@opengsd/gsd-pi 1.9.0 → 1.10.0-dev.a5e42e99

package/src/resources/extensions/gsd/state-reconciliation/drift/external-markdown-edit.ts

@@ -9,13 +9,15 @@
 // refreshes the marker entry, closing the loop.
 
 import { existsSync, readFileSync } from "node:fs";
-import { join } from "node:path";
+import { basename, join } from "node:path";
 
 import {
   computeProjectionSha,
   readCompatMarker,
   writeCompatMarker,
+  type ProjectionEntry,
 } from "../../compat/compat-marker.js";
+import { _getAdapter } from "../../gsd-db.js";
 import { logWarning } from "../../workflow-logger.js";
 import type { GSDState } from "../../types.js";
 import type { DriftContext, DriftHandler, DriftRecord } from "../types.js";
@@ -25,6 +27,102 @@ type ExternalMarkdownEditDrift = Extract<
   { kind: "external-markdown-edit" }
 >;
 
+type ArtifactContentRow = {
+  full_content: string;
+};
+
+type ProjectionScope = {
+  milestoneId: string;
+  sliceId: string | null;
+  taskId: string | null;
+};
+
+function deepestProjectionScope(entities: readonly string[]): ProjectionScope | null {
+  let best: string[] | null = null;
+  for (const entity of entities) {
+    const parts = entity.split("/").filter(Boolean);
+    if (parts.length === 0) continue;
+    if (!best || parts.length > best.length) best = parts;
+  }
+  if (!best) return null;
+  return {
+    milestoneId: best[0]!,
+    sliceId: best[1] ?? null,
+    taskId: best[2] ?? null,
+  };
+}
+
+function artifactTypeFromProjectionPath(projectionPath: string): string | null {
+  const name = basename(projectionPath);
+  const match = name.match(/^(?:[A-Za-z]\d+|\d+(?:-\d+)?)-(.+)\.md$/);
+  return match?.[1]?.toUpperCase() ?? null;
+}
+
+function dbProjectionMatches(
+  projectionPath: string,
+  entry: ProjectionEntry,
+  actualSha: string,
+): boolean {
+  const adapter = _getAdapter();
+  if (!adapter) return false;
+
+  try {
+    const rows: ArtifactContentRow[] = adapter
+      .prepare("SELECT full_content FROM artifacts WHERE path = :path")
+      .all({ ":path": projectionPath }) as ArtifactContentRow[];
+
+    const scope = deepestProjectionScope(entry.entities);
+    const artifactType = artifactTypeFromProjectionPath(projectionPath);
+    if (scope?.taskId) {
+      rows.push(
+        ...(adapter
+          .prepare(
+            `SELECT full_content
+             FROM artifacts
+             WHERE milestone_id = :mid AND slice_id = :sid AND task_id = :tid
+               AND (:type IS NULL OR artifact_type = :type)`,
+          )
+          .all({
+            ":mid": scope.milestoneId,
+            ":sid": scope.sliceId,
+            ":tid": scope.taskId,
+            ":type": artifactType,
+          }) as ArtifactContentRow[]),
+      );
+    } else if (scope?.sliceId) {
+      rows.push(
+        ...(adapter
+          .prepare(
+            `SELECT full_content
+             FROM artifacts
+             WHERE milestone_id = :mid AND slice_id = :sid AND task_id IS NULL
+               AND (:type IS NULL OR artifact_type = :type)`,
+          )
+          .all({
+            ":mid": scope.milestoneId,
+            ":sid": scope.sliceId,
+            ":type": artifactType,
+          }) as ArtifactContentRow[]),
+      );
+    } else if (scope) {
+      rows.push(
+        ...(adapter
+          .prepare(
+            `SELECT full_content
+             FROM artifacts
+             WHERE milestone_id = :mid AND slice_id IS NULL AND task_id IS NULL
+               AND (:type IS NULL OR artifact_type = :type)`,
+          )
+          .all({ ":mid": scope.milestoneId, ":type": artifactType }) as ArtifactContentRow[]),
+      );
+    }
+
+    return rows.some((row) => computeProjectionSha(row.full_content) === actualSha);
+  } catch {
+    return false;
+  }
+}
+
 /**
  * Detect sha drift between the marker baseline and current file contents.
  *
@@ -38,16 +136,43 @@ function detectExternalMarkdownEdit(
   _state: GSDState,
   ctx: DriftContext,
 ): ExternalMarkdownEditDrift[] {
-  const marker = readCompatMarker(ctx.basePath);
+  const marker = readCompatMarker(ctx.basePath, {
+    healInvalidKeys: !ctx.dryRun,
+    quarantineInvalid: !ctx.dryRun,
+  });
   const entries = Object.entries(marker.projections);
   if (entries.length === 0) return [];
 
   const records: ExternalMarkdownEditDrift[] = [];
+  let markerChanged = false;
   for (const [projectionPath, entry] of entries) {
     const abs = join(ctx.basePath, ".gsd", projectionPath);
     if (!existsSync(abs)) continue;
     const actual = computeProjectionSha(readFileSync(abs, "utf-8"));
     if (actual === entry.sha) continue;
+    if (dbProjectionMatches(projectionPath, entry, actual)) {
+      if (!ctx.dryRun) {
+        marker.projections[projectionPath] = {
+          sha: actual,
+          entities: entry.entities,
+        };
+        markerChanged = true;
+        logWarning(
+          "reconcile",
+          `external-markdown-edit: refreshed stale compat marker for ${projectionPath} (expectedSha=${entry.sha}, actualSha=${actual})`,
+        );
+      } else {
+        logWarning(
+          "reconcile",
+          `external-markdown-edit: stale compat marker for ${projectionPath} (expectedSha=${entry.sha}, actualSha=${actual})`,
+        );
+      }
+      continue;
+    }
+    logWarning(
+      "reconcile",
+      `external-markdown-edit drift for ${projectionPath} (expectedSha=${entry.sha}, actualSha=${actual})`,
+    );
     records.push({
       kind: "external-markdown-edit",
       projectionPath,
@@ -56,6 +181,11 @@ function detectExternalMarkdownEdit(
       entities: entry.entities,
     });
   }
+  if (markerChanged) {
+    marker.lastWriter = "gsd-pi";
+    marker.lastProjectedAt = new Date().toISOString();
+    writeCompatMarker(ctx.basePath, marker);
+  }
   return records;
 }
 

package/src/resources/extensions/gsd/state-reconciliation/drift/external-planning-edit.ts

@@ -82,7 +82,10 @@ async function detectExternalPlanningEdit(
   _state: GSDState,
   ctx: DriftContext,
 ): Promise<ExternalPlanningEditDrift[]> {
-  const marker = readCompatMarker(ctx.basePath);
+  const marker = readCompatMarker(ctx.basePath, {
+    healInvalidKeys: !ctx.dryRun,
+    quarantineInvalid: !ctx.dryRun,
+  });
   if (!marker.planning?.active) {
     // Not yet activated. Activation (layout parse + DB import) is owned by
     // capturePlanningCompatIfNeeded, called from reconcileBeforeDispatch when

package/src/resources/extensions/gsd/state-reconciliation/drift/roadmap.ts

@@ -16,7 +16,7 @@ import { renderRoadmapFromDb } from "../../markdown-renderer.js";
 import { findMilestoneIds } from "../../milestone-ids.js";
 import { parseRoadmap } from "../../parsers-legacy.js";
 import { resolveMilestoneFile } from "../../paths.js";
-import { isClosedStatus } from "../../status-guards.js";
+import { isClosedStatus, isSkippedForDispatch } from "../../status-guards.js";
 import type { GSDState } from "../../types.js";
 import type { DriftContext, DriftHandler, DriftRecord } from "../types.js";
 
@@ -94,7 +94,9 @@ export function detectRoadmapDivergenceDrift(
   for (const milestoneId of findMilestoneIds(ctx.basePath)) {
     // Skip milestones that don't yet have a DB row — that's the
     // unregistered-milestone drift handler's responsibility.
-    if (!getMilestone(milestoneId)) continue;
+    const milestone = getMilestone(milestoneId);
+    if (!milestone) continue;
+    if (isSkippedForDispatch(milestone.status)) continue;
     if (milestoneHasDivergence(ctx.basePath, milestoneId)) {
       drifts.push({ kind: "roadmap-divergence", milestoneId });
     }

package/src/resources/extensions/gsd/stop-notice.ts

@@ -29,6 +29,14 @@ export function formatStopNoticePrefix(reason?: string | null): string {
   return displayReason ? `${prefix} — ${displayReason}` : prefix;
 }
 
+export function formatVerdictRecordedNotice(message: string): string {
+  return `Verdict recorded: ${message}`;
+}
+
+export function formatVerdictRejectedNotice(message: string): string {
+  return `Verdict rejected: ${message}`;
+}
+
 // ─── Classification (headless host side) ────────────────────────────────
 // The canonical lowercase prefixes the headless event loop recognizes in
 // notify messages. Emitters above and ad-hoc emitters elsewhere must start
@@ -42,6 +50,8 @@ export const TERMINAL_NOTICE_PREFIXES = [
   "auto-mode complete",
   "no active milestone",
   "auto-mode idle",
+  "verdict recorded",
+  "verdict rejected",
 ] as const;
 
 /** Manual-resolution notices emitted before auto-mode can formally pause/stop. */
@@ -85,6 +95,7 @@ export function isInteractiveMenuUnavailableNotice(message: string): boolean {
 export function isBlockedNoticeMessage(message: string): boolean {
   return (
     message.includes("blocked:") ||
+    message.startsWith("verdict rejected") ||
     (isPauseNotice(message) && !isNonBlockingPauseNotice(message)) ||
     isManualResolutionNotice(message) ||
     isInteractiveMenuUnavailableNotice(message)

package/src/resources/extensions/gsd/templates/PREFERENCES.md

@@ -92,6 +92,14 @@ remote_questions:
 uat_dispatch:
 post_unit_hooks: []
 pre_dispatch_hooks: []
+# planning_subagent_registry:
+#   my-custom-planner:
+#     read_only_specialist: true
+# planning_subagents:
+#   plan-milestone:
+#     allowed: [scout, planner, security]
+#   plan-slice:
+#     allowed: [scout, planner, reviewer, security]
 # language:
 # experimental:
 #   rtk: false

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

@@ -442,11 +442,11 @@ test("bare milestone id still falls back to a lone team-suffix projection (#1195
 // ── #1208: flat-phase execute-task summaries resolve at the phase root ────────
 //
 // In flat-phase layout task summaries are written beside the plan files
-// (phases/<phase>/T##-SUMMARY.md), NOT under a tasks/ subdir. A tasks/ dir may
-// still exist to hold auxiliary task-scoped gate artifacts (e.g. T01-VERIFY.json).
+// (phases/<phase>/S##-T##-SUMMARY.md), NOT under a tasks/ subdir. A tasks/ dir
+// may still exist to hold auxiliary task-scoped gate artifacts (e.g. T01-VERIFY.json).
 // Before the fix, the mere existence of that tasks/ dir redirected summary
-// verification into tasks/T##-SUMMARY.md — which never exists — trapping auto-mode
-// in a false verification retry that eventually paused after a successful task.
+// verification into tasks/S##-T##-SUMMARY.md — which never exists — trapping
+// auto-mode in a false verification retry after a successful task.
 
 test("flat-phase execute-task summary resolves at phase root when no tasks/ dir exists (#1208)", () => {
   const root = realpathSync(mkdtempSync(join(tmpdir(), "gsd-flat-task-summary-")));
@@ -461,7 +461,7 @@ test("flat-phase execute-task summary resolves at phase root when no tasks/ dir
 
     assert.equal(
       resolveExpectedArtifactPath("execute-task", "M049/S03/T02", root),
-      join(phaseDir, "T02-SUMMARY.md"),
+      join(phaseDir, "S03-T02-SUMMARY.md"),
     );
   } finally {
     _clearGsdRootCache();
@@ -489,12 +489,12 @@ test("flat-phase execute-task summary resolves at phase root even when tasks/ di
     const resolved = resolveExpectedArtifactPath("execute-task", "M049/S03/T02", root);
     assert.equal(
       resolved,
-      join(phaseDir, "T02-SUMMARY.md"),
+      join(phaseDir, "S03-T02-SUMMARY.md"),
       "flat-phase task summary must resolve at the phase root, not tasks/",
     );
     assert.notEqual(
       resolved,
-      join(tasksDir, "T02-SUMMARY.md"),
+      join(tasksDir, "S03-T02-SUMMARY.md"),
       "a tasks/ dir with gate artifacts must NOT redirect summary resolution into tasks/",
     );
   } finally {

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

@@ -41,6 +41,7 @@ import { claimMilestoneLease } from "../db/milestone-leases.js";
 import { recordDispatchClaim, markCanceled } from "../db/unit-dispatches.js";
 import { setRuntimeKv, getRuntimeKv } from "../db/runtime-kv.js";
 import { SourceObservationStore } from "../source-observations.js";
+import { autoCommitCurrentBranch } from "../worktree.js";
 
 // ─── Helpers ─────────────────────────────────────────────────────────────────
 
@@ -832,6 +833,37 @@ test("runUnit returns cancelled when session creation fails", async () => {
   assert.equal(pi.calls.length, 0);
 });
 
+test("runUnit returns cancelled when command context lacks newSession", async () => {
+  _resetPendingResolve();
+
+  const ctx = {
+    ...makeMockCtx(),
+    ui: {
+      notify: () => {},
+      setStatus: () => {},
+      setWorkingMessage: () => {},
+    },
+    sessionManager: {
+      getEntries: () => [],
+    },
+    modelRegistry: {
+      getProviderAuthMode: () => undefined,
+      isProviderRequestReady: () => true,
+    },
+  } as any;
+  const pi = makeMockPi();
+  const s = makeMockSession();
+  s.cmdCtx = {} as any;
+
+  const result = await runUnit(ctx, pi, s, "task", "T01", "prompt");
+
+  assert.equal(result.status, "cancelled");
+  assert.equal(result.errorContext?.category, "session-failed");
+  assert.equal(result.errorContext?.isTransient, false);
+  assert.match(result.errorContext?.message ?? "", /missing newSession/);
+  assert.equal(pi.calls.length, 0);
+});
+
 test("runUnit: TypeError from newSession is classified as structural (isTransient: false)", async () => {
   // Regression for #572: a TypeError thrown from newSession (e.g. "something is
   // not a function") indicates a programming error, not a transient provider
@@ -1582,6 +1614,130 @@ test("autoLoop exits when s.active is set to false", async (t) => {
   );
 });
 
+test("autoLoop stops before dispatch when command context lacks newSession", async () => {
+  _resetPendingResolve();
+
+  const ctx = makeMockCtx();
+  ctx.ui.setStatus = () => {};
+  const pi = makeMockPi();
+  const s = makeLoopSession({
+    cmdCtx: {
+      getContextUsage: () => ({ percent: 10, tokens: 1000, limit: 10000 }),
+    },
+  });
+  let stopReason: string | undefined;
+  let preserveWorktree: boolean | undefined;
+  let deriveCalled = false;
+
+  try {
+    const deps = makeMockDeps({
+      stopAuto: async (_ctx, _pi, reason, options) => {
+        stopReason = reason;
+        preserveWorktree = options?.preserveWorktree;
+        s.active = false;
+      },
+      deriveState: async () => {
+        deriveCalled = true;
+        throw new Error("deriveState should not run without command session support");
+      },
+    });
+
+    await autoLoop(ctx, pi, s, deps);
+
+    assert.equal(stopReason, "Auto-mode has no command context for dispatch.");
+    assert.equal(preserveWorktree, true);
+    assert.equal(deriveCalled, false);
+    assert.equal(pi.calls.length, 0);
+  } finally {
+    rmSync(s.basePath, { recursive: true, force: true });
+  }
+});
+
+test("autoLoop commits open unit work when command context lacks newSession", async () => {
+  _resetPendingResolve();
+
+  const ctx = makeMockCtx();
+  ctx.ui.setStatus = () => {};
+  const pi = makeMockPi();
+  const s = makeLoopSession({
+    cmdCtx: {
+      getContextUsage: () => ({ percent: 10, tokens: 1000, limit: 10000 }),
+    },
+    currentUnit: { type: "execute-task", id: "T01", startedAt: Date.now() },
+  });
+  const outputPath = join(s.basePath, "executor-output.txt");
+  let autoCommitArgs: { unitType: string; unitId: string } | undefined;
+  let preserveWorktree: boolean | undefined;
+
+  try {
+    const deps = makeMockDeps({
+      autoCommitUnit: async (basePath, unitType, unitId) => {
+        autoCommitArgs = { unitType, unitId };
+        return autoCommitCurrentBranch(basePath, unitType, unitId);
+      },
+      stopAuto: async (_ctx, _pi, _reason, options) => {
+        preserveWorktree = options?.preserveWorktree;
+        s.active = false;
+      },
+      deriveState: async () => {
+        throw new Error("deriveState should not run without command session support");
+      },
+    });
+
+    writeFileSync(outputPath, "open unit work before stop\n", "utf-8");
+
+    await autoLoop(ctx, pi, s, deps);
+
+    assert.deepEqual(autoCommitArgs, { unitType: "execute-task", unitId: "T01" });
+    assert.equal(preserveWorktree, true);
+    const committed = execSync("git show HEAD:executor-output.txt", {
+      cwd: s.basePath,
+      encoding: "utf-8",
+    });
+    assert.equal(committed, "open unit work before stop\n");
+    assert.equal(pi.calls.length, 0);
+  } finally {
+    rmSync(s.basePath, { recursive: true, force: true });
+  }
+});
+
+test("autoLoop snapshots dirty work when unit dispatch crashes", async () => {
+  _resetPendingResolve();
+
+  const ctx = makeMockCtx();
+  ctx.ui.setStatus = () => {};
+  const pi = makeMockPi();
+  const s = makeLoopSession();
+  const outputPath = join(s.basePath, "executor-output.txt");
+  let autoCommitCalls = 0;
+
+  try {
+    const deps = makeMockDeps({
+      ensurePreconditions: () => {
+        writeFileSync(outputPath, "saved before crash\n", "utf-8");
+        throw new Error("dispatch crash after file write");
+      },
+      autoCommitUnit: async (basePath, unitType, unitId) => {
+        autoCommitCalls++;
+        const commitMsg = autoCommitCurrentBranch(basePath, unitType, unitId);
+        s.active = false;
+        return commitMsg;
+      },
+    });
+
+    await autoLoop(ctx, pi, s, deps);
+
+    assert.equal(autoCommitCalls, 1);
+    const committed = execSync("git show HEAD:executor-output.txt", {
+      cwd: s.basePath,
+      encoding: "utf-8",
+    });
+    assert.equal(committed, "saved before crash\n");
+  } finally {
+    rmSync(s.basePath, { recursive: true, force: true });
+  }
+});
+
 test("autoLoop pauses visibly when Auto Orchestration Module is not wired", async () => {
   _resetPendingResolve();
 
@@ -5421,6 +5577,114 @@ test("runUnitPhase retries 0-tool units with ordinary network-related assistant
   assert.equal(deps.callLog.includes("pauseAuto"), false);
 });
 
+test("runUnitPhase pauses 0-tool units with pseudo tool-call text as serialization drift", async (t) => {
+  _resetPendingResolve();
+
+  const basePath = makeLoopTestBase("gsd-zero-tool-pseudo-tool-");
+  t.after(() => {
+    rmSync(basePath, { recursive: true, force: true });
+  });
+
+  const notifications: string[] = [];
+  const ctx = {
+    ...makeMockCtx(),
+    ui: {
+      notify: (msg: string) => { notifications.push(msg); },
+      setStatus: () => {},
+      setWorkingMessage: () => {},
+    },
+    sessionManager: {
+      getEntries: () => [],
+    },
+    modelRegistry: {
+      getProviderAuthMode: () => undefined,
+      isProviderRequestReady: () => true,
+    },
+  } as any;
+  const pi = {
+    ...makeMockPi(),
+    sendMessage: () => {
+      queueMicrotask(() => resolveAgentEnd(makeEvent([
+        {
+          role: "assistant",
+          content: [
+            {
+              type: "text",
+              text: 'bash<arg_key>command</arg_key><arg_value>ls -la /tmp && echo "---SRC---"</arg_value></tool_call>',
+            },
+          ],
+        },
+      ])));
+    },
+  } as any;
+  const s = makeLoopSession({
+    basePath,
+    canonicalProjectRoot: basePath,
+    originalBasePath: basePath,
+  });
+  const mockLedger = {
+    version: 1,
+    projectStartedAt: Date.now(),
+    units: [] as any[],
+  };
+  const deps = makeMockDeps({
+    closeoutUnit: async () => {
+      mockLedger.units.push({
+        type: "execute-task",
+        id: "M001/S01/T01",
+        startedAt: s.currentUnit?.startedAt ?? Date.now(),
+        toolCalls: 0,
+        assistantMessages: 1,
+        tokens: { input: 100, output: 20, total: 120, cacheRead: 0, cacheWrite: 0 },
+        cost: 0.01,
+      });
+    },
+    getLedger: () => mockLedger,
+  });
+  let seq = 0;
+
+  const result = await runUnitPhase(
+    { ctx, pi, s, deps, prefs: undefined, iteration: 1, flowId: "flow-zero-tool-pseudo-tool", nextSeq: () => ++seq },
+    {
+      unitType: "execute-task",
+      unitId: "M001/S01/T01",
+      prompt: "do work",
+      finalPrompt: "do work",
+      pauseAfterUatDispatch: false,
+      state: {
+        phase: "executing",
+        activeMilestone: { id: "M001", title: "Milestone" },
+        activeSlice: { id: "S01", title: "Slice" },
+        activeTask: { id: "T01", title: "Task" },
+        registry: [{ id: "M001", title: "Milestone", status: "active" }],
+        recentDecisions: [],
+        blockers: [],
+        nextAction: "",
+        progress: { milestones: { done: 0, total: 1 } },
+        requirements: { active: 0, validated: 0, deferred: 0, outOfScope: 0, blocked: 0, total: 0 },
+      } as any,
+      mid: "M001",
+      midTitle: "Milestone",
+      isRetry: false,
+      previousTier: undefined,
+    },
+    { recentUnits: [{ key: "execute-task/M001/S01/T01" }], stuckRecoveryAttempts: 0, consecutiveFinalizeTimeouts: 0 },
+  );
+
+  assert.equal(result.action, "break");
+  assert.equal((result as any).reason, "zero-tool-serialization-drift");
+  assert.equal(deps.callLog.includes("pauseAuto"), true);
+  assert.equal(s.zeroToolRetryCount.has("execute-task/M001/S01/T01"), false);
+  assert.ok(
+    notifications.some((msg) => msg.includes("serialization drift") && msg.includes("bash<arg_key>command")),
+    "serialization drift notification should include a snippet of the pseudo tool-call text",
+  );
+  assert.ok(
+    !notifications.some((msg) => msg.includes("context exhaustion")),
+    "pseudo tool-call text should not be labeled as context exhaustion",
+  );
+});
+
 test("runUnitPhase pauses auto-mode when zero-tool-call retry is exhausted", async (t) => {
   _resetPendingResolve();
 

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

@@ -1218,6 +1218,10 @@ test("hasImplementationArtifacts finds integration implementation-only commits w
 
     execFileSync("git", ["checkout", "-b", "milestone/M001"], { cwd: base, stdio: "ignore" });
     writeIntegrationBranch(base, "M001", "main");
+    // ADR-045: writeIntegrationBranch no longer creates milestones/<MID>/ as a
+    // side effect (META is now flat at .gsd/<MID>-META.json), so scaffold the
+    // legacy summary dir explicitly.
+    mkdirSync(join(base, ".gsd", "milestones", "M001"), { recursive: true });
     writeFileSync(join(base, ".gsd", "milestones", "M001", "M001-SUMMARY.md"), "# Milestone Summary\nDone.");
     execFileSync("git", ["add", "."], { cwd: base, stdio: "ignore" });
     execFileSync("git", ["commit", "-m", "chore: auto-commit after complete-milestone\n\nGSD-Unit: M001"], { cwd: base, stdio: "ignore" });
@@ -2063,6 +2067,41 @@ test("#1088: writeReactiveExecuteBlocker preserves deferred batch task statuses"
   }
 });
 
+test("#1343: writeReactiveExecuteBlocker uses slice-qualified summaries in flat-phase (no cross-slice collision)", () => {
+  const base = join(tmpdir(), `gsd-test-${randomUUID()}`);
+  try {
+    // Flat-phase layout: slices S01 and S02 share the phase dir and reuse task
+    // id T03. Only the sibling slice S01 produced a summary.
+    const phaseDir = join(base, ".gsd", "phases", "01-test");
+    mkdirSync(phaseDir, { recursive: true });
+    writeFileSync(join(phaseDir, "01-01-PLAN.md"), "# S01: First\n", "utf-8");
+    writeFileSync(join(phaseDir, "01-02-PLAN.md"), "# S02: Second\n", "utf-8");
+    writeFileSync(join(phaseDir, "S01-T03-SUMMARY.md"), "# S01 T03 Summary\n", "utf-8");
+
+    openDatabase(join(base, ".gsd", "gsd.db"));
+    insertMilestone({ id: "M001", title: "Milestone", status: "active" });
+    insertSlice({ id: "S01", milestoneId: "M001", title: "First", status: "pending" });
+    insertSlice({ id: "S02", milestoneId: "M001", title: "Second", status: "pending" });
+    insertTask({ id: "T03", milestoneId: "M001", sliceId: "S02", title: "Shared id", status: "pending" });
+
+    const recovery = writeReactiveExecuteBlocker(
+      "M001/S02/reactive+T03",
+      base,
+      "verification retries exhausted",
+    );
+
+    assert.ok(recovery, "recovery should run with DB available");
+    // S02 never wrote S02-T03-SUMMARY.md; the sibling S01-T03-SUMMARY.md must
+    // NOT make S02/T03 look complete. It has to be skipped.
+    assert.deepEqual(recovery!.completedTaskIds, []);
+    assert.deepEqual(recovery!.skippedTaskIds, ["T03"]);
+    assert.equal(getTask("M001", "S02", "T03")?.status, "skipped");
+  } finally {
+    closeDatabase();
+    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-index-lock.test.ts

@@ -4,7 +4,7 @@ import { existsSync, mkdtempSync, mkdirSync, rmSync, writeFileSync } from "node:
 import { join } from "node:path";
 import { tmpdir } from "node:os";
 
-import { hasGitIndexLockForTest } from "../auto-start.ts";
+import { hasGitIndexLockForTest, reconcileFlatPhaseBootstrapLayout } from "../auto-start.ts";
 
 test("bootstrapAutoSession detects .git/index.lock without deleting it", (t) => {
   const base = mkdtempSync(join(tmpdir(), "gsd-index-lock-"));
@@ -24,3 +24,23 @@ test("bootstrapAutoSession reports no index lock when the lock file is absent",
 
   assert.equal(hasGitIndexLockForTest(base), false);
 });
+
+test("bootstrapAutoSession removes empty legacy milestones dir in flat-phase projects", (t) => {
+  const base = mkdtempSync(join(tmpdir(), "gsd-flat-bootstrap-clean-"));
+  t.after(() => rmSync(base, { recursive: true, force: true }));
+  mkdirSync(join(base, ".gsd", "phases"), { recursive: true });
+  mkdirSync(join(base, ".gsd", "milestones"), { recursive: true });
+
+  assert.equal(reconcileFlatPhaseBootstrapLayout(base), true);
+  assert.equal(existsSync(join(base, ".gsd", "phases")), true);
+  assert.equal(existsSync(join(base, ".gsd", "milestones")), false);
+});
+
+test("bootstrapAutoSession creates phases instead of milestones for empty GSD state", (t) => {
+  const base = mkdtempSync(join(tmpdir(), "gsd-flat-bootstrap-init-"));
+  t.after(() => rmSync(base, { recursive: true, force: true }));
+
+  assert.equal(reconcileFlatPhaseBootstrapLayout(base), true);
+  assert.equal(existsSync(join(base, ".gsd", "phases")), true);
+  assert.equal(existsSync(join(base, ".gsd", "milestones")), false);
+});