gsd-pi 2.77.0-dev.c9c8932c6 → 2.77.0-dev.eaa4973bc

package/src/resources/extensions/claude-code-cli/stream-adapter.ts

@@ -723,7 +723,7 @@ export async function resolveClaudePermissionMode(
 		);
 		return "bypassPermissions";
 	}
-	return "acceptEdits";
+	return "bypassPermissions";
 }
 
 // NOTE: These helpers intentionally mirror @gsd/pi-ai anthropic-shared
@@ -783,21 +783,25 @@ export function buildSdkOptions(
 ): Record<string, unknown> {
 	const { reasoning, ...sdkExtraOptions } = extraOptions;
 	const mcpServers = buildWorkflowMcpServers();
-	const permissionMode = overrides?.permissionMode ?? "acceptEdits";
-	const disallowedTools = ["AskUserQuestion"];
-	// Pre-authorize the safe built-ins and every registered workflow MCP
-	// server's tools. `acceptEdits` mode (the interactive default) only
-	// auto-approves file edits — Read/Glob/Grep, basic shell inspection, and
-	// every `mcp__gsd-workflow__*` call still surface as "This command
-	// requires approval" and block GSD actions (#4099).
+	const permissionMode = overrides?.permissionMode ?? "bypassPermissions";
+	// Globally unblock all tools. Users reported that the `acceptEdits` default
+	// plus a narrow allowlist silently declined most Bash/Agent/WebFetch calls
+	// when `extensionUIContext` wasn't threaded through. Default to
+	// bypassPermissions and an empty disallow list so every tool — including
+	// `AskUserQuestion` and every `mcp__*` workflow tool — is auto-approved.
+	// Opt back into gated mode with GSD_CLAUDE_CODE_PERMISSION_MODE=acceptEdits.
+	const disallowedTools: string[] = [];
 	const allowedTools = [
 		"Read",
 		"Write",
 		"Edit",
 		"Glob",
 		"Grep",
-		"Bash(ls:*)",
-		"Bash(pwd)",
+		"Bash",
+		"Agent",
+		"WebFetch",
+		"WebSearch",
+		"AskUserQuestion",
 		...(mcpServers ? Object.keys(mcpServers).map((serverName) => `mcp__${serverName}__*`) : []),
 	];
 	const supportsAdaptive = modelSupportsAdaptiveThinking(modelId);

package/src/resources/extensions/claude-code-cli/tests/stream-adapter.test.ts

@@ -761,15 +761,18 @@ describe("stream-adapter — session persistence (#2859)", () => {
 			assert.equal(srv.env.GSD_CLI_PATH, "/tmp/gsd");
 			assert.equal(srv.env.GSD_PERSIST_WRITE_GATE_STATE, "1");
 			assert.equal(srv.env.GSD_WORKFLOW_PROJECT_ROOT, "/tmp/project");
-			assert.deepEqual(options.disallowedTools, ["AskUserQuestion"]);
+			assert.deepEqual(options.disallowedTools, []);
 			assert.deepEqual(options.allowedTools, [
 				"Read",
 				"Write",
 				"Edit",
 				"Glob",
 				"Grep",
-				"Bash(ls:*)",
-				"Bash(pwd)",
+				"Bash",
+				"Agent",
+				"WebFetch",
+				"WebSearch",
+				"AskUserQuestion",
 				"mcp__gsd-workflow__*",
 			]);
 		} finally {
@@ -781,7 +784,7 @@ describe("stream-adapter — session persistence (#2859)", () => {
 		}
 	});
 
-	test("buildSdkOptions disables AskUserQuestion for custom workflow MCP server names", () => {
+	test("buildSdkOptions auto-approves every tool for custom workflow MCP server names", () => {
 		const prev = {
 			GSD_WORKFLOW_MCP_COMMAND: process.env.GSD_WORKFLOW_MCP_COMMAND,
 			GSD_WORKFLOW_MCP_NAME: process.env.GSD_WORKFLOW_MCP_NAME,
@@ -799,15 +802,18 @@ describe("stream-adapter — session persistence (#2859)", () => {
 			const options = buildSdkOptions("claude-sonnet-4-20250514", "test");
 			const mcpServers = options.mcpServers as Record<string, any>;
 			assert.ok(mcpServers?.["custom-workflow"], "expected custom workflow server config");
-			assert.deepEqual(options.disallowedTools, ["AskUserQuestion"]);
+			assert.deepEqual(options.disallowedTools, []);
 			assert.deepEqual(options.allowedTools, [
 				"Read",
 				"Write",
 				"Edit",
 				"Glob",
 				"Grep",
-				"Bash(ls:*)",
-				"Bash(pwd)",
+				"Bash",
+				"Agent",
+				"WebFetch",
+				"WebSearch",
+				"AskUserQuestion",
 				"mcp__custom-workflow__*",
 			]);
 		} finally {
@@ -845,9 +851,9 @@ describe("stream-adapter — session persistence (#2859)", () => {
 			const mcpServers = (options as any).mcpServers;
 			if (mcpServers) {
 				assert.ok(mcpServers["gsd-workflow"], "if present, must be gsd-workflow");
-				assert.deepEqual((options as any).disallowedTools, ["AskUserQuestion"]);
+				assert.deepEqual((options as any).disallowedTools, []);
 			} else {
-				assert.deepEqual((options as any).disallowedTools, ["AskUserQuestion"]);
+				assert.deepEqual((options as any).disallowedTools, []);
 			}
 			rmSync(emptyDir, { recursive: true, force: true });
 		} finally {
@@ -894,7 +900,7 @@ describe("stream-adapter — session persistence (#2859)", () => {
 			assert.equal(srv.env.GSD_CLI_PATH, "/tmp/gsd");
 			assert.equal(srv.env.GSD_PERSIST_WRITE_GATE_STATE, "1");
 			assert.equal(srv.env.GSD_WORKFLOW_PROJECT_ROOT, resolvedRepoDir);
-			assert.deepEqual(options.disallowedTools, ["AskUserQuestion"]);
+			assert.deepEqual(options.disallowedTools, []);
 		} finally {
 			process.chdir(originalCwd);
 			rmSync(repoDir, { recursive: true, force: true });
@@ -1282,14 +1288,14 @@ describe("stream-adapter — permission mode (F10)", () => {
 		}
 	}
 
-	test("buildSdkOptions defaults to acceptEdits (#4383)", () => {
+	test("buildSdkOptions defaults to bypassPermissions (globally unblocks all tools)", () => {
 		clearWorkflowMcpEnv();
 		const opts = buildSdkOptions("claude-sonnet-4-6", "test");
-		assert.equal(opts.permissionMode, "acceptEdits");
+		assert.equal(opts.permissionMode, "bypassPermissions");
 		assert.equal(
 			opts.allowDangerouslySkipPermissions,
-			false,
-			"allowDangerouslySkipPermissions must be false when permissionMode is acceptEdits",
+			true,
+			"allowDangerouslySkipPermissions must be true when permissionMode is bypassPermissions",
 		);
 	});
 
@@ -1304,9 +1310,9 @@ describe("stream-adapter — permission mode (F10)", () => {
 		);
 	});
 
-	test("resolveClaudePermissionMode defaults to acceptEdits when no env var is set (#4383)", async () => {
+	test("resolveClaudePermissionMode defaults to bypassPermissions when no env var is set (globally unblocks all tools)", async () => {
 		const mode = await resolveClaudePermissionMode({});
-		assert.equal(mode, "acceptEdits");
+		assert.equal(mode, "bypassPermissions");
 	});
 
 	test("resolveClaudePermissionMode honours the GSD_CLAUDE_CODE_PERMISSION_MODE env override", async () => {

package/src/resources/extensions/gsd/abandon-detect.ts

@@ -0,0 +1,62 @@
+/**
+ * Abandon-milestone detection for rewrite-docs overrides (#3490).
+ *
+ * Isolated from auto-post-unit.ts so behavioral tests can import this module
+ * without pulling in the full post-unit handler graph (which transitively
+ * loads model-router, workflow engine, etc.).
+ */
+
+import type { Override } from "./files.js";
+
+// Detect when a rewrite-docs override is about abandoning THE CURRENT
+// MILESTONE — not just any override containing an abandon verb. Naively
+// matching `/\b(abandon|cancel|drop|...)\b/` against override text produces
+// false positives on scope-change prose ("cancel the standup reminder",
+// "drop the dependency on X", "scrap the v1 design for the landing page").
+//
+// To qualify as an abandon-milestone signal, an override must contain both:
+//   1. An abandon-family verb (abandon|descope|cancel|shelve|drop|scrap)
+//   2. A milestone reference — either the literal word "milestone" or the
+//      current milestone ID — in the same override text.
+
+// Verb variants cover both US and UK inflections:
+//   cancel / canceled / canceling / cancelled / cancelling / cancels
+//   travel-style "l"-doubling also applies to shelve/drop/scrap.
+// "descope" also accepts "de-scope" and "de scope" (hyphen / space forms).
+const ABANDON_VERB_RE = /\b(abandon(?:ed|ing|s)?|de[-\s]?scope(?:d|s|ing)?|cancel(?:led|ling|ed|ing|s)?|shelve(?:d|s)?|shelving|drop(?:ped|ping|s)?|scrap(?:ped|ping|s)?)\b/i;
+
+export interface AbandonDecision {
+  shouldPark: boolean;
+  reason: string;
+  matched: string[];
+}
+
+/**
+ * Decide whether a set of active overrides indicates the current milestone
+ * should be parked. Pure function — no I/O, no imports beyond types.
+ */
+export function detectAbandonMilestone(
+  overrides: Override[],
+  currentMilestoneId: string | null | undefined,
+): AbandonDecision {
+  if (!currentMilestoneId) {
+    return { shouldPark: false, reason: "", matched: [] };
+  }
+
+  const escapedId = currentMilestoneId.replace(/[.*+?^${}()|[\]\\]/g, "\\$&");
+  const milestoneRefRe = new RegExp(`\\b(?:milestone|${escapedId})\\b`, "i");
+
+  const matched = overrides
+    .filter(o => ABANDON_VERB_RE.test(o.change) && milestoneRefRe.test(o.change))
+    .map(o => o.change);
+
+  if (matched.length === 0) {
+    return { shouldPark: false, reason: "", matched: [] };
+  }
+
+  return {
+    shouldPark: true,
+    reason: matched.join("; "),
+    matched,
+  };
+}

package/src/resources/extensions/gsd/auto/resolve.ts

@@ -11,6 +11,7 @@
 import type { UnitResult, AgentEndEvent, ErrorContext } from "./types.js";
 import type { AutoSession } from "./session.js";
 import { debugLog } from "../debug-logger.js";
+import { bumpTurnGeneration } from "./turn-epoch.js";
 
 // ─── Per-unit one-shot promise state ────────────────────────────────────────
 //
@@ -68,6 +69,24 @@ export function isSessionSwitchInFlight(): boolean {
   return _sessionSwitchInFlight;
 }
 
+// ─── bumpAndResolveSynthetic ────────────────────────────────────────────────
+
+/**
+ * Bump the turn epoch and synthetically resolve the pending unit promise —
+ * the exact sequence timeout recovery must perform when it advances past a
+ * timed-out unit. Using this helper enforces the invariant "bump iff we are
+ * actually superseding the turn" so a future caller cannot resolve without
+ * bumping (orphaned writes leak) or bump without resolving (next turn starts
+ * already stale).
+ *
+ * NOT to be used for steering retries that keep the same turn alive — those
+ * do not supersede the turn and must not bump.
+ */
+export function bumpAndResolveSynthetic(reason: string): void {
+  bumpTurnGeneration(reason);
+  resolveAgentEnd({ messages: [], _synthetic: reason } as unknown as AgentEndEvent);
+}
+
 // ─── resolveAgentEndCancelled ─────────────────────────────────────────────────
 
 /**
@@ -79,6 +98,12 @@ export function isSessionSwitchInFlight(): boolean {
  */
 export function resolveAgentEndCancelled(errorContext?: ErrorContext): void {
   if (_currentResolve) {
+    // Cancellation supersedes the in-flight turn the same way timeout
+    // recovery does — bump the turn epoch so any lingering writes from the
+    // cancelled turn drop themselves.
+    bumpTurnGeneration(
+      `cancelled:${errorContext?.category ?? "unknown"}`,
+    );
     debugLog("resolveAgentEndCancelled", { status: "resolving-cancelled" });
     const r = _currentResolve;
     _currentResolve = null;
@@ -97,6 +122,10 @@ export function _resetPendingResolve(): void {
   _sessionSwitchInFlight = false;
 }
 
+export function _hasPendingResolveForTest(): boolean {
+  return _currentResolve !== null;
+}
+
 /**
  * No-op for backward compatibility with tests that previously set the
  * active session. The module no longer holds a session reference.

package/src/resources/extensions/gsd/auto/run-unit.ts

@@ -9,7 +9,11 @@ import type { ExtensionAPI, ExtensionContext } from "@gsd/pi-coding-agent";
 import type { AutoSession } from "./session.js";
 import { NEW_SESSION_TIMEOUT_MS } from "./session.js";
 import type { UnitResult } from "./types.js";
-import { _setCurrentResolve, _setSessionSwitchInFlight } from "./resolve.js";
+import { _clearCurrentResolve, _setCurrentResolve, _setSessionSwitchInFlight } from "./resolve.js";
+import {
+  getCurrentTurnGeneration,
+  runWithTurnGeneration,
+} from "./turn-epoch.js";
 import { debugLog } from "../debug-logger.js";
 import { logWarning, logError } from "../workflow-logger.js";
 import { resolveAutoSupervisorConfig } from "../preferences.js";
@@ -144,6 +148,7 @@ export async function runUnit(
       }
 
       if (!ready) {
+        _clearCurrentResolve();
         return {
           status: "cancelled",
           errorContext: {
@@ -156,6 +161,13 @@ export async function runUnit(
     }
   }
 
+  // ── Capture turn generation for stale-write detection ──
+  // Any write site reached via the sendMessage → tool-call → await chain
+  // below sees this generation via AsyncLocalStorage. If a timeout recovery
+  // or cancellation bumps the generation while this turn is in flight, those
+  // writes see themselves as stale and self-drop.
+  const capturedTurnGen = getCurrentTurnGeneration();
+
   // ── Send the prompt ──
   debugLog("runUnit", { phase: "send-message", unitType, unitId });
 
@@ -179,7 +191,9 @@ export async function runUnit(
       resolve({ status: "cancelled", errorContext: { message: "Unit hard timeout — supervision may have failed", category: "timeout", isTransient: true } });
     }, UNIT_HARD_TIMEOUT_MS);
   });
-  const result = await Promise.race([unitPromise, timeoutResult]);
+  const result = await runWithTurnGeneration(capturedTurnGen, () =>
+    Promise.race([unitPromise, timeoutResult]),
+  );
   if (unitTimeoutHandle) clearTimeout(unitTimeoutHandle);
   debugLog("runUnit", {
     phase: "agent-end-received",

package/src/resources/extensions/gsd/auto/turn-epoch.ts

@@ -0,0 +1,108 @@
+/**
+ * auto/turn-epoch.ts — Turn generation counter + AsyncLocalStorage-backed
+ * capture for stale-turn write dropping.
+ *
+ * Problem: when auto-timeout-recovery synthetically resolves a timed-out
+ * unit so the loop can advance, the original LLM turn keeps running in the
+ * background. Its subsequent writes (journal events, audit events, tool
+ * calls that flow through closeout) then race the replacement unit's
+ * writes. DB-level guards (complete-task/complete-slice) block double
+ * state transitions, but journal/audit/closeout side-effects still fire
+ * with fresh identifiers and pollute forensics.
+ *
+ * Containment: every time we decide a turn is done (timeout recovery,
+ * explicit cancellation), bump a module-level generation counter.
+ * Turn-aware call sites wrap their body in `runWithTurnGeneration`, which
+ * captures the generation into AsyncLocalStorage. Write sites deep in the
+ * stack call `isStaleWrite` — if the captured generation is older than
+ * current, the turn has been superseded and the write is dropped.
+ *
+ * Failure mode: if AsyncLocalStorage context is lost across some exotic
+ * async boundary (e.g. a native-side worker callback), the write site sees
+ * `no-store` and falls through to current behavior — the write proceeds
+ * normally. That is a safe default; the correctness regression is only
+ * "noisier forensics under rare boundary loss," not duplicated state.
+ */
+
+import { AsyncLocalStorage } from "node:async_hooks";
+
+import { debugLog } from "../debug-logger.js";
+
+let _currentGeneration = 0;
+
+const turnContext = new AsyncLocalStorage<{ capturedGen: number }>();
+
+/** Current turn generation. Mutated only by bumpTurnGeneration. */
+export function getCurrentTurnGeneration(): number {
+  return _currentGeneration;
+}
+
+/**
+ * Bump the turn generation and return the new value. Every caller should
+ * pass a short `reason` string so forensics can reconstruct why a given
+ * turn was marked stale.
+ */
+export function bumpTurnGeneration(reason: string): number {
+  _currentGeneration += 1;
+  debugLog("turnEpoch.bump", { reason, newGeneration: _currentGeneration });
+  return _currentGeneration;
+}
+
+/**
+ * Run fn() with `capturedGen` attached to AsyncLocalStorage so that any
+ * write site reached from within fn() can check for staleness without
+ * parameter threading.
+ */
+export function runWithTurnGeneration<T>(capturedGen: number, fn: () => T): T {
+  return turnContext.run({ capturedGen }, fn);
+}
+
+/**
+ * True when the current async context was started at a turn generation
+ * older than the current one — meaning the turn has been superseded by
+ * recovery/cancellation since it began.
+ *
+ * Returns false when there is no captured generation (e.g. the write is
+ * happening outside any wrapped turn). That is the safe default: writes
+ * proceed as they did before this epoch was introduced.
+ */
+export function isStaleWrite(component?: string): boolean {
+  const store = turnContext.getStore();
+  if (!store) return false;
+  const captured = store.capturedGen;
+  const current = _currentGeneration;
+  if (captured < current) {
+    debugLog("turnEpoch.stale", {
+      component: component ?? "unknown",
+      captured,
+      current,
+    });
+    return true;
+  }
+  return false;
+}
+
+/**
+ * Snapshot of both the captured turn generation and the current one.
+ * Used by closeoutUnit to persist an orphan-marker entry instead of
+ * silently skipping the full closeout on a stale turn.
+ */
+export function describeTurnEpoch(): {
+  captured: number | null;
+  current: number;
+  stale: boolean;
+} {
+  const store = turnContext.getStore();
+  const captured = store?.capturedGen ?? null;
+  const current = _currentGeneration;
+  return {
+    captured,
+    current,
+    stale: captured !== null && captured < current,
+  };
+}
+
+/** Test helper — resets module state so tests start from a known baseline. */
+export function _resetTurnEpoch(): void {
+  _currentGeneration = 0;
+}

package/src/resources/extensions/gsd/auto-dispatch.ts

@@ -421,6 +421,26 @@ export const DISPATCH_RULES: DispatchRule[] = [
       };
     },
   },
+  {
+    name: "planning (require_slice_discussion) → pause for discussion (#3454)",
+    match: async ({ state, mid, basePath, prefs }) => {
+      if (state.phase !== "planning") return null;
+      if (!prefs?.phases?.require_slice_discussion) return null;
+      if (!state.activeSlice) return null;
+      // Only pause if the slice has no context file yet (discussion not done).
+      // resolveSliceFile returns null when the file does not exist on disk,
+      // but cachedReaddir could return a stale hit — verify with existsSync
+      // so the guard is defence-in-depth and the contract is explicit at the
+      // call site.
+      const sliceContextFile = resolveSliceFile(basePath, mid, state.activeSlice.id, "CONTEXT");
+      if (sliceContextFile && existsSync(sliceContextFile)) return null; // discussion already done, proceed
+      return {
+        action: "stop" as const,
+        reason: `Slice ${state.activeSlice.id} requires discussion before planning (require_slice_discussion is enabled). Run /gsd discuss to discuss this slice, then /gsd auto to resume.`,
+        level: "info" as const,
+      };
+    },
+  },
   {
     // Keep this rule before the single-slice research rule so the multi-slice
     // path wins whenever 2+ slices are ready.

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

@@ -9,7 +9,7 @@
 
 export { autoLoop, runUokKernelLoop, runLegacyAutoLoop } from "./auto/loop.js";
 export { isInfrastructureError, INFRA_ERROR_CODES } from "./auto/infra-errors.js";
-export { resolveAgentEnd, resolveAgentEndCancelled, isSessionSwitchInFlight, _resetPendingResolve, _setActiveSession } from "./auto/resolve.js";
+export { resolveAgentEnd, resolveAgentEndCancelled, isSessionSwitchInFlight, _hasPendingResolveForTest, _resetPendingResolve, _setActiveSession } from "./auto/resolve.js";
 export { detectStuck } from "./auto/detect-stuck.js";
 export { runUnit } from "./auto/run-unit.js";
 export type { LoopDeps } from "./auto/loop-deps.js";

package/src/resources/extensions/gsd/auto-post-unit.ts

@@ -71,6 +71,7 @@ import { resolveUokFlags } from "./uok/flags.js";
 import { UokGateRunner } from "./uok/gate-runner.js";
 import { writeTurnGitTransaction } from "./uok/gitops.js";
 import { isClosedStatus } from "./status-guards.js";
+import { detectAbandonMilestone } from "./abandon-detect.js";
 
 /** Maximum verification retry attempts before escalating to blocker placeholder (#2653). */
 const MAX_VERIFICATION_RETRIES = 3;
@@ -569,6 +570,35 @@ export async function postUnitPreVerification(pctx: PostUnitContext, opts?: PreV
     // Rewrite-docs completion
     if (s.currentUnit.type === "rewrite-docs") {
       await runSafely("postUnit", "rewrite-docs-resolve", async () => {
+        // Detect abandon/descope overrides BEFORE resolving them (#3490).
+        // If an override is about abandoning the milestone, park it so the
+        // state engine skips it. Without this, rewrite-docs only edits
+        // markdown but the DB still has the milestone as active.
+        try {
+          const { loadActiveOverrides } = await import("./files.js");
+          const overrides = await loadActiveOverrides(s.basePath);
+          const decision = detectAbandonMilestone(overrides, s.currentMilestoneId);
+          if (decision.shouldPark && s.currentMilestoneId) {
+            const { parkMilestone } = await import("./milestone-actions.js");
+            const parked = parkMilestone(s.basePath, s.currentMilestoneId, decision.reason);
+            if (parked) {
+              ctx.ui.notify(`Milestone ${s.currentMilestoneId} parked: "${decision.reason}"`, "info");
+            } else {
+              // Park refused: milestone directory missing, milestone already
+              // completed (SUMMARY present), or PARKED.md already exists.
+              // resolveAllOverrides below will still consume the override —
+              // surface this loudly so the user notices state drift rather
+              // than silently losing the abandon directive.
+              const msg = `Abandon detected for ${s.currentMilestoneId} but park refused (milestone is completed, already parked, or missing). Override will be resolved anyway — verify state is correct.`;
+              logError("engine", msg);
+              ctx.ui.notify(msg, "warning");
+            }
+          }
+        } catch (err) {
+          logError("engine", `abandon-detect failed: ${(err as Error).message}`);
+          ctx.ui.notify(`Abandon detection failed — check logs. Overrides will still be resolved.`, "warning");
+        }
+
         await resolveAllOverrides(s.basePath);
         // Reset both disk and in-memory counters. Disk counter is authoritative
         // (survives restarts); in-memory is kept in sync for the current session.

package/src/resources/extensions/gsd/auto-timeout-recovery.ts

@@ -18,7 +18,7 @@ import {
 } from "./auto-recovery.js";
 import { existsSync } from "node:fs";
 
-import { resolveAgentEnd } from "./auto-loop.js";
+import { bumpAndResolveSynthetic } from "./auto/resolve.js";
 
 export interface RecoveryContext {
   basePath: string;
@@ -35,6 +35,13 @@ export async function recoverTimedOutUnit(
   reason: "idle" | "hard",
   rctx: RecoveryContext,
 ): Promise<"recovered" | "paused"> {
+  // Note on turn epoch: the bump is intentionally NOT unconditional at
+  // function entry. Two branches below (the "steering retry" paths) keep
+  // the same LLM turn alive and let it try again — they must NOT bump,
+  // otherwise the retry's legitimate writes get marked stale and drop.
+  // Each advance branch calls `bumpAndResolveSynthetic` to bump+resolve
+  // atomically. Search for that helper to find all supersede sites.
+
   const { basePath, verbose, currentUnitStartedAt, unitRecoveryCount } = rctx;
 
   const runtime = readUnitRuntimeRecord(basePath, unitType, unitId);
@@ -74,7 +81,7 @@ export async function recoverTimedOutUnit(
         "info",
       );
       unitRecoveryCount.delete(recoveryKey);
-      resolveAgentEnd({ messages: [], _synthetic: "timeout-recovery" } as any);
+      bumpAndResolveSynthetic(`timeout-recovery:${reason}:${unitType}/${unitId}`);
       return "recovered";
     }
 
@@ -145,7 +152,7 @@ export async function recoverTimedOutUnit(
         "warning",
       );
       unitRecoveryCount.delete(recoveryKey);
-      resolveAgentEnd({ messages: [], _synthetic: "timeout-recovery" } as any);
+      bumpAndResolveSynthetic(`timeout-recovery:${reason}:${unitType}/${unitId}`);
       return "recovered";
     }
 
@@ -179,7 +186,7 @@ export async function recoverTimedOutUnit(
       "info",
     );
     unitRecoveryCount.delete(recoveryKey);
-    resolveAgentEnd({ messages: [], _synthetic: "timeout-recovery" } as any);
+    bumpAndResolveSynthetic(`timeout-recovery:${reason}:${unitType}/${unitId}`);
     return "recovered";
   }
 
@@ -265,7 +272,7 @@ export async function recoverTimedOutUnit(
       "warning",
     );
     unitRecoveryCount.delete(recoveryKey);
-    resolveAgentEnd({ messages: [], _synthetic: "timeout-recovery" } as any);
+    bumpAndResolveSynthetic(`timeout-recovery:${reason}:${unitType}/${unitId}`);
     return "recovered";
   }
 

package/src/resources/extensions/gsd/auto-unit-closeout.ts

@@ -45,9 +45,20 @@ export async function closeoutUnit(
       const { buildMemoryLLMCall, extractMemoriesFromUnit } = await import('./memory-extractor.js');
       const llmCallFn = buildMemoryLLMCall(ctx);
       if (llmCallFn) {
-        extractMemoriesFromUnit(activityFile, unitType, unitId, llmCallFn).catch((err) => {
-          logWarning("engine", `memory extraction failed for ${unitType}/${unitId}: ${(err as Error).message}`);
-        });
+        // Awaited: a fire-and-forget here lets memory-extractor writes land in
+        // .gsd/ after closeoutUnit returns but before the milestone merge
+        // runs, which made the working tree appear dirty to `git merge
+        // --squash` (root cause class of #4704). Completion latency is now
+        // bounded by the extractor's LLM call, which is the acceptable price
+        // for not racing the merge boundary.
+        try {
+          await extractMemoriesFromUnit(activityFile, unitType, unitId, llmCallFn);
+        } catch (err) {
+          logWarning(
+            "engine",
+            `memory extraction failed for ${unitType}/${unitId}: ${(err as Error).message}`,
+          );
+        }
       }
     } catch (err) { /* non-fatal */
       logWarning("engine", `operation failed: ${err instanceof Error ? err.message : String(err)}`);