Merge pull request #4042 from mastertyko/fix/3760-forensics-session-aware-loops

fix(gsd): scope stuck-loop forensics to auto sessions

src/resources/extensions/gsd/auto.ts

@@ -549,11 +549,13 @@ function buildSnapshotOpts(
   _unitType: string,
   _unitId: string,
 ): {
+  autoSessionKey?: string;
   continueHereFired?: boolean;
   promptCharCount?: number;
   baselineCharCount?: number;
 } & Record<string, unknown> {
   return {
+    ...(s.autoStartTime > 0 ? { autoSessionKey: String(s.autoStartTime) } : {}),
     promptCharCount: s.lastPromptCharCount,
     baselineCharCount: s.lastBaselineCharCount,
     ...(s.currentUnitRouting ?? {}),

src/resources/extensions/gsd/forensics.ts

@@ -650,19 +650,33 @@ function getDbCompletionCounts(): DbCompletionCounts | null {
  * Exported for testability.
  */
 export function detectStuckLoops(units: UnitMetrics[], anomalies: ForensicAnomaly[]): void {
-  // First, collect unique startedAt values per type/id key
-  const dispatchMap = new Map<string, Set<number>>();
+  // First, collect unique startedAt values per type/id key, bucketed by
+  // autoSessionKey when available so cross-session recovery does not look
+  // like a within-session stuck loop.
+  const dispatchMap = new Map<string, Map<string, Set<number>>>();
   for (const u of units) {
     const key = `${u.type}/${u.id}`;
-    let starts = dispatchMap.get(key);
+    let sessionBuckets = dispatchMap.get(key);
+    if (!sessionBuckets) {
+      sessionBuckets = new Map();
+      dispatchMap.set(key, sessionBuckets);
+    }
+
+    const sessionKey = u.autoSessionKey ?? "__legacy__";
+    let starts = sessionBuckets.get(sessionKey);
     if (!starts) {
       starts = new Set();
-      dispatchMap.set(key, starts);
+      sessionBuckets.set(sessionKey, starts);
     }
     starts.add(u.startedAt);
   }
-  for (const [key, starts] of dispatchMap) {
-    const count = starts.size;
+
+  for (const [key, sessionBuckets] of dispatchMap) {
+    const hasSessionAwareData = Array.from(sessionBuckets.keys()).some((sessionKey) => sessionKey !== "__legacy__");
+    const count = hasSessionAwareData
+      ? Math.max(...Array.from(sessionBuckets.values(), (starts) => starts.size))
+      : (sessionBuckets.get("__legacy__")?.size ?? 0);
+
     if (count > 1) {
       const [unitType, ...idParts] = key.split("/");
       anomalies.push({
@@ -671,7 +685,9 @@ export function detectStuckLoops(units: UnitMetrics[], anomalies: ForensicAnomal
         unitType,
         unitId: idParts.join("/"),
         summary: `Unit ${key} was dispatched ${count} times`,
-        details: `Repeated dispatch suggests the unit completed but its artifacts weren't verified, or the state machine kept returning it.`,
+        details: hasSessionAwareData
+          ? `Repeated dispatch within the same auto session suggests the unit completed but its artifacts were not verified, or the state machine kept returning it. Cross-session recovery runs are ignored.`
+          : `Repeated dispatch suggests the unit completed but its artifacts weren't verified, or the state machine kept returning it.`,
       });
     }
   }

src/resources/extensions/gsd/metrics.ts

@@ -41,6 +41,7 @@ export interface UnitMetrics {
   model: string;           // model ID used
   startedAt: number;       // ms timestamp
   finishedAt: number;      // ms timestamp
+  autoSessionKey?: string; // identifies one auto-mode run across pause/resume
   tokens: TokenCounts;
   cost: number;            // total USD cost
   toolCalls: number;
@@ -133,7 +134,16 @@ export function snapshotUnitMetrics(
   unitId: string,
   startedAt: number,
   model: string,
-  opts?: { tier?: string; modelDowngraded?: boolean; contextWindowTokens?: number; truncationSections?: number; continueHereFired?: boolean; promptCharCount?: number; baselineCharCount?: number },
+  opts?: {
+    tier?: string;
+    modelDowngraded?: boolean;
+    contextWindowTokens?: number;
+    truncationSections?: number;
+    continueHereFired?: boolean;
+    promptCharCount?: number;
+    baselineCharCount?: number;
+    autoSessionKey?: string;
+  },
 ): UnitMetrics | null {
   if (!ledger) return null;
 
@@ -181,6 +191,7 @@ export function snapshotUnitMetrics(
     model,
     startedAt,
     finishedAt: Date.now(),
+    ...(opts?.autoSessionKey ? { autoSessionKey: opts.autoSessionKey } : {}),
     tokens,
     cost,
     toolCalls,

src/resources/extensions/gsd/tests/forensics-stuck-loops.test.ts

@@ -101,3 +101,65 @@ test("#1943 detectStuckLoops ignores watchdog duplicates but flags real re-dispa
   assert.equal(anomalies.length, 1, `expected 1 anomaly (for the 3x dispatched task), got ${anomalies.length}`);
   assert.ok(anomalies[0].summary.includes("3 times"));
 });
+
+test("#3760 detectStuckLoops ignores cross-session recovery re-dispatches", () => {
+  const anomalies: ForensicAnomaly[] = [];
+
+  const units: UnitMetrics[] = [
+    makeUnit({
+      type: "plan-slice",
+      id: "M001/S02",
+      startedAt: 1000,
+      finishedAt: 2000,
+      autoSessionKey: "session-a",
+    }),
+    makeUnit({
+      type: "plan-slice",
+      id: "M001/S02",
+      startedAt: 5000,
+      finishedAt: 6000,
+      autoSessionKey: "session-b",
+    }),
+  ];
+
+  detectStuckLoops(units, anomalies);
+
+  assert.equal(anomalies.length, 0, "cross-session recovery should not be flagged as a stuck loop");
+});
+
+test("#3760 detectStuckLoops still flags repeated dispatches within one auto session", () => {
+  const anomalies: ForensicAnomaly[] = [];
+
+  const units: UnitMetrics[] = [
+    makeUnit({
+      type: "complete-slice",
+      id: "M011/S02",
+      startedAt: 1000,
+      finishedAt: 2000,
+      autoSessionKey: "session-a",
+    }),
+    makeUnit({
+      type: "complete-slice",
+      id: "M011/S02",
+      startedAt: 5000,
+      finishedAt: 6000,
+      autoSessionKey: "session-a",
+    }),
+    makeUnit({
+      type: "complete-slice",
+      id: "M011/S02",
+      startedAt: 9000,
+      finishedAt: 10000,
+      autoSessionKey: "session-b",
+    }),
+  ];
+
+  detectStuckLoops(units, anomalies);
+
+  assert.equal(anomalies.length, 1, "within-session retries should still be flagged");
+  assert.ok(anomalies[0].summary.includes("2 times"), `summary should reflect the worst same-session loop: ${anomalies[0].summary}`);
+  assert.ok(
+    anomalies[0].details.includes("Cross-session recovery runs are ignored"),
+    `details should explain the session-aware rule: ${anomalies[0].details}`,
+  );
+});