gsd-pi 2.76.0 → 2.77.0-dev.2daa994b6

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

@@ -32,11 +32,12 @@ import { isInfrastructureError, isTransientCooldownError, getCooldownRetryAfterM
 import { resolveEngine } from "../engine-resolver.js";
 import { logWarning } from "../workflow-logger.js";
 import { gsdRoot } from "../paths.js";
+import { atomicWriteSync } from "../atomic-write.js";
 import { resolveUokFlags } from "../uok/flags.js";
 import { scheduleSidecarQueue } from "../uok/execution-graph.js";
 import { ExecutionGraphScheduler } from "../uok/execution-graph.js";
 import type { UokGraphNode } from "../uok/contracts.js";
-import { readFileSync, writeFileSync, mkdirSync } from "node:fs";
+import { readFileSync, writeFileSync, mkdirSync, unlinkSync } from "node:fs";
 import { join } from "node:path";
 
 // ── Stuck detection persistence (#3704) ──────────────────────────────────
@@ -50,6 +51,13 @@ function stuckStatePath(basePath: string): string {
 function loadStuckState(basePath: string): { recentUnits: Array<{ key: string }>; stuckRecoveryAttempts: number } {
   try {
     const data = JSON.parse(readFileSync(stuckStatePath(basePath), "utf-8"));
+    // Only load state written by a DIFFERENT process (real session restart).
+    // If the PID matches the current process, this state was written by an earlier
+    // autoLoop call in the same process (e.g., a test that completed before this
+    // one), not by a crashed session — skip it to prevent test state pollution.
+    if (data.pid === process.pid) {
+      return { recentUnits: [], stuckRecoveryAttempts: 0 };
+    }
     return {
       recentUnits: Array.isArray(data.recentUnits) ? data.recentUnits : [],
       stuckRecoveryAttempts: typeof data.stuckRecoveryAttempts === "number" ? data.stuckRecoveryAttempts : 0,
@@ -65,6 +73,7 @@ function saveStuckState(basePath: string, state: LoopState): void {
     const filePath = stuckStatePath(basePath);
     mkdirSync(join(gsdRoot(basePath), "runtime"), { recursive: true });
     writeFileSync(filePath, JSON.stringify({
+      pid: process.pid,
       recentUnits: state.recentUnits.slice(-20), // keep last 20 entries
       stuckRecoveryAttempts: state.stuckRecoveryAttempts,
       updatedAt: new Date().toISOString(),
@@ -74,12 +83,69 @@ function saveStuckState(basePath: string, state: LoopState): void {
   }
 }
 
+// ── Custom workflow verification retry persistence ───────────────────────
+// Custom workflows can request verification retries after a step runs. The
+// retry budget must survive an auto-mode restart or a failing verifier can
+// consume a fresh retry budget every session.
+function customVerifyRetryStateDir(s: Pick<AutoSession, "activeRunDir" | "basePath">): string {
+  return s.activeRunDir ? join(s.activeRunDir, "runtime") : join(gsdRoot(s.basePath), "runtime");
+}
+
+function customVerifyRetryStatePath(s: Pick<AutoSession, "activeRunDir" | "basePath">): string {
+  return join(customVerifyRetryStateDir(s), "custom-verify-retries.json");
+}
+
+function hydrateCustomVerifyRetryCounts(s: AutoSession): Map<string, number> {
+  if (s.verificationRetryCount.size > 0) {
+    return s.verificationRetryCount;
+  }
+
+  try {
+    const raw = JSON.parse(readFileSync(customVerifyRetryStatePath(s), "utf-8"));
+    const counts = raw && typeof raw === "object" && raw.counts && typeof raw.counts === "object"
+      ? raw.counts as Record<string, unknown>
+      : {};
+    for (const [key, value] of Object.entries(counts)) {
+      if (typeof value === "number" && Number.isFinite(value) && value > 0) {
+        s.verificationRetryCount.set(key, Math.floor(value));
+      }
+    }
+  } catch (err) {
+    debugLog("autoLoop", { phase: "load-custom-verify-retries-failed", error: err instanceof Error ? err.message : String(err) });
+  }
+
+  return s.verificationRetryCount;
+}
+
+function saveCustomVerifyRetryCounts(s: AutoSession): void {
+  const retryCounts = s.verificationRetryCount;
+  const filePath = customVerifyRetryStatePath(s);
+
+  try {
+    if (!retryCounts || retryCounts.size === 0) {
+      unlinkSync(filePath);
+      return;
+    }
+    mkdirSync(customVerifyRetryStateDir(s), { recursive: true });
+    atomicWriteSync(filePath, JSON.stringify({
+      counts: Object.fromEntries(retryCounts),
+      updatedAt: new Date().toISOString(),
+    }) + "\n");
+  } catch (err) {
+    const code = err && typeof err === "object" && "code" in err ? (err as { code?: string }).code : undefined;
+    if (code !== "ENOENT") {
+      debugLog("autoLoop", { phase: "save-custom-verify-retries-failed", error: err instanceof Error ? err.message : String(err) });
+    }
+  }
+}
+
 // ── Memory pressure monitoring (#3331) ──────────────────────────────────
 // Check heap usage every N iterations and trigger graceful shutdown before
 // the OS OOM killer sends SIGKILL. The threshold is 90% of the V8 heap
 // limit (--max-old-space-size or default ~1.5-4GB depending on platform).
 const MEMORY_CHECK_INTERVAL = 5; // check every 5 iterations
 const MEMORY_PRESSURE_THRESHOLD = 0.85; // 85% of heap limit
+const MAX_CUSTOM_ENGINE_VERIFY_RETRIES = 3;
 
 type DispatchContract = "legacy-direct" | "uok-scheduler";
 
@@ -429,16 +495,51 @@ export async function autoLoop(
           break;
         }
         if (verifyResult === "retry") {
-          debugLog("autoLoop", { phase: "custom-engine-verify-retry", iteration, unitId: iterData.unitId });
+          const recoveryKey = `${iterData.unitType}/${iterData.unitId}`;
+          const retryCounts = hydrateCustomVerifyRetryCounts(s);
+          const attempts = (retryCounts.get(recoveryKey) ?? 0) + 1;
+          retryCounts.set(recoveryKey, attempts);
+          saveCustomVerifyRetryCounts(s);
+          debugLog("autoLoop", { phase: "custom-engine-verify-retry", iteration, unitId: iterData.unitId, attempts });
           deps.uokObserver?.onPhaseResult("custom-engine", "retry", {
             unitType: iterData.unitType,
             unitId: iterData.unitId,
+            attempts,
           });
+          if (attempts > MAX_CUSTOM_ENGINE_VERIFY_RETRIES) {
+            const recovery = await policy.recover(iterData.unitType, iterData.unitId, { basePath: s.basePath });
+            if (recovery.outcome === "pause") {
+              await deps.pauseAuto(ctx, pi);
+              finishTurn("paused", "manual-attention", recovery.reason ?? "custom-engine-verify-retry-exhausted");
+              break;
+            }
+            if (recovery.outcome === "skip") {
+              await deps.stopAuto(
+                ctx,
+                pi,
+                recovery.reason ??
+                  `Custom workflow verification for ${iterData.unitId} requested skip after retry exhaustion, but the custom engine cannot reconcile skipped steps.`,
+              );
+              finishTurn("stopped", "manual-attention", "custom-engine-verify-retry-exhausted");
+              break;
+            }
+            const exhaustedReason =
+              `Custom workflow verification for ${iterData.unitId} requested retry ${attempts} times without passing.`;
+            await deps.stopAuto(
+              ctx,
+              pi,
+              recovery.outcome === "stop" && recovery.reason ? recovery.reason : exhaustedReason,
+            );
+            finishTurn("stopped", "manual-attention", "custom-engine-verify-retry-exhausted");
+            break;
+          }
           finishTurn("retry");
           continue;
         }
 
         // Verification passed — mark step complete
+        s.verificationRetryCount?.delete(`${iterData.unitType}/${iterData.unitId}`);
+        saveCustomVerifyRetryCounts(s);
         debugLog("autoLoop", { phase: "custom-engine-reconcile", iteration, unitId: iterData.unitId });
         const reconcileResult = await engine.reconcile(engineState, {
           unitType: iterData.unitType,
@@ -590,6 +691,7 @@ export async function autoLoop(
       consecutiveCooldowns = 0;
       recentErrorMessages.length = 0;
       deps.emitJournalEvent({ ts: new Date().toISOString(), flowId, seq: nextSeq(), eventType: "iteration-end", data: { iteration } });
+      saveStuckState(s.basePath, loopState); // persist across session restarts (#4382)
       debugLog("autoLoop", { phase: "iteration-complete", iteration });
       finishTurn("completed");
     } catch (loopErr) {

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

@@ -26,12 +26,12 @@ import {
 import { detectStuck } from "./detect-stuck.js";
 import { runUnit } from "./run-unit.js";
 import { debugLog } from "../debug-logger.js";
-import { PROJECT_FILES } from "../detection.js";
+import { PROJECT_FILES, hasProjectFileInAncestor } from "../detection.js";
 import { MergeConflictError } from "../git-service.js";
 import { setCurrentPhase, clearCurrentPhase } from "../../shared/gsd-phase-state.js";
 import { pauseAutoForProviderError } from "../provider-error-pause.js";
 import { resumeAutoAfterProviderDelay } from "../bootstrap/provider-error-resume.js";
-import { join, basename, dirname, parse as parsePath } from "node:path";
+import { join, basename } from "node:path";
 import { existsSync, cpSync, readdirSync } from "node:fs";
 import {
   logWarning,
@@ -51,10 +51,11 @@ import { getEligibleSlices } from "../slice-parallel-eligibility.js";
 import { startSliceParallel } from "../slice-parallel-orchestrator.js";
 import { isDbAvailable, getMilestoneSlices } from "../gsd-db.js";
 import type { MinimalModelRegistry } from "../context-budget.js";
-import { ensurePlanV2Graph } from "../uok/plan-v2.js";
+import { ensurePlanV2Graph, isMissingFinalizedContextResult } from "../uok/plan-v2.js";
 import { resolveUokFlags } from "../uok/flags.js";
 import { UokGateRunner } from "../uok/gate-runner.js";
-import { resetEvidence } from "../safety/evidence-collector.js";
+import { resetEvidence, loadEvidenceFromDisk } from "../safety/evidence-collector.js";
+import { parseUnitId } from "../unit-id.js";
 import { createCheckpoint, cleanupCheckpoint, rollbackToCheckpoint } from "../safety/git-checkpoint.js";
 import { resolveSafetyHarnessConfig } from "../safety/safety-harness.js";
 import {
@@ -80,7 +81,11 @@ export function resetSessionTimeoutState(): void {
  * Exported for testing as _resolveReportBasePath.
  */
 export function _resolveReportBasePath(s: Pick<AutoSession, "originalBasePath" | "basePath">): string {
-  return s.originalBasePath || s.basePath;
+  // Strip /.gsd/worktrees/ suffix when basePath is itself a worktree path and
+  // originalBasePath is falsy — prevents reports landing in the worktree (#3729).
+  const resolved = s.originalBasePath || s.basePath;
+  const markerIdx = resolved.indexOf("/.gsd/worktrees/");
+  return markerIdx !== -1 ? resolved.slice(0, markerIdx) : resolved;
 }
 
 /**
@@ -91,7 +96,12 @@ export function _resolveReportBasePath(s: Pick<AutoSession, "originalBasePath" |
 export function _resolveDispatchGuardBasePath(
   s: Pick<AutoSession, "originalBasePath" | "basePath">,
 ): string {
-  return s.originalBasePath || s.basePath;
+  // Strip /.gsd/worktrees/ suffix when basePath is itself a worktree path and
+  // originalBasePath is falsy — prevents guard checks running against the
+  // worktree instead of the project root (#3729).
+  const resolved = s.originalBasePath || s.basePath;
+  const markerIdx = resolved.indexOf("/.gsd/worktrees/");
+  return markerIdx !== -1 ? resolved.slice(0, markerIdx) : resolved;
 }
 
 const PLAN_V2_GATE_PHASES: ReadonlySet<Phase> = new Set([
@@ -410,27 +420,41 @@ export async function runPreDispatch(
     const compiled = ensurePlanV2Graph(s.basePath, state);
     if (!compiled.ok) {
       const reason = compiled.reason ?? "Plan v2 compilation failed";
+      if (isMissingFinalizedContextResult(compiled)) {
+        await runPreDispatchGate({
+          gateId: "plan-v2-gate",
+          gateType: "policy",
+          outcome: "pass",
+          failureClass: "none",
+          rationale: "plan v2 missing context recovery deferred to dispatch",
+          findings: reason,
+          milestoneId: state.activeMilestone?.id ?? undefined,
+        });
+      } else {
+        await runPreDispatchGate({
+          gateId: "plan-v2-gate",
+          gateType: "policy",
+          outcome: "manual-attention",
+          failureClass: "manual-attention",
+          rationale: "plan v2 compile gate failed",
+          findings: reason,
+          milestoneId: state.activeMilestone?.id ?? undefined,
+        });
+        ctx.ui.notify(`Plan gate failed-closed: ${reason}\n\nIf this keeps happening, try: /gsd doctor heal`, "error");
+        await deps.pauseAuto(ctx, pi);
+        return { action: "break", reason: "plan-v2-gate-failed" };
+      }
+    }
+    if (compiled.ok) {
       await runPreDispatchGate({
         gateId: "plan-v2-gate",
         gateType: "policy",
-        outcome: "manual-attention",
-        failureClass: "manual-attention",
-        rationale: "plan v2 compile gate failed",
-        findings: reason,
+        outcome: "pass",
+        failureClass: "none",
+        rationale: "plan v2 compile gate passed",
         milestoneId: state.activeMilestone?.id ?? undefined,
       });
-      ctx.ui.notify(`Plan gate failed-closed: ${reason}`, "error");
-      await deps.pauseAuto(ctx, pi);
-      return { action: "break", reason: "plan-v2-gate-failed" };
     }
-    await runPreDispatchGate({
-      gateId: "plan-v2-gate",
-      gateType: "policy",
-      outcome: "pass",
-      failureClass: "none",
-      rationale: "plan v2 compile gate passed",
-      milestoneId: state.activeMilestone?.id ?? undefined,
-    });
   }
   deps.syncCmuxSidebar(prefs, state);
   let mid = state.activeMilestone?.id;
@@ -545,6 +569,12 @@ export async function runPreDispatch(
     loopState.stuckRecoveryAttempts = 0;
 
     // Worktree lifecycle on milestone transition — merge current, enter next
+    // #2909: preflight — warn + stash dirty working tree before merge
+    const preflightTransition = deps.preflightCleanRoot(
+      s.originalBasePath || s.basePath,
+      s.currentMilestoneId!,
+      ctx.ui.notify.bind(ctx.ui),
+    );
     try {
       deps.resolver.mergeAndExit(s.currentMilestoneId!, ctx.ui);
     } catch (mergeErr) {
@@ -566,6 +596,14 @@ export async function runPreDispatch(
       await deps.stopAuto(ctx, pi, `Merge error on milestone ${s.currentMilestoneId}: ${String(mergeErr)}`);
       return { action: "break", reason: "merge-failed" };
     }
+    // #2909: postflight — restore stashed changes after successful merge
+    if (preflightTransition.stashPushed) {
+      deps.postflightPopStash(
+        s.originalBasePath || s.basePath,
+        s.currentMilestoneId!,
+        ctx.ui.notify.bind(ctx.ui),
+      );
+    }
 
     // PR creation (auto_pr) is handled inside mergeMilestoneToMain (#2302)
 
@@ -644,6 +682,12 @@ export async function runPreDispatch(
     if (incomplete.length === 0 && state.registry.length > 0) {
       // All milestones complete — merge milestone branch before stopping
       if (s.currentMilestoneId) {
+        // #2909: preflight — warn + stash dirty working tree before merge
+        const preflightAllComplete = deps.preflightCleanRoot(
+          s.originalBasePath || s.basePath,
+          s.currentMilestoneId,
+          ctx.ui.notify.bind(ctx.ui),
+        );
         try {
           deps.resolver.mergeAndExit(s.currentMilestoneId, ctx.ui);
           // Prevent stopAuto from attempting the same merge (#2645)
@@ -665,6 +709,14 @@ export async function runPreDispatch(
           await deps.stopAuto(ctx, pi, `Merge error on milestone ${s.currentMilestoneId}: ${String(mergeErr)}`);
           return { action: "break", reason: "merge-failed" };
         }
+        // #2909: postflight — restore stashed changes after successful merge
+        if (preflightAllComplete.stashPushed) {
+          deps.postflightPopStash(
+            s.originalBasePath || s.basePath,
+            s.currentMilestoneId,
+            ctx.ui.notify.bind(ctx.ui),
+          );
+        }
 
         // PR creation (auto_pr) is handled inside mergeMilestoneToMain (#2302)
       }
@@ -758,6 +810,12 @@ export async function runPreDispatch(
   if (state.phase === "complete") {
     // Milestone merge on complete (before closeout so branch state is clean)
     if (s.currentMilestoneId) {
+      // #2909: preflight — warn + stash dirty working tree before merge
+      const preflightComplete = deps.preflightCleanRoot(
+        s.originalBasePath || s.basePath,
+        s.currentMilestoneId,
+        ctx.ui.notify.bind(ctx.ui),
+      );
       try {
         deps.resolver.mergeAndExit(s.currentMilestoneId, ctx.ui);
         // Prevent stopAuto from attempting the same merge (#2645)
@@ -779,6 +837,14 @@ export async function runPreDispatch(
         await deps.stopAuto(ctx, pi, `Merge error on milestone ${s.currentMilestoneId}: ${String(mergeErr)}`);
         return { action: "break", reason: "merge-failed" };
       }
+      // #2909: postflight — restore stashed changes after successful merge
+      if (preflightComplete.stashPushed) {
+        deps.postflightPopStash(
+          s.originalBasePath || s.basePath,
+          s.currentMilestoneId,
+          ctx.ui.notify.bind(ctx.ui),
+        );
+      }
 
       // PR creation (auto_pr) is handled inside mergeMilestoneToMain (#2302)
     }
@@ -1131,16 +1197,42 @@ export async function runGuards(
       s.lastBudgetAlertLevel =
         newBudgetAlertLevel as AutoSession["lastBudgetAlertLevel"];
 
-      if (threshold.pct === 100 && budgetEnforcementAction !== "none") {
+      // Emit Layer 2 budget_threshold event (post-plan hook recommendation).
+      // Extensions / Layer 0 shell hooks may return an action override.
+      let hookAction: "pause" | "downgrade" | "continue" | undefined;
+      try {
+        const { emitBudgetThreshold } = await import("../hook-emitter.js");
+        const hookResult = await emitBudgetThreshold({
+          fraction: budgetPct,
+          spent: totalCost,
+          limit: budgetCeiling,
+        });
+        if (hookResult?.action) hookAction = hookResult.action;
+      } catch (hookErr) {
+        logWarning("engine", `budget_threshold hook emission failed: ${(hookErr as Error).message}`);
+      }
+
+      // Apply hook override to enforcement action. "continue" → "none" (no enforcement),
+      // "pause" and "downgrade" map to the matching enforcement path below.
+      let effectiveAction = budgetEnforcementAction;
+      if (hookAction === "continue") {
+        effectiveAction = "none";
+      } else if (hookAction === "pause") {
+        effectiveAction = "pause";
+      } else if (hookAction === "downgrade") {
+        effectiveAction = "warn";
+      }
+
+      if (threshold.pct === 100 && effectiveAction !== "none") {
         // 100% — special enforcement logic (halt/pause/warn)
         const msg = `Budget ceiling ${deps.formatCost(budgetCeiling)} reached (spent ${deps.formatCost(totalCost)}).`;
-        if (budgetEnforcementAction === "halt") {
+        if (effectiveAction === "halt") {
           deps.sendDesktopNotification("GSD", msg, "error", "budget", basename(s.originalBasePath || s.basePath));
           await deps.stopAuto(ctx, pi, "Budget ceiling reached");
           debugLog("autoLoop", { phase: "exit", reason: "budget-halt" });
           return { action: "break", reason: "budget-halt" };
         }
-        if (budgetEnforcementAction === "pause") {
+        if (effectiveAction === "pause") {
           ctx.ui.notify(
             `${msg} Pausing auto-mode — /gsd auto to override and continue.`,
             "warning",
@@ -1286,21 +1378,10 @@ export async function runUnitPhase(
     // Monorepo support (#2347): if no project files in the worktree directory,
     // walk parent directories up to the filesystem root. In monorepos,
     // package.json / Cargo.toml etc. live in a parent directory.
-    let hasProjectFileInParent = false;
-    if (!hasProjectFile && !hasSrcDir && !hasXcodeBundle) {
-      let checkDir = dirname(s.basePath);
-      const { root } = parsePath(checkDir);
-      while (checkDir !== root) {
-        // Stop at git repository boundary — ancestors above the repo root
-        // (e.g. ~ or /usr/local) may contain unrelated project files.
-        if (deps.existsSync(join(checkDir, ".git"))) break;
-        if (PROJECT_FILES.some((f) => deps.existsSync(join(checkDir, f)))) {
-          hasProjectFileInParent = true;
-          break;
-        }
-        checkDir = dirname(checkDir);
-      }
-    }
+    const hasProjectFileInParent =
+      !hasProjectFile && !hasSrcDir && !hasXcodeBundle
+        ? hasProjectFileInAncestor(s.basePath, deps.existsSync)
+        : false;
     if (!hasProjectFile && !hasSrcDir && !hasXcodeBundle && !hasProjectFileInParent) {
       // Greenfield projects won't have project files yet — the first task creates them.
       // Log a warning but allow execution to proceed. The .git check above is sufficient
@@ -1359,6 +1440,14 @@ export async function runUnitPhase(
   );
   if (safetyConfig.enabled && safetyConfig.evidence_collection) {
     resetEvidence();
+    // Restore persisted evidence so session-restart resumes don't produce
+    // false-positive "no bash calls" warnings (Bug #4385).
+    if (s.basePath && unitType === "execute-task") {
+      const { milestone: eMid, slice: eSid, task: eTid } = parseUnitId(unitId);
+      if (eMid && eSid && eTid) {
+        loadEvidenceFromDisk(s.basePath, eMid, eSid, eTid);
+      }
+    }
   }
   // Only checkpoint code-executing units (not lifecycle/planning units)
   if (safetyConfig.enabled && safetyConfig.checkpoints && unitType === "execute-task") {
@@ -1445,6 +1534,7 @@ export async function runUnitPhase(
     sidecarItem ? undefined : { isRetry, previousTier },
     undefined,
     s.manualSessionModelOverride,
+    s.autoModeStartThinkingLevel,
   );
   s.currentUnitRouting =
     modelResult.routing as AutoSession["currentUnitRouting"];
@@ -1459,6 +1549,9 @@ export async function runUnitPhase(
     if (match) {
       const ok = await pi.setModel(match, { persist: false });
       if (ok) {
+        if (s.autoModeStartThinkingLevel) {
+          pi.setThinkingLevel(s.autoModeStartThinkingLevel);
+        }
         s.currentUnitModel = match as AutoSession["currentUnitModel"];
         ctx.ui.notify(`Hook model override: ${match.provider}/${match.id}`, "info");
       } else {
@@ -1589,21 +1682,38 @@ export async function runUnitPhase(
   }
 
   if (unitResult.status === "cancelled") {
-    // Provider-error pause: pauseAuto already handled cleanup and scheduled
-    // recovery. Don't hard-stop — just break out of the loop (#2762).
-    if (unitResult.errorContext?.category === "provider") {
+    const errorCategory = unitResult.errorContext?.category;
+    // Provider-error pause: agent_end recovery normally pauses before this
+    // branch. Provider readiness failures happen before dispatch, so pause here
+    // if nothing upstream already did.
+    if (errorCategory === "provider") {
+      if (!s.paused) {
+        const detail = unitResult.errorContext?.message ?? `Provider unavailable for ${unitType} ${unitId}`;
+        await pauseAutoForProviderError(
+          ctx.ui,
+          detail,
+          () => deps.pauseAuto(ctx, pi),
+          {
+            isRateLimit: false,
+            isTransient: Boolean(unitResult.errorContext?.isTransient),
+            retryAfterMs: unitResult.errorContext?.retryAfterMs,
+          },
+        );
+      }
       await emitCancelledUnitEnd(ic, unitType, unitId, unitStartSeq, unitResult.errorContext);
-      debugLog("autoLoop", { phase: "exit", reason: "provider-pause", isTransient: unitResult.errorContext.isTransient });
+      debugLog("autoLoop", { phase: "exit", reason: "provider-pause", isTransient: unitResult.errorContext?.isTransient });
       return { action: "break", reason: "provider-pause" };
     }
     // Timeout category covers two distinct scenarios:
     //   1. Session creation timeout (120s) — transient, auto-resume with backoff
     //   2. Unit hard timeout (30min+) — stuck agent, pause for manual review
+    // Transient session-failed covers recoverable newSession failures and should
+    // pause instead of hard-stopping.
     // Structural errors (TypeError, is not a function) are NOT transient
     // and must hard-stop to avoid infinite retry loops.
     if (
       unitResult.errorContext?.isTransient &&
-      unitResult.errorContext?.category === "timeout"
+      errorCategory === "timeout"
     ) {
       const isSessionCreationTimeout = unitResult.errorContext.message?.includes("Session creation timed out");
 
@@ -1666,6 +1776,20 @@ export async function runUnitPhase(
       await emitCancelledUnitEnd(ic, unitType, unitId, unitStartSeq, unitResult.errorContext);
       return { action: "break", reason: "unit-hard-timeout" };
     }
+    if (
+      unitResult.errorContext?.isTransient &&
+      errorCategory === "session-failed"
+    ) {
+      ctx.ui.notify(
+        `Session creation failed transiently for ${unitType} ${unitId}: ${unitResult.errorContext?.message ?? "unknown"}. Pausing auto-mode (recoverable).`,
+        "warning",
+      );
+      debugLog("autoLoop", { phase: "session-start-transient-pause", unitType, unitId, category: errorCategory });
+      await deps.pauseAuto(ctx, pi);
+      await deps.autoCommitUnit?.(s.basePath, unitType, unitId, ctx);
+      await emitCancelledUnitEnd(ic, unitType, unitId, unitStartSeq, unitResult.errorContext);
+      return { action: "break", reason: "session-timeout" };
+    }
     // All other cancelled states (structural errors, non-transient failures): hard stop
     if (s.currentUnit) {
       await deps.closeoutUnit(
@@ -1881,6 +2005,20 @@ export async function runFinalize(
       debugLog("autoLoop", { phase: "sidecar-artifact-retry-skipped", iteration: ic.iteration });
     } else {
       // s.pendingVerificationRetry was set by postUnitPreVerification.
+      // Emit a dedicated journal event so forensics can distinguish bounded
+      // verification retries from genuine stuck-loop dispatch repetitions (#4540).
+      const retryInfo = s.pendingVerificationRetry;
+      deps.emitJournalEvent({
+        ts: new Date().toISOString(),
+        flowId: ic.flowId,
+        seq: ic.nextSeq(),
+        eventType: "artifact-verification-retry",
+        data: {
+          unitType: preUnitSnapshot?.type,
+          unitId: retryInfo?.unitId,
+          attempt: retryInfo?.attempt,
+        },
+      });
       // Continue the loop — next iteration will inject the retry context into the prompt.
       debugLog("autoLoop", { phase: "artifact-verification-retry", iteration: ic.iteration });
       return { action: "continue" };

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.