fix(planning): keep the selected model pair on rebuilt planning agents

ensureSessionAgent rebuilt the agent with an empty provider/model pair while
preserving draftThinkingLevel, so resumed turns (respond, retry, rewind, drafts
resumed after the in-memory agent was dropped) fell through to the runtime's
built-in default model (anthropic/claude-opus-4-8) and hit api.anthropic.com
with a key the operator never configured. The non-streaming start had the same
hole. Resolve the pair from the persisted draft, then the lane's settings, on
every rebuild and start.

Also route planning through createResolvedAgentSession like chat/executor/merger
so CLI and plugin runtimes can own their own auth and planning emits
session:runtime-resolved.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

.changeset/planning-model-pair-rehydration.md

@@ -0,0 +1,7 @@
+---
+"@runfusion/fusion": patch
+---
+
+summary: Fix Planning Mode failing mid-interview with a provider auth error on a model you never selected.
+category: fix
+dev: `ensureSessionAgent` rebuilt the planning agent with an empty provider/model pair, so resumed turns (`/planning/respond`, `/planning/:id/retry`, rewind, drafts resumed after the in-memory agent was dropped) fell through to the runtime's built-in default model (`anthropic/claude-opus-4-8`) and hit api.anthropic.com with a key the operator never configured. The pair is now resolved from the persisted draft, then the lane's `resolvePlanningSettingsModel` result, on every rebuild and on the non-streaming start. Planning also now constructs sessions through `createResolvedAgentSession` (`sessionPurpose: "executor"`) like chat/executor/merger, so CLI and plugin runtimes can own their own auth and planning emits `session:runtime-resolved`.

packages/dashboard/src/__tests__/mcp-lane-forwarding.test.ts

@@ -2,7 +2,7 @@
 
 import { beforeEach, describe, expect, it, vi } from "vitest";
 
-const { createFnAgentMock, resolveMcpServersForStoreMock } = vi.hoisted(() => ({
+const { createFnAgentMock, createResolvedAgentSessionMock, resolveMcpServersForStoreMock } = vi.hoisted(() => {
   /*
   FNXC:DashboardTests 2026-07-18-12:20:
   Planning defaults clarificationEnabled=false, so createSession forces a summary after the
@@ -10,7 +10,7 @@ const { createFnAgentMock, resolveMcpServersForStoreMock } = vi.hoisted(() => ({
   complete payload so MCP-forwarding assertions exercise the default product path instead of
   throwing Clarification-disabled follow-up did not produce a summary.
   */
-  createFnAgentMock: vi.fn(async () => ({
+  const makeScriptedAgent = () => ({
     session: {
       state: { messages: [] as Array<{ role: string; content: string }> },
       prompt: vi.fn(async function (this: { state: { messages: Array<{ role: string; content: string }> } }, _message: string) {
@@ -42,12 +42,23 @@ const { createFnAgentMock, resolveMcpServersForStoreMock } = vi.hoisted(() => ({
       }),
       dispose: vi.fn(),
     },
-  })),
-  resolveMcpServersForStoreMock: vi.fn(async () => ({
-    servers: [{ name: "docs", transport: "stdio", command: "node", env: { TOKEN: "materialized-secret" } }],
-    errors: [],
-  })),
-}));
+  });
+  return {
+    createFnAgentMock: vi.fn(makeScriptedAgent),
+    /*
+    FNXC:PlanningRuntimeResolution 2026-07-24-16:20:
+    Planning now builds its session through the shared runtime-resolving seam
+    (`createResolvedAgentSession`), not a bare `createFnAgent` call, so the MCP-forwarding
+    contract for the planning lanes is asserted on that seam. The mission/target interview
+    lanes still construct through `createFnAgent` and keep their own mock.
+    */
+    createResolvedAgentSessionMock: vi.fn(makeScriptedAgent),
+    resolveMcpServersForStoreMock: vi.fn(async () => ({
+      servers: [{ name: "docs", transport: "stdio", command: "node", env: { TOKEN: "materialized-secret" } }],
+      errors: [],
+    })),
+  };
+});
 
 vi.mock("@fusion/core", async (importOriginal) => {
   const actual = await importOriginal<typeof import("@fusion/core")>();
@@ -68,6 +79,7 @@ vi.mock("@fusion/engine", async (importOriginal) => {
     createChatTaskDocumentTools: vi.fn(() => []),
     createWorkflowAuthoringTools: vi.fn(() => []),
     createFnAgent: createFnAgentMock,
+    createResolvedAgentSession: createResolvedAgentSessionMock,
     resolveMcpServersForStore: resolveMcpServersForStoreMock,
   };
 });
@@ -99,16 +111,17 @@ describe("dashboard MCP lane forwarding", () => {
   beforeEach(() => {
     __resetPlanningState();
     createFnAgentMock.mockClear();
+    createResolvedAgentSessionMock.mockClear();
     resolveMcpServersForStoreMock.mockClear();
   });
 
-  it("forwards the materialized MCP set to chat/planning createFnAgent sessions", async () => {
+  it("forwards the materialized MCP set to chat/planning agent sessions", async () => {
     const store = makePlanningStore();
 
     await createSession("127.0.0.1", "Build a feature", store, "/tmp/fusion-dashboard-test");
 
     expect(resolveMcpServersForStoreMock).toHaveBeenCalledWith(store);
-    expect(createFnAgentMock).toHaveBeenCalledWith(expect.objectContaining({
+    expect(createResolvedAgentSessionMock).toHaveBeenCalledWith(expect.objectContaining({
       cwd: "/tmp/fusion-dashboard-test",
       tools: "readonly",
       allowMcpToolsInReadonly: true,
@@ -121,7 +134,7 @@ describe("dashboard MCP lane forwarding", () => {
 
     await createSession("127.0.0.1", "Build without MCP", makePlanningStore(), "/tmp/fusion-dashboard-test");
 
-    expect(createFnAgentMock).toHaveBeenCalledWith(expect.objectContaining({
+    expect(createResolvedAgentSessionMock).toHaveBeenCalledWith(expect.objectContaining({
       tools: "readonly",
       allowMcpToolsInReadonly: true,
       mcpServers: [],
@@ -142,7 +155,7 @@ describe("dashboard MCP lane forwarding", () => {
     expect(startInitialTurn).toBeTypeOf("function");
     startInitialTurn?.();
 
-    await vi.waitFor(() => expect(createFnAgentMock).toHaveBeenCalledWith(expect.objectContaining({
+    await vi.waitFor(() => expect(createResolvedAgentSessionMock).toHaveBeenCalledWith(expect.objectContaining({
       tools: "readonly",
       allowMcpToolsInReadonly: true,
       mcpServers: [],

packages/dashboard/src/__tests__/planning-answered-question-reemit.test.ts

@@ -25,6 +25,10 @@ vi.mock("@fusion/engine", () => ({
     skillSource: "role-fallback" as const,
   }),
   createFnAgent: vi.fn(),
+  // FNXC:PlanningRuntimeResolution 2026-07-24-16:20: planning builds sessions through the
+  // shared runtime-resolving seam; these suites drive it via __setCreateFnAgent, so the
+  // engine export only needs to exist on the mock factory.
+  createResolvedAgentSession: vi.fn(),
   createWorkflowAuthoringTools: () => [],
   createChatTaskDocumentTools: () => [],
   createChatTaskLogsReadTool: () => ({}),

packages/dashboard/src/__tests__/planning-context-compaction.test.ts

@@ -27,6 +27,10 @@ vi.mock("@fusion/engine", () => ({
     skillSource: "role-fallback" as const,
   }),
   createFnAgent: vi.fn(),
+  // FNXC:PlanningRuntimeResolution 2026-07-24-16:20: planning builds sessions through the
+  // shared runtime-resolving seam; these suites drive it via __setCreateFnAgent, so the
+  // engine export only needs to exist on the mock factory.
+  createResolvedAgentSession: vi.fn(),
   createWorkflowAuthoringTools: () => [],
   createChatTaskDocumentTools: () => [],
   createChatTaskLogsReadTool: () => ({}),

packages/dashboard/src/__tests__/planning-e2e-plan-creation.test.ts

@@ -9,6 +9,10 @@ vi.mock("@fusion/engine", () => ({
   resolveMcpServersForStore: async () => ({ servers: [] }),
   buildSessionSkillContextSync: () => ({ skillSelectionContext: undefined, resolvedSkillNames: [], skillSource: "role-fallback" as const }),
   createFnAgent: vi.fn(),
+  // FNXC:PlanningRuntimeResolution 2026-07-24-16:20: planning builds sessions through the
+  // shared runtime-resolving seam; these suites drive it via __setCreateFnAgent, so the
+  // engine export only needs to exist on the mock factory.
+  createResolvedAgentSession: vi.fn(),
   createAgentTask: vi.fn(),
   createWorkflowAuthoringTools: () => [],
   createChatTaskDocumentTools: () => [],

packages/dashboard/src/__tests__/planning-generation-cancellation.test.ts

@@ -13,6 +13,10 @@ vi.mock("@fusion/engine", () => ({
     skillSource: "role-fallback" as const,
   }),
   createFnAgent: vi.fn(),
+  // FNXC:PlanningRuntimeResolution 2026-07-24-16:20: planning builds sessions through the
+  // shared runtime-resolving seam; these suites drive it via __setCreateFnAgent, so the
+  // engine export only needs to exist on the mock factory.
+  createResolvedAgentSession: vi.fn(),
   createWorkflowAuthoringTools: () => [],
   createChatTaskDocumentTools: () => [],
   createChatTaskLogsReadTool: () => ({}),

packages/dashboard/src/__tests__/planning-infinite-interview.test.ts

@@ -10,6 +10,10 @@ vi.mock("@fusion/engine", () => ({
   resolveMcpServersForStore: async () => ({ servers: [] }),
   buildSessionSkillContextSync: () => ({ skillSelectionContext: undefined, resolvedSkillNames: [], skillSource: "role-fallback" as const }),
   createFnAgent: vi.fn(),
+  // FNXC:PlanningRuntimeResolution 2026-07-24-16:20: planning builds sessions through the
+  // shared runtime-resolving seam; these suites drive it via __setCreateFnAgent, so the
+  // engine export only needs to exist on the mock factory.
+  createResolvedAgentSession: vi.fn(),
   createWorkflowAuthoringTools: () => [],
   createChatTaskDocumentTools: () => [],
   createChatTaskLogsReadTool: () => ({}),

packages/dashboard/src/__tests__/planning-model-pair-rehydration.test.ts

@@ -0,0 +1,244 @@
+// @vitest-environment node
+
+/*
+FNXC:PlanningModelRehydration 2026-07-24-16:20:
+Regression tests for the planning lane's provider/model pair.
+
+Reported symptom: a planning interview streamed Q1 and Q2 normally on the operator's
+configured (custom) provider, then a later turn died with
+`401 {"type":"error","error":{"type":"authentication_error","message":"invalid x-api-key"}}`
+from api.anthropic.com — a provider the operator never selected for planning. Every other
+Fusion lane using the same model worked.
+
+Cause: `ensureSessionAgent` rebuilt the agent with `undefined, undefined` for the model pair
+(while still preserving draftThinkingLevel), and the non-streaming start never passed one at
+all. With an incomplete pair the runtime forwards no `model`, so pi-coding-agent selects its
+OWN built-in default (`anthropic/claude-opus-4-8`) — the resumed turn silently left the
+operator's provider and hit Anthropic with a key they never configured.
+
+Invariant asserted here across ALL rebuild/start surfaces, not just the reported one:
+every planning agent construction receives a complete `defaultProvider`/`defaultModelId`
+pair — the pair persisted on the draft when present, otherwise the lane's settings-resolved
+pair — and planning is created through the shared runtime-resolving seam.
+*/
+
+import { beforeEach, describe, expect, it, vi } from "vitest";
+import { EventEmitter } from "node:events";
+import type { TaskStore } from "@fusion/core";
+
+const { createResolvedAgentSessionMock } = vi.hoisted(() => ({
+  createResolvedAgentSessionMock: vi.fn(),
+}));
+
+vi.mock("@fusion/engine", () => ({
+  listCliAdapterDescriptors: () => [],
+  resolveMcpServersForStore: async () => ({ servers: [] }),
+  buildSessionSkillContextSync: () => ({
+    skillSelectionContext: undefined,
+    resolvedSkillNames: ["fusion"],
+    skillSource: "role-fallback" as const,
+  }),
+  createResolvedAgentSession: createResolvedAgentSessionMock,
+  createWorkflowAuthoringTools: () => [],
+  createChatTaskDocumentTools: () => [],
+  createChatTaskLogsReadTool: () => ({}),
+}));
+
+import {
+  __resetPlanningState,
+  __setCreateFnAgent,
+  createSession,
+  createSessionWithAgent,
+  getSession,
+  planningStreamManager,
+  retrySession,
+  setAiSessionStore,
+  submitResponse,
+} from "../planning.js";
+
+const SETTINGS_PAIR = { planningProvider: "acme-custom", planningModelId: "acme-large" };
+
+function taskStore(settings: Record<string, unknown> = SETTINGS_PAIR): TaskStore {
+  return {
+    listTasks: vi.fn(async () => []),
+    getSettings: vi.fn(async () => settings),
+    getTask: vi.fn(async () => {
+      throw new Error("not found");
+    }),
+  } as unknown as TaskStore;
+}
+
+const QUESTION_JSON = JSON.stringify({
+  type: "question",
+  data: { id: "q-next", type: "single_select", question: "What next?" },
+});
+
+function scriptedAgent() {
+  const messages: Array<{ role: string; content: string }> = [];
+  return {
+    session: {
+      state: { messages },
+      prompt: vi.fn(async () => {
+        messages.push({ role: "assistant", content: QUESTION_JSON });
+      }),
+      dispose: vi.fn(),
+    },
+  };
+}
+
+function lastPair(factory: ReturnType<typeof vi.fn>): { provider?: string; model?: string } | undefined {
+  const pairs = capturedPairs(factory);
+  return pairs[pairs.length - 1];
+}
+
+async function waitFor(predicate: () => Promise<boolean> | boolean, attempts = 50): Promise<void> {
+  for (let i = 0; i < attempts; i++) {
+    if (await predicate()) return;
+    await new Promise((resolve) => setTimeout(resolve, 5));
+  }
+  throw new Error("condition not reached");
+}
+
+/** Every provider/model pair the planning lane handed to the session factory. */
+function capturedPairs(factory: ReturnType<typeof vi.fn>): Array<{ provider?: string; model?: string }> {
+  return factory.mock.calls.map(([options]) => ({
+    provider: options?.defaultProvider,
+    model: options?.defaultModelId,
+  }));
+}
+
+describe("planning provider/model pair is never dropped", () => {
+  let factory: ReturnType<typeof vi.fn>;
+
+  beforeEach(() => {
+    __resetPlanningState();
+    createResolvedAgentSessionMock.mockReset();
+    setAiSessionStore(Object.assign(new EventEmitter(), {
+      upsert: vi.fn(async () => {}),
+      get: vi.fn(async () => null),
+      updateThinking: vi.fn(),
+    }) as never);
+    factory = vi.fn(async () => scriptedAgent());
+  });
+
+  /*
+  Ordered first deliberately: it exercises the module's DEFAULT session factory, which the
+  `__setCreateFnAgent` seam in the later tests permanently replaces for this module instance.
+  */
+  it("creates planning sessions through the shared runtime-resolving seam", async () => {
+    createResolvedAgentSessionMock.mockImplementation(async () => scriptedAgent());
+
+    const sessionId = await createSessionWithAgent(
+      "10.9.0.1",
+      "Plan something small",
+      "/tmp/project",
+      taskStore(),
+      "acme-custom",
+      "acme-large",
+    );
+    planningStreamManager.consumeInitialTurn(sessionId)?.();
+    await waitFor(async () => Boolean((await getSession(sessionId))?.currentQuestion));
+
+    expect(createResolvedAgentSessionMock).toHaveBeenCalled();
+    const [options] = createResolvedAgentSessionMock.mock.calls[0];
+    // Runtime resolution is what lets a CLI/plugin runtime own its own auth instead of
+    // forcing every planning turn onto a direct, key-requiring provider call.
+    expect(options.sessionPurpose).toBe("executor");
+    expect(options.defaultProvider).toBe("acme-custom");
+    expect(options.defaultModelId).toBe("acme-large");
+  });
+
+  it("rebuilds a resumed agent on the pair the session started with (reported symptom)", async () => {
+    __setCreateFnAgent(factory as never);
+    const store = taskStore();
+    const sessionId = await createSessionWithAgent(
+      "10.9.0.2",
+      "Plan something small",
+      "/tmp/project",
+      store,
+      "acme-custom",
+      "acme-large",
+    );
+    planningStreamManager.consumeInitialTurn(sessionId)?.();
+    await waitFor(async () => Boolean((await getSession(sessionId))?.currentQuestion));
+
+    const session = (await getSession(sessionId))!;
+    const question = session.currentQuestion!;
+    // Exactly the reported state: the in-memory agent is gone (restart / eviction /
+    // draft resumed from History) and the next turn must rebuild it.
+    session.agent = undefined;
+
+    await submitResponse(sessionId, { [question.id]: "option-1" }, "/tmp/project", undefined, store);
+
+    expect(factory.mock.calls.length).toBeGreaterThan(1);
+    // The original failure: the rebuild call carried no pair at all, so the runtime fell
+    // through to its built-in anthropic default and 401'd on a raw x-api-key.
+    for (const pair of capturedPairs(factory)) {
+      expect(pair).toEqual({ provider: "acme-custom", model: "acme-large" });
+    }
+  });
+
+  it("falls back to the settings-resolved pair when the draft carries none", async () => {
+    __setCreateFnAgent(factory as never);
+    const store = taskStore();
+    // No explicit pair on start — the draft has nothing to replay from.
+    const sessionId = await createSessionWithAgent("10.9.0.3", "Plan something small", "/tmp/project", store);
+    planningStreamManager.consumeInitialTurn(sessionId)?.();
+    await waitFor(async () => Boolean((await getSession(sessionId))?.currentQuestion));
+
+    const session = (await getSession(sessionId))!;
+    const question = session.currentQuestion!;
+    session.agent = undefined;
+    session.draftModelProvider = undefined;
+    session.draftModelId = undefined;
+
+    await submitResponse(sessionId, { [question.id]: "option-1" }, "/tmp/project", undefined, store);
+
+    const rebuild = lastPair(factory);
+    expect(rebuild).toEqual({ provider: "acme-custom", model: "acme-large" });
+  });
+
+  it("keeps the pair on the retry rebuild surface", async () => {
+    __setCreateFnAgent(factory as never);
+    const store = taskStore();
+    const sessionId = await createSessionWithAgent(
+      "10.9.0.4",
+      "Plan something small",
+      "/tmp/project",
+      store,
+      "acme-custom",
+      "acme-large",
+    );
+    planningStreamManager.consumeInitialTurn(sessionId)?.();
+    await waitFor(async () => Boolean((await getSession(sessionId))?.currentQuestion));
+
+    const session = (await getSession(sessionId))!;
+    session.error = "AI returned no valid JSON. Retry this planning session or start a new one.";
+    session.agent = undefined;
+
+    await retrySession(sessionId, "/tmp/project", undefined, store);
+
+    const rebuild = lastPair(factory);
+    expect(rebuild).toEqual({ provider: "acme-custom", model: "acme-large" });
+  });
+
+  it("resolves a pair on the non-streaming start surface too", async () => {
+    __setCreateFnAgent(factory as never);
+    const store = taskStore();
+
+    await createSession("10.9.0.5", "Plan something small", store, "/tmp/project");
+
+    expect(capturedPairs(factory)[0]).toEqual({ provider: "acme-custom", model: "acme-large" });
+  });
+
+  it("still constructs an agent when no pair can be resolved anywhere", async () => {
+    __setCreateFnAgent(factory as never);
+    // Empty settings: no lane pair, no project default. The session must still start
+    // (runtime built-in default) rather than throwing at the operator mid-interview.
+    const store = taskStore({});
+
+    await createSession("10.9.0.6", "Plan something small", store, "/tmp/project");
+
+    expect(capturedPairs(factory)[0]).toEqual({ provider: undefined, model: undefined });
+  });
+});

packages/dashboard/src/__tests__/planning-provider-error-recovery.test.ts

@@ -26,6 +26,10 @@ vi.mock("@fusion/engine", () => ({
     skillSource: "role-fallback" as const,
   }),
   createFnAgent: vi.fn(),
+  // FNXC:PlanningRuntimeResolution 2026-07-24-16:20: planning builds sessions through the
+  // shared runtime-resolving seam; these suites drive it via __setCreateFnAgent, so the
+  // engine export only needs to exist on the mock factory.
+  createResolvedAgentSession: vi.fn(),
   createWorkflowAuthoringTools: () => [],
   createChatTaskDocumentTools: () => [],
   createChatTaskLogsReadTool: () => ({}),

packages/dashboard/src/__tests__/planning-question-regeneration.test.ts

@@ -30,6 +30,10 @@ vi.mock("@fusion/engine", () => ({
     skillSource: "role-fallback" as const,
   }),
   createFnAgent: vi.fn(),
+  // FNXC:PlanningRuntimeResolution 2026-07-24-16:20: planning builds sessions through the
+  // shared runtime-resolving seam; these suites drive it via __setCreateFnAgent, so the
+  // engine export only needs to exist on the mock factory.
+  createResolvedAgentSession: vi.fn(),
   createWorkflowAuthoringTools: () => [],
   createChatTaskDocumentTools: () => [],
   createChatTaskLogsReadTool: () => ({}),

packages/dashboard/src/__tests__/planning-turn-admission.test.ts

@@ -24,6 +24,10 @@ vi.mock("@fusion/engine", () => ({
     skillSource: "role-fallback" as const,
   }),
   createFnAgent: vi.fn(),
+  // FNXC:PlanningRuntimeResolution 2026-07-24-16:20: planning builds sessions through the
+  // shared runtime-resolving seam; these suites drive it via __setCreateFnAgent, so the
+  // engine export only needs to exist on the mock factory.
+  createResolvedAgentSession: vi.fn(),
   createWorkflowAuthoringTools: () => [],
   createChatTaskDocumentTools: () => [],
   createChatTaskLogsReadTool: () => ({}),

packages/dashboard/src/planning.ts

@@ -28,6 +28,7 @@ import {
   TASK_PRIORITIES,
   THINKING_LEVELS,
   formatPlanningPlanMd,
+  resolvePlanningSettingsModel,
   summarizeTitle,
   type PromptOverrideMap,
 } from "@fusion/core";
@@ -47,7 +48,7 @@ import {
   buildSessionSkillContextSync,
   createChatTaskDocumentTools,
   createChatTaskLogsReadTool,
-  createFnAgent as engineCreateFnAgent,
+  createResolvedAgentSession,
   createWorkflowAuthoringTools,
   resolveMcpServersForStore,
 } from "@fusion/engine";
@@ -81,8 +82,31 @@ type PlanningSessionOptions = {
   messageStore?: MessageStore;
   pluginRunner?: SkillPluginRunner;
 };
+/*
+FNXC:PlanningRuntimeResolution 2026-07-24-16:20:
+Planning sessions must be created through the shared `createResolvedAgentSession` seam that chat, executor, reviewer, merger, and heartbeat already use — NOT through a bare `createFnAgent` call. `createFnAgent` pins the session to the default pi runtime, so a planning selection could never route to a CLI/plugin runtime (claude-local/ACP, grok, omp) that owns its own auth; it always issued a direct HTTPS call to the provider endpoint with a Fusion-resolved key. That divergence is why a subscription/CLI-authenticated operator saw planning fail with a raw-key `401 invalid x-api-key` while every other lane worked. Routing here also emits the `session:runtime-resolved` run-audit event, so planning's resolved runtime and post-transform model pair are finally visible.
+
+Planning has no ambient task or bound agent, so it carries no runtimeHint of its own; `resolveRuntime` falls back to the default pi runtime exactly as before when no plugin runtime claims the purpose. `sessionPurpose: "executor"` matches the role planning already requests for skill selection (`buildSessionSkillContextSync(null, "executor", ...)`) and the purpose chat/QuickChat pass for the same reason.
+*/
 // eslint-disable-next-line @typescript-eslint/no-explicit-any
-let createFnAgent: any = engineCreateFnAgent;
+async function createPlanningRuntimeSession(options: any): Promise<AgentResult> {
+  const { pluginRunner, runtimeHint, settings, ...runtimeOptions } = options ?? {};
+  // Prefer the live engine binding so `ensureEngineReady()`-driven late loading
+  // still resolves, mirroring how this module already reaches engine helpers.
+  const create = (engineModule as unknown as {
+    createResolvedAgentSession?: typeof createResolvedAgentSession;
+  }).createResolvedAgentSession ?? createResolvedAgentSession;
+  return create({
+    sessionPurpose: "executor",
+    ...(runtimeHint ? { runtimeHint } : {}),
+    ...(pluginRunner ? { pluginRunner } : {}),
+    ...(settings ? { settings } : {}),
+    ...runtimeOptions,
+  });
+}
+
+// eslint-disable-next-line @typescript-eslint/no-explicit-any
+let createFnAgent: any = createPlanningRuntimeSession;
 
 function isThinkingLevel(value: unknown): value is ThinkingLevel {
   return THINKING_LEVELS.includes(value as ThinkingLevel);
@@ -1324,6 +1348,14 @@ async function runCreateSessionFirstTurn(
 
   const skillContext = buildSessionSkillContextSync(null, "executor", rootDir, pluginRunner);
 
+  /*
+  FNXC:PlanningModelRehydration 2026-07-24-16:20:
+  The non-streaming start is the same surface as the streaming start and the rebuild path:
+  it must resolve an explicit provider/model pair instead of leaving the runtime on its
+  built-in default. See resolveSessionPlanningModel for why an unset pair is a live bug.
+  */
+  const { provider: startProvider, modelId: startModelId } = await resolveSessionPlanningModel(session, store);
+
   /*
   FNXC:PlanningSkills 2026-06-17-19:33:
   Planning sessions are agent-acting lanes with planning and workflow tools, so they must request the same executor role fallback plus enabled plugin skills (for example ce-debug) as task execution sessions.
@@ -1351,6 +1383,10 @@ async function runCreateSessionFirstTurn(
       ...createChatTaskDocumentTools(store),
       createChatTaskLogsReadTool(store),
     ],
+    ...(startProvider && startModelId
+      ? { defaultProvider: startProvider, defaultModelId: startModelId }
+      : {}),
+    ...(pluginRunner ? { pluginRunner } : {}),
     onThinking: () => {
       // Non-streaming path ignores thinking output
     },
@@ -2103,6 +2139,8 @@ async function createPlanningAgent(
         }
       : {}),
     ...(thinkingLevel ? { defaultThinkingLevel: thinkingLevel } : {}),
+    // Runtime resolution needs the plugin runner to see plugin-provided runtimes.
+    ...(pluginRunner ? { pluginRunner } : {}),
     onThinking: (delta: string) => {
       if (callbacksInvalidated()) return;
       markPlanningGenerationProgress(session.id, delta);
@@ -2144,6 +2182,50 @@ function buildHistoryReplayPrompt(
   ].join("\n\n");
 }
 
+/*
+FNXC:PlanningModelRehydration 2026-07-24-16:20:
+A rebuilt planning agent MUST be given the same provider/model pair the session started on. `ensureSessionAgent` previously passed `undefined, undefined` while still preserving `draftThinkingLevel`, so any rebuild — `/planning/respond`, `/planning/:id/retry`, rewind, or a draft resumed after the in-memory agent was dropped — silently discarded the model selection. `createFnAgent`/`createResolvedAgentSession` forward NO `model` when the pair is incomplete (pi.ts `createSessionWithModel`), so pi-coding-agent then picked its OWN built-in default (`anthropic/claude-opus-4-8`). For an operator on a custom provider or a CLI/subscription runtime that means the resumed turn hit `api.anthropic.com` with a raw key they never configured and died on `401 invalid x-api-key` — mid-interview, after earlier turns on the correct model had already streamed. That is the exact reported symptom: the first turns work, the resumed turn does not.
+
+Resolution order mirrors the start route (`register-planning-subtask-routes.ts` → `resolvePlanningSettingsModel`): the pair persisted on the draft wins, then the lane's settings-resolved pair. Both halves must be present — a half-set pair is treated as unset, matching the runtime's own pair semantics.
+*/
+async function resolveSessionPlanningModel(
+  session: Session,
+  store: TaskStore,
+): Promise<{ provider?: string; modelId?: string }> {
+  if (session.draftModelProvider && session.draftModelId) {
+    return { provider: session.draftModelProvider, modelId: session.draftModelId };
+  }
+
+  try {
+    const settings = await store.getSettings();
+    const resolved = resolvePlanningSettingsModel(settings);
+    if (resolved.provider && resolved.modelId) {
+      // Cache onto the session so later rebuilds in this process stay on the
+      // same pair even if settings change mid-interview.
+      session.draftModelProvider = resolved.provider;
+      session.draftModelId = resolved.modelId;
+      return { provider: resolved.provider, modelId: resolved.modelId };
+    }
+  } catch (err) {
+    diagnostics.warn("Failed to resolve planning model for rebuilt agent", {
+      sessionId: session.id,
+      operation: "resolve-session-planning-model",
+      error: String(err),
+    });
+  }
+
+  /*
+  No complete pair resolved. The runtime falls back to its built-in default model,
+  which is provider-specific and may not match the operator's configured provider —
+  warn loudly rather than let that surface as an opaque provider auth error.
+  */
+  diagnostics.warn(
+    "Rebuilt planning agent has no resolved provider/model pair; the runtime will use its built-in default model",
+    { sessionId: session.id, operation: "resolve-session-planning-model" },
+  );
+  return {};
+}
+
 async function ensureSessionAgent(
   session: Session,
   rootDir: string | undefined,
@@ -2172,7 +2254,21 @@ async function ensureSessionAgent(
     );
   }
 
-  session.agent = await createPlanningAgent(session, effectiveRootDir, effectiveStore, undefined, undefined, session.draftThinkingLevel, promptOverrides, session.pluginRunner);
+  const { provider: resumeProvider, modelId: resumeModelId } = await resolveSessionPlanningModel(
+    session,
+    effectiveStore,
+  );
+
+  session.agent = await createPlanningAgent(
+    session,
+    effectiveRootDir,
+    effectiveStore,
+    resumeProvider,
+    resumeModelId,
+    session.draftThinkingLevel,
+    promptOverrides,
+    session.pluginRunner,
+  );
 
   if (historyForReplay.length === 0) {
     return;