string) => void }).onText;
+      return {
+        session: { state: { messages }, prompt, dispose },
+      };
+    });
+
+    const sessionId = await startAgentOnboardingSession(
+      "127.0.0.1",
+      { intent: "stalled reformat response", existingAgents: [], templates: [] },
+      process.cwd(),
+    );
+    await vi.advanceTimersByTimeAsync(0);
+    expect(prompt).toHaveBeenCalledTimes(2);
+
+    await vi.advanceTimersByTimeAsync(120_000);
+
+    const session = getAgentOnboardingSession(sessionId) as { error?: string; agent?: unknown; thinkingOutput: string } | undefined;
+    expect(session?.error).toBe("AI generation timed out. You can retry.");
+    expect(session?.agent).toBeUndefined();
+    expect(dispose).toHaveBeenCalledTimes(1);
+    expiredOnText?.("stale expired output");
+    expect(session?.thinkingOutput).toBe("");
+    expect(agentOnboardingStreamManager.getBufferedEvents(sessionId, 0)).toContainEqual(
+      expect.objectContaining({ event: "error", data: JSON.stringify("AI generation timed out. You can retry.") }),
+    );
+  });
+
+  it("settles a stopped generation without letting late cleanup detach a newer retry", async () => {
+    const response = JSON.stringify({
+      type: "question",
+      data: { id: "stale", type: "text", question: "Stale question?" },
+    });
+    const expiredMessages: Array<{ role: string; content: string }> = [];
+    let resolveExpiredPrompt!: () => void;
+    const expiredPrompt = vi.fn(() => new Promise<void>((resolve) => {
+      resolveExpiredPrompt = () => {
+        expiredMessages.push({ role: "assistant", content: response });
+        resolve();
+      };
+    }));
+    let signalRetryPromptStarted!: () => void;
+    const retryPromptStarted = new Promise<void>((resolve) => {
+      signalRetryPromptStarted = resolve;
+    });
+    const retryPrompt = vi.fn(() => {
+      signalRetryPromptStarted();
+      return new Promise<void>(() => {});
+    });
+    mockCreateFnAgent
+      .mockResolvedValueOnce({ session: { state: { messages: expiredMessages }, prompt: expiredPrompt, dispose: vi.fn() } })
+      .mockResolvedValueOnce({ session: { state: { messages: [] }, prompt: retryPrompt, dispose: vi.fn() } });
+
+    const sessionId = await startAgentOnboardingSession(
+      "127.0.0.1",
+      { intent: "stop and retry", existingAgents: [], templates: [] },
+      process.cwd(),
+    );
+    expect(expiredPrompt).toHaveBeenCalledTimes(1);
+    expect(stopAgentOnboardingGeneration(sessionId)).toBe(true);
+
+    const retry = retryAgentOnboardingSession(sessionId);
+    await retryPromptStarted;
+    expect(retryPrompt).toHaveBeenCalledTimes(1);
+
+    resolveExpiredPrompt();
+    await Promise.resolve();
+    await Promise.resolve();
+    await Promise.resolve();
+
+    expect(getAgentOnboardingSession(sessionId)?.currentQuestion).toBeUndefined();
+    expect(stopAgentOnboardingGeneration(sessionId)).toBe(true);
+    await retry;
+    expect(getAgentOnboardingSession(sessionId)?.error).toBe("Generation stopped by user. You can retry.");
+  });
+
   it("progresses through start -> question -> response -> final summary", async () => {
     mockCreateFnAgent.mockResolvedValueOnce(
       createMockAgent([

packages/dashboard/src/__tests__/chat.test.ts

@@ -39,11 +39,13 @@ vi.mock("@fusion/core", () => ({
 }));
 /*
 FNXC:DashboardChatTests 2026-07-12-08:15:
-chat.ts has 14 named runtime imports + `import * as engineModule` from @fusion/engine (lines 43-59).
 The engine module transitively imports many @fusion/core exports (AWAITING_APPROVAL_PAUSE_REASON,
 THINKING_LEVELS, etc.), so loading the real engine against a partial @fusion/core mock throws.
 Since this test only exercises resolveFileReferences (no real AI calls), stub every engine export
 chat.ts references so the real engine module never loads.
+
+FNXC:DashboardChatTests 2026-07-15-16:15:
+chat.ts now has 25 named runtime imports plus `import * as engineModule` from @fusion/engine. Keep this isolated mock complete as chat gains tool factories so the Gate's mock-completeness invariant does not regress.
 */
 vi.mock("@fusion/engine", () => ({
   createFnAgent: vi.fn(),
@@ -58,6 +60,17 @@ vi.mock("@fusion/engine", () => ({
   createChatArtifactTools: vi.fn(() => []),
   createChatTaskDocumentTools: vi.fn(() => []),
   createWorkflowAuthoringTools: vi.fn(() => []),
+  createTaskCreateTool: vi.fn(),
+  createTaskListTool: vi.fn(),
+  createTaskShowTool: vi.fn(),
+  createTaskSearchTool: vi.fn(),
+  createListAgentsTool: vi.fn(),
+  createDelegateTaskTool: vi.fn(),
+  createGetAgentConfigTool: vi.fn(),
+  createWebFetchTool: vi.fn(),
+  createGoalRetrievalTools: vi.fn(() => []),
+  createMemoryTools: vi.fn(() => []),
+  createResearchTools: vi.fn(() => []),
   resolveMcpServersForStore: vi.fn(async () => ({ servers: [], errors: [] })),
   resolveExecutorThinkingLevel: vi.fn(() => undefined),
   /*

packages/dashboard/src/agent-onboarding.ts

@@ -65,7 +65,27 @@ type SkillSelectionPluginRunner = Parameters<typeof buildSessionSkillContextSync
 const SESSION_TTL_MS = 30 * 60 * 1000;
 const CLEANUP_INTERVAL_MS = 5 * 60 * 1000;
 const GENERATION_TIMEOUT_MS = 120_000;
+class AgentOnboardingGenerationTimeoutError extends Error {
+  constructor() {
+    super("AI generation timed out. You can retry.");
+    this.name = "AgentOnboardingGenerationTimeoutError";
+  }
+}
+class AgentOnboardingGenerationStoppedError extends Error {
+  constructor() {
+    super("Generation stopped by user. You can retry.");
+    this.name = "AgentOnboardingGenerationStoppedError";
+  }
+}
+const REFORMAT_PROMPT =
+  "Your previous response could not be parsed as JSON. " +
+  'Respond with ONLY one valid JSON object using either {"type":"question","data":{...}} or {"type":"complete","data":{...}}. ' +
+  "No markdown, no explanation, just the JSON.";
 
+/*
+FNXC:AgentOnboarding 2026-07-15-16:15:
+Requests for Hermes, computer use, desktop automation, or UI testing must use the exact runtimeHint "hermes". Descriptive runtime names are not valid routing identifiers.
+*/
 export const AGENT_ONBOARDING_SYSTEM_PROMPT = `You are an agent onboarding assistant for the fn task board system.
 
 Your job is to guide users through creating a new agent with a short interview.
@@ -85,6 +105,7 @@ Rules:
 - Prefer structuring instructionsText with these markdown sections when drafting: ## Description, ## Expertise, ## Priorities, ## Boundaries, ## Communication, ## Collaboration & Escalation
 - Freeform instructionsText is still acceptable for compatibility; sectioned structure is preferred for new agents
 - modelHint and runtimeHint are optional draft suggestions only (not final runtime selection)
+- When the user requests Hermes, computer use, desktop automation, or UI testing, use the exact runtimeHint "hermes"; never invent a descriptive runtime name
 - heartbeatProcedurePath, heartbeatIntervalMs, and heartbeatEnabled are optional draft hints only.`;
 
 type OnboardingAgent = Awaited<ReturnType<typeof engineCreateFnAgent>>;
@@ -99,6 +120,7 @@ interface Session {
   error?: string;
   history: Array<{ question: PlanningQuestion; response: Record<string, unknown> }>;
   thinkingOutput: string;
+  agentEpoch: number;
   agent?: OnboardingAgent;
   rootDir: string;
   modelProvider?: string;
@@ -110,7 +132,12 @@ interface Session {
 }
 
 const sessions = new Map<string, Session>();
-const activeGenerations = new Map<string, { abortController: AbortController; timer: NodeJS.Timeout }>();
+type ActiveGeneration = {
+  abortController: AbortController;
+  timer: NodeJS.Timeout;
+  reject: (reason?: unknown) => void;
+};
+const activeGenerations = new Map<string, ActiveGeneration>();
 
 export class AgentOnboardingStreamManager extends EventEmitter {
   private readonly sessions = new Map<string, Set<AgentOnboardingStreamCallback>>();
@@ -171,6 +198,13 @@ function repairJson(text: string): string {
   return text.replace(/,\s*([}\]])/g, "$1");
 }
 
+function normalizeOptionalSummaryString(value: unknown, field: string): string | undefined {
+  if (value === undefined || value === null) return undefined;
+  if (typeof value !== "string") throw new Error(`Invalid summary.${field}`);
+  const trimmed = value.trim();
+  return trimmed || undefined;
+}
+
 export function parseAgentOnboardingResponse(text: string): { type: "question"; data: PlanningQuestion } | { type: "complete"; data: AgentOnboardingSummary } {
   const candidate = extractJsonCandidate(text);
   if (!candidate) throw new Error("AI returned no valid JSON");
@@ -207,12 +241,10 @@ export function parseAgentOnboardingResponse(text: string): { type: "question";
       throw new Error("Invalid summary.maxTurns");
     }
 
-    if (data.heartbeatProcedurePath !== undefined) {
-      if (typeof data.heartbeatProcedurePath !== "string" || !data.heartbeatProcedurePath.trim()) {
-        throw new Error("Invalid summary.heartbeatProcedurePath");
-      }
-      data.heartbeatProcedurePath = data.heartbeatProcedurePath.trim();
-    }
+    data.heartbeatProcedurePath = normalizeOptionalSummaryString(
+      data.heartbeatProcedurePath,
+      "heartbeatProcedurePath",
+    );
 
     if (data.heartbeatIntervalMs !== undefined) {
       if (typeof data.heartbeatIntervalMs !== "number" || !Number.isInteger(data.heartbeatIntervalMs) || data.heartbeatIntervalMs <= 0) {
@@ -224,13 +256,8 @@ export function parseAgentOnboardingResponse(text: string): { type: "question";
       throw new Error("Invalid summary.heartbeatEnabled");
     }
 
-    if (data.modelHint !== undefined && typeof data.modelHint !== "string") {
-      throw new Error("Invalid summary.modelHint");
-    }
-
-    if (data.runtimeHint !== undefined && typeof data.runtimeHint !== "string") {
-      throw new Error("Invalid summary.runtimeHint");
-    }
+    data.modelHint = normalizeOptionalSummaryString(data.modelHint, "modelHint");
+    data.runtimeHint = normalizeOptionalSummaryString(data.runtimeHint, "runtimeHint");
 
     return { type: "complete", data: data as AgentOnboardingSummary };
   }
@@ -307,6 +334,7 @@ export async function startAgentOnboardingSession(
     }),
     history: [],
     thinkingOutput: "",
+    agentEpoch: 0,
     rootDir,
     modelProvider,
     modelId,
@@ -324,6 +352,7 @@ export async function startAgentOnboardingSession(
 }
 
 async function createAgentOnboardingAgent(session: Session, store?: TaskStore): Promise<OnboardingAgent> {
+  const agentEpoch = ++session.agentEpoch;
   const systemPrompt = resolvePrompt("agent-onboarding-system", session.promptOverrides) || AGENT_ONBOARDING_SYSTEM_PROMPT;
   const skillContext = buildSessionSkillContextSync(null, "executor", session.rootDir, session.pluginRunner);
   const mcpServers = (await resolveMcpServersForStore(store ?? {})).servers;
@@ -346,10 +375,16 @@ async function createAgentOnboardingAgent(session: Session, store?: TaskStore):
     */
     ...(skillContext.skillSelectionContext ? { skillSelection: skillContext.skillSelectionContext } : {}),
     onThinking: (delta: string) => {
+      /*
+      FNXC:AgentOnboarding 2026-07-15-16:48:
+      Provider cancellation is best-effort. Ignore callbacks from an invalidated agent epoch so an expired prompt cannot contaminate a fresh retry's streamed output or SSE timeline.
+      */
+      if (session.agentEpoch !== agentEpoch) return;
       session.thinkingOutput += delta;
       agentOnboardingStreamManager.broadcast(session.id, { type: "thinking", data: delta });
     },
     onText: (delta: string) => {
+      if (session.agentEpoch !== agentEpoch) return;
       session.thinkingOutput += delta;
       agentOnboardingStreamManager.broadcast(session.id, { type: "thinking", data: delta });
     },
@@ -363,47 +398,122 @@ async function runGenerationWithTimeout<T>(session: Session, operation: () => Pr
     existing.abortController.abort();
   }
   const abortController = new AbortController();
+  /*
+  FNXC:AgentOnboarding 2026-07-15-16:36:
+  A generation timeout must settle the wrapper even when the provider ignores cancellation and its prompt promise never resolves. Race the operation against an explicit rejection; abort remains best-effort cleanup, while continueConversation owns the single terminal error event.
+  */
+  let rejectGeneration!: (reason?: unknown) => void;
+  const interruption = new Promise<never>((_, reject) => {
+    rejectGeneration = reject;
+  });
   const timer = setTimeout(() => {
-    session.error = "AI generation timed out. You can retry.";
-    agentOnboardingStreamManager.broadcast(session.id, { type: "error", data: session.error });
     abortController.abort();
+    rejectGeneration(new AgentOnboardingGenerationTimeoutError());
   }, GENERATION_TIMEOUT_MS);
-  activeGenerations.set(session.id, { abortController, timer });
+  /*
+  FNXC:AgentOnboarding 2026-07-15-17:32:
+  User stop must settle the active prompt race even when the provider ignores cancellation. Keep a per-generation reject handle and identity-guard cleanup so a late stopped prompt cannot publish stale state or delete a newer retry's registration.
+  */
+  const activeGeneration: ActiveGeneration = { abortController, timer, reject: rejectGeneration };
+  activeGenerations.set(session.id, activeGeneration);
   try {
-    return await operation();
+    return await Promise.race([operation(), interruption]);
   } finally {
     clearTimeout(timer);
-    activeGenerations.delete(session.id);
+    if (activeGenerations.get(session.id) === activeGeneration) {
+      activeGenerations.delete(session.id);
+    }
   }
 }
 
+type OnboardingMessage = {
+  role: string;
+  content?: string | Array<
+    | { type: "text"; text: string }
+    | { type: "thinking"; thinking: string }
+    | { type: string }
+  >;
+};
+
+function extractLastAssistantResponse(messages: unknown, streamedOutput: string): string {
+  const assistant = (Array.isArray(messages) ? messages : [])
+    .filter((message): message is OnboardingMessage => (
+      typeof message === "object"
+      && message !== null
+      && "role" in message
+      && (message as { role?: unknown }).role === "assistant"
+    ))
+    .pop();
+  if (typeof assistant?.content === "string") return assistant.content.trim() || streamedOutput;
+  if (!Array.isArray(assistant?.content)) return streamedOutput;
+
+  const textContent = assistant.content
+    .filter((block): block is { type: "text"; text: string } => block.type === "text" && "text" in block && typeof block.text === "string")
+    .map((block) => block.text)
+    .join("");
+  const thinkingContent = assistant.content
+    .filter((block): block is { type: "thinking"; thinking: string } => block.type === "thinking" && "thinking" in block && typeof block.thinking === "string")
+    .map((block) => block.thinking)
+    .join("");
+  /*
+  FNXC:AgentOnboarding 2026-07-15-16:15:
+  Pi can emit valid onboarding JSON in a thinking block alongside explanatory text. Select the first parseable text, thinking, or streamed candidate; only preserve the historical text-first fallback when none parses so the bounded reformat turn still receives the model's visible response.
+  */
+  const candidates = [textContent, thinkingContent, streamedOutput]
+    .map((candidate) => candidate.trim())
+    .filter((candidate) => candidate.length > 0);
+  for (const candidate of candidates) {
+    try {
+      parseAgentOnboardingResponse(candidate);
+      return candidate;
+    } catch {
+      // Try the next model-output surface before invoking the bounded recovery turn.
+    }
+  }
+  return candidates[0] ?? streamedOutput;
+}
+
 async function continueConversation(session: Session, message: string): Promise<void> {
   if (!session.agent) throw new Error("Session agent not initialized");
   const agent = session.agent;
   session.thinkingOutput = "";
   try {
-    await runGenerationWithTimeout(session, async () => {
-      await agent.session.prompt(message);
-      const assistant = (agent.session.state.messages as Array<{ role: string; content?: string | Array<{ type: string; text: string }> }>).filter((m) => m.role === "assistant").pop();
-      let responseText = session.thinkingOutput;
-      if (assistant?.content) {
-        if (typeof assistant.content === "string") responseText = assistant.content;
-        else responseText = assistant.content.filter((c) => c.type === "text").map((c) => c.text).join("");
-      }
-      const parsed = parseAgentOnboardingResponse(responseText);
-      session.error = undefined;
-      session.updatedAt = new Date();
-      if (parsed.type === "question") {
-        session.currentQuestion = parsed.data;
-        agentOnboardingStreamManager.broadcast(session.id, { type: "question", data: parsed.data });
-      } else {
-        session.summary = parsed.data;
-        session.currentQuestion = undefined;
-        agentOnboardingStreamManager.broadcast(session.id, { type: "summary", data: parsed.data });
-        agentOnboardingStreamManager.broadcast(session.id, { type: "complete" });
-      }
-    });
+    await runGenerationWithTimeout(session, () => agent.session.prompt(message));
+    let responseText = extractLastAssistantResponse(agent.session.state.messages, session.thinkingOutput);
+    let parsed: ReturnType<typeof parseAgentOnboardingResponse>;
+    try {
+      parsed = parseAgentOnboardingResponse(responseText);
+    } catch {
+      /*
+      FNXC:AgentOnboarding 2026-07-15-14:32:
+      A malformed first interview response must not consume the recovery turn's generation budget. Run the reformat prompt as a distinct timed generation so it receives the full timeout and remains independently stoppable.
+      */
+      session.thinkingOutput = "";
+      await runGenerationWithTimeout(session, () => agent.session.prompt(REFORMAT_PROMPT));
+      responseText = extractLastAssistantResponse(agent.session.state.messages, session.thinkingOutput);
+      parsed = parseAgentOnboardingResponse(responseText);
+    }
+    session.error = undefined;
+    session.updatedAt = new Date();
+    if (parsed.type === "question") {
+      session.currentQuestion = parsed.data;
+      agentOnboardingStreamManager.broadcast(session.id, { type: "question", data: parsed.data });
+    } else {
+      session.summary = parsed.data;
+      session.currentQuestion = undefined;
+      agentOnboardingStreamManager.broadcast(session.id, { type: "summary", data: parsed.data });
+      agentOnboardingStreamManager.broadcast(session.id, { type: "complete" });
+    }
   } catch (err) {
+    if (err instanceof AgentOnboardingGenerationStoppedError) {
+      return;
+    }
+    if (err instanceof AgentOnboardingGenerationTimeoutError) {
+      const expiredAgent = session.agent;
+      session.agentEpoch += 1;
+      session.agent = undefined;
+      try { expiredAgent?.session.dispose?.(); } catch { /* best-effort provider cancellation */ }
+    }
     session.error = err instanceof Error ? err.message : String(err);
     agentOnboardingStreamManager.broadcast(session.id, { type: "error", data: session.error });
   }
@@ -481,9 +591,14 @@ export function stopAgentOnboardingGeneration(sessionId: string): boolean {
   activeGenerations.delete(sessionId);
   const session = sessions.get(sessionId);
   if (session) {
+    const expiredAgent = session.agent;
+    session.agentEpoch += 1;
+    session.agent = undefined;
+    try { expiredAgent?.session.dispose?.(); } catch { /* best-effort provider cancellation */ }
     session.error = "Generation stopped by user. You can retry.";
     agentOnboardingStreamManager.broadcast(session.id, { type: "error", data: session.error });
   }
+  active.reject(new AgentOnboardingGenerationStoppedError());
   return true;
 }