gsd-pi 2.82.0 → 3.0.0-dev.8b8d129d7

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

@@ -58,6 +58,7 @@ import {
 import {
   writeLock,
   clearLock,
+  clearStaleWorkerLock,
   readCrashLock,
   isLockProcessAlive,
   formatCrashInfo,
@@ -240,12 +241,14 @@ import { runAutoLoopWithUok } from "./uok/kernel.js";
 import { resolveUokFlags } from "./uok/flags.js";
 import { validateDirectory } from "./validate-directory.js";
 import { createAutoOrchestrator } from "./auto/orchestrator.js";
-import type { AutoOrchestrationModule, AutoOrchestratorDeps } from "./auto/contracts.js";
+import type { AutoAdvanceResult, AutoOrchestrationModule, AutoOrchestratorDeps, DispatchAdapter } from "./auto/contracts.js";
 import { reconcileBeforeDispatch } from "./state-reconciliation.js";
 import { compileUnitToolContract } from "./tool-contract.js";
 import { createWorktreeSafetyModule } from "./worktree-safety.js";
 import { resolveManifest } from "./unit-context-manifest.js";
 import { classifyFailure } from "./recovery-classification.js";
+import { supportsStructuredQuestions } from "./workflow-mcp.js";
+import type { MinimalModelRegistry } from "./context-budget.js";
 // Slice-level parallelism (#2340)
 import { getEligibleSlices } from "./slice-parallel-eligibility.js";
 import { startSliceParallel } from "./slice-parallel-orchestrator.js";
@@ -281,6 +284,7 @@ import type {
   UnitRouting,
   StartModel,
   AutoSession,
+  PendingOrchestrationDispatch,
 } from "./auto/session.js";
 export {
   STUB_RECOVERY_THRESHOLD,
@@ -314,6 +318,13 @@ import { normalizeRealPath } from "./paths.js";
 /** Throttle STATE.md rebuilds — at most once per 30 seconds */
 const STATE_REBUILD_MIN_INTERVAL_MS = 30_000;
 
+export function formatAutoStopNotification(prefix: string, totals: { cost: number; tokens: { total: number } }, unitCount: number): string {
+  return [
+    `${prefix}.`,
+    `Session: ${formatCost(totals.cost)} · ${formatTokenCount(totals.tokens.total)} tokens · ${unitCount} units`,
+  ].join("\n");
+}
+
 /**
  * Phase B — register this auto-mode process in the workers table so other
  * workers and janitors can detect liveness via heartbeat. Best-effort: if
@@ -1019,6 +1030,8 @@ export async function rerootCommandSession(
 }
 
 export async function cleanupAfterLoopExit(ctx: ExtensionContext): Promise<void> {
+  const preserveStepSurface = s.preserveStepSurfaceAfterLoopExit;
+  const preservePausedSurface = s.paused;
   s.currentUnit = null;
   s.active = false;
   deactivateGSD();
@@ -1041,21 +1054,24 @@ export async function cleanupAfterLoopExit(ctx: ExtensionContext): Promise<void>
   // A transient provider-error pause intentionally leaves the paused badge
   // visible so the user still has a resumable auto-mode signal on screen.
   if (!s.paused) {
-    ctx.ui.setStatus("gsd-auto", undefined);
-    if (s.completionStopInProgress) {
-      s.completionStopInProgress = false;
+    if (preserveStepSurface) {
+      s.preserveStepSurfaceAfterLoopExit = false;
+    } else {
+      ctx.ui.setStatus("gsd-auto", undefined);
+      ctx.ui.setWidget("gsd-progress", undefined);
+      if (s.completionStopInProgress) {
+        s.completionStopInProgress = false;
+      }
+      initHealthWidget(ctx);
     }
-    initHealthWidget(ctx);
   }
 
-  // ADR-016 phase 3 (#5693): the stop-path basePath restore routes through
-  // `Lifecycle.restoreToProjectRoot()`, the sole owner of `s.basePath`
-  // mutation. The verb assigns `s.basePath` before any throwable work
-  // (rebuildGitService, cache invalidation), so a thrown error still leaves
-  // basePath restored — no fallback assignment needed at the call site.
-  // The chdir stays here because `restoreToProjectRoot` is a pure
-  // session-state mutation.
-  if (s.originalBasePath) {
+  // ADR-016 phase 3 (#5693): the stop-path basePath restore + chdir routes
+  // through `Lifecycle.restoreToProjectRoot()`, the sole owner of both
+  // `s.basePath` mutation and the paired `process.chdir` for auto-loop
+  // transitions. The verb assigns `s.basePath` before any throwable work, so
+  // a thrown error still leaves basePath restored.
+  if (s.originalBasePath && !preserveStepSurface && !preservePausedSurface) {
     try {
       buildLifecycle().restoreToProjectRoot();
     } catch (err) {
@@ -1065,14 +1081,9 @@ export async function cleanupAfterLoopExit(ctx: ExtensionContext): Promise<void>
         { file: "auto.ts" },
       );
     }
-    try {
-      process.chdir(s.originalBasePath);
-    } catch (err) {
-      logWarning("engine", `basePath restore/chdir failed: ${err instanceof Error ? err.message : String(err)}`, { file: "auto.ts" });
-    }
   }
 
-  if (s.originalBasePath && s.cmdCtx) {
+  if (s.originalBasePath && s.cmdCtx && !preserveStepSurface && !preservePausedSurface) {
     const result = await rerootCommandSession(s.cmdCtx, s.originalBasePath);
     if (result.status === "cancelled") {
       logWarning("engine", "post-loop session re-root was cancelled", { file: "auto.ts", basePath: s.originalBasePath });
@@ -1183,6 +1194,8 @@ export async function stopAuto(
         ? "merge-conflict"
         : rawReason.startsWith("Merge error") || rawReason.startsWith("Merge failed")
           ? "merge-failed"
+          : rawReason.startsWith("Slice-parallel dispatched")
+            ? "stop"
           : rawReason.startsWith("slice-merge-conflict")
             ? "slice-merge-conflict"
             : rawReason === "All milestones complete"
@@ -1384,20 +1397,14 @@ export async function stopAuto(
     }
 
     // ── Step 7: Restore basePath and chdir (ADR-016 phase 3, #5693) ──
-    // `restoreToProjectRoot` assigns s.basePath before any throwable work;
-    // no fallback assignment is needed at the call site.
+    // `restoreToProjectRoot` owns both s.basePath restore and process.chdir;
+    // no paired chdir is needed at the call site.
     if (s.originalBasePath) {
       try {
         buildLifecycle().restoreToProjectRoot();
       } catch (e) {
         debugLog("stop-cleanup-basepath", { error: e instanceof Error ? e.message : String(e) });
       }
-      try {
-        process.chdir(s.basePath);
-      } catch (err) {
-        /* best-effort */
-        logWarning("engine", `chdir failed: ${err instanceof Error ? err.message : String(err)}`, { file: "auto.ts" });
-      }
     }
 
     // Re-root the active command session/tool runtime after worktree teardown.
@@ -1426,7 +1433,7 @@ export async function stopAuto(
       if (ledger && ledger.units.length > 0) {
         const totals = getProjectTotals(ledger.units);
         ctx?.ui.notify(
-          `${notificationPrefix}. Session: ${formatCost(totals.cost)} · ${formatTokenCount(totals.tokens.total)} tokens · ${ledger.units.length} units`,
+          formatAutoStopNotification(notificationPrefix, totals, ledger.units.length),
           "info",
         );
       } else {
@@ -1733,7 +1740,6 @@ export async function pauseAuto(
   restoreProjectRootEnv();
   restoreMilestoneLockEnv();
   s.pendingVerificationRetry = null;
-  s.verificationRetryCount.clear();
   ctx?.ui.setStatus("gsd-auto", "paused");
   ctx?.ui.setWidget("gsd-progress", undefined);
   const resumeCmd = s.stepMode ? "/gsd next" : "/gsd auto";
@@ -1792,6 +1798,100 @@ function buildLifecycle(): WorktreeLifecycle {
   return new WorktreeLifecycle(s, buildWorktreeLifecycleDeps());
 }
 
+/**
+ * Build the production `DispatchAdapter` used by `createWiredAutoOrchestrationModule`.
+ *
+ * Exported so tests can verify parity with `runDispatch`'s `resolveDispatch` call —
+ * the wired adapter must derive `structuredQuestionsAvailable`, `sessionContextWindow`,
+ * `sessionProvider`, and `modelRegistry` the same way phases.ts:runDispatch does.
+ */
+export function createWiredDispatchAdapter(
+  ctx: ExtensionContext,
+  pi: ExtensionAPI,
+  dispatchBasePath: string,
+  session?: AutoSession,
+): DispatchAdapter {
+  return {
+    async decideNextUnit(input) {
+      const state = input.stateSnapshot;
+      const active = state.activeMilestone;
+      if (!active) return null;
+
+      const prefs = loadEffectiveGSDPreferences(dispatchBasePath)?.preferences;
+
+      // Derive session-derived dispatch inputs the same way phases.ts:runDispatch does
+      // (#5789). Prefer caller-supplied values when present so test harnesses and
+      // alternative wirings can inject deterministic snapshots; otherwise pull from
+      // the captured pi/ctx references.
+      const sessionProvider = input.sessionProvider ?? ctx.model?.provider;
+      const sessionContextWindow = input.sessionContextWindow ?? ctx.model?.contextWindow;
+      const modelRegistry = input.modelRegistry ?? (ctx.modelRegistry as MinimalModelRegistry | undefined);
+      const authMode =
+        sessionProvider && typeof ctx.modelRegistry?.getProviderAuthMode === "function"
+          ? ctx.modelRegistry.getProviderAuthMode(sessionProvider)
+          : undefined;
+      const activeTools = typeof pi.getActiveTools === "function" ? pi.getActiveTools() : [];
+      // Mirrors runDispatch: deep-planning keeps approval gates in plain chat
+      // because structured questions can be cancelled outside the chat turn on
+      // some transports.
+      const structuredQuestionsAvailable =
+        input.structuredQuestionsAvailable ??
+        (prefs?.planning_depth === "deep"
+          ? "false"
+          : supportsStructuredQuestions(activeTools, {
+              authMode,
+              baseUrl: ctx.model?.baseUrl,
+            })
+            ? "true"
+            : "false");
+
+      const action = await resolveDispatch({
+        basePath: dispatchBasePath,
+        mid: active.id,
+        midTitle: active.title,
+        state,
+        prefs,
+        session: input.session,
+        structuredQuestionsAvailable,
+        sessionContextWindow,
+        sessionProvider,
+        modelRegistry,
+      });
+
+      if (action.action === "stop") {
+        if (session) session.pendingOrchestrationDispatch = null;
+        return {
+          kind: "blocked",
+          reason: action.reason,
+          action: action.level === "warning" ? "pause" : "stop",
+        };
+      }
+      if (action.action !== "dispatch") {
+        if (session) session.pendingOrchestrationDispatch = null;
+        return null;
+      }
+      if (session) {
+        const pending: PendingOrchestrationDispatch = {
+          unitType: action.unitType,
+          unitId: action.unitId,
+          prompt: action.prompt,
+          pauseAfterUatDispatch: action.pauseAfterDispatch ?? false,
+          state,
+          mid: active.id,
+          midTitle: active.title,
+        };
+        session.pendingOrchestrationDispatch = pending;
+      }
+      return {
+        unitType: action.unitType,
+        unitId: action.unitId,
+        reason: action.matchedRule ?? "dispatch",
+        preconditions: [],
+      };
+    },
+  };
+}
+
 /**
  * Thin entry glue for the new Auto Orchestration module.
  *
@@ -1801,7 +1901,7 @@ function buildLifecycle(): WorktreeLifecycle {
  */
 export function createWiredAutoOrchestrationModule(
   ctx: ExtensionContext,
-  _pi: ExtensionAPI,
+  pi: ExtensionAPI,
   dispatchBasePath: string,
   runtimeBasePath = resolveProjectRoot(dispatchBasePath),
 ): AutoOrchestrationModule {
@@ -1830,30 +1930,7 @@ export function createWiredAutoOrchestrationModule(
         };
       },
     },
-    dispatch: {
-      async decideNextUnit(input) {
-        const state = input.stateSnapshot;
-        const active = state.activeMilestone;
-        if (!active) return null;
-
-        const prefs = loadEffectiveGSDPreferences(dispatchBasePath)?.preferences;
-        const action = await resolveDispatch({
-          basePath: dispatchBasePath,
-          mid: active.id,
-          midTitle: active.title,
-          state,
-          prefs,
-        });
-
-        if (action.action !== "dispatch") return null;
-        return {
-          unitType: action.unitType,
-          unitId: action.unitId,
-          reason: action.matchedRule ?? "dispatch",
-          preconditions: [],
-        };
-      },
-    },
+    dispatch: createWiredDispatchAdapter(ctx, pi, dispatchBasePath, s),
     recovery: {
       async classifyAndRecover(input) {
         const recovery = classifyFailure(input);
@@ -1902,12 +1979,25 @@ export function createWiredAutoOrchestrationModule(
       async cleanupOnStop() {},
     },
     health: {
+      checkResourcesStale() {
+        return checkResourcesStale(s.resourceVersionOnStart);
+      },
       async preAdvanceGate() {
-        const gate = await preDispatchHealthGate(dispatchBasePath);
-        return {
-          allow: gate.proceed,
-          reason: gate.reason,
-        };
+        try {
+          const gate = await preDispatchHealthGate(dispatchBasePath);
+          if (gate.proceed) {
+            return {
+              kind: "pass",
+              fixesApplied: gate.fixesApplied,
+            };
+          }
+          return {
+            kind: "fail",
+            reason: gate.reason ?? "Pre-dispatch health check failed — run /gsd doctor for details.",
+          };
+        } catch (error) {
+          return { kind: "threw", error };
+        }
       },
       async postAdvanceRecord(result) {
         if (result.kind === "error") {
@@ -1936,22 +2026,22 @@ export function createWiredAutoOrchestrationModule(
       },
       async journalTransition(event) {
         const eventType = event.name === "start"
-          ? "iteration-start"
+          ? "orchestrator-iteration-start"
           : event.name === "resume"
-            ? "iteration-start"
+            ? "orchestrator-iteration-start"
             : event.name === "advance"
-              ? "dispatch-match"
+              ? "orchestrator-dispatch-match"
               : event.name === "advance-blocked"
-                ? "guard-block"
+                ? "orchestrator-guard-block"
                 : event.name === "advance-stopped"
-                  ? "dispatch-stop"
+                  ? "orchestrator-dispatch-stop"
                   : event.name === "advance-error"
-                    ? "iteration-end"
+                    ? "orchestrator-iteration-end"
                     : event.name === "advance-paused" || event.name === "advance-retry"
-                      ? "guard-block"
+                      ? "orchestrator-guard-block"
                       : event.name === "stop"
-                      ? "terminal"
-                      : "iteration-end";
+                      ? "orchestrator-terminal"
+                      : "orchestrator-iteration-end";
 
         _emitJournalEvent(runtimeBasePath, {
           ts: new Date().toISOString(),
@@ -1975,11 +2065,60 @@ export function createWiredAutoOrchestrationModule(
         }
       },
     },
+    uokGate: {
+      async emit(input) {
+        const prefs = loadEffectiveGSDPreferences(dispatchBasePath)?.preferences;
+        const uokFlags = resolveUokFlags(prefs);
+        if (!uokFlags.gates) return;
+        const milestoneId = input.milestoneId ?? s.currentMilestoneId ?? undefined;
+        try {
+          const { UokGateRunner } = await import("./uok/gate-runner.js");
+          const runner = new UokGateRunner();
+          runner.register({
+            id: input.gateId,
+            type: input.gateType,
+            execute: async () => ({
+              outcome: input.outcome,
+              failureClass: input.failureClass,
+              rationale: input.rationale,
+              findings: input.findings ?? "",
+            }),
+          });
+          await runner.run(input.gateId, {
+            basePath: dispatchBasePath,
+            traceId: `pre-dispatch:${flowId}`,
+            turnId: `orch-${seq}`,
+            milestoneId,
+            unitType: "pre-dispatch",
+            unitId: `orch-${seq}`,
+          });
+        } catch (err) {
+          logWarning("engine", `uok gate emit failed: ${getErrorMessage(err)}`, {
+            file: "auto.ts",
+            gateId: input.gateId,
+            gateType: input.gateType,
+            ...(milestoneId ? { milestoneId } : {}),
+          });
+        }
+      },
+    },
   };
 
   return createAutoOrchestrator(deps);
 }
 
+function notifyResumeBlocked(ctx: ExtensionContext, result: Extract<AutoAdvanceResult, { kind: "blocked" }>): void {
+  const resumeCmd = s.stepMode ? "/gsd next" : "/gsd auto";
+  ctx.ui.notify(`Auto-mode blocked: ${result.reason}. Fix and run ${resumeCmd} to resume.`, "warning");
+  setLifecycleOutcome(ctx, {
+    status: "blocked",
+    title: "Auto-mode blocked",
+    detail: result.reason,
+    nextAction: `Fix the blocker, then run ${resumeCmd} to resume.`,
+    commands: ["/gsd status for overview", `${resumeCmd} to resume`, "/gsd doctor to diagnose"],
+  });
+}
+
 function ensureOrchestrationModule(ctx: ExtensionContext, pi: ExtensionAPI, basePath: string): void {
   s.orchestration = createWiredAutoOrchestrationModule(ctx, pi, basePath, lockBase());
 }
@@ -2317,7 +2456,7 @@ export async function startAuto(
     // This closes the journal gap reported in #3348 where the worker wrote side
     // effects (SUMMARY.md, DB updates) but died before emitting unit-end.
     emitCrashRecoveredUnitEnd(base, freshStartAssessment.lock);
-    clearLock(base);
+    clearStaleWorkerLock(base);
   }
 
   if (!s.paused) {
@@ -2388,6 +2527,8 @@ export async function startAuto(
     s.unitLifetimeDispatches.clear();
     if (!getLedger()) initMetrics(base);
     if (s.currentMilestoneId) setActiveMilestoneId(base, s.currentMilestoneId);
+    await openProjectDbIfPresent(base);
+    registerAutoWorkerForSession(s, base);
 
     // Re-register health level notification callback lost across process restart
     setLevelChangeCallback((_from, to, summary) => {
@@ -2495,7 +2636,12 @@ export async function startAuto(
     pi.events.emit(CMUX_CHANNELS.LOG, { preferences: loadEffectiveGSDPreferences(s.basePath || undefined)?.preferences, message: s.stepMode ? "Step-mode resumed." : "Auto-mode resumed.", level: "progress" });
 
     try {
-      await s.orchestration?.resume();
+      const resumeResult = await s.orchestration?.resume();
+      if (resumeResult?.kind === "blocked") {
+        notifyResumeBlocked(ctx, resumeResult);
+        await cleanupAfterLoopExit(ctx);
+        return;
+      }
     } catch (err) {
       debugLog("resume-orchestration-resume", { error: err instanceof Error ? err.message : String(err) });
     }
@@ -2516,6 +2662,7 @@ export async function startAuto(
   const bootstrapDeps: BootstrapDeps = {
     shouldUseWorktreeIsolation,
     registerSigtermHandler,
+    registerAutoWorkerForSession: (projectRoot) => registerAutoWorkerForSession(s, projectRoot),
     lockBase,
     buildLifecycle,
   };

package/src/resources/extensions/gsd/bootstrap/agent-end-recovery.ts

@@ -32,6 +32,7 @@ import { shouldIgnoreAgentEndForActiveUnit } from "../auto/unit-runner-events.js
 import { resolveModelId } from "../auto-model-selection.js";
 import { resolveProjectRoot } from "../worktree.js";
 import { clearDiscussionFlowState } from "./write-gate.js";
+import { clearGuidedUnitContext } from "../guided-unit-context.js";
 import { resumeAutoAfterProviderDelay } from "./provider-error-resume.js";
 import {
   classifyError,
@@ -48,6 +49,11 @@ const MAX_NETWORK_RETRIES = 2;
 function isObjectRecord(value: unknown): value is Record<string, unknown> {
   return !!value && typeof value === "object";
 }
+
+export function _hasEmptyAgentEndContent(content: unknown): boolean {
+  return content == null || (Array.isArray(content) && content.length === 0);
+}
+
 /**
  * Cap on auto-resume attempts for sustained transient-provider errors.
  *
@@ -94,6 +100,14 @@ export function isUserInitiatedAbortMessage(message: string | undefined | null):
   return /\b(?:claude code process aborted by user|request aborted by user|process aborted by user)\b/i.test(message);
 }
 
+export function shouldDeferTransientErrorToCoreRetry(
+  cls: ErrorClass,
+  rawErrorMsg: string,
+): boolean {
+  if (!isTransient(cls) || cls.kind === "rate-limit") return false;
+  return !/retry failed after \d+ attempts:/i.test(rawErrorMsg);
+}
+
 function isBareClaudeCodeSessionSwitchAbortMarker(message: string | undefined | null): boolean {
   if (!message) return false;
   const normalized = message.trim().replace(/\s+/g, " ").toLowerCase();
@@ -196,6 +210,14 @@ export function resolveAgentEndErrorDisplay(
   return rawErrorMsg;
 }
 
+export function isTerminalDeletedWorktreeProviderError(
+  message: string | undefined | null,
+): boolean {
+  if (!message) return false;
+  if (!/\bdoes not exist\b/i.test(message)) return false;
+  return /[/\\]\.gsd[/\\](?:projects[/\\][^/\\]+[/\\])?worktrees[/\\][^/\\\s"']+/i.test(message);
+}
+
 async function pauseTransientWithBackoff(
   cls: ErrorClass,
   pi: ExtensionAPI,
@@ -244,9 +266,11 @@ export async function handleAgentEnd(
   // falsely report files as missing — producing a spurious "ready signal
   // rejected" loop even though the files are on disk.
   clearPathCache();
+  const basePath = resolveAgentEndBasePath();
+  clearGuidedUnitContext(basePath);
 
   try {
-    if (await checkDeepProjectSetupAfterTurn(event, ctx, resolveAgentEndBasePath())) {
+    if (await checkDeepProjectSetupAfterTurn(event, ctx, basePath)) {
       return;
     }
   } catch (err) {
@@ -254,8 +278,8 @@ export async function handleAgentEnd(
     logWarning("bootstrap", `checkDeepProjectSetupAfterTurn failed: ${message}`);
   }
 
-  if (checkAutoStartAfterDiscuss()) {
-    clearDiscussionFlowState(resolveAgentEndBasePath() ?? process.cwd());
+  if (checkAutoStartAfterDiscuss(basePath)) {
+    clearDiscussionFlowState(basePath ?? process.cwd());
     return;
   }
 
@@ -263,14 +287,14 @@ export async function handleAgentEnd(
   // are missing, `checkAutoStartAfterDiscuss` returns false silently. Surface
   // that and nudge the LLM to complete the writes before the user hits the
   // downstream "All milestones complete" warning loop.
-  if (maybeHandleReadyPhraseWithoutFiles(event)) return;
+  if (maybeHandleReadyPhraseWithoutFiles(event, basePath)) return;
 
   // #4573 — Empty-turn recovery: if the LLM announced intent in prose but
   // emitted no tool calls, nudge it to execute. Fires only when auto-mode is
   // active or a discussion autostart is pending (non-auto interactive discuss
   // is user-driven). Runs before `isAutoActive` early return so pending
   // discussions (where isAutoActive may be false) still get recovered.
-  if (maybeHandleEmptyIntentTurn(event, isAutoActive())) return;
+  if (maybeHandleEmptyIntentTurn(event, isAutoActive(), basePath)) return;
 
   if (!isAutoActive()) return;
 
@@ -310,7 +334,7 @@ export async function handleAgentEnd(
     // that carry error context — e.g. errorMessage field or non-empty content
     // indicating a mid-stream failure. (#2695)
     const content = "content" in lastMsg ? lastMsg.content : undefined;
-    const hasEmptyContent = Array.isArray(content) && content.length === 0;
+    const hasEmptyContent = _hasEmptyAgentEndContent(content);
     const hasErrorMessage = "errorMessage" in lastMsg && !!lastMsg.errorMessage;
 
     if (hasEmptyContent && !hasErrorMessage) {
@@ -355,6 +379,17 @@ export async function handleAgentEnd(
       rawErrorMsg,
       "content" in lastMsg ? lastMsg.content : undefined,
     );
+    if (
+      isAutoCompletionStopInProgress() &&
+      isTerminalDeletedWorktreeProviderError(`${rawErrorMsg}\n${displayMsg}`)
+    ) {
+      resetRetryState(retryState);
+      logWarning(
+        "bootstrap",
+        `Ignoring stale deleted-worktree provider error during terminal completion reroot: ${displayMsg || rawErrorMsg}`,
+      );
+      return;
+    }
     const errorDetail = displayMsg ? `: ${displayMsg}` : "";
     const explicitRetryAfterMs = ("retryAfterMs" in lastMsg && typeof lastMsg.retryAfterMs === "number") ? lastMsg.retryAfterMs : undefined;
 
@@ -461,7 +496,7 @@ export async function handleAgentEnd(
     // Core retries transient failures in-session after this handler.
     // Keep that behavior for non-rate-limit classes to avoid pause/retry races,
     // but let rate-limit continue into model fallback logic below (#4373).
-    if (isTransient(cls) && cls.kind !== "rate-limit") {
+    if (shouldDeferTransientErrorToCoreRetry(cls, rawErrorMsg)) {
       return;
     }
 

package/src/resources/extensions/gsd/bootstrap/db-tools.ts

@@ -487,17 +487,18 @@ export function registerDbTools(pi: ExtensionAPI): void {
     promptGuidelines: [
       "Use gsd_plan_milestone for milestone planning instead of writing ROADMAP.md directly.",
       "Keep parameters flat and provide the full milestone planning payload, including slices.",
+      "Milestone and slice titles must not contain forward slash (/), en dash, or em dash characters.",
       "The tool validates input, writes milestone and slice planning data transactionally, renders ROADMAP.md from DB, and clears both state and parse caches after success.",
       "Use the canonical name gsd_plan_milestone; gsd_milestone_plan is only an alias.",
     ],
     parameters: Type.Object({
       // ── Core identification + content (required) ──────────────────────
       milestoneId: Type.String({ description: "Milestone ID (e.g. M001)" }),
-      title: Type.String({ description: "Milestone title" }),
+      title: Type.String({ description: "Milestone title; must not contain forward slash (/), en dash, or em dash characters" }),
       vision: Type.String({ description: "Milestone vision" }),
       slices: Type.Array(Type.Object({
         sliceId: Type.String({ description: "Slice ID (e.g. S01)" }),
-        title: Type.String({ description: "Slice title" }),
+        title: Type.String({ description: "Slice title; must not contain forward slash (/), en dash, or em dash characters" }),
         risk: Type.String({ description: "Slice risk" }),
         depends: Type.Array(Type.String(), { description: "Slice dependency IDs" }),
         demo: Type.String({ description: "Roadmap demo text / After this" }),
@@ -570,10 +571,10 @@ export function registerDbTools(pi: ExtensionAPI): void {
         title: Type.String({ description: "Task title" }),
         description: Type.String({ description: "Task description / steps block" }),
         estimate: Type.String({ description: "Task estimate string" }),
-        files: Type.Array(Type.String(), { description: "Files likely touched" }),
+        files: Type.Array(Type.String(), { description: "Array<string> of files likely touched; pass [\"path\"] or [], never a single string" }),
         verify: Type.String({ description: "Verification command or block" }),
-        inputs: Type.Array(Type.String(), { description: "Input files or references" }),
-        expectedOutput: Type.Array(Type.String(), { description: "Expected output files or artifacts" }),
+        inputs: Type.Array(Type.String(), { description: "Array<string> of input files or references; pass [\"path\"] or [], never a single string" }),
+        expectedOutput: Type.Array(Type.String(), { description: "Array<string> of expected output files or artifacts; pass [\"path\"] or [], never a single string" }),
         observabilityImpact: Type.Optional(Type.String({ description: "Task observability impact" })),
       }), { description: "Planned tasks for the slice" }),
       // ── Enrichment metadata (optional — defaults to empty) ────────────
@@ -650,10 +651,10 @@ export function registerDbTools(pi: ExtensionAPI): void {
       title: Type.String({ description: "Task title" }),
       description: Type.String({ description: "Task description / steps block" }),
       estimate: Type.String({ description: "Task estimate string" }),
-      files: Type.Array(Type.String(), { description: "Files likely touched" }),
+      files: Type.Array(Type.String(), { description: "Array<string> of files likely touched; pass [\"path\"] or [], never a single string" }),
       verify: Type.String({ description: "Verification command or block" }),
-      inputs: Type.Array(Type.String(), { description: "Input files or references" }),
-      expectedOutput: Type.Array(Type.String(), { description: "Expected output files or artifacts" }),
+      inputs: Type.Array(Type.String(), { description: "Array<string> of input files or references; pass [\"path\"] or [], never a single string" }),
+      expectedOutput: Type.Array(Type.String(), { description: "Array<string> of expected output files or artifacts; pass [\"path\"] or [], never a single string" }),
       observabilityImpact: Type.Optional(Type.String({ description: "Task observability impact" })),
       // Single-writer v3 audit trail (Stream 2): caller-provided actor identity + causation.
       actorName: Type.Optional(Type.String({ description: "Caller-provided actor identity for the audit trail (e.g. 'executor-01', 'gsd-orchestrator')" })),
@@ -912,7 +913,7 @@ export function registerDbTools(pi: ExtensionAPI): void {
     label: "Skip Slice",
     description:
       "Mark a slice as skipped so auto-mode advances past it without executing. " +
-      "Non-closed tasks within the slice are cascaded to skipped so milestone completion is not blocked by leftover pending tasks (#4375). " +
+      "Non-closed tasks within the slice are cascaded to skipped so milestone completion is not blocked by leftover pending tasks. " +
       "The slice data is preserved for reference. The state machine treats skipped slices like completed ones for dependency satisfaction.",
     promptSnippet: "Skip a GSD slice (mark as skipped, auto-mode will advance past it)",
     promptGuidelines: [