gsd-pi 2.82.0-dev.2841a1e44 → 2.82.0-dev.9d5798940

package/src/resources/extensions/gsd/tests/unit-context-manifest.test.ts

@@ -14,7 +14,10 @@ import {
   type SkillsPolicy,
   type UnitContextManifest,
 } from "../unit-context-manifest.ts";
-import { ALLOWED_PLANNING_DISPATCH_AGENTS } from "../bootstrap/write-gate.ts";
+import {
+  ALLOWED_PLANNING_DISPATCH_AGENTS,
+  shouldBlockPlanningUnit,
+} from "../bootstrap/write-gate.ts";
 import {
   getRequiredWorkflowToolsForAutoUnit,
   getRequiredWorkflowToolsForGuidedUnit,
@@ -217,7 +220,7 @@ test("#4934: every manifest declares a tools policy", () => {
 });
 
 test("#4934: tools.mode is one of the declared policies", () => {
-  const validModes = new Set(["all", "read-only", "planning", "planning-dispatch", "docs"]);
+  const validModes = new Set(["all", "read-only", "planning", "planning-dispatch", "docs", "verification"]);
   for (const [unitType, manifest] of Object.entries(UNIT_MANIFESTS)) {
     const mode = (manifest as { tools: { mode: string } }).tools.mode;
     assert.ok(
@@ -227,28 +230,83 @@ test("#4934: tools.mode is one of the declared policies", () => {
   }
 });
 
-test('#4934: only execute-task and reactive-execute may use tools.mode "all" (full source-tree write access)', () => {
-  const allowedAllUnits = new Set(["execute-task", "reactive-execute"]);
+test('#4934: only execution units and complete-milestone may use tools.mode "all"', () => {
+  const allowedAllUnits = new Set(["execute-task", "reactive-execute", "complete-milestone"]);
   for (const [unitType, manifest] of Object.entries(UNIT_MANIFESTS)) {
     const mode = (manifest as { tools: { mode: string } }).tools.mode;
     if (mode === "all") {
       assert.ok(
         allowedAllUnits.has(unitType),
-        `manifest "${unitType}" declares tools.mode = "all" but is not on the execute-track. ` +
-        'Only execute-task and reactive-execute should have full source write access; ' +
+        `manifest "${unitType}" declares tools.mode = "all" but is not explicitly allowed. ` +
+        'Only execute-task, reactive-execute, and complete-milestone should have full source write access; ' +
         'planning/discuss/research units must use "planning" or "planning-dispatch" (or "docs" for rewrite-docs).',
       );
     }
   }
 });
 
+test("#5453: complete-milestone uses all tools so bash verification is not planning-dispatch blocked", () => {
+  const manifest = UNIT_MANIFESTS["complete-milestone"];
+
+  assert.strictEqual(manifest.tools.mode, "all");
+  assert.deepEqual(resolveSubagentPermissionContract("complete-milestone"), {
+    allowed: true,
+    allowedSubagents: ["*"],
+    toolsMode: "all",
+  });
+  // Runtime gate-level regression: these verification commands were blocked
+  // under planning-dispatch in #5453; complete-milestone must bypass that gate.
+  for (const cmd of ["git diff --name-only HEAD~1", "git log -n1 --oneline"]) {
+    const result = shouldBlockPlanningUnit(
+      "bash",
+      cmd,
+      process.cwd(),
+      "complete-milestone",
+      manifest.tools,
+    );
+    assert.strictEqual(
+      result.block,
+      false,
+      `shouldBlockPlanningUnit must not block ${cmd} for complete-milestone: ${result.reason}`,
+    );
+  }
+});
+
+test("#5843: run-uat uses verification tools policy so build/test commands can run", () => {
+  const manifest = UNIT_MANIFESTS["run-uat"];
+
+  assert.strictEqual(manifest.tools.mode, "verification");
+
+  const buildResult = shouldBlockPlanningUnit(
+    "bash",
+    "npm run build 2>&1",
+    process.cwd(),
+    "run-uat",
+    manifest.tools,
+  );
+  assert.strictEqual(
+    buildResult.block,
+    false,
+    `run-uat must allow build verification commands: ${buildResult.reason}`,
+  );
+
+  const sourceWriteResult = shouldBlockPlanningUnit(
+    "edit",
+    "src/main.ts",
+    process.cwd(),
+    "run-uat",
+    manifest.tools,
+  );
+  assert.strictEqual(sourceWriteResult.block, true);
+  assert.match(sourceWriteResult.reason!, /tools-policy "verification"/);
+});
+
 test('planning-dispatch mode is reserved for slice-level decomposition and completion units', () => {
   const allowedDispatchUnits = new Set([
     "plan-slice",
     "research-slice",
     "refine-slice",
     "complete-slice",
-    "complete-milestone",
     "gate-evaluate",
     // Deep planning mode: research-project orchestrates 4 parallel research
     // subagents (stack/features/architecture/pitfalls). Subagent dispatch is

package/src/resources/extensions/gsd/tests/worktree-journal-events.test.ts

@@ -30,11 +30,13 @@ function initGitRepoIn(base: string, isolation: "worktree" | "branch" | "none"):
 }
 import {
   WorktreeLifecycle,
+  resetRecentWorktreeMergeFailuresForTest,
   type WorktreeLifecycleDeps,
   type NotifyCtx,
 } from "../worktree-lifecycle.js";
 import { WorktreeStateProjection } from "../worktree-state-projection.js";
 import { type TaskCommitContext } from "../worktree.js";
+import { MergeConflictError } from "../git-service.js";
 
 // ADR-016 phase 2 / C-track retired all worktree-manager + cache + prefs
 // fields from `WorktreeLifecycleDeps`. Tests still pass them as overrides
@@ -141,6 +143,20 @@ function readJournalEntries(basePath: string): JournalEntry[] {
   }
 }
 
+function setupMergeWorktree(basePath: string, milestoneId: string): string {
+  initGitRepoIn(basePath, "worktree");
+  execFileSync("git", ["checkout", "-b", `milestone/${milestoneId}`], { cwd: basePath, stdio: "pipe" });
+  execFileSync("git", ["checkout", "main"], { cwd: basePath, stdio: "pipe" });
+  const wt = join(basePath, ".gsd", "worktrees", milestoneId);
+  execFileSync("git", ["worktree", "add", wt, `milestone/${milestoneId}`], { cwd: basePath, stdio: "pipe" });
+  mkdirSync(join(basePath, ".gsd", "milestones", milestoneId), { recursive: true });
+  writeFileSync(
+    join(basePath, ".gsd", "milestones", milestoneId, `${milestoneId}-ROADMAP.md`),
+    `# ${milestoneId}\n- [x] S01: Slice one\n`,
+  );
+  return wt;
+}
+
 // ─── Tests ───────────────────────────────────────────────────────────────────
 
 describe("worktree journal events", () => {
@@ -148,6 +164,7 @@ describe("worktree journal events", () => {
   const originalCwd = process.cwd();
 
   beforeEach(() => {
+    resetRecentWorktreeMergeFailuresForTest();
     // realpathSync to match what `auto-worktree.ts` returns from
     // `resolveWorktreeProjectRoot` (macOS resolves `/var` → `/private/var`).
     tmp = realpathSync(mkdtempSync(join(tmpdir(), "wt-journal-")));
@@ -284,17 +301,7 @@ describe("worktree journal events", () => {
   });
 
   test("mergeAndExit emits worktree-merge-failed on error", () => {
-    initGitRepoIn(tmp, "worktree");
-    execFileSync("git", ["checkout", "-b", "milestone/M001"], { cwd: tmp, stdio: "pipe" });
-    execFileSync("git", ["checkout", "main"], { cwd: tmp, stdio: "pipe" });
-    const wt = join(tmp, ".gsd", "worktrees", "M001");
-    execFileSync("git", ["worktree", "add", wt, "milestone/M001"], { cwd: tmp, stdio: "pipe" });
-    mkdirSync(join(tmp, ".gsd", "milestones", "M001"), { recursive: true });
-    writeFileSync(
-      join(tmp, ".gsd", "milestones", "M001", "M001-ROADMAP.md"),
-      "# M001\n- [x] S01: Slice one\n",
-    );
-
+    const wt = setupMergeWorktree(tmp, "M001");
     const s = makeSession({ basePath: wt, originalBasePath: tmp });
     const deps = makeDeps({
       mergeMilestoneToMain: () => { throw new Error("conflict in main"); },
@@ -318,11 +325,56 @@ describe("worktree journal events", () => {
       );
     }
 
+    new WorktreeLifecycle(s, deps).exitMilestone(
+      "M001",
+      { merge: true },
+      makeNotifyCtx(),
+    );
+
     const entries = readJournalEntries(tmp);
-    const failed = entries.find(e => e.eventType === "worktree-merge-failed");
+    const failures = entries.filter(e => e.eventType === "worktree-merge-failed");
+    const failed = failures[0];
     assert.ok(failed, "worktree-merge-failed event should be emitted");
     assert.equal(failed!.data?.milestoneId, "M001");
     assert.equal(failed!.data?.error, "conflict in main");
+    assert.equal(failures.length, 1, "duplicate merge failures are journaled once");
+  });
+
+  test("merge failure dedupe uses stable conflict category and expires", (t) => {
+    let now = 1_000_000;
+    t.mock.method(Date, "now", () => now);
+    const wt = setupMergeWorktree(tmp, "M001");
+    const s = makeSession({ basePath: wt, originalBasePath: tmp });
+    let attempt = 0;
+    const deps = makeDeps({
+      mergeMilestoneToMain: () => {
+        attempt += 1;
+        throw new MergeConflictError(
+          attempt === 1 ? ["src/a.ts"] : ["src/b.ts", "src/c.ts"],
+          "squash",
+          "milestone/M001",
+          "main",
+        );
+      },
+    });
+    const lifecycle = new WorktreeLifecycle(s, deps);
+
+    lifecycle.exitMilestone("M001", { merge: true }, makeNotifyCtx());
+    lifecycle.exitMilestone("M001", { merge: true }, makeNotifyCtx());
+
+    let failures = readJournalEntries(tmp).filter(e => e.eventType === "worktree-merge-failed");
+    assert.equal(failures.length, 1, "variable conflict filenames should not bypass dedupe");
+    assert.match(
+      String(failures[0]!.data?.error),
+      /src\/a\.ts/,
+      "journal payload keeps the original error message",
+    );
+
+    now += 60_001;
+    lifecycle.exitMilestone("M001", { merge: true }, makeNotifyCtx());
+
+    failures = readJournalEntries(tmp).filter(e => e.eventType === "worktree-merge-failed");
+    assert.equal(failures.length, 2, "same merge failure is journaled again after dedupe expiry");
   });
 
   test("journal entries have valid flowId, seq, and ts fields", () => {

package/src/resources/extensions/gsd/tests/write-gate-planning-unit.test.ts

@@ -26,6 +26,7 @@ const PLANNING_DISPATCH_REVIEW: ToolsPolicy = {
 };
 const READ_ONLY: ToolsPolicy = { mode: 'read-only' };
 const ALL: ToolsPolicy = { mode: 'all' };
+const VERIFICATION: ToolsPolicy = { mode: 'verification' };
 const DOCS: ToolsPolicy = {
   mode: 'docs',
   allowedPathGlobs: ['docs/**', 'README.md', 'README.*.md', 'CHANGELOG.md', '*.md'],
@@ -331,6 +332,36 @@ test('all-mode: execute-task can dispatch subagents', () => {
   assert.strictEqual(r.block, false);
 });
 
+// ─── verification mode: bash allowed, writes still scoped ─────────────────
+
+test('verification-mode: run-uat can run build commands', () => {
+  const r = shouldBlockPlanningUnit('bash', 'npm run build 2>&1', BASE, 'run-uat', VERIFICATION);
+  assert.strictEqual(r.block, false);
+});
+
+test('verification-mode: run-uat blocks destructive bash (rm -rf)', () => {
+  const r = shouldBlockPlanningUnit('bash', 'rm -rf dist', BASE, 'run-uat', VERIFICATION);
+  assert.strictEqual(r.block, true);
+  assert.match(r.reason!, /bash is restricted to build\/test verification commands/);
+});
+
+test('verification-mode: run-uat allows read-only investigative bash (git status)', () => {
+  const r = shouldBlockPlanningUnit('bash', 'git status', BASE, 'run-uat', VERIFICATION);
+  assert.strictEqual(r.block, false);
+});
+
+test('verification-mode: run-uat still blocks user source edits', () => {
+  const r = shouldBlockPlanningUnit('edit', join(BASE, 'src', 'main.ts'), BASE, 'run-uat', VERIFICATION);
+  assert.strictEqual(r.block, true);
+  assert.match(r.reason!, /tools-policy "verification"/);
+});
+
+test('verification-mode: run-uat still blocks subagent dispatch', () => {
+  const r = shouldBlockPlanningUnit('subagent', '', BASE, 'run-uat', VERIFICATION);
+  assert.strictEqual(r.block, true);
+  assert.match(r.reason!, /subagent dispatch is not permitted/);
+});
+
 // ─── read-only mode ───────────────────────────────────────────────────────
 
 test('read-only: blocks any edit even to .gsd/', () => {

package/src/resources/extensions/gsd/tools/plan-slice.ts

@@ -10,6 +10,7 @@ import {
   upsertTaskPlanning,
   insertGateRow,
   updateSliceStatus,
+  setSliceSketchFlag,
 } from "../gsd-db.js";
 import type { GateId } from "../types.js";
 import { invalidateStateCache } from "../state.js";
@@ -184,6 +185,7 @@ export async function handlePlanSlice(
       if (isDeferredStatus(parentSlice.status)) {
         updateSliceStatus(params.milestoneId, params.sliceId, "pending");
       }
+      setSliceSketchFlag(params.milestoneId, params.sliceId, false);
 
       upsertSlicePlanning(params.milestoneId, params.sliceId, {
         goal: params.goal,

package/src/resources/extensions/gsd/unit-context-manifest.ts

@@ -132,6 +132,9 @@ export type ContextModePolicy =
  *                    the explicit `allowedPathGlobs` set; Bash safe-allowlist;
  *                    no subagents. Reserved for rewrite-docs, which legitimately
  *                    edits project markdown outside .gsd/.
+ *   - "verification"
+ *                  — Read tools + Bash for verification commands, writes
+ *                    restricted to .gsd/**, no subagents.
  *
  * The allowlist for "docs" is declared per-manifest rather than hardcoded so
  * projects with non-standard doc layouts can extend it without forking the
@@ -143,7 +146,8 @@ export type ToolsPolicy =
   | { readonly mode: "read-only" }
   | { readonly mode: "planning" }
   | { readonly mode: "planning-dispatch"; readonly allowedSubagents: readonly string[] }
-  | { readonly mode: "docs"; readonly allowedPathGlobs: readonly string[] };
+  | { readonly mode: "docs"; readonly allowedPathGlobs: readonly string[] }
+  | { readonly mode: "verification" };
 
 // ─── Computed-artifact registry (#4924 v2 contract) ───────────────────────
 
@@ -288,6 +292,7 @@ const COMMON_BUDGET_SMALL = 250_000;    // ~65K tokens
 
 const TOOLS_ALL: ToolsPolicy = { mode: "all" };
 const TOOLS_PLANNING: ToolsPolicy = { mode: "planning" };
+const TOOLS_VERIFICATION: ToolsPolicy = { mode: "verification" };
 // Like TOOLS_PLANNING but permits dispatch to read-only recon/planning
 // specialists. Runtime-enforced by write-gate.ts before the subagent tool runs.
 const TOOLS_PLANNING_DISPATCH_RECON: ToolsPolicy = {
@@ -406,8 +411,8 @@ export const UNIT_MANIFESTS: Record<UnitType, UnitContextManifest> = {
     contextMode: "verification",
     // planning-dispatch: validation is a verification-fan-out unit. It reads
     // the milestone surface and dispatches reviewer/security/tester subagents
-    // to report findings without touching user source. Mirrors
-    // complete-milestone's policy. Write isolation to .gsd/ is preserved.
+    // to report findings without touching user source. Write isolation to
+    // .gsd/ is preserved.
     tools: TOOLS_PLANNING_DISPATCH_REVIEW,
     artifacts: {
       inline: ["roadmap", "slice-summary", "slice-uat", "requirements", "decisions", "templates"],
@@ -423,11 +428,9 @@ export const UNIT_MANIFESTS: Record<UnitType, UnitContextManifest> = {
     codebaseMap: false,
     preferences: "active-only",
     contextMode: "verification",
-    // planning-dispatch: completion is a high-leverage place to fan out to
-    // reviewer / security / tester subagents. They read the diff and report
-    // findings; they do not write user source. Write isolation to .gsd/ is
-    // preserved.
-    tools: TOOLS_PLANNING_DISPATCH_REVIEW,
+    // Milestone closeout must run unrestricted shell verification commands
+    // against the final diff before recording completion.
+    tools: TOOLS_ALL,
     artifacts: {
       // #4780 landed slice-summary as excerpt for this unit; phase 2 of
       // the architecture will read this manifest as the source of truth
@@ -589,7 +592,7 @@ export const UNIT_MANIFESTS: Record<UnitType, UnitContextManifest> = {
     codebaseMap: false,
     preferences: "active-only",
     contextMode: "verification",
-    tools: TOOLS_PLANNING,
+    tools: TOOLS_VERIFICATION,
     artifacts: {
       // Phase 3 migration (#4782): manifest matches today's actual
       // buildRunUatPrompt inlining. Prior phase-1 entry listed

package/src/resources/extensions/gsd/worktree-lifecycle.ts

@@ -76,6 +76,13 @@ import {
   teardownAutoWorktree,
 } from "./auto-worktree.js";
 
+const recentWorktreeMergeFailures = new Map<string, number>();
+const MERGE_FAILURE_DEDUPE_MS = 60_000;
+
+export function resetRecentWorktreeMergeFailuresForTest(): void {
+  recentWorktreeMergeFailures.clear();
+}
+
 // ─── Types ───────────────────────────────────────────────────────────────
 
 export interface NotifyCtx {
@@ -836,6 +843,32 @@ function rebuildGitService(
   s.gitService = deps.gitServiceFactory(s.basePath);
 }
 
+function emitWorktreeMergeFailedOnce(
+  basePath: string,
+  milestoneId: string,
+  err: unknown,
+): void {
+  const msg = err instanceof Error ? err.message : String(err);
+  const errorCategory = err instanceof Error ? err.name : "Error";
+  const now = Date.now();
+  const key = `${basePath}\0${milestoneId}\0${errorCategory}`;
+  const previous = recentWorktreeMergeFailures.get(key);
+  if (previous && now - previous < MERGE_FAILURE_DEDUPE_MS) return;
+  for (const [candidate, ts] of recentWorktreeMergeFailures) {
+    if (now - ts >= MERGE_FAILURE_DEDUPE_MS) {
+      recentWorktreeMergeFailures.delete(candidate);
+    }
+  }
+  emitJournalEvent(basePath, {
+    ts: new Date().toISOString(),
+    flowId: randomUUID(),
+    seq: 0,
+    eventType: "worktree-merge-failed",
+    data: { milestoneId, error: msg },
+  });
+  recentWorktreeMergeFailures.set(key, now);
+}
+
 // ─── Session-less merge entry (ADR-016 phase 2 / A1) ─────────────────────
 
 /**
@@ -985,13 +1018,7 @@ function _mergeWorktreeModeImpl(
       error: msg,
       fallback: "chdir-to-project-root",
     });
-    emitJournalEvent(originalBasePath || worktreeBasePath, {
-      ts: new Date().toISOString(),
-      flowId: randomUUID(),
-      seq: 0,
-      eventType: "worktree-merge-failed",
-      data: { milestoneId, error: msg },
-    });
+    emitWorktreeMergeFailedOnce(originalBasePath || worktreeBasePath, milestoneId, err);
     // Surface a clear, actionable error. Worktree and milestone branch
     // are intentionally preserved — nothing has been deleted. User can
     // retry /gsd dispatch complete-milestone or merge manually once the