gsd-pi 2.77.0-dev.eaa4973bc → 2.78.0-dev.aeeb2ca00

package/src/resources/extensions/gsd/tests/ready-phrase-no-files-4573.test.ts

@@ -60,10 +60,24 @@ function mkBase(): string {
   return base;
 }
 
-function assistantMsg(text: string, opts: { toolUse?: boolean } = {}): any {
+function assistantMsg(
+  text: string,
+  opts: { toolUse?: boolean | "toolCall" | "serverToolUse" } = {},
+): any {
   const content: any[] = [];
   if (text) content.push({ type: "text", text });
-  if (opts.toolUse) content.push({ type: "tool_use", name: "whatever", input: {} });
+  if (opts.toolUse) {
+    // The canonical pi-ai AssistantMessage uses "toolCall" / "serverToolUse"
+    // (see packages/pi-ai/src/types.ts). Every provider — anthropic-direct,
+    // claude-code-cli, openai — normalizes incoming tool blocks into these
+    // shapes before they reach guided-flow. The Anthropic-wire literal
+    // "tool_use" never appears here.
+    if (opts.toolUse === "serverToolUse") {
+      content.push({ type: "serverToolUse", id: "test-id", name: "web_search", input: {} });
+    } else {
+      content.push({ type: "toolCall", id: "test-id", name: "whatever", arguments: {} });
+    }
+  }
   return { role: "assistant", content };
 }
 
@@ -314,6 +328,65 @@ describe("#4573 maybeHandleEmptyIntentTurn", () => {
     }
   });
 
+  // Regression for #4658 — under claude-code-cli, MCP tool calls (e.g.
+  // ask_user_questions) reach guided-flow as canonical "toolCall" / "serverToolUse"
+  // blocks. Pre-fix, hasToolUse only matched the Anthropic-wire literal "tool_use",
+  // so the empty-turn nudge fired during pre-question narration and pre-empted the
+  // user's chance to answer.
+  test("cc-cli MCP tool call surfaced as canonical toolCall → not flagged", () => {
+    const base = mkBase();
+    try {
+      const cap = mkCapture();
+      setPendingAutoStart(base, {
+        basePath: base,
+        milestoneId: "M001",
+        ctx: mkCtx(cap),
+        pi: mkPi(cap),
+      });
+      const handled = maybeHandleEmptyIntentTurn(
+        {
+          messages: [
+            assistantMsg("Let me call ask_user_questions to gather your preferences.", {
+              toolUse: "toolCall",
+            }),
+          ],
+        },
+        false,
+      );
+      assert.equal(handled, false);
+      assert.equal(cap.messages.length, 0);
+    } finally {
+      clearPendingAutoStart();
+    }
+  });
+
+  test("serverToolUse block (cc-cli web search etc.) → not flagged", () => {
+    const base = mkBase();
+    try {
+      const cap = mkCapture();
+      setPendingAutoStart(base, {
+        basePath: base,
+        milestoneId: "M001",
+        ctx: mkCtx(cap),
+        pi: mkPi(cap),
+      });
+      const handled = maybeHandleEmptyIntentTurn(
+        {
+          messages: [
+            assistantMsg("Let me invoke the search tool now.", {
+              toolUse: "serverToolUse",
+            }),
+          ],
+        },
+        false,
+      );
+      assert.equal(handled, false);
+      assert.equal(cap.messages.length, 0);
+    } finally {
+      clearPendingAutoStart();
+    }
+  });
+
   test("ready phrase is NOT treated as empty-turn (handled by other recovery path)", () => {
     const base = mkBase();
     try {

package/src/resources/extensions/gsd/tests/reassess-default-optin.test.ts

@@ -0,0 +1,132 @@
+// GSD-2 — ADR-003 §4 behavior contract: reassess-roadmap is opt-in.
+// Companion to (eventually replacing) the source-grep assertions in
+// token-profile.test.ts. This file verifies the dispatch rule's guard
+// behavior directly rather than inspecting source text.
+
+import test from "node:test";
+import assert from "node:assert/strict";
+import { mkdirSync, readFileSync, rmSync } from "node:fs";
+import { tmpdir } from "node:os";
+import { join } from "node:path";
+import { randomUUID } from "node:crypto";
+
+import {
+  DISPATCH_RULES,
+  setReassessmentCheckerForTest,
+  type DispatchAction,
+  type DispatchContext,
+} from "../auto-dispatch.ts";
+import { resolveProfileDefaults } from "../preferences-models.ts";
+import type { GSDState } from "../types.ts";
+import type { GSDPreferences } from "../preferences.ts";
+
+const REASSESS_RULE_NAME = "reassess-roadmap (post-completion)";
+
+function makeIsolatedBase(): string {
+  const base = join(tmpdir(), `gsd-reassess-default-${randomUUID()}`);
+  mkdirSync(join(base, ".gsd", "milestones", "M001", "slices", "S01", "tasks"), { recursive: true });
+  return base;
+}
+
+function makeCtx(prefs: GSDPreferences | undefined, basePath: string): DispatchContext {
+  const state: GSDState = {
+    phase: "executing",
+    activeMilestone: { id: "M001", title: "Test" },
+    activeSlice: { id: "S01", title: "First" },
+    activeTask: null,
+    recentDecisions: [],
+    blockers: [],
+    nextAction: "",
+    registry: [{ id: "M001", title: "Test", status: "active" }],
+  };
+  return { basePath, mid: "M001", midTitle: "Test", state, prefs };
+}
+
+function reassessRule() {
+  const rule = DISPATCH_RULES.find(r => r.name === REASSESS_RULE_NAME);
+  assert.ok(rule, `dispatch rule "${REASSESS_RULE_NAME}" must exist`);
+  return rule!;
+}
+
+const guardCases: Array<{ name: string; prefs: GSDPreferences | undefined; message?: string }> = [
+  {
+    name: "prefs is undefined (new default)",
+    prefs: undefined,
+    message: "default behavior must be opt-in — no prefs means no reassess dispatch",
+  },
+  {
+    name: "prefs.phases is undefined",
+    prefs: {} as GSDPreferences,
+  },
+  {
+    name: "phases.reassess_after_slice is explicitly false",
+    prefs: { phases: { reassess_after_slice: false } } as unknown as GSDPreferences,
+  },
+  {
+    name: "phases.skip_reassess is true (short-circuit guard preserved)",
+    prefs: { phases: { skip_reassess: true, reassess_after_slice: true } } as unknown as GSDPreferences,
+    message: "skip_reassess must win over reassess_after_slice",
+  },
+];
+
+for (const { name, prefs, message } of guardCases) {
+  test(`ADR-003 §4: reassess-roadmap does NOT dispatch when ${name}`, async (t) => {
+    const base = makeIsolatedBase();
+    t.after(() => { try { rmSync(base, { recursive: true, force: true }); } catch {} });
+
+    let checkerCalls = 0;
+    const restoreChecker = setReassessmentCheckerForTest(async () => {
+      checkerCalls++;
+      return { sliceId: "S01" };
+    });
+    t.after(restoreChecker);
+
+    const result = await reassessRule().match(makeCtx(prefs, base));
+    assert.strictEqual(result, null, message);
+    assert.strictEqual(checkerCalls, 0, "preference guards must short-circuit before reassessment detection");
+  });
+}
+
+test("ADR-003 §4: reassess-roadmap opt-in path dispatches after reassessment detection", async (t) => {
+  const base = makeIsolatedBase();
+  t.after(() => { try { rmSync(base, { recursive: true, force: true }); } catch {} });
+
+  const checkerCalls: Array<{ basePath: string; mid: string; activeSliceId: string | undefined }> = [];
+  const restoreChecker = setReassessmentCheckerForTest(async (checkerBase, checkerMid, state) => {
+    checkerCalls.push({ basePath: checkerBase, mid: checkerMid, activeSliceId: state.activeSlice?.id });
+    return { sliceId: "S01" };
+  });
+  t.after(restoreChecker);
+
+  const prefs = { phases: { reassess_after_slice: true } } as unknown as GSDPreferences;
+  const result: DispatchAction | null = await reassessRule().match(makeCtx(prefs, base));
+
+  assert.deepStrictEqual(checkerCalls, [{ basePath: base, mid: "M001", activeSliceId: "S01" }]);
+  assert.ok(result, "opt-in path should return a dispatch action when reassessment is needed");
+  assert.strictEqual(result.action, "dispatch");
+  if (result.action !== "dispatch") assert.fail("expected dispatch action");
+  assert.strictEqual(result.unitType, "reassess-roadmap");
+  assert.strictEqual(result.unitId, "M001/S01");
+  assert.match(result.prompt, /reassess/i);
+});
+
+test("ADR-003 §4: burn-max profile opts into dedicated reassess-roadmap dispatch", () => {
+  const defaults = resolveProfileDefaults("burn-max");
+  assert.strictEqual(defaults.phases?.reassess_after_slice, true);
+  assert.strictEqual(defaults.phases?.skip_reassess, false);
+});
+
+test("ADR-003 §4: plan-slice prompt and MCP tool agree on reassess sliceChanges shape", () => {
+  const planSlicePrompt = readFileSync(new URL("../prompts/plan-slice.md", import.meta.url), "utf8");
+  assert.match(planSlicePrompt, /gsd_reassess_roadmap/);
+  assert.match(planSlicePrompt, /sliceChanges\.modified/);
+  assert.match(planSlicePrompt, /sliceChanges\.added/);
+  assert.match(planSlicePrompt, /sliceChanges\.removed/);
+
+  const dbToolsSource = readFileSync(new URL("../bootstrap/db-tools.ts", import.meta.url), "utf8");
+  assert.match(dbToolsSource, /name:\s*"gsd_reassess_roadmap"/);
+  assert.match(dbToolsSource, /sliceChanges:\s*Type\.Object/);
+  assert.match(dbToolsSource, /modified:\s*Type\.Array/);
+  assert.match(dbToolsSource, /added:\s*Type\.Array/);
+  assert.match(dbToolsSource, /removed:\s*Type\.Array/);
+});

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

@@ -81,46 +81,14 @@ try {
     );
   }
 
-  // ═══ Verify the BUG scenario: omitting recoveryAttempts carries it over ═══
-
-  {
-    console.log("\n=== #2322: demonstrates bug — omitting recoveryAttempts carries it over ===");
-
-    const unitType = "execute-task";
-    const unitId = "M001/S01/T02";
-    const startedAt1 = Date.now() - 10000;
-
-    // First dispatch
-    writeUnitRuntimeRecord(base, unitType, unitId, startedAt1, {
-      phase: "dispatched",
-    });
-
-    // Timeout bumps recoveryAttempts to 1
-    writeUnitRuntimeRecord(base, unitType, unitId, startedAt1, {
-      recoveryAttempts: 1,
-      lastRecoveryReason: "hard",
-    });
-
-    // Re-dispatch WITHOUT resetting recoveryAttempts (the bug)
-    const startedAt2 = Date.now();
-    writeUnitRuntimeRecord(base, unitType, unitId, startedAt2, {
-      phase: "dispatched",
-      wrapupWarningSent: false,
-      timeoutAt: null,
-      lastProgressAt: startedAt2,
-      progressCount: 0,
-      lastProgressKind: "dispatch",
-      // recoveryAttempts: NOT included — this is the bug
-    });
-
-    const afterBuggyRedispatch = readUnitRuntimeRecord(base, unitType, unitId);
-    // This DEMONSTRATES the bug: recoveryAttempts is still 1
-    assertEq(
-      afterBuggyRedispatch?.recoveryAttempts,
-      1,
-      "BUG DEMO: recoveryAttempts carries over when not explicitly reset",
-    );
-  }
+  // Note (#4837): the previous "BUG DEMO" block was tautological — it only
+  // asserted that writeUnitRuntimeRecord merges partial updates into the
+  // stored record (which is the function's documented behaviour and is
+  // already covered by unit-runtime.test.ts). Proving merge semantics here
+  // added zero coverage of the #2322 fix. The first block above verifies the
+  // fix itself (explicit recoveryAttempts: 0 on re-dispatch resets the
+  // counter); the block below verifies the end-to-end budget-reset
+  // invariant.
 
   // ═══ Hard timeout maxRecoveryAttempts=1 — second dispatch must get full budget ═══