@seanyao/roll 4.721.1 → 4.721.2

package/CHANGELOG.md

@@ -2,6 +2,11 @@
 
 ## Unreleased
 
+## v4.721.2 — 2026-07-21
+
+### 稳定性
+- loop 的构建子进程失联或被系统杀掉后，现在会在限定时间内察觉、收尾并写入终态（标记为 aborted），不再让整个受监督周期永久挂住；新增子进程存活/超时检测（FIX-1474）[loop]
+
 ## v4.721.1 — 2026-07-21
 
 ### 稳定性

package/dist/roll.mjs

@@ -14639,6 +14639,19 @@ function timeoutTeardownCommands(ctx) {
     { kind: "release_lock" }
   ];
 }
+function abortedTeardownCommands(ctx) {
+  return [
+    { kind: "kill_agent", graceSec: WATCHDOG_KILL_GRACE_SEC },
+    { kind: "measure_worktree" },
+    { kind: "emit_event", event: cycleEndEvent(ctx, "aborted") },
+    { kind: "append_run", status: "aborted", outcome: mapV2Status("aborted"), cycleId: ctx.cycleId },
+    {
+      kind: "append_alert",
+      message: `cycle ${ctx.cycleId}: builder child process lost (killed out-of-band) \u2014 cycle aborted; worktree preserved`
+    },
+    { kind: "release_lock" }
+  ];
+}
 function cycleTimeoutVerdict(input) {
   const wall = input.wallLimitSec ?? CYCLE_WALL_TIMEOUT_SEC;
   const idle = input.noProgressLimitSec ?? CYCLE_NO_PROGRESS_SEC;
@@ -14833,6 +14846,12 @@ function cycleStep(state, event) {
         { kind: "cleanup_worktree", branch: state.ctx.branch }
       ]);
     case "agent_exited": {
+      if (event.lost === true) {
+        return {
+          state: { ...state, phase: "execute", terminal: "aborted", done: true, ctx: { ...state.ctx, agentExitCode: event.exit, agentTimedOut: event.timedOut } },
+          commands: abortedTeardownCommands(terminalCtx(state))
+        };
+      }
       const plan = retryPlan({ attempt: state.attempt, exit: event.exit, timedOut: event.timedOut });
       if (plan.action === "abort_timeout") {
         return {
@@ -246187,6 +246206,10 @@ function spawnAndWait(bin, args, opts, pty = false, profile) {
       stderr += d.toString("utf8");
     });
     liveAgents.add(child);
+    try {
+      opts.onSpawn?.(child);
+    } catch {
+    }
     let settled = false;
     let exitDrainFallback;
     const settle = (result2) => {
@@ -284694,6 +284717,68 @@ function startStallDetector(opts) {
     }
   };
 }
+var LIVENESS_POLL_MS = 5e3;
+var LIVENESS_CONFIRM_TICKS = 2;
+function defaultIsAlive(pid) {
+  try {
+    process.kill(pid, 0);
+    return true;
+  } catch (e) {
+    return e.code === "EPERM";
+  }
+}
+function startBuilderLivenessProbe(opts) {
+  const { cycleId, agent, appendEvent: appendEvent3 } = opts;
+  const isAlive = opts.isAlive ?? defaultIsAlive;
+  const kill = opts.kill ?? (() => killLiveAgents("SIGKILL"));
+  const pollMs = opts.pollMs ?? (Number((process.env["ROLL_LIVENESS_POLL_MS"] ?? "").trim()) || LIVENESS_POLL_MS);
+  const confirmTicks = opts.confirmTicks ?? LIVENESS_CONFIRM_TICKS;
+  let lost = false;
+  let deadStreak = 0;
+  const tick = () => {
+    if (lost || !opts.spawnPending())
+      return;
+    const pid = opts.pid();
+    if (pid === void 0) {
+      deadStreak = 0;
+      return;
+    }
+    let alive = true;
+    try {
+      alive = isAlive(pid);
+    } catch {
+    }
+    if (alive) {
+      deadStreak = 0;
+      return;
+    }
+    deadStreak += 1;
+    if (deadStreak < confirmTicks)
+      return;
+    lost = true;
+    clearInterval(timer);
+    try {
+      appendEvent3({ type: "cycle:agent_lost", cycleId, agent, pid, ts: Date.now() });
+    } catch {
+    }
+    try {
+      kill();
+    } catch {
+    }
+    try {
+      opts.onLost?.({ pid });
+    } catch {
+    }
+  };
+  const timer = setInterval(tick, pollMs);
+  timer.unref?.();
+  return {
+    stop: () => {
+      clearInterval(timer);
+      return { lost };
+    }
+  };
+}
 function createCaptureMarkerSink(runDir, capture) {
   let buf = "";
   const pending = [];
@@ -284817,6 +284902,24 @@ ${skillBodyForSpawn}` : skillBodyForSpawn;
       let timeoutFired = null;
       let activeMainLeak = { detected: false, files: [] };
       let mainLeakWatchdog;
+      let spawnedPid;
+      let spawnSettled = false;
+      let agentLost = false;
+      let resolveLostRace;
+      const lostRace = new Promise((resolve17) => {
+        resolveLostRace = () => resolve17({ stdout: "", stderr: "", exitCode: 137, timedOut: false });
+      });
+      const livenessProbe = startBuilderLivenessProbe({
+        cycleId: ctx.cycleId ?? "",
+        agent: cmd.agent,
+        pid: () => spawnedPid,
+        spawnPending: () => !spawnSettled,
+        appendEvent: (ev) => ports.events.appendEvent(ports.paths.eventsPath, ev),
+        onLost: () => {
+          agentLost = true;
+          resolveLostRace?.();
+        }
+      });
       try {
         appendWriteProtectionEvent(ports, applyMainCheckoutWriteProtection({
           repoCwd: ports.repoCwd,
@@ -284825,37 +284928,47 @@ ${skillBodyForSpawn}` : skillBodyForSpawn;
           nowMs: () => eventTs2(ports)
         }));
         mainLeakWatchdog = startMainCheckoutLeakWatchdog(ports, ctx);
-        res = await ports.agentSpawn(cmd.agent, {
-          purpose: "builder",
-          // E4: the builder runs in the submodule cycle worktree for a submodule
-          // story (execCwd); its git env + writable roots target the submodule's
-          // own repo/object store so its commits actually land there.
-          cwd: execCwd,
-          skillBody: finalSkillBody,
-          ...ctx.evidenceRunDir !== void 0 ? { runDir: ctx.evidenceRunDir } : {},
-          writableRoots: submoduleAgentWritableRoots(ports.repoCwd, execRepoCwd, ports.paths.alertsPath),
-          ...ctx.model !== void 0 && ctx.model !== "" ? { model: ctx.model } : {},
-          env: {
-            ...process.env,
-            ROLL_LOOP_ALERT: ports.paths.alertsPath,
-            ...worktreeGitEnv(execCwd, ports.repoCwd),
-            ...agentSpawnEnvironment(cmd.agent)
-          },
-          // FIX-204B: pin the executor-picked story into the agent prompt — the
-          // claim (pick_story → 🔨) and the work must be the same story.
-          ...ctx.storyId !== void 0 && ctx.storyId !== "" ? { storyId: ctx.storyId } : {},
-          onChunk: (d) => {
-            timeoutWatchdog.markProgress();
-            stallDetector.markProgress();
-            captureSink?.onChunk(d);
-            try {
-              appendFileSync16(livePath, d);
-            } catch {
+        res = await Promise.race([
+          ports.agentSpawn(cmd.agent, {
+            purpose: "builder",
+            // E4: the builder runs in the submodule cycle worktree for a submodule
+            // story (execCwd); its git env + writable roots target the submodule's
+            // own repo/object store so its commits actually land there.
+            cwd: execCwd,
+            skillBody: finalSkillBody,
+            ...ctx.evidenceRunDir !== void 0 ? { runDir: ctx.evidenceRunDir } : {},
+            writableRoots: submoduleAgentWritableRoots(ports.repoCwd, execRepoCwd, ports.paths.alertsPath),
+            ...ctx.model !== void 0 && ctx.model !== "" ? { model: ctx.model } : {},
+            env: {
+              ...process.env,
+              ROLL_LOOP_ALERT: ports.paths.alertsPath,
+              ...worktreeGitEnv(execCwd, ports.repoCwd),
+              ...agentSpawnEnvironment(cmd.agent)
+            },
+            // FIX-204B: pin the executor-picked story into the agent prompt — the
+            // claim (pick_story → 🔨) and the work must be the same story.
+            ...ctx.storyId !== void 0 && ctx.storyId !== "" ? { storyId: ctx.storyId } : {},
+            onChunk: (d) => {
+              timeoutWatchdog.markProgress();
+              stallDetector.markProgress();
+              captureSink?.onChunk(d);
+              try {
+                appendFileSync16(livePath, d);
+              } catch {
+              }
+              signalRecorder.accept(d);
+            },
+            // FIX-1474: hand the live child to the liveness probe so an
+            // out-of-band death is detected within a bounded window.
+            onSpawn: (child) => {
+              spawnedPid = child.pid;
             }
-            signalRecorder.accept(d);
-          }
-        });
+          }),
+          lostRace
+        ]);
       } finally {
+        spawnSettled = true;
+        livenessProbe.stop();
         if (mainLeakWatchdog !== void 0) {
           activeMainLeak = await mainLeakWatchdog.stop();
         }
@@ -284963,7 +285076,7 @@ ${res.stderr}`.split("\n").find((l) => l.trim() !== "") ?? "").slice(0, 200);
           // FIX-343 (step ①): the builder session id is part of the auditable
           // header — the attest gate's scorer≠builder-session invariant is then
           // traceable to a recorded build-session id, not asserted.
-          `# exit=${res.exitCode} timedOut=${res.timedOut} build-session=${builderSessionId}
+          `# exit=${res.exitCode} timedOut=${res.timedOut} lost=${agentLost} build-session=${builderSessionId}
 --- stdout ---
 ${res.stdout}
 --- stderr ---
@@ -285011,7 +285124,10 @@ ${res.stderr}
         cycleStartSec: ctx.startSec
       }) ?? void 0 : void 0;
       return {
-        event: { type: "agent_exited", exit: res.exitCode, timedOut: res.timedOut },
+        // FIX-1474: a lost child converges the cycle to the explicit `aborted`
+        // terminal in the orchestrator (no retry, no blocked) — the death was
+        // environmental, recorded fail-loud via cycle:agent_lost.
+        event: { type: "agent_exited", exit: res.exitCode, timedOut: res.timedOut, ...agentLost ? { lost: true } : {} },
         // FIX-343 (step ①): persist the builder session id on the cycle context so
         // it survives to the attest gate (the scorer≠builder-session invariant is then
         // traceable to a recorded build-session id, not asserted).

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@seanyao/roll",
-  "version": "4.721.1",
+  "version": "4.721.2",
   "description": "Roll — Roll out features with AI agents",
   "packageManager": "pnpm@11.1.3",
   "scripts": {