gsd-pi 2.81.0-dev.72a81bdf3 → 2.82.0-dev.20138ae42

package/src/resources/extensions/gsd/tests/auto-orchestrator.test.ts

@@ -4,9 +4,14 @@
 import test from "node:test";
 import assert from "node:assert/strict";
 
-import { createAutoOrchestrator } from "../auto/orchestrator.js";
+import { createAutoOrchestrator, STUCK_WINDOW_SIZE } from "../auto/orchestrator.js";
 import type { AutoOrchestratorDeps } from "../auto/contracts.js";
 import type { GSDState } from "../types.js";
+import { createWiredDispatchAdapter } from "../auto.js";
+import { resolveDispatch, type DispatchContext } from "../auto-dispatch.js";
+import { RuleRegistry, setRegistry, resetRegistry } from "../rule-registry.js";
+import type { UnifiedRule } from "../rule-types.js";
+import { supportsStructuredQuestions } from "../workflow-mcp.js";
 
 function makeState(): GSDState {
   return {
@@ -62,9 +67,13 @@ function makeDeps(overrides: Partial<AutoOrchestratorDeps> = {}): { deps: AutoOr
       async cleanupOnStop() { calls.push("worktree.cleanup"); },
     },
     health: {
+      checkResourcesStale() {
+        calls.push("health.stale");
+        return null;
+      },
       async preAdvanceGate() {
         calls.push("health.pre");
-        return { allow: true };
+        return { kind: "pass" };
       },
       async postAdvanceRecord() { calls.push("health.post"); },
     },
@@ -75,6 +84,9 @@ function makeDeps(overrides: Partial<AutoOrchestratorDeps> = {}): { deps: AutoOr
     notifications: {
       async notifyLifecycle(event) { calls.push(`notify:${event.name}`); },
     },
+    uokGate: {
+      async emit(input) { calls.push(`gate:${input.gateId}:${input.outcome}`); },
+    },
   };
 
   return { deps: { ...deps, ...overrides }, calls };
@@ -87,6 +99,7 @@ test("start() advances and records active unit", async () => {
   const result = await orchestrator.start({ basePath: "/tmp/project", trigger: "manual" });
 
   assert.equal(result.kind, "advanced");
+  assert.deepEqual(result.unit, { unitType: "execute-task", unitId: "T01" });
   const status = orchestrator.getStatus();
   assert.equal(status.phase, "running");
   assert.deepEqual(status.activeUnit, { unitType: "execute-task", unitId: "T01" });
@@ -95,9 +108,10 @@ test("start() advances and records active unit", async () => {
 });
 
 test("advance() returns blocked when health gate denies", async () => {
-  const { deps } = makeDeps({
+  const { deps, calls } = makeDeps({
     health: {
-      async preAdvanceGate() { return { allow: false, reason: "doctor-block" }; },
+      checkResourcesStale: () => null,
+      async preAdvanceGate() { return { kind: "fail", reason: "doctor-block" }; },
       async postAdvanceRecord() {},
     },
   });
@@ -107,6 +121,69 @@ test("advance() returns blocked when health gate denies", async () => {
 
   assert.equal(result.kind, "blocked");
   assert.equal(result.reason, "doctor-block");
+  assert.equal(result.action, "pause");
+  assert.ok(calls.includes("gate:pre-dispatch-health-gate:manual-attention"));
+});
+
+test("advance() returns blocked stop when resources are stale", async () => {
+  const { deps, calls } = makeDeps({
+    health: {
+      checkResourcesStale: () => "resources changed since session start",
+      async preAdvanceGate() { return { kind: "pass" }; },
+      async postAdvanceRecord() {},
+    },
+  });
+  const orchestrator = createAutoOrchestrator(deps);
+
+  const result = await orchestrator.advance();
+
+  assert.equal(result.kind, "blocked");
+  assert.equal(result.reason, "resources changed since session start");
+  assert.equal(result.action, "stop");
+  assert.ok(calls.includes("gate:resource-version-guard:fail"));
+  assert.ok(!calls.includes("health.pre"));
+  assert.ok(!calls.includes("state.reconcile"));
+});
+
+test("advance() continues past pre-dispatch health gate when it throws", async () => {
+  const { deps, calls } = makeDeps({
+    health: {
+      checkResourcesStale: () => null,
+      async preAdvanceGate() { return { kind: "threw", error: new Error("boom") }; },
+      async postAdvanceRecord() {},
+    },
+  });
+  const orchestrator = createAutoOrchestrator(deps);
+
+  const result = await orchestrator.advance();
+
+  assert.equal(result.kind, "advanced");
+  assert.ok(calls.includes("gate:pre-dispatch-health-gate:manual-attention"));
+  assert.ok(calls.includes("state.reconcile"));
+  assert.ok(calls.includes("dispatch.decide"));
+});
+
+test("advance() forwards fixesApplied into pre-dispatch-health-gate pass findings", async () => {
+  let observed = "";
+  const { deps } = makeDeps({
+    health: {
+      checkResourcesStale: () => null,
+      async preAdvanceGate() { return { kind: "pass", fixesApplied: ["fix-a", "fix-b"] }; },
+      async postAdvanceRecord() {},
+    },
+    uokGate: {
+      async emit(input) {
+        if (input.gateId === "pre-dispatch-health-gate" && input.outcome === "pass") {
+          observed = input.findings ?? "";
+        }
+      },
+    },
+  });
+  const orchestrator = createAutoOrchestrator(deps);
+
+  await orchestrator.advance();
+
+  assert.equal(observed, "fix-a, fix-b");
 });
 
 test("advance() follows the ADR-015 invariant sequence before journaling advance", async () => {
@@ -116,9 +193,13 @@ test("advance() follows the ADR-015 invariant sequence before journaling advance
   const result = await orchestrator.advance();
 
   assert.equal(result.kind, "advanced");
+  assert.deepEqual(result.unit, { unitType: "execute-task", unitId: "T01" });
   assert.deepEqual(calls, [
     "runtime.lock",
+    "health.stale",
+    "gate:resource-version-guard:pass",
     "health.pre",
+    "gate:pre-dispatch-health-gate:pass",
     "state.reconcile",
     "dispatch.decide",
     "tool.compile",
@@ -144,6 +225,7 @@ test("advance() blocks before dispatch when State Reconciliation blocks", async
 
   assert.equal(result.kind, "blocked");
   assert.equal(result.reason, "state drift blocked");
+  assert.equal(result.action, "pause");
   assert.ok(!calls.includes("dispatch.decide"));
   assert.ok(calls.includes("journal:advance-blocked"));
 });
@@ -163,6 +245,7 @@ test("advance() blocks before Runtime persistence when Tool Contract fails", asy
 
   assert.equal(result.kind, "blocked");
   assert.equal(result.reason, "unknown Unit");
+  assert.equal(result.action, "pause");
   assert.ok(!calls.includes("worktree.prepare"));
   assert.ok(!calls.includes("journal:advance"));
   assert.ok(calls.includes("journal:advance-blocked"));
@@ -185,6 +268,7 @@ test("advance() blocks before Runtime persistence when Worktree Safety fails", a
 
   assert.equal(result.kind, "blocked");
   assert.equal(result.reason, "worktree invalid");
+  assert.equal(result.action, "pause");
   assert.ok(!calls.includes("journal:advance"));
   assert.ok(!calls.includes("worktree.sync"));
   assert.ok(calls.includes("journal:advance-blocked"));
@@ -204,6 +288,51 @@ test("advance() stops when dispatch has no next unit", async () => {
   assert.equal(orchestrator.getStatus().phase, "stopped");
 });
 
+test("advance() surfaces dispatch blocker reason instead of generic no remaining units", async () => {
+  const { deps, calls } = makeDeps({
+    dispatch: {
+      async decideNextUnit() {
+        return {
+          kind: "blocked",
+          reason: "Milestone M001 validation verdict is needs-remediation but all slices are complete.",
+          action: "pause",
+        };
+      },
+    },
+  });
+  const orchestrator = createAutoOrchestrator(deps);
+
+  const result = await orchestrator.advance();
+
+  assert.equal(result.kind, "blocked");
+  if (result.kind !== "blocked") return;
+  assert.equal(result.reason, "Milestone M001 validation verdict is needs-remediation but all slices are complete.");
+  assert.equal(result.action, "pause");
+  assert.ok(calls.includes("journal:advance-blocked"));
+  assert.ok(!calls.includes("journal:advance-stopped"));
+});
+
+test("resume() returns blocked when advance detects a dispatch blocker", async () => {
+  const { deps } = makeDeps({
+    dispatch: {
+      async decideNextUnit() {
+        return {
+          kind: "blocked",
+          reason: "remediation required",
+          action: "pause",
+        };
+      },
+    },
+  });
+  const orchestrator = createAutoOrchestrator(deps);
+
+  const result = await orchestrator.resume();
+
+  assert.equal(result.kind, "blocked");
+  if (result.kind !== "blocked") return;
+  assert.equal(result.reason, "remediation required");
+});
+
 test("advance() uses recovery on error", async () => {
   const { deps, calls } = makeDeps({
     runtime: {
@@ -232,8 +361,10 @@ test("advance() is idempotent for the same active unit", async () => {
   const second = await orchestrator.advance();
 
   assert.equal(first.kind, "advanced");
+  assert.deepEqual(first.unit, { unitType: "execute-task", unitId: "T01" });
   assert.equal(second.kind, "blocked");
   assert.equal(second.reason, "idempotent advance: unit already active");
+  assert.equal(second.action, "stop");
 
   const prepareCalls = calls.filter((c) => c === "worktree.prepare").length;
   assert.equal(prepareCalls, 1);
@@ -468,3 +599,355 @@ test("stop() cleans up worktree and transitions to stopped", async () => {
   assert.ok(calls.includes("journal:stop"));
   assert.ok(calls.includes("notify:stop"));
 });
+
+// ────────────────────────────────────────────────────────────────────────
+// Stuck-loop ring buffer (issue #5787)
+// ────────────────────────────────────────────────────────────────────────
+
+test("STUCK_WINDOW_SIZE matches the legacy auto/phases.ts constant", () => {
+  assert.equal(STUCK_WINDOW_SIZE, 6);
+});
+
+test("stuck-loop: empty ring on a freshly constructed orchestrator advances normally", async () => {
+  const { deps } = makeDeps();
+  const orchestrator = createAutoOrchestrator(deps);
+
+  const result = await orchestrator.advance();
+
+  assert.equal(result.kind, "advanced");
+});
+
+test("stuck-loop: partial fill of mixed units does not block", async () => {
+  // Alternate A/B for STUCK_WINDOW_SIZE rounds. No single key saturates the
+  // window, so neither idempotency nor stuck-loop should fire.
+  let i = 0;
+  const sequence = ["A", "B", "A", "B", "A", "B"];
+  const { deps } = makeDeps({
+    dispatch: {
+      async decideNextUnit() {
+        const id = sequence[i++ % sequence.length];
+        return { unitType: "execute-task", unitId: id, reason: "ready", preconditions: [] };
+      },
+    },
+  });
+  const orchestrator = createAutoOrchestrator(deps);
+
+  for (let round = 0; round < STUCK_WINDOW_SIZE; round++) {
+    const result = await orchestrator.advance();
+    assert.equal(result.kind, "advanced", `round ${round} should advance, got ${result.kind}`);
+  }
+});
+
+test("stuck-loop: ring saturated with same unit blocks with action 'stop' and stuck-loop reason", async () => {
+  // Dispatch picks the same unit every time. The first advance succeeds.
+  // Calls 2..STUCK_WINDOW_SIZE-1 are idempotency-blocked while the ring fills.
+  // The STUCK_WINDOW_SIZE'th call sees a saturated ring and returns stuck-loop.
+  const { deps } = makeDeps();
+  const orchestrator = createAutoOrchestrator(deps);
+
+  const results: Awaited<ReturnType<typeof orchestrator.advance>>[] = [];
+  for (let i = 0; i < STUCK_WINDOW_SIZE; i++) {
+    results.push(await orchestrator.advance());
+  }
+
+  // First call advances.
+  assert.equal(results[0].kind, "advanced");
+
+  // Intermediate calls are blocked by idempotency (not stuck-loop yet).
+  for (let i = 1; i < STUCK_WINDOW_SIZE - 1; i++) {
+    const r = results[i];
+    assert.equal(r.kind, "blocked", `round ${i} should be blocked`);
+    if (r.kind !== "blocked") return;
+    assert.equal(r.reason, "idempotent advance: unit already active");
+    assert.equal(r.action, "stop");
+  }
+
+  // The final call (ring now holds STUCK_WINDOW_SIZE copies) returns stuck-loop.
+  const last = results[STUCK_WINDOW_SIZE - 1];
+  assert.equal(last.kind, "blocked");
+  if (last.kind !== "blocked") return;
+  assert.equal(last.action, "stop");
+  assert.equal(last.reason, `stuck-loop: execute-task:T01 picked ${STUCK_WINDOW_SIZE} times`);
+});
+
+test("stuck-loop: idempotency block continues to fire with its own reason before saturation", async () => {
+  // Two identical calls should produce idempotent (not stuck-loop). Ensures the
+  // existing idempotency block is not absorbed by the new check.
+  const { deps } = makeDeps();
+  const orchestrator = createAutoOrchestrator(deps);
+
+  const first = await orchestrator.advance();
+  const second = await orchestrator.advance();
+
+  assert.equal(first.kind, "advanced");
+  assert.equal(second.kind, "blocked");
+  assert.equal(second.reason, "idempotent advance: unit already active");
+  assert.equal(second.action, "stop");
+});
+
+test("stuck-loop: start() resets the ring so a fresh saturation cycle is required", async () => {
+  // Fill the ring to one short of saturation, then start() — the ring should
+  // be cleared, and the next advance must succeed instead of going stuck.
+  const { deps } = makeDeps();
+  const orchestrator = createAutoOrchestrator(deps);
+
+  for (let i = 0; i < STUCK_WINDOW_SIZE - 1; i++) {
+    await orchestrator.advance();
+  }
+
+  const restarted = await orchestrator.start({ basePath: "/tmp/project", trigger: "manual" });
+  assert.equal(restarted.kind, "advanced");
+
+  // Immediately after start(), the next advance is idempotent (one element in
+  // ring), not stuck-loop, confirming the ring was reset.
+  const next = await orchestrator.advance();
+  assert.equal(next.kind, "blocked");
+  assert.equal(next.reason, "idempotent advance: unit already active");
+});
+
+test("stuck-loop: resume() resets the ring", async () => {
+  const { deps } = makeDeps();
+  const orchestrator = createAutoOrchestrator(deps);
+
+  for (let i = 0; i < STUCK_WINDOW_SIZE - 1; i++) {
+    await orchestrator.advance();
+  }
+
+  const resumed = await orchestrator.resume();
+  assert.equal(resumed.kind, "advanced");
+
+  const next = await orchestrator.advance();
+  assert.equal(next.kind, "blocked");
+  assert.equal(next.reason, "idempotent advance: unit already active");
+});
+
+test("stuck-loop: stop() resets the ring", async () => {
+  const { deps } = makeDeps();
+  const orchestrator = createAutoOrchestrator(deps);
+
+  for (let i = 0; i < STUCK_WINDOW_SIZE - 1; i++) {
+    await orchestrator.advance();
+  }
+
+  const stopped = await orchestrator.stop("user-request");
+  assert.equal(stopped.kind, "stopped");
+
+  // Ring is cleared by stop(). A subsequent advance is a fresh first-touch.
+  const next = await orchestrator.advance();
+  assert.equal(next.kind, "advanced");
+});
+
+test("stuck-loop: journal records the stuck-loop reason on advance-blocked", async () => {
+  const { deps, calls } = makeDeps();
+  const orchestrator = createAutoOrchestrator(deps);
+
+  for (let i = 0; i < STUCK_WINDOW_SIZE; i++) {
+    await orchestrator.advance();
+  }
+
+  assert.ok(calls.includes("journal:advance-blocked"));
+});
+
+// ─── #5789 parity: wired dispatch adapter mirrors runDispatch's resolveDispatch call ───
+
+test("wired DispatchAdapter forwards session-derived dispatch inputs identically to runDispatch", async () => {
+  const stateSnapshot = makeState();
+
+  // Install a capturing registry so we observe the DispatchContext both code paths
+  // build, and force a deterministic dispatch action so the parity assertion is
+  // about *inputs*, not rule evaluation.
+  const captured: DispatchContext[] = [];
+  const captureRule: UnifiedRule = {
+    name: "test-capture",
+    when: "dispatch",
+    evaluation: "first-match",
+    where: async (ctx: DispatchContext) => {
+      captured.push(ctx);
+      return {
+        action: "dispatch" as const,
+        unitType: "execute-task",
+        unitId: "T01",
+        prompt: "parity-fixture",
+      };
+    },
+    then: (r: unknown) => r,
+  };
+  setRegistry(new RuleRegistry([captureRule]));
+
+  try {
+    // Mock ExtensionContext + ExtensionAPI with the surface the wired adapter touches.
+    const fakeModelRegistry = {
+      getAll: () => [],
+      getProviderAuthMode: (_provider: string) => "apiKey" as const,
+    };
+    const ctx = {
+      model: {
+        provider: "anthropic",
+        baseUrl: "https://api.anthropic.com",
+        contextWindow: 200_000,
+      },
+      modelRegistry: fakeModelRegistry,
+    } as any;
+    const pi = {
+      getActiveTools: () => ["read_file", "write_file"],
+    } as any;
+    const basePath = "/tmp/parity-fixture";
+
+    // Path A — wired adapter (what createWiredAutoOrchestrationModule uses).
+    const adapter = createWiredDispatchAdapter(ctx, pi, basePath);
+    const adapterResult = await adapter.decideNextUnit({ stateSnapshot });
+
+    // Path B — direct resolveDispatch call mirroring phases.ts:runDispatch.
+    // Inline the same derivations runDispatch uses so any drift here is a parity break.
+    const prefs = undefined; // loadEffectiveGSDPreferences returns null for /tmp/parity-fixture.
+    const provider = ctx.model?.provider;
+    const authMode = provider && typeof ctx.modelRegistry?.getProviderAuthMode === "function"
+      ? ctx.modelRegistry.getProviderAuthMode(provider)
+      : undefined;
+    const activeTools = typeof pi.getActiveTools === "function" ? pi.getActiveTools() : [];
+    const structuredQuestionsAvailable: "true" | "false" =
+      prefs !== undefined && (prefs as { planning_depth?: string }).planning_depth === "deep"
+        ? "false"
+        : supportsStructuredQuestions(activeTools, {
+            authMode,
+            baseUrl: ctx.model?.baseUrl,
+          })
+          ? "true"
+          : "false";
+
+    const builtDirectCtx: DispatchContext = {
+      basePath,
+      mid: stateSnapshot.activeMilestone!.id,
+      midTitle: stateSnapshot.activeMilestone!.title,
+      state: stateSnapshot,
+      prefs,
+      structuredQuestionsAvailable,
+      sessionContextWindow: ctx.model?.contextWindow,
+      sessionProvider: ctx.model?.provider,
+      modelRegistry: ctx.modelRegistry,
+    };
+    const directAction = await resolveDispatch(builtDirectCtx);
+
+    // Two contexts captured: one per resolveDispatch call.
+    assert.equal(captured.length, 2, "expected two captured dispatch contexts");
+    const [adapterCtx, directCtx] = captured;
+
+    // Parity assertion: session-derived fields are identical.
+    assert.equal(adapterCtx.structuredQuestionsAvailable, directCtx.structuredQuestionsAvailable);
+    assert.equal(adapterCtx.sessionContextWindow, directCtx.sessionContextWindow);
+    assert.equal(adapterCtx.sessionProvider, directCtx.sessionProvider);
+    assert.equal(adapterCtx.modelRegistry, directCtx.modelRegistry);
+    assert.equal(adapterCtx.basePath, directCtx.basePath);
+    assert.equal(adapterCtx.mid, directCtx.mid);
+    assert.equal(adapterCtx.midTitle, directCtx.midTitle);
+
+    // Dispatch action equality: both flows reach the same dispatch decision.
+    if (!adapterResult || !("unitType" in adapterResult)) {
+      assert.fail("expected adapter result to be a dispatch decision");
+    }
+    assert.equal(adapterResult.unitType, "execute-task");
+    assert.equal(adapterResult.unitId, "T01");
+    assert.equal(adapterResult.reason, "test-capture");
+    assert.equal(directAction.action, "dispatch");
+    if (directAction.action === "dispatch") {
+      assert.equal(directAction.unitType, adapterResult.unitType);
+      assert.equal(directAction.unitId, adapterResult.unitId);
+      assert.equal(directAction.matchedRule, adapterResult.reason);
+    }
+  } finally {
+    resetRegistry();
+  }
+});
+
+test("wired DispatchAdapter prefers caller-supplied dispatch inputs over ctx-derived values", async () => {
+  const stateSnapshot = makeState();
+  const captured: DispatchContext[] = [];
+  const captureRule: UnifiedRule = {
+    name: "test-capture-overrides",
+    when: "dispatch",
+    evaluation: "first-match",
+    where: async (ctx: DispatchContext) => {
+      captured.push(ctx);
+      return {
+        action: "dispatch" as const,
+        unitType: "execute-task",
+        unitId: "T01",
+        prompt: "override-fixture",
+      };
+    },
+    then: (r: unknown) => r,
+  };
+  setRegistry(new RuleRegistry([captureRule]));
+
+  try {
+    const ctxModelRegistry = {
+      getAll: () => [],
+      getProviderAuthMode: (_provider: string) => "apiKey" as const,
+    };
+    const overrideModelRegistry = {
+      getAll: () => [],
+      getProviderAuthMode: (_provider: string) => "oauth" as const,
+    };
+    const ctx = {
+      model: {
+        provider: "anthropic",
+        baseUrl: "https://api.anthropic.com",
+        contextWindow: 200_000,
+      },
+      modelRegistry: ctxModelRegistry,
+    } as any;
+    const pi = {
+      getActiveTools: () => [],
+    } as any;
+    const adapter = createWiredDispatchAdapter(ctx, pi, "/tmp/parity-fixture");
+
+    const result = await adapter.decideNextUnit({
+      stateSnapshot,
+      structuredQuestionsAvailable: "true",
+      sessionContextWindow: 500_000,
+      sessionProvider: "openai",
+      modelRegistry: overrideModelRegistry,
+    });
+
+    assert.ok(result);
+    assert.equal(captured.length, 1, "expected one captured dispatch context");
+    assert.equal(captured[0].structuredQuestionsAvailable, "true");
+    assert.equal(captured[0].sessionContextWindow, 500_000);
+    assert.equal(captured[0].sessionProvider, "openai");
+    assert.equal(captured[0].modelRegistry, overrideModelRegistry);
+  } finally {
+    resetRegistry();
+  }
+});
+
+test("wired DispatchAdapter preserves stop reason as a blocked decision", async () => {
+  const stateSnapshot = makeState();
+  const stopRule: UnifiedRule = {
+    name: "test-stop",
+    when: "dispatch",
+    evaluation: "first-match",
+    where: async () => ({
+      action: "stop" as const,
+      reason: "remediation blocker",
+      level: "warning" as const,
+    }),
+    then: (r: unknown) => r,
+  };
+  setRegistry(new RuleRegistry([stopRule]));
+
+  try {
+    const ctx = { model: {}, modelRegistry: { getAll: () => [] } } as any;
+    const pi = { getActiveTools: () => [] } as any;
+    const adapter = createWiredDispatchAdapter(ctx, pi, "/tmp/parity-fixture");
+
+    const result = await adapter.decideNextUnit({ stateSnapshot });
+
+    assert.deepEqual(result, {
+      kind: "blocked",
+      reason: "remediation blocker",
+      action: "pause",
+    });
+  } finally {
+    resetRegistry();
+  }
+});

package/src/resources/extensions/gsd/tests/auto-paused-ui-cleanup.test.ts

@@ -43,7 +43,7 @@ test("cleanupAfterLoopExit preserves paused auto badge after provider pause", as
   }
 });
 
-test("cleanupAfterLoopExit clears status without replacing the last auto surface", async () => {
+test("cleanupAfterLoopExit clears status and progress widget without replacing outcome surface", async () => {
   const statusCalls: unknown[] = [];
   const widgetCalls: unknown[] = [];
 
@@ -64,8 +64,8 @@ test("cleanupAfterLoopExit clears status without replacing the last auto surface
     assert.deepEqual(statusCalls, [["gsd-auto", undefined]]);
     assert.equal(
       widgetCalls.some((args) => Array.isArray(args) && args[0] === "gsd-progress" && args[1] === undefined),
-      false,
-      "cleanup must not clear the last meaningful auto progress surface",
+      true,
+      "cleanup must clear the stale auto progress widget",
     );
     assert.equal(
       widgetCalls.some((args) => Array.isArray(args) && args[0] === "gsd-outcome"),
@@ -79,7 +79,7 @@ test("cleanupAfterLoopExit clears status without replacing the last auto surface
   }
 });
 
-test("cleanupAfterLoopExit preserves completion roll-up after stopAuto reset", async () => {
+test("cleanupAfterLoopExit clears progress widget after stopAuto reset", async () => {
   const statusCalls: unknown[] = [];
   const widgetCalls: unknown[] = [];
 
@@ -103,8 +103,8 @@ test("cleanupAfterLoopExit preserves completion roll-up after stopAuto reset", a
     assert.deepEqual(statusCalls, [["gsd-auto", undefined]]);
     assert.equal(
       widgetCalls.some((args) => Array.isArray(args) && args[0] === "gsd-progress" && args[1] === undefined),
-      false,
-      "completion cleanup must not clear the roll-up progress widget",
+      true,
+      "completion cleanup must clear the stale progress widget",
     );
     assert.equal(
       widgetCalls.some((args) => Array.isArray(args) && args[0] === "gsd-outcome"),
@@ -159,14 +159,15 @@ test("cleanupAfterLoopExit keeps cleanup best-effort when lifecycle restore thro
   const previousCwd = process.cwd();
   let restoreCalls = 0;
   // ADR-016 phase 3 (#5693): the real `restoreToProjectRoot` assigns
-  // `s.basePath = s.originalBasePath` BEFORE any throwable work
+  // `s.basePath = s.originalBasePath` AND chdir's BEFORE any throwable work
   // (rebuildGitService, cache invalidation). Mirror that ordering in the
-  // mock so the throw scenario reflects production: basePath is restored
-  // even when the verb throws partway through.
+  // mock so the throw scenario reflects production: basePath and cwd are
+  // restored even when the verb throws partway through.
   t.mock.method(WorktreeLifecycle.prototype, "restoreToProjectRoot", function (this: WorktreeLifecycle) {
     restoreCalls += 1;
-    (this as unknown as { s: { basePath: string; originalBasePath: string } })
-      .s.basePath = (this as unknown as { s: { originalBasePath: string } }).s.originalBasePath;
+    const sRef = this as unknown as { s: { basePath: string; originalBasePath: string } };
+    sRef.s.basePath = sRef.s.originalBasePath;
+    try { process.chdir(sRef.s.basePath); } catch { /* mirror real verb's best-effort */ }
     throw new Error("restore failed");
   });