gsd-pi 2.82.0-dev.ed17d078d → 3.0.0-dev.1b44e695b

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

@@ -58,6 +58,7 @@ import {
 import {
   writeLock,
   clearLock,
+  clearStaleWorkerLock,
   readCrashLock,
   isLockProcessAlive,
   formatCrashInfo,
@@ -178,6 +179,7 @@ import { debugLog, isDebugEnabled, writeDebugSummary } from "./debug-logger.js";
 import {
   buildLoopRemediationSteps,
   reconcileMergeState,
+  verifyExpectedArtifact,
 } from "./auto-recovery.js";
 import { classifyMilestoneSummaryContent } from "./milestone-summary-classifier.js";
 import { resolveDispatch, DISPATCH_RULES } from "./auto-dispatch.js";
@@ -240,7 +242,7 @@ 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, DispatchAdapter } 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";
@@ -283,6 +285,7 @@ import type {
   UnitRouting,
   StartModel,
   AutoSession,
+  PendingOrchestrationDispatch,
 } from "./auto/session.js";
 export {
   STUB_RECOVERY_THRESHOLD,
@@ -296,7 +299,10 @@ export type {
 import { autoSession as s } from "./auto-runtime-state.js";
 import { gsdHome } from "./gsd-home.js";
 import { createWorkspace, scopeMilestone } from "./workspace.js";
-import { registerAutoWorker, markWorkerStopping } from "./db/auto-workers.js";
+import {
+  registerAutoWorker,
+  markWorkerStopping,
+} from "./db/auto-workers.js";
 import { releaseMilestoneLease } from "./db/milestone-leases.js";
 import { normalizeRealPath } from "./paths.js";
 
@@ -316,6 +322,13 @@ 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
@@ -461,6 +474,60 @@ export function _synthesizePausedSessionRecoveryForTest(
   return synthesizePausedSessionRecovery(basePath, unitType, unitId, sessionFile);
 }
 
+type PausedResumeRecoverySessionState = {
+  pausedSessionFile: string | null;
+  currentUnit: { type: string; id: string } | null;
+  pausedUnitType: string | null;
+  pausedUnitId: string | null;
+  pendingCrashRecovery: string | null;
+};
+
+function handlePausedSessionResumeRecovery(
+  basePath: string,
+  state: PausedResumeRecoverySessionState,
+  notify: (message: string) => void,
+): { skippedReplay: boolean } {
+  if (!state.pausedSessionFile) return { skippedReplay: false };
+
+  const pausedRecoveryUnitType = state.currentUnit?.type ?? state.pausedUnitType ?? "unknown";
+  const pausedRecoveryUnitId = state.currentUnit?.id ?? state.pausedUnitId ?? "unknown";
+  const completedPausedUnit = verifyExpectedArtifact(
+    pausedRecoveryUnitType,
+    pausedRecoveryUnitId,
+    basePath,
+  );
+
+  if (completedPausedUnit) {
+    state.pausedSessionFile = null;
+    state.pausedUnitType = null;
+    state.pausedUnitId = null;
+    state.pendingCrashRecovery = null;
+    return { skippedReplay: true };
+  }
+
+  const recovery = synthesizePausedSessionRecovery(
+    basePath,
+    pausedRecoveryUnitType,
+    pausedRecoveryUnitId,
+    state.pausedSessionFile,
+  );
+  if (recovery && recovery.trace.toolCallCount > 0) {
+    state.pendingCrashRecovery = recovery.prompt;
+    notify(`Recovered ${recovery.trace.toolCallCount} tool calls from paused session. Resuming with context.`);
+  }
+  state.pausedSessionFile = null;
+  state.pausedUnitType = null;
+  state.pausedUnitId = null;
+  return { skippedReplay: false };
+}
+
+export function _handlePausedSessionResumeRecoveryForTest(
+  basePath: string,
+  state: PausedResumeRecoverySessionState,
+): { skippedReplay: boolean } {
+  return handlePausedSessionResumeRecovery(basePath, state, () => {});
+}
+
 // `_resolvePausedResumeBasePathForTest` was retired in ADR-016 phase 2 / B3
 // (#5621). Production callers go through
 // `WorktreeLifecycle.resumeFromPausedSession`; the pure helper for tests is
@@ -860,6 +927,7 @@ export function checkRemoteAutoSession(projectRoot: string): {
 
   if (!isLockProcessAlive(lock)) {
     // Stale lock from a dead process — not a live remote session
+    clearLock(projectRoot);
     return { running: false };
   }
 
@@ -1021,6 +1089,8 @@ 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();
@@ -1043,21 +1113,24 @@ 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) {
@@ -1067,14 +1140,9 @@ 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 });
@@ -1171,32 +1239,16 @@ export async function stopAuto(
   // worktree was active, and whether the current milestone was merged before
   // stopAuto. The unmerged-work warning is only meaningful for real worktrees.
   try {
-    const { emitAutoExit } = await import("./worktree-telemetry.js");
-    type AutoExitReason =
-      | "pause" | "stop" | "blocked" | "merge-conflict" | "merge-failed"
-      | "slice-merge-conflict" | "all-complete" | "no-active-milestone" | "other";
+    const { emitAutoExit, normalizeAutoExitReason } = await import("./worktree-telemetry.js");
     // Normalize the free-form reason to a closed set so the telemetry
     // aggregator buckets stably. Raw detail is preserved in the phases.ts
     // notification and the notify'd error string.
     const rawReason = reason ?? "stop";
-    const normalizedReason: AutoExitReason = rawReason.startsWith("Blocked:")
-      ? "blocked"
-      : rawReason.startsWith("Merge conflict")
-        ? "merge-conflict"
-        : rawReason.startsWith("Merge error") || rawReason.startsWith("Merge failed")
-          ? "merge-failed"
-          : rawReason.startsWith("slice-merge-conflict")
-            ? "slice-merge-conflict"
-            : rawReason === "All milestones complete"
-              ? "all-complete"
-              : rawReason === "No active milestone"
-                ? "no-active-milestone"
-                : rawReason === "stop" || rawReason === "pause"
-                  ? rawReason
-                  : "other";
+    const normalizedReason = normalizeAutoExitReason(rawReason);
     const telemetryBase = s.originalBasePath || s.basePath;
     emitAutoExit(telemetryBase, {
       reason: normalizedReason,
+      rawReason,
       milestoneId: s.currentMilestoneId ?? undefined,
       milestoneMerged: s.milestoneMergedInPhases === true,
       isolationMode: getIsolationMode(telemetryBase),
@@ -1386,20 +1438,14 @@ 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.
@@ -1428,7 +1474,7 @@ 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 {
@@ -1735,7 +1781,6 @@ 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";
@@ -1805,6 +1850,7 @@ export function createWiredDispatchAdapter(
   ctx: ExtensionContext,
   pi: ExtensionAPI,
   dispatchBasePath: string,
+  session?: AutoSession,
 ): DispatchAdapter {
   return {
     async decideNextUnit(input) {
@@ -1812,7 +1858,9 @@ export function createWiredDispatchAdapter(
       const active = state.activeMilestone;
       if (!active) return null;
 
-      const prefs = loadEffectiveGSDPreferences(dispatchBasePath)?.preferences;
+      const activeSession = input.session ?? session;
+      const activeDispatchBasePath = activeSession?.basePath || dispatchBasePath;
+      const prefs = loadEffectiveGSDPreferences(activeDispatchBasePath)?.preferences;
 
       // Derive session-derived dispatch inputs the same way phases.ts:runDispatch does
       // (#5789). Prefer caller-supplied values when present so test harnesses and
@@ -1841,18 +1889,42 @@ export function createWiredDispatchAdapter(
             : "false");
 
       const action = await resolveDispatch({
-        basePath: dispatchBasePath,
+        basePath: activeDispatchBasePath,
         mid: active.id,
         midTitle: active.title,
         state,
         prefs,
+        session: activeSession,
         structuredQuestionsAvailable,
         sessionContextWindow,
         sessionProvider,
         modelRegistry,
       });
 
-      if (action.action !== "dispatch") return null;
+      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,
@@ -1901,7 +1973,7 @@ export function createWiredAutoOrchestrationModule(
         };
       },
     },
-    dispatch: createWiredDispatchAdapter(ctx, pi, dispatchBasePath),
+    dispatch: createWiredDispatchAdapter(ctx, pi, dispatchBasePath, s),
     recovery: {
       async classifyAndRecover(input) {
         const recovery = classifyFailure(input);
@@ -1924,10 +1996,16 @@ export function createWiredAutoOrchestrationModule(
             reason: `No Unit manifest is registered for ${unitType}`,
           };
         }
+        if (getIsolationMode(runtimeBasePath) !== "worktree") {
+          return { ok: true, reason: "not-required" };
+        }
         const writeScope =
           manifest.tools.mode === "all" || manifest.tools.mode === "docs"
             ? "source-writing"
             : "planning-only";
+        if (getIsolationMode(runtimeBasePath) !== "worktree") {
+          return { ok: true, reason: "isolation-not-worktree" };
+        }
         const safety = createWorktreeSafetyModule();
         const snapshot = await deriveState(dispatchBasePath);
         const milestoneId = snapshot.activeMilestone?.id ?? null;
@@ -1939,6 +2017,7 @@ export function createWiredAutoOrchestrationModule(
           projectRoot: runtimeBasePath,
           unitRoot: dispatchBasePath,
           milestoneId,
+          isolationMode: getIsolationMode(runtimeBasePath),
           expectedBranch,
         });
         if (!result.ok) {
@@ -1997,22 +2076,22 @@ 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(),
@@ -2078,6 +2157,18 @@ export function createWiredAutoOrchestrationModule(
   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());
 }
@@ -2415,7 +2506,7 @@ 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) {
@@ -2486,6 +2577,8 @@ 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) => {
@@ -2558,22 +2651,11 @@ export async function startAuto(
     }
     invalidateAllCaches();
 
-    if (s.pausedSessionFile) {
-      const recovery = synthesizePausedSessionRecovery(
-        s.basePath,
-        s.currentUnit?.type ?? s.pausedUnitType ?? "unknown",
-        s.currentUnit?.id ?? s.pausedUnitId ?? "unknown",
-        s.pausedSessionFile,
-      );
-      if (recovery && recovery.trace.toolCallCount > 0) {
-        s.pendingCrashRecovery = recovery.prompt;
-        ctx.ui.notify(
-          `Recovered ${recovery.trace.toolCallCount} tool calls from paused session. Resuming with context.`,
-          "info",
-        );
-      }
-      s.pausedSessionFile = null;
-    }
+    handlePausedSessionResumeRecovery(
+      s.basePath,
+      s,
+      (message) => ctx.ui.notify(message, "info"),
+    );
 
     captureProjectRootEnv(s.originalBasePath || s.basePath);
     registerAutoWorkerForSession(s);
@@ -2588,12 +2670,20 @@ export async function startAuto(
         "resuming",
         s.currentMilestoneId ?? "unknown",
       );
-      clearPausedSession("paused-session DB cleanup failed (resume activation)");
     }
+    clearPausedSession("paused-session DB cleanup failed (resume activation)");
     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" && resumeResult.action === "stop") {
+        notifyResumeBlocked(ctx, resumeResult);
+        await cleanupAfterLoopExit(ctx);
+        return;
+      }
+      if (resumeResult?.kind === "blocked") {
+        notifyResumeBlocked(ctx, resumeResult);
+      }
     } catch (err) {
       debugLog("resume-orchestration-resume", { error: err instanceof Error ? err.message : String(err) });
     }
@@ -2614,6 +2704,7 @@ 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 +49,19 @@ 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 {
+  if (content == null) return true;
+  if (!Array.isArray(content)) return false;
+  if (content.length === 0) return true;
+  return content.every((block) => {
+    if (!block || typeof block !== "object") return true;
+    const typedBlock = block as { type?: unknown; text?: unknown };
+    if (typedBlock.type !== "text") return false;
+    return typeof typedBlock.text !== "string" || typedBlock.text.trim() === "";
+  });
+}
+
 /**
  * Cap on auto-resume attempts for sustained transient-provider errors.
  *
@@ -94,6 +108,14 @@ 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 +218,14 @@ 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 +274,11 @@ 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 +286,8 @@ 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 +295,14 @@ 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;
 
@@ -292,9 +324,22 @@ export async function handleAgentEnd(
   }
 
   if (isBareClaudeCodeStreamAbortPlaceholder(lastMsg)) {
-    // The Claude Code adapter can emit this placeholder after a prior turn has
-    // already completed and the next unit is active. It has no user/provider
-    // diagnostic value and must not cancel the newly-dispatched unit.
+    if (isSessionSwitchAbortGraceActive()) {
+      // Old turn leaking through after a session switch — drop it.
+      return;
+    }
+
+    // Mid-unit stream abort with no diagnostic. Treat as non-fatal so the loop
+    // can continue, but surface it to the user and resolve the in-flight unit.
+    ctx.ui.notify("Claude Code stream aborted mid-unit (no diagnostic). Continuing.", "warning");
+    try {
+      resetRetryState(retryState);
+      resolveAgentEnd(event);
+    } catch (err) {
+      const message = err instanceof Error ? err.message : String(err);
+      ctx.ui.notify(`Auto-mode error after stream-abort placeholder: ${message}. Stopping auto-mode.`, "error");
+      try { await pauseAuto(ctx, pi); } catch (e) { logWarning("bootstrap", `pauseAuto failed after stream-abort placeholder: ${(e as Error).message}`); }
+    }
     return;
   }
 
@@ -310,7 +355,7 @@ 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 +400,17 @@ 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 +517,7 @@ 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;
     }