gsd-pi 2.76.0-dev.fe143342a → 2.77.0-dev.538325aea

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

@@ -35,12 +35,15 @@ function makeMockSession(opts?: {
   newSessionDelayMs?: number;
   onNewSessionStart?: (session: any) => void;
   onNewSessionSettle?: (session: any) => void;
+  /** Called after the delay with the aborted state of any passed abortSignal.
+   *  Used to verify that runUnit passes an aborted signal on late resolution (#3731). */
+  onSignalCheck?: (aborted: boolean) => void;
 }) {
   const session = {
     active: true,
     verbose: false,
     cmdCtx: {
-      newSession: () => {
+      newSession: (options?: { abortSignal?: AbortSignal }) => {
         opts?.onNewSessionStart?.(session);
         if (opts?.newSessionThrows) {
           return Promise.reject(new Error(opts.newSessionThrows));
@@ -50,11 +53,17 @@ function makeMockSession(opts?: {
         if (delay > 0) {
           return new Promise<{ cancelled: boolean }>((res) =>
             setTimeout(() => {
+              // Simulate AgentSession.newSession() checking abortSignal after
+              // its internal async work (abort()) completes — this is where the
+              // real code captures process.cwd() and rebuilds the tool runtime.
+              // If the signal is aborted, the real code discards the session.
+              opts?.onSignalCheck?.(options?.abortSignal?.aborted ?? false);
               opts?.onNewSessionSettle?.(session);
               res(result);
             }, delay),
           );
         }
+        opts?.onSignalCheck?.(options?.abortSignal?.aborted ?? false);
         opts?.onNewSessionSettle?.(session);
         return Promise.resolve(result);
       },
@@ -349,6 +358,181 @@ test("runUnit cancels before dispatch when model restore fails after newSession"
   ]);
 });
 
+test("runUnit cancels before dispatch when provider is not request-ready (#4555)", async () => {
+  _resetPendingResolve();
+
+  const ctx = makeMockCtx();
+  ctx.model = { provider: "anthropic", id: "claude-opus-4-6" };
+  ctx.modelRegistry = {
+    isProviderRequestReady: (_provider: string) => false,
+  };
+
+  const pi = makeMockPi();
+  const s = makeMockSession();
+
+  const result = await runUnit(ctx, pi, s, "task", "T01", "prompt");
+
+  assert.equal(result.status, "cancelled");
+  assert.equal(result.errorContext?.category, "provider");
+  assert.match(
+    result.errorContext?.message ?? "",
+    /Provider anthropic is not request-ready/,
+  );
+  assert.equal(pi.calls.length, 0, "sendMessage must not be called when provider is not ready");
+});
+
+test("runUnit cancels before dispatch using currentUnitModel provider when set (#4555)", async () => {
+  _resetPendingResolve();
+
+  const ctx = makeMockCtx();
+  // ctx.model uses "openai" which IS ready — if the code ignores currentUnitModel
+  // and falls back to ctx.model.provider, the unit would NOT be cancelled. The
+  // test therefore differentiates: only a bug (wrong provider lookup) would pass.
+  ctx.model = { provider: "openai", id: "gpt-4o" };
+  // modelRegistry says anthropic is not ready but openai is
+  ctx.modelRegistry = {
+    isProviderRequestReady: (provider: string) => provider === "openai",
+  };
+
+  const pi = makeMockPi();
+  const s = makeMockSession();
+  // currentUnitModel overrides the provider used in the readiness check
+  s.currentUnitModel = { provider: "anthropic", id: "claude-opus-4-6" };
+
+  const result = await runUnit(ctx, pi, s, "task", "T01", "prompt");
+
+  assert.equal(result.status, "cancelled");
+  assert.equal(result.errorContext?.category, "provider");
+  assert.match(
+    result.errorContext?.message ?? "",
+    /Provider anthropic is not request-ready/,
+  );
+  assert.equal(pi.calls.length, 0, "sendMessage must not be called — anthropic (currentUnitModel) is not ready");
+});
+
+test("runUnit does not cancel before dispatch when provider is request-ready (#4555)", async () => {
+  _resetPendingResolve();
+
+  const ctx = makeMockCtx();
+  ctx.model = { provider: "anthropic", id: "claude-opus-4-6" };
+  ctx.modelRegistry = {
+    isProviderRequestReady: (_provider: string) => true,
+  };
+
+  const pi = makeMockPi();
+  const s = makeMockSession();
+
+  const resultPromise = runUnit(ctx, pi, s, "task", "T01", "prompt");
+
+  await new Promise((r) => setTimeout(r, 10));
+  resolveAgentEnd(makeEvent());
+
+  const result = await resultPromise;
+  assert.equal(result.status, "completed");
+  assert.equal(pi.calls.length, 1, "sendMessage must be called when provider is ready");
+});
+
+test("runUnit proceeds when modelRegistry is absent (no readiness check available) (#4555)", async () => {
+  _resetPendingResolve();
+
+  const ctx = makeMockCtx();
+  ctx.model = { provider: "anthropic", id: "claude-opus-4-6" };
+  // No modelRegistry on ctx — pre-check should be skipped
+
+  const pi = makeMockPi();
+  const s = makeMockSession();
+
+  const resultPromise = runUnit(ctx, pi, s, "task", "T01", "prompt");
+
+  await new Promise((r) => setTimeout(r, 10));
+  resolveAgentEnd(makeEvent());
+
+  const result = await resultPromise;
+  assert.equal(result.status, "completed");
+  assert.equal(pi.calls.length, 1);
+});
+
+test("runUnit proceeds when isProviderRequestReady throws (defensive) (#4555)", async () => {
+  _resetPendingResolve();
+
+  const ctx = makeMockCtx();
+  ctx.model = { provider: "anthropic", id: "claude-opus-4-6" };
+  ctx.modelRegistry = {
+    isProviderRequestReady: (_provider: string) => {
+      throw new Error("registry error");
+    },
+  };
+
+  const pi = makeMockPi();
+  const s = makeMockSession();
+
+  const result = await runUnit(ctx, pi, s, "task", "T01", "prompt");
+
+  // When the readyCheck throws, ready=false → unit cancelled
+  assert.equal(result.status, "cancelled");
+  assert.equal(result.errorContext?.category, "provider");
+  assert.equal(pi.calls.length, 0);
+});
+
+test("late-resolving newSession() after timeout receives aborted signal so tool runtime is not configured with root cwd (#3731)", async () => {
+  // When newSession() times out in runUnit(), auto-mode restores cwd to project
+  // root. If newSession() later resolves, it must NOT use process.cwd() to
+  // configure the tool runtime (which would give it root cwd, not worktree cwd).
+  //
+  // The fix: runUnit creates an AbortController, aborts it on timeout, and passes
+  // the signal to newSession(). AgentSession.newSession() checks the signal after
+  // its internal await this.abort() completes and returns early (discards) if aborted.
+  //
+  // This test uses mock.timers to control timing precisely.
+  _resetPendingResolve();
+  mock.timers.enable();
+
+  try {
+    let abortedWhenLateSessionSettled: boolean | null = null;
+
+    // newSession mock simulates AgentSession.newSession() behavior:
+    // after an internal delay (representing await this.abort()), it checks the
+    // abortSignal — that's where the real code would capture process.cwd() and
+    // call _buildRuntime. If aborted, the real code must discard the session.
+    const s = makeMockSession({
+      newSessionDelayMs: 200_000, // longer than NEW_SESSION_TIMEOUT_MS (120s)
+      onSignalCheck: (aborted) => {
+        abortedWhenLateSessionSettled = aborted;
+      },
+    });
+
+    const ctx = makeMockCtx();
+    const pi = makeMockPi();
+
+    const resultPromise = runUnit(ctx, pi, s, "task", "T01", "prompt");
+
+    // Tick past the 120s NEW_SESSION_TIMEOUT_MS — runUnit returns cancelled
+    mock.timers.tick(121_000);
+    await Promise.resolve();
+
+    const result = await resultPromise;
+    assert.equal(result.status, "cancelled", "runUnit must return cancelled on session timeout");
+
+    // Tick past the delayed newSession (200s total) — the late newSession resolves
+    mock.timers.tick(80_000);
+    // Drain microtask queue so the .finally() and setTimeout callbacks run
+    await Promise.resolve();
+    await Promise.resolve();
+
+    // The key assertion: when the late newSession() resolves, runUnit must have
+    // passed an aborted AbortSignal. Without the fix, no signal is passed and
+    // abortedWhenLateSessionSettled would be false (or null, if signal not passed at all).
+    assert.equal(
+      abortedWhenLateSessionSettled,
+      true,
+      "runUnit must pass an aborted AbortSignal to newSession() when it resolves after the session-creation timeout (#3731). " +
+      "Without this, AgentSession.newSession() captures root process.cwd() and rebuilds the tool runtime with wrong cwd.",
+    );
+  } finally {
+    mock.timers.reset();
+  }
+});
+
 // ─── Structural assertions ───────────────────────────────────────────────────
 
 test("auto-loop.ts exports autoLoop, runUnit, resolveAgentEnd", async () => {
@@ -409,7 +593,7 @@ test("auto/phases.ts: selectAndApplyModel called exactly once and before updateP
   // Extract the runUnitPhase function body
   const fnStart = src.indexOf("export async function runUnitPhase");
   assert.ok(fnStart > 0, "runUnitPhase should exist in phases.ts");
-  const fnBody = src.slice(fnStart, fnStart + 12000);
+  const fnBody = src.slice(fnStart, fnStart + 16000);
 
   // selectAndApplyModel must appear exactly once
   const allOccurrences = [...fnBody.matchAll(/selectAndApplyModel\(/g)];
@@ -497,6 +681,8 @@ function makeMockDeps(
     autoWorktreeBranch: () => "auto/M001",
     resolveMilestoneFile: () => null,
     reconcileMergeState: () => "clean",
+    preflightCleanRoot: () => ({ stashPushed: false, summary: "" }),
+    postflightPopStash: () => {},
     getLedger: () => null,
     getProjectTotals: () => ({ cost: 0 }),
     formatCost: (c: number) => `$${c.toFixed(2)}`,

package/src/resources/extensions/gsd/tests/auto-paused-session-validation.test.ts

@@ -39,6 +39,18 @@ test("auto.ts validates milestone before restoring paused session (#1664)", () =
     source.includes('resolveMilestoneFile(base, meta.milestoneId, "SUMMARY")'),
     "auto.ts must check for SUMMARY file to detect completed milestones",
   );
+
+  // Resume path must sanitize paused session file metadata before unlink/recovery.
+  assert.ok(
+    source.includes("normalizeSessionFilePath(meta.sessionFile ?? null)"),
+    "auto.ts must sanitize paused-session metadata sessionFile before using it",
+  );
+
+  // Pause path must sanitize live session file path before persisting metadata.
+  assert.ok(
+    source.includes("normalizeSessionFilePath(ctx?.sessionManager?.getSessionFile() ?? null)"),
+    "auto.ts must sanitize sessionManager getSessionFile output before persisting",
+  );
 });
 
 // ─── Filesystem validation unit tests ───────────────────────────────────────

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

@@ -706,6 +706,79 @@ test("verifyExpectedArtifact complete-milestone passes with impl files (#1703)",
   }
 });
 
+test("verifyExpectedArtifact complete-milestone fails when DB milestone is not complete (#4658)", () => {
+  const base = makeGitBase();
+  try {
+    execFileSync("git", ["checkout", "-b", "feat/ms-db-active"], { cwd: base, stdio: "ignore" });
+    mkdirSync(join(base, ".gsd", "milestones", "M001"), { recursive: true });
+    writeFileSync(join(base, ".gsd", "milestones", "M001", "M001-SUMMARY.md"), "# Milestone Summary\nverification FAILED — not complete.");
+    mkdirSync(join(base, "src"), { recursive: true });
+    writeFileSync(join(base, "src", "app.ts"), "console.log('hello');");
+    execFileSync("git", ["add", "."], { cwd: base, stdio: "ignore" });
+    execFileSync("git", ["commit", "-m", "feat: implementation with failed summary"], { cwd: base, stdio: "ignore" });
+
+    openDatabase(join(base, ".gsd", "gsd.db"));
+    insertMilestone({ id: "M001", title: "Milestone One", status: "active" });
+
+    const result = verifyExpectedArtifact("complete-milestone", "M001", base);
+    assert.equal(result, false, "complete-milestone must fail when DB status is not complete");
+  } finally {
+    cleanup(base);
+  }
+});
+
+test("verifyExpectedArtifact complete-milestone passes when DB milestone is complete (#4658)", () => {
+  const base = makeGitBase();
+  try {
+    execFileSync("git", ["checkout", "-b", "feat/ms-db-complete"], { cwd: base, stdio: "ignore" });
+    mkdirSync(join(base, ".gsd", "milestones", "M001"), { recursive: true });
+    writeFileSync(join(base, ".gsd", "milestones", "M001", "M001-SUMMARY.md"), "# Milestone Summary\nDone.");
+    mkdirSync(join(base, "src"), { recursive: true });
+    writeFileSync(join(base, "src", "app.ts"), "console.log('hello');");
+    execFileSync("git", ["add", "."], { cwd: base, stdio: "ignore" });
+    execFileSync("git", ["commit", "-m", "feat: implementation complete"], { cwd: base, stdio: "ignore" });
+
+    openDatabase(join(base, ".gsd", "gsd.db"));
+    insertMilestone({ id: "M001", title: "Milestone One", status: "complete" });
+
+    const result = verifyExpectedArtifact("complete-milestone", "M001", base);
+    assert.equal(result, true, "complete-milestone should pass when DB status is complete");
+  } finally {
+    cleanup(base);
+  }
+});
+
+test("verifyExpectedArtifact complete-milestone tolerates transient DB lag when SUMMARY is canonical success (#4658)", () => {
+  const base = makeGitBase();
+  try {
+    execFileSync("git", ["checkout", "-b", "feat/ms-db-lag-success"], { cwd: base, stdio: "ignore" });
+    mkdirSync(join(base, ".gsd", "milestones", "M001"), { recursive: true });
+    writeFileSync(
+      join(base, ".gsd", "milestones", "M001", "M001-SUMMARY.md"),
+      [
+        "---",
+        "id: M001",
+        "status: complete",
+        "---",
+        "",
+        "# M001: Success",
+      ].join("\n"),
+    );
+    mkdirSync(join(base, "src"), { recursive: true });
+    writeFileSync(join(base, "src", "app.ts"), "console.log('hello');");
+    execFileSync("git", ["add", "."], { cwd: base, stdio: "ignore" });
+    execFileSync("git", ["commit", "-m", "feat: implementation with stale db"], { cwd: base, stdio: "ignore" });
+
+    openDatabase(join(base, ".gsd", "gsd.db"));
+    insertMilestone({ id: "M001", title: "Milestone One", status: "active" });
+
+    const result = verifyExpectedArtifact("complete-milestone", "M001", base);
+    assert.equal(result, true, "canonical success SUMMARY should pass verification during transient DB lag");
+  } finally {
+    cleanup(base);
+  }
+});
+
 test("verifyExpectedArtifact checks pending gate-evaluate artifacts without ESM require failures", () => {
   const base = makeTmpProject();
 
@@ -775,6 +848,55 @@ test("#4414: parallel-research sentinel path does not collide with RESEARCH suff
   }
 });
 
+test("#4068: verifyExpectedArtifact parallel-research treats PARALLEL-BLOCKER as terminal completion", () => {
+  // Regression: when a parallel-research unit times out and the timeout-recovery
+  // machinery writes a PARALLEL-BLOCKER placeholder, verifyExpectedArtifact must
+  // return true so the dispatch loop can advance.  Previously it only returned
+  // true when every slice had a RESEARCH file — meaning a timeout always left
+  // verifyExpectedArtifact returning false, the unit was never cleared from
+  // unitDispatchCount, and the dispatch rule re-fired on the next iteration
+  // (infinite loop, issue #4068 / #4355).
+  const base = makeTmpBase();
+  try {
+    // Write a minimal roadmap
+    writeFileSync(
+      join(base, ".gsd", "milestones", "M001", "M001-ROADMAP.md"),
+      [
+        "# M001: Timeout Test",
+        "",
+        "## Slices",
+        "",
+        "- [ ] **S01: Alpha** `risk:low` `depends:[]`",
+        "- [ ] **S02: Beta** `risk:low` `depends:[]`",
+        "",
+      ].join("\n"),
+      "utf-8",
+    );
+
+    // No RESEARCH files written — subagents timed out
+    clearParseCache();
+    invalidateAllCaches();
+
+    // Simulate timeout-recovery writing the PARALLEL-BLOCKER placeholder
+    const blockerPath = resolveExpectedArtifactPath("research-slice", "M001/parallel-research", base);
+    assert.ok(blockerPath, "PARALLEL-BLOCKER path must resolve for parallel-research unit");
+    writeFileSync(blockerPath!, "# BLOCKER — timeout recovery\n\n**Reason**: hard timeout.\n", "utf-8");
+
+    clearParseCache();
+    invalidateAllCaches();
+
+    // After blocker is written, verifyExpectedArtifact must return true
+    // so the dispatch loop treats this unit as complete and moves on.
+    assert.equal(
+      verifyExpectedArtifact("research-slice", "M001/parallel-research", base),
+      true,
+      "#4068: PARALLEL-BLOCKER on disk must satisfy verifyExpectedArtifact so the loop does not re-dispatch",
+    );
+  } finally {
+    cleanup(base);
+  }
+});
+
 test("#4414: verifyExpectedArtifact parallel-research succeeds when all research-ready slices have RESEARCH", () => {
   const base = makeTmpBase();
   try {

package/src/resources/extensions/gsd/tests/auto-start-bootstrap-await-3420.test.ts

@@ -0,0 +1,141 @@
+/**
+ * auto-start-bootstrap-await-3420.test.ts — Regression test for #3420.
+ *
+ * Bug: In bootstrapAutoSession, when state.phase === "pre-planning" and no
+ * context file exists, showSmartEntry is called to dispatch a discuss workflow.
+ * showSmartEntry calls dispatchWorkflow which calls pi.sendMessage() — a
+ * fire-and-forget call that returns immediately. The LLM discussion runs
+ * asynchronously in a separate turn.
+ *
+ * The bug: after showSmartEntry returns (before the LLM has run), the code
+ * immediately calls invalidateAllCaches() + deriveState() + checks postState.
+ * Since the discussion hasn't run yet, postState.phase is still "pre-planning"
+ * and the context check fails, producing the warning:
+ *   "Discussion completed but milestone context is still missing. Run /gsd to try again."
+ *
+ * The discussion never ran — the warning fires immediately.
+ *
+ * Fix: bootstrapAutoSession must return false (release lock) after showSmartEntry
+ * dispatches the workflow. The checkAutoStartAfterDiscuss callback in guided-flow.ts
+ * already handles re-entering auto-mode when the discussion completes.
+ *
+ * This test verifies the fix by asserting that the pre-planning !hasContext block
+ * does NOT contain a postState phase check after showSmartEntry — it must
+ * return false immediately to let the async dispatch complete.
+ */
+
+import test from "node:test";
+import assert from "node:assert/strict";
+import { readFileSync } from "node:fs";
+import { join } from "node:path";
+
+const sourcePath = join(import.meta.dirname, "..", "auto-start.ts");
+const source = readFileSync(sourcePath, "utf-8");
+
+test("bootstrapAutoSession: pre-planning no-context path does NOT check postState immediately after showSmartEntry (#3420)", () => {
+  // Find the pre-planning block that handles the case where context is missing.
+  // This block dispatches showSmartEntry which is async (fire-and-forget via pi.sendMessage).
+  // After the dispatch, checking postState.phase immediately is premature — the
+  // LLM discussion hasn't run yet. The block should return false instead.
+  const prePlanningNoContextBlock = source.match(
+    /\/\/ Active milestone exists but has no roadmap\s*\n\s*if\s*\(\s*state\.phase\s*===\s*"pre-planning"\s*\)([\s\S]*?)\/\/ Active milestone has CONTEXT-DRAFT/,
+  );
+  assert.ok(
+    !!prePlanningNoContextBlock,
+    "auto-start.ts must have the pre-planning handler block before needs-discussion",
+  );
+
+  const block = prePlanningNoContextBlock![1];
+
+  // The block must call showSmartEntry when !hasContext
+  assert.ok(
+    block.includes("showSmartEntry"),
+    "pre-planning !hasContext block must call showSmartEntry to dispatch the discuss workflow",
+  );
+
+  // FAILING ASSERTION (before fix): after showSmartEntry, the block must NOT
+  // immediately check postState.phase — that check fires before the LLM runs.
+  // Instead, it must return false (release lock) so the async dispatch can complete.
+  // The warning "Discussion completed but milestone context is still missing"
+  // fires prematurely when this postState check exists.
+  assert.ok(
+    !block.includes("Discussion completed but milestone context is still missing"),
+    "pre-planning !hasContext block must NOT check postState.phase immediately after showSmartEntry — " +
+    "the dispatch is async (pi.sendMessage is fire-and-forget) and the discussion hasn't run yet; " +
+    "return false instead so checkAutoStartAfterDiscuss can re-enter auto-mode after discussion completes (#3420)",
+  );
+});
+
+test("bootstrapAutoSession: complete/no-milestone path does NOT check postState immediately after showSmartEntry (#3420)", () => {
+  // Find the complete/no-milestone block
+  const completeBlock = source.match(
+    /\/\/ No active work — start a new milestone via discuss flow\s*\n\s*if\s*\(!state\.activeMilestone\s*\|\|\s*state\.phase\s*===\s*"complete"\s*\)([\s\S]*?)\/\/ Active milestone exists but has no roadmap/,
+  );
+  assert.ok(
+    !!completeBlock,
+    "auto-start.ts must have the complete/no-milestone handler block",
+  );
+
+  const block = completeBlock![1];
+
+  // The block must call showSmartEntry
+  assert.ok(
+    block.includes("showSmartEntry"),
+    "complete/no-milestone block must call showSmartEntry",
+  );
+
+  // After showSmartEntry dispatches, checking postState.phase is premature —
+  // the LLM hasn't had a turn yet. The block should return false.
+  // Specifically, the "no milestone context was written" warning fires too early.
+  assert.ok(
+    !block.includes("Discussion completed but no milestone context was written"),
+    "complete/no-milestone block must NOT check postState.phase immediately after showSmartEntry dispatch — " +
+    "return false instead so the async LLM turn can complete (#3420)",
+  );
+});
+
+test("bootstrapAutoSession: showSmartEntry in pre-planning path is followed by releaseLockAndReturn, not postState check (#3420)", () => {
+  // After the fix, the pre-planning !hasContext branch should call showSmartEntry
+  // and then immediately return releaseLockAndReturn() — not check postState.
+  const prePlanningNoContextBlock = source.match(
+    /\/\/ Active milestone exists but has no roadmap\s*\n\s*if\s*\(\s*state\.phase\s*===\s*"pre-planning"\s*\)([\s\S]*?)\/\/ Active milestone has CONTEXT-DRAFT/,
+  );
+  assert.ok(!!prePlanningNoContextBlock, "pre-planning handler block found");
+
+  const block = prePlanningNoContextBlock![1];
+
+  // After the fix, the !hasContext branch ends with releaseLockAndReturn
+  assert.ok(
+    block.includes("releaseLockAndReturn"),
+    "pre-planning !hasContext block must call releaseLockAndReturn() after showSmartEntry dispatch (#3420)",
+  );
+
+  // The showSmartEntry call must appear before releaseLockAndReturn
+  const showSmartEntryIdx = block.indexOf("showSmartEntry");
+  const releaseLockIdx = block.indexOf("releaseLockAndReturn");
+  assert.ok(
+    showSmartEntryIdx > -1 && releaseLockIdx > -1,
+    "both showSmartEntry and releaseLockAndReturn must appear in the block",
+  );
+  assert.ok(
+    showSmartEntryIdx < releaseLockIdx,
+    "showSmartEntry must appear before releaseLockAndReturn in pre-planning !hasContext block",
+  );
+
+  // There must be NO invalidateAllCaches between showSmartEntry and releaseLockAndReturn
+  // (invalidateAllCaches + deriveState after showSmartEntry is the buggy premature check)
+  const afterShowSmartEntry = block.substring(showSmartEntryIdx);
+  const cacheInvalidateIdx = afterShowSmartEntry.indexOf("invalidateAllCaches");
+  const releaseFromShowIdx = afterShowSmartEntry.indexOf("releaseLockAndReturn");
+
+  // If invalidateAllCaches appears, it must appear AFTER releaseLockAndReturn
+  // (which is impossible since releaseLockAndReturn returns) — so invalidateAllCaches
+  // must not appear at all between showSmartEntry and the end of the !hasContext block.
+  if (cacheInvalidateIdx !== -1) {
+    assert.ok(
+      cacheInvalidateIdx > releaseFromShowIdx,
+      "invalidateAllCaches must NOT appear between showSmartEntry and releaseLockAndReturn — " +
+      "this is the premature postState check that causes #3420",
+    );
+  }
+});

package/src/resources/extensions/gsd/tests/auto-start-clean-runtime-db-gated.test.ts

@@ -0,0 +1,63 @@
+// GSD-2 auto-start regression test: cleanStaleRuntimeUnits is DB-gated (#4663)
+//
+// Source-level structural check that the stale-runtime-cleanup predicate in
+// auto-start.ts consults DB status when available instead of treating a
+// SUMMARY-file on disk as proof of milestone completion. Pairs with #4658 /
+// PR #4660 fix in auto-dispatch + auto-recovery.
+
+import { describe, test } from "node:test";
+import assert from "node:assert/strict";
+import { readFileSync } from "node:fs";
+import { join, dirname } from "node:path";
+import { fileURLToPath } from "node:url";
+
+const __dirname = dirname(fileURLToPath(import.meta.url));
+const sourceFile = join(__dirname, "..", "auto-start.ts");
+
+describe("auto-start cleanStaleRuntimeUnits DB gating (#4663)", () => {
+  const source = readFileSync(sourceFile, "utf-8");
+
+  test("imports isClosedStatus for DB-status check", () => {
+    assert.match(
+      source,
+      /import\s*\{\s*isClosedStatus\s*\}\s*from\s*["']\.\/status-guards/,
+    );
+  });
+
+  test("cleanStaleRuntimeUnits is called after openProjectDbIfPresent", () => {
+    const openDbIdx = source.indexOf("await openProjectDbIfPresent(base)");
+    assert.ok(openDbIdx > -1, "openProjectDbIfPresent should be called in auto-start");
+    const cleanIdx = source.indexOf("cleanStaleRuntimeUnits(", openDbIdx);
+    assert.ok(
+      cleanIdx > -1,
+      "cleanStaleRuntimeUnits must run AFTER openProjectDbIfPresent so predicate can consult DB",
+    );
+  });
+
+  test("cleanStaleRuntimeUnits predicate consults DB status when available", () => {
+    const cleanIdx = source.indexOf("cleanStaleRuntimeUnits(");
+    assert.ok(cleanIdx > -1);
+    const snippet = source.slice(cleanIdx, cleanIdx + 600);
+    assert.match(
+      snippet,
+      /isDbAvailable\(\)/,
+      "predicate must branch on DB availability",
+    );
+    assert.match(
+      snippet,
+      /isClosedStatus\(/,
+      "predicate must check DB status via isClosedStatus",
+    );
+  });
+
+  test("cleanStaleRuntimeUnits predicate still falls back to SUMMARY-file when DB unavailable", () => {
+    const cleanIdx = source.indexOf("cleanStaleRuntimeUnits(");
+    assert.ok(cleanIdx > -1);
+    const snippet = source.slice(cleanIdx, cleanIdx + 600);
+    assert.match(
+      snippet,
+      /resolveMilestoneFile\(base,\s*mid,\s*["']SUMMARY["']\)/,
+      "legacy FS-fallback branch should still use SUMMARY file presence",
+    );
+  });
+});

package/src/resources/extensions/gsd/tests/auto-wrapup-inflight-guard.test.ts

@@ -144,3 +144,26 @@ describe("#3512: pauseAuto and stopAuto must flush queued follow-up messages", (
     );
   });
 });
+
+describe("#4365: tool_execution_start hook must pass toolName to markToolStart", () => {
+  test("tool_execution_start handler passes event.toolName to markToolStart", () => {
+    // The tool_execution_start handler must forward toolName so that
+    // hasInteractiveToolInFlight() can correctly identify ask_user_questions
+    // and prevent the idle watchdog from firing during interactive tool calls.
+    const startMarker = 'pi.on("tool_execution_start", async (event) => {';
+    const endMarker = 'pi.on("tool_execution_end", async (event) => {';
+    const toolExecutionStartSection = registerHooksSrc.slice(
+      registerHooksSrc.indexOf(startMarker),
+      registerHooksSrc.indexOf(endMarker),
+    );
+
+    assert.ok(
+      toolExecutionStartSection.length > 0,
+      "Could not locate tool_execution_start handler section",
+    );
+    assert.ok(
+      toolExecutionStartSection.includes("markToolStart(event.toolCallId, event.toolName)"),
+      "tool_execution_start handler must pass event.toolName to markToolStart so hasInteractiveToolInFlight() works correctly",
+    );
+  });
+});