@opengsd/gsd-pi 1.2.0-dev.844675c9 → 1.2.0-dev.8e6112e9

package/src/resources/extensions/gsd/tests/commands-dispatcher-workspace-git.test.ts

@@ -64,3 +64,14 @@ test("getWorkspaceGitBlockMessageForBase allows doctor on product conflicts", as
     cleanup(base);
   }
 });
+
+test("getWorkspaceGitBlockMessageForBase allows forensics on product conflicts", async () => {
+  const base = makeTempRepo("gsd-dispatch-ws-git-forensics-");
+  try {
+    seedProductConflict(base);
+    const blocked = await getWorkspaceGitBlockMessageForBase(base, "forensics");
+    assert.equal(blocked, null);
+  } finally {
+    cleanup(base);
+  }
+});

package/src/resources/extensions/gsd/tests/commands-verdict.test.ts

@@ -10,7 +10,7 @@
  */
 import test from "node:test";
 import assert from "node:assert/strict";
-import { mkdirSync, mkdtempSync, readFileSync, rmSync, writeFileSync } from "node:fs";
+import { existsSync, mkdirSync, mkdtempSync, readFileSync, rmSync, statSync, writeFileSync } from "node:fs";
 import { join } from "node:path";
 import { tmpdir } from "node:os";
 import { randomUUID } from "node:crypto";
@@ -472,17 +472,55 @@ test("auto-dispatch needs-attention pause message references /gsd verdict", asyn
   }
 });
 
-test("state.ts needs-remediation blocker messages reference /gsd verdict", async () => {
+test("guidance.ts needs-remediation blocker messages reference /gsd verdict", async () => {
   // We don't need to invoke deriveState — just assert the substring is in the
-  // source. The blocker strings are constructed inline and shipped to the user
-  // verbatim, so a static check is sufficient and avoids fragile DB setup.
-  const stateSource = readFileSync(
-    new URL("../state.ts", import.meta.url).pathname,
+  // source. The blocker strings are constructed in the guidance catalog and
+  // shipped to the user verbatim, so a static check is sufficient and avoids
+  // fragile DB setup.
+  const guidanceSource = readFileSync(
+    new URL("../guidance.ts", import.meta.url).pathname,
     "utf-8",
   );
-  const occurrences = stateSource.match(/`\/gsd verdict /g) ?? [];
+  const occurrences = guidanceSource.match(/`\/gsd verdict /g) ?? [];
   assert.ok(
     occurrences.length >= 2,
-    `expected at least 2 references to /gsd verdict in state.ts blockers, found ${occurrences.length}`,
+    `expected at least 2 references to /gsd verdict in guidance.ts blockers, found ${occurrences.length}`,
   );
 });
+
+// ─── WAL checkpoint regression (#563) ───────────────────────────────────
+
+test("handleVerdict pass checkpoints WAL so subsequent processes see the updated verdict (#563)", async () => {
+  // Regression: before the fix, executeValidateMilestone wrote the assessment
+  // only into the WAL file.  The next /gsd auto invocation (a new process)
+  // opened gsd.db directly and saw stale pre-verdict data because the WAL
+  // had not been flushed.  The fix adds checkpointDatabase() after a
+  // successful verdict so the WAL is truncated and gsd.db is self-contained.
+  const base = makeBase();
+  try {
+    openTestDb(base);
+    seedMilestone("M001", "Checkpoint Test Milestone");
+    seedSlice("M001", "S01", "complete");
+    writeValidation(base, "M001", "needs-attention");
+
+    const { ctx } = makeMockCtx();
+    await handleVerdict("pass --milestone M001", ctx, base);
+
+    // After handleVerdict succeeds, PRAGMA wal_checkpoint(TRUNCATE) must have
+    // run.  The WAL file is truncated to zero bytes — a new process reading
+    // gsd.db directly sees the updated verdict without needing the WAL.
+    const walPath = join(base, ".gsd", "gsd.db-wal");
+    if (existsSync(walPath)) {
+      const { size } = statSync(walPath);
+      assert.equal(
+        size,
+        0,
+        "WAL must be zero bytes after handleVerdict pass — checkpointDatabase() was not called (#563)",
+      );
+    }
+  } finally {
+    closeDatabase();
+    invalidateStateCache();
+    cleanup(base);
+  }
+});

package/src/resources/extensions/gsd/tests/complete-slice-verification-gate.test.ts

@@ -258,6 +258,48 @@ describe('complete-slice verification gate (#3580)', () => {
     }
   });
 
+  test('allows a browser UAT declared as a bare keyword under ## UAT Type (M006/S01 format drift)', async () => {
+    // Regression: the agent wrote `## UAT Type\nbrowser-executable` (no
+    // `- UAT mode:` bullet). The old parser defaulted that to artifact-driven
+    // and the gate rejected the slice in a loop.
+    const body = [
+      '## UAT Type',
+      'browser-executable',
+      '',
+      '## Smoke Test',
+      '1. Open the page in a browser and perform add/edit/complete/delete once.',
+    ].join('\n');
+    const result = await handleCompleteSlice(
+      makeParams({ uatContent: body }),
+      basePath,
+    );
+    if ('error' in result) {
+      assert.doesNotMatch(
+        result.error,
+        /requires browser verification/i,
+        `bare browser-executable declaration must pass the browser gate, got: ${result.error}`,
+      );
+    }
+  });
+
+  test('explains the missing declaration when a browser UAT has no parseable UAT mode', async () => {
+    // When nothing was declared, the error must not claim the agent
+    // "declared artifact-driven" — it must show the expected bullet format.
+    const body = [
+      '## Smoke Test',
+      '1. Open the page in a browser and perform add/edit/complete/delete once.',
+    ].join('\n');
+    const result = await handleCompleteSlice(
+      makeParams({ uatContent: body }),
+      basePath,
+    );
+    assert.ok('error' in result, 'expected handler to reject an undeclared browser UAT');
+    const error = (result as { error: string }).error;
+    assert.match(error, /no parseable UAT mode declaration/i);
+    assert.match(error, /- UAT mode: browser-executable/);
+    assert.doesNotMatch(error, /but declares "UAT mode: artifact-driven"/);
+  });
+
   test('allows a browser UAT when it is declared mixed (mixed receives browser tools)', async () => {
     const body = BROWSER_UAT_BODY.replace('artifact-driven', 'mixed (artifact-driven + browser)');
     const result = await handleCompleteSlice(

package/src/resources/extensions/gsd/tests/complete-task.test.ts

@@ -80,6 +80,20 @@ function createTempProject(): { basePath: string; planPath: string } {
   return { basePath, planPath };
 }
 
+function writeProjectPreferences(basePath: string, yaml: string): void {
+  fs.writeFileSync(path.join(basePath, '.gsd', 'PREFERENCES.md'), `---\n${yaml}---\n`);
+}
+
+async function withWorkingDirectory<T>(cwd: string, action: () => Promise<T>): Promise<T> {
+  const previousCwd = process.cwd();
+  process.chdir(cwd);
+  try {
+    return await action();
+  } finally {
+    process.chdir(previousCwd);
+  }
+}
+
 function makeValidParams() {
   return {
     taskId: 'T01',
@@ -104,6 +118,13 @@ function makeValidParams() {
   };
 }
 
+function makeEscalationOptions() {
+  return [
+    { id: 'continue', label: 'Continue', tradeoffs: 'Keeps execution moving with the default path.' },
+    { id: 'pause', label: 'Pause', tradeoffs: 'Stops execution until the blocker is reviewed.' },
+  ];
+}
+
 // ═══════════════════════════════════════════════════════════════════════════
 // complete-task: Fresh DB is migrated to the current schema version
 // ═══════════════════════════════════════════════════════════════════════════
@@ -359,6 +380,105 @@ console.log('\n=== complete-task: handler happy path ===');
   cleanup(dbPath);
 }
 
+// ═══════════════════════════════════════════════════════════════════════════
+// complete-task: hard-blocker escalation with mid-execution escalation disabled
+// ═══════════════════════════════════════════════════════════════════════════
+
+console.log('\n=== complete-task: disabled hard-blocker escalation rolls back completion ===');
+{
+  const dbPath = tempDbPath();
+  openDatabase(dbPath);
+
+  const { basePath } = createTempProject();
+  writeProjectPreferences(basePath, 'phases:\n  mid_execution_escalation: false\n');
+
+  insertMilestone({ id: 'M001', title: 'Test Milestone' });
+  insertSlice({ id: 'S01', milestoneId: 'M001', title: 'Test Slice' });
+
+  const params = {
+    ...makeValidParams(),
+    blockerDiscovered: true,
+    escalation: {
+      question: 'Should execution pause for the hard blocker?',
+      options: makeEscalationOptions(),
+      recommendation: 'pause',
+      recommendationRationale: 'The blocker should not be silently advanced.',
+      continueWithDefault: false,
+    },
+  };
+
+  const result = await withWorkingDirectory(basePath, () => handleCompleteTask(params, basePath));
+
+  assertTrue('error' in result, 'hard-blocker escalation should fail when escalation handling is disabled');
+  if ('error' in result) {
+    assertMatch(result.error, /hard-blocker escalation/, 'error should mention hard-blocker escalation');
+    assertMatch(result.error, /mid_execution_escalation is disabled/, 'error should mention disabled preference');
+  }
+
+  const task = getTask('M001', 'S01', 'T01');
+  assertTrue(task !== null, 'task row should remain after rollback');
+  assertEq(task!.status, 'pending', 'task status should be rolled back to pending');
+  assertEq(task!.blocker_discovered, true, 'blocker flag should remain recorded for visibility');
+  assertEq(task!.escalation_pending, 0, 'disabled preference should not create a pending escalation flag');
+  assertEq(task!.escalation_awaiting_review, 0, 'disabled preference should not create an awaiting-review flag');
+
+  const adapter = _getAdapter()!;
+  const evRows = adapter.prepare(
+    "SELECT * FROM verification_evidence WHERE task_id = 'T01' AND slice_id = 'S01' AND milestone_id = 'M001'"
+  ).all();
+  assertEq(evRows.length, 0, 'verification evidence should be deleted when completion rolls back');
+
+  cleanupDir(basePath);
+  cleanup(dbPath);
+}
+
+// ═══════════════════════════════════════════════════════════════════════════
+// complete-task: soft escalation with mid-execution escalation disabled
+// ═══════════════════════════════════════════════════════════════════════════
+
+console.log('\n=== complete-task: disabled soft escalation still completes ===');
+{
+  const dbPath = tempDbPath();
+  openDatabase(dbPath);
+
+  const { basePath } = createTempProject();
+  writeProjectPreferences(basePath, 'phases:\n  mid_execution_escalation: false\n');
+
+  insertMilestone({ id: 'M001', title: 'Test Milestone' });
+  insertSlice({ id: 'S01', milestoneId: 'M001', title: 'Test Slice' });
+
+  const params = {
+    ...makeValidParams(),
+    escalation: {
+      question: 'Should execution continue with the default?',
+      options: makeEscalationOptions(),
+      recommendation: 'continue',
+      recommendationRationale: 'The default path is safe enough to continue.',
+      continueWithDefault: true,
+    },
+  };
+
+  const result = await withWorkingDirectory(basePath, () => handleCompleteTask(params, basePath));
+
+  assertTrue(!('error' in result), 'soft escalation should still complete when escalation handling is disabled');
+  if (!('error' in result)) {
+    assertTrue(!result.escalation, 'disabled preference should not return escalation metadata');
+  }
+
+  const task = getTask('M001', 'S01', 'T01');
+  assertTrue(task !== null, 'task row should exist');
+  assertEq(task!.status, 'complete', 'soft escalation should leave task complete');
+
+  const adapter = _getAdapter()!;
+  const evRows = adapter.prepare(
+    "SELECT * FROM verification_evidence WHERE task_id = 'T01' AND slice_id = 'S01' AND milestone_id = 'M001'"
+  ).all();
+  assertEq(evRows.length, 1, 'verification evidence should remain for soft escalation completion');
+
+  cleanupDir(basePath);
+  cleanup(dbPath);
+}
+
 // ═══════════════════════════════════════════════════════════════════════════
 // complete-task: Handler validation errors
 // ═══════════════════════════════════════════════════════════════════════════
@@ -414,6 +534,9 @@ console.log('\n=== complete-task: handler idempotency ===');
   // First call should succeed
   const r1 = await handleCompleteTask(params, basePath);
   assertTrue(!('error' in r1), 'first call should succeed');
+  if ('error' in r1) {
+    throw new Error(r1.error);
+  }
 
   // Verify complete-task did not duplicate T01. S01-PLAN.md is a projection,
   // so the remaining plan task is not imported implicitly.
@@ -421,16 +544,29 @@ console.log('\n=== complete-task: handler idempotency ===');
   assertEq(tasks.length, 1, 'should only have the completed DB task after first call');
   assertEq(tasks.filter(t => t.id === 'T01').length, 1, 'should have exactly one T01 row after first call');
 
-  // Second call with same params — state machine guard rejects (task is already complete)
+  // If the DB row is complete but the projection was lost, the duplicate call
+  // should repair the missing summary from full_summary_md instead of forcing a
+  // reopen/re-complete loop.
+  fs.unlinkSync(r1.summaryPath);
+  assertTrue(!fs.existsSync(r1.summaryPath), 'fixture should remove the task summary before repair');
   const r2 = await handleCompleteTask(params, basePath);
-  assertTrue('error' in r2, 'second call should return error (task already complete)');
+  assertTrue(!('error' in r2), 'second call should repair missing summary for DB-complete task');
   if ('error' in r2) {
-    assertMatch(r2.error, /already complete/, 'error should mention already complete');
+    throw new Error(r2.error);
+  }
+  assertTrue(fs.existsSync(r2.summaryPath), 'missing summary should be restored on disk');
+  assertEq(r2.duplicate, true, 'repair should be reported as a duplicate/no-op state mutation');
+
+  // Third call with the summary present — state machine guard rejects (task is already complete)
+  const r3 = await handleCompleteTask(params, basePath);
+  assertTrue('error' in r3, 'third call should return error (task already complete)');
+  if ('error' in r3) {
+    assertMatch(r3.error, /already complete/, 'error should mention already complete');
   }
 
-  // Still no duplicate rows from the rejected second call.
+  // Still no duplicate rows from the repair or rejected third call.
   const tasksAfter = getSliceTasks('M001', 'S01');
-  assertEq(tasksAfter.length, 1, 'should still only have T01 after rejected second call');
+  assertEq(tasksAfter.length, 1, 'should still only have T01 after duplicate repair and rejected third call');
   assertEq(tasksAfter.filter(t => t.id === 'T01').length, 1, 'should still have exactly one T01 row');
 
   cleanupDir(basePath);

package/src/resources/extensions/gsd/tests/consent-question.test.ts

@@ -0,0 +1,351 @@
+import test from "node:test";
+import assert from "node:assert/strict";
+
+import {
+  type AnswerOutcome,
+  classifyQuestion,
+  evaluateAnswer,
+  evaluateAskUserQuestionsRound,
+  failPolicyForKind,
+  formatUnansweredConsentQuestionMessage,
+  isAwaitingUserInput,
+  lastAssistantText,
+  messageHasPendingAskUserQuestionsTool,
+  shouldPauseForQuestion,
+} from "../consent-question.ts";
+
+const GATE_OPTIONS = [
+  { label: "Yes, you got it (Recommended)" },
+  { label: "Needs adjustment" },
+];
+
+const CONSENT_OPTIONS = [
+  { label: "Option A (Recommended)" },
+  { label: "Option B" },
+];
+
+function assistantMessage(text: string): unknown {
+  return { role: "assistant", content: [{ type: "text", text }] };
+}
+
+// ── Classification ──────────────────────────────────────────────────────────
+
+test("classifyQuestion: gate ids, consent defaults, prose kinds", () => {
+  assert.deepEqual(
+    classifyQuestion({ id: "depth_verification_M001_confirm" }),
+    { kind: "gate", gateSubKind: "approval" },
+  );
+  assert.deepEqual(
+    classifyQuestion({ id: "depth_verification" }),
+    { kind: "gate", gateSubKind: "depth-verification" },
+  );
+  assert.deepEqual(
+    classifyQuestion({ id: "destructive_confirm_rm" }),
+    { kind: "gate", gateSubKind: "destructive-confirm" },
+  );
+  // Any other structured question is a real elicitation → consent, fail-closed.
+  assert.equal(classifyQuestion({ id: "m004_shape", options: CONSENT_OPTIONS }).kind, "consent");
+  assert.equal(classifyQuestion({ id: "free_text_q" }).kind, "consent");
+  // Prose classification.
+  assert.equal(classifyQuestion({ text: "Should I proceed with the write?" }).kind, "consent");
+  assert.equal(
+    classifyQuestion({ text: "Want me to research this or skip?", unitType: "research-decision" }).kind,
+    "decision",
+  );
+  assert.equal(classifyQuestion({ text: "Here is the summary of changes." }).kind, "informational");
+});
+
+test("failPolicyForKind: only informational is fail-open", () => {
+  assert.equal(failPolicyForKind("gate"), "closed");
+  assert.equal(failPolicyForKind("consent"), "closed");
+  assert.equal(failPolicyForKind("decision"), "closed");
+  assert.equal(failPolicyForKind("informational"), "open");
+});
+
+// ── Policy table: (kind × answer shape) → outcome ───────────────────────────
+
+type Row = {
+  name: string;
+  question: { id: string; options?: Array<{ label?: string }> };
+  details: Parameters<typeof evaluateAnswer>[0]["details"];
+  expected: AnswerOutcome;
+};
+
+const gateQ = { id: "depth_verification_M001_confirm", options: GATE_OPTIONS };
+const consentQ = { id: "boundary_choice", options: CONSENT_OPTIONS };
+
+const POLICY_TABLE: Row[] = [
+  // gate
+  { name: "gate × confirm selected → verified", question: gateQ, details: { response: { answers: { [gateQ.id]: { selected: "Yes, you got it (Recommended)" } } } }, expected: "verified" },
+  { name: "gate × decline selected → declined", question: gateQ, details: { response: { answers: { [gateQ.id]: { selected: "Needs adjustment" } } } }, expected: "declined" },
+  { name: "gate × empty selected → waiting", question: gateQ, details: { response: { answers: { [gateQ.id]: { selected: "" } } } }, expected: "waiting" },
+  { name: "gate × empty selected array → waiting", question: gateQ, details: { response: { answers: { [gateQ.id]: { selected: [] } } } }, expected: "waiting" },
+  { name: "gate × missing response → waiting", question: gateQ, details: {}, expected: "waiting" },
+  { name: "gate × notes-only → waiting (notes never satisfy a gate)", question: gateQ, details: { response: { answers: { [gateQ.id]: { notes: "looks fine" } } } }, expected: "waiting" },
+  { name: "gate × cancelled → cancelled", question: gateQ, details: { cancelled: true, response: null }, expected: "cancelled" },
+  // consent
+  { name: "consent × valid selected → answered", question: consentQ, details: { response: { answers: { [consentQ.id]: { selected: "Option B" } } } }, expected: "answered" },
+  { name: "consent × empty selected → waiting (#528)", question: consentQ, details: { response: { answers: { [consentQ.id]: { selected: "" } } } }, expected: "waiting" },
+  { name: "consent × missing answer → waiting", question: consentQ, details: { response: { answers: {} } }, expected: "waiting" },
+  { name: "consent × missing response → waiting", question: consentQ, details: {}, expected: "waiting" },
+  { name: "consent × notes-only → answered (a real user utterance)", question: consentQ, details: { response: { answers: { [consentQ.id]: { notes: "neither — keep it simple" } } } }, expected: "answered" },
+  { name: "consent × cancelled → cancelled", question: consentQ, details: { cancelled: true, response: null }, expected: "cancelled" },
+];
+
+for (const row of POLICY_TABLE) {
+  test(`policy table: ${row.name}`, () => {
+    assert.equal(evaluateAnswer({ question: row.question, details: row.details }), row.expected);
+  });
+}
+
+// ── Round evaluation ────────────────────────────────────────────────────────
+
+test("evaluateAskUserQuestionsRound: most blocking outcome wins", () => {
+  const details = {
+    response: {
+      answers: {
+        [consentQ.id]: { selected: "Option A (Recommended)" },
+        second: { selected: "" },
+      },
+    },
+  };
+  assert.equal(
+    evaluateAskUserQuestionsRound([consentQ, { id: "second", options: CONSENT_OPTIONS }], details),
+    "waiting",
+  );
+  assert.equal(evaluateAskUserQuestionsRound([consentQ], { cancelled: true }), "cancelled");
+  assert.equal(
+    evaluateAskUserQuestionsRound([consentQ], {
+      response: { answers: { [consentQ.id]: { selected: "Option A (Recommended)" } } },
+    }),
+    "answered",
+  );
+  // Empty round with no response is a no-op for callers.
+  assert.equal(evaluateAskUserQuestionsRound([], {}), "answered");
+});
+
+// ── Regression: #528 empty selected on a consent question never answers ─────
+
+test("#528: empty selected on a NON-gate question is waiting, not answered", () => {
+  const outcome = evaluateAskUserQuestionsRound(
+    [{ id: "scope_check", options: CONSENT_OPTIONS }],
+    { response: { answers: { scope_check: { selected: "" } } } },
+  );
+  assert.equal(outcome, "waiting");
+  const message = formatUnansweredConsentQuestionMessage([{ id: "scope_check" }]);
+  assert.match(message, /scope_check/);
+  assert.match(message, /not consent/i);
+});
+
+// ── Regression: #682 pause promotion outside the old unit-type allowlist ────
+
+test("#682: consent question in a non-allowlisted unit type pauses", () => {
+  const messages = [assistantMessage("The slice is staged. Ready to proceed with the merge?")];
+  assert.equal(shouldPauseForQuestion("execute-task", messages), true);
+});
+
+test("#682: consent question in interactive mode (no unit type) pauses", () => {
+  const messages = [assistantMessage("Does this capture your intent correctly?")];
+  assert.equal(shouldPauseForQuestion(undefined, messages), true);
+});
+
+test("#682: cancellation marker pauses regardless of unit type", () => {
+  const messages = [assistantMessage("ask_user_questions was cancelled before receiving a response")];
+  assert.equal(shouldPauseForQuestion(undefined, messages), true);
+  assert.equal(shouldPauseForQuestion("execute-task", messages), true);
+});
+
+test("shouldPauseForQuestion: informational prose never pauses", () => {
+  const messages = [assistantMessage("All tests pass. Moving on to the next task.")];
+  assert.equal(shouldPauseForQuestion(undefined, messages), false);
+  assert.equal(shouldPauseForQuestion("discuss-project", messages), false);
+});
+
+test("shouldPauseForQuestion: plain non-approval question does not pause", () => {
+  // Generic questions (no approval/consent language) stay informational —
+  // matches the historical contract in deep-project-auto-loop tests.
+  const messages = [assistantMessage("Which file did you mean?")];
+  assert.equal(shouldPauseForQuestion("discuss-project", messages), false);
+});
+
+test("shouldPauseForQuestion: explicit waiting-for-approval phrase pauses without a question mark", () => {
+  // The streaming pre-filter (no "?" → bail) must not skip the explicit
+  // wait-phrase boundary, which carries no question mark.
+  const messages = [assistantMessage("I am waiting for your approval before writing the file.")];
+  assert.equal(shouldPauseForQuestion(undefined, messages), true);
+});
+
+// ── Message-text extraction (merged from user-input-boundary tests) ─────────
+
+test("lastAssistantText extracts the latest assistant text block content", () => {
+  assert.equal(
+    lastAssistantText([
+      { role: "assistant", content: "Older message" },
+      {
+        role: "assistant",
+        content: [
+          { type: "text", text: "First line" },
+          { type: "text", text: "Second line" },
+        ],
+      },
+    ]),
+    "First line\nSecond line",
+  );
+  assert.equal(lastAssistantText(null), "");
+});
+
+test("lastAssistantText includes thinking blocks so rate-limit notices are not dropped", () => {
+  // Turn with only a thinking block (no text block) — must not return ""
+  const result = lastAssistantText([
+    {
+      role: "assistant",
+      content: [
+        { type: "thinking", thinking: "You've hit your limit · resets in 2h" },
+      ],
+    },
+  ]);
+  assert.ok(result.includes("You've hit your limit"), `expected rate-limit text, got: ${JSON.stringify(result)}`);
+});
+
+// ── Awaiting-input boundaries (merged from user-input-boundary tests) ───────
+
+test("isAwaitingUserInput does not trigger on thinking-block question marks", () => {
+  // A thinking block with a question mark must NOT pause auto-mode —
+  // it's internal reasoning, not a user-visible prompt.
+  const messages = [
+    {
+      role: "assistant",
+      content: [
+        { type: "thinking", thinking: "Should I skip research? Let me check the config." },
+      ],
+    },
+  ];
+  assert.equal(isAwaitingUserInput(messages), false);
+  assert.equal(shouldPauseForQuestion("discuss-project", messages), false);
+});
+
+test("isAwaitingUserInput does not trigger on thinking-block approval phrases", () => {
+  // A thinking block with approval phrases must NOT pause auto-mode.
+  const messages = [
+    {
+      role: "assistant",
+      content: [
+        { type: "thinking", thinking: "The user confirmed and approved the plan. Should I proceed?" },
+      ],
+    },
+  ];
+  assert.equal(isAwaitingUserInput(messages), false);
+  assert.equal(shouldPauseForQuestion("discuss-requirements", messages), false);
+});
+
+test("isAwaitingUserInput treats plain-text next steps menus as waiting for the user (#454)", () => {
+  const messages = [
+    {
+      role: "assistant",
+      content: [
+        "Next steps:",
+        "1. Walk through the runtime placement check above.",
+        "2. Build a release once you're satisfied.",
+        "3. Other.",
+      ].join("\n"),
+    },
+  ];
+  assert.equal(isAwaitingUserInput(messages), true);
+});
+
+test("isAwaitingUserInput still triggers on text-block question marks when thinking is also present", () => {
+  // When thinking + text are both present and the text asks a question, it should still pause.
+  const messages = [
+    {
+      role: "assistant",
+      content: [
+        { type: "thinking", thinking: "Internal reasoning without questions." },
+        { type: "text", text: "Does this look correct?" },
+      ],
+    },
+  ];
+  assert.equal(isAwaitingUserInput(messages), true);
+});
+
+// ── In-flight ask_user_questions detection (merged from user-input-boundary) ─
+
+test("messageHasPendingAskUserQuestionsTool detects in-flight structured question tools", () => {
+  // state: "running" with no externalResult → still in-flight
+  assert.equal(
+    messageHasPendingAskUserQuestionsTool({
+      role: "assistant",
+      content: [
+        { type: "text", text: "Which direction?" },
+        { type: "toolCall", name: "mcp__gsd-workflow__ask_user_questions", state: "running" },
+      ],
+    }),
+    true,
+  );
+
+  // no state, no externalResult — streaming block that hasn't completed yet
+  assert.equal(
+    messageHasPendingAskUserQuestionsTool({
+      role: "assistant",
+      content: [
+        { type: "toolCall", name: "ask_user_questions" },
+      ],
+    }),
+    true,
+  );
+
+  // state: "completed" — legacy state-based completion
+  assert.equal(
+    messageHasPendingAskUserQuestionsTool({
+      role: "assistant",
+      content: [
+        { type: "toolCall", name: "ask_user_questions", state: "completed" },
+      ],
+    }),
+    false,
+  );
+
+  // externalResult present — Claude Code signals completion via externalResult, not state
+  assert.equal(
+    messageHasPendingAskUserQuestionsTool({
+      role: "assistant",
+      content: [
+        { type: "toolCall", name: "ask_user_questions", externalResult: { content: [], isError: false } },
+      ],
+    }),
+    false,
+  );
+  assert.equal(
+    messageHasPendingAskUserQuestionsTool({
+      role: "assistant",
+      content: [
+        {
+          type: "toolCall",
+          name: "ask_user_questions",
+          externalResult: { content: [{ type: "text", text: "answer" }], isError: false },
+        },
+      ],
+    }),
+    false,
+  );
+
+  // serverToolUse shape (claude-code-cli MCP path) — no externalResult → in-flight
+  assert.equal(
+    messageHasPendingAskUserQuestionsTool({
+      role: "assistant",
+      content: [
+        { type: "serverToolUse", name: "mcp__gsd-workflow__ask_user_questions" },
+      ],
+    }),
+    true,
+  );
+  // serverToolUse shape — externalResult present → completed
+  assert.equal(
+    messageHasPendingAskUserQuestionsTool({
+      role: "assistant",
+      content: [
+        { type: "serverToolUse", name: "ask_user_questions", externalResult: { content: [], isError: false } },
+      ],
+    }),
+    false,
+  );
+});