@lumi.ai/runner 0.6.0 → 0.6.2

package/dist/cli.js

@@ -764,7 +764,7 @@ function mcpUrl(config2) {
 }
 
 // src/version.ts
-var RUNNER_VERSION = true ? "0.6.0" : "0.0.0-dev";
+var RUNNER_VERSION = true ? "0.6.2" : "0.0.0-dev";
 
 // src/auth.ts
 import { signInWithCustomToken } from "firebase/auth";
@@ -2328,6 +2328,48 @@ function subscribeEngineLimits(db, shipId, cb, onError) {
   );
 }
 
+// src/jobs/claimBackoff.ts
+var BACKOFF_BASE_MS = 3e3;
+var BACKOFF_MAX_MS = 6e4;
+var BACKOFF_RETENTION_MS = 10 * 6e4;
+function backoffFor(releases) {
+  return Math.min(BACKOFF_MAX_MS, BACKOFF_BASE_MS * 2 ** Math.max(0, releases - 1));
+}
+function noteTransientRelease(backoff, jobId, now) {
+  const releases = (backoff.get(jobId)?.releases ?? 0) + 1;
+  const entry = { releases, eligibleAt: now + backoffFor(releases), poked: false };
+  backoff.set(jobId, entry);
+  return entry;
+}
+function isBackedOff(backoff, jobId, now) {
+  const entry = backoff.get(jobId);
+  return !!entry && entry.eligibleAt > now;
+}
+function clearBackoff(backoff, jobId) {
+  backoff.delete(jobId);
+}
+function pruneExpired2(backoff, now) {
+  const cleared = [];
+  for (const [key, entry] of backoff) {
+    if (entry.eligibleAt > now) continue;
+    if (!entry.poked) {
+      cleared.push(key);
+      entry.poked = true;
+    }
+    if (now - entry.eligibleAt >= BACKOFF_RETENTION_MS) backoff.delete(key);
+  }
+  return cleared;
+}
+function nextEligibleAt(backoff, now) {
+  let soonest = null;
+  for (const entry of backoff.values()) {
+    if (entry.eligibleAt > now && (soonest === null || entry.eligibleAt < soonest)) {
+      soonest = entry.eligibleAt;
+    }
+  }
+  return soonest;
+}
+
 // src/jobs/capacity.ts
 var DEFAULT_PARALLEL_JOBS = 1;
 var PARALLEL_MAX = 8;
@@ -3445,6 +3487,7 @@ async function startDaemon() {
   }
   const pending = /* @__PURE__ */ new Map();
   const engineLimits = /* @__PURE__ */ new Map();
+  const claimBackoff = /* @__PURE__ */ new Map();
   const agentEngines = /* @__PURE__ */ new Map();
   const unsubsByShip = /* @__PURE__ */ new Map();
   const listenerError = (shipId, what) => (e) => {
@@ -3627,6 +3670,23 @@ async function startDaemon() {
     limitTimer = setTimeout(sweepLimits, Math.max(0, at - Date.now()) + 1e3);
     limitTimer.unref();
   }
+  let backoffTimer = null;
+  function armBackoffTimer() {
+    if (backoffTimer) {
+      clearTimeout(backoffTimer);
+      backoffTimer = null;
+    }
+    const at = nextEligibleAt(claimBackoff, Date.now());
+    if (at === null) return;
+    backoffTimer = setTimeout(sweepBackoff, Math.max(0, at - Date.now()) + 1e3);
+    backoffTimer.unref();
+  }
+  function sweepBackoff() {
+    backoffTimer = null;
+    const cleared = pruneExpired2(claimBackoff, Date.now());
+    armBackoffTimer();
+    if (cleared.length > 0) poke();
+  }
   function sweepLimits() {
     limitTimer = null;
     const cleared = pruneExpired(engineLimits, Date.now());
@@ -3658,7 +3718,11 @@ async function startDaemon() {
           // the limit map is keyed by (Ship, engine) and the engine comes from the registry via
           // the agent, so the job loop still knows nothing about Claude. Skipped entries STAY in
           // `pending`, which is what lets a reset resume them with only a poke.
-          eligible: (p) => approved.get(p.shipId) === true && !isLimited(engineLimits, p.shipId, engineForJob(p.shipId, p.job), now)
+          eligible: (p) => approved.get(p.shipId) === true && !isLimited(engineLimits, p.shipId, engineForJob(p.shipId, p.job), now) && // The connection cooldown. Same "skip, don't stop" shape as the limit gate above, and
+          // skipped entries STAY in `pending` for the same reason — `sweepBackoff` resumes them
+          // with a poke. Without this the `transient` release below re-claims what it just put
+          // down, every three seconds, forever (jobs/claimBackoff.ts).
+          !isBackedOff(claimBackoff, p.job.id, now)
         });
         for (const pick of picks) {
           pending.delete(`${pick.shipId}/${pick.job.id}`);
@@ -4014,14 +4078,16 @@ async function startDaemon() {
           target.kind === "task" ? `Task ${target.taskId} is done.` : delivery === "agent-replied" ? "Agent replied in a chat." : "Agent finished a chat run without replying \u2014 its answer was posted for it."
         );
       } else if (transient) {
+        const held = noteTransientRelease(claimBackoff, job.id, Date.now());
         await releaseJob(
           sess(shipId).fb.db,
           shipId,
           job,
           `Lost the connection to Firestore \u2014 released without consuming a retry: ${failure.slice(0, 120)}`
         );
+        armBackoffTimer();
         log2(
-          `Job ${job.id} released: lost the connection to Firestore. Attempt ${job.attempt} preserved \u2014 ${failure.slice(0, 120)}`
+          `Job ${job.id} released: lost the connection to Firestore. Attempt ${job.attempt} preserved \u2014 ${failure.slice(0, 120)}. Not re-claiming it here for ${Math.round((held.eligibleAt - Date.now()) / 1e3)}s (drop ${held.releases} on this machine); another daemon may take it meanwhile.`
         );
       } else if (!terminal && job.attempt < MAX_ATTEMPTS) {
         log2(`Job ${job.id} failed (attempt ${job.attempt}) \u2014 re-queueing: ${failure.slice(0, 120)}`);
@@ -4043,6 +4109,7 @@ async function startDaemon() {
         log2(`Job ${job.id} FAILED terminally: ${failure.slice(0, 120)}`);
         notify("Crew job failed", `${targetLabel}: ${failure.slice(0, 120)}`);
       }
+      if (!transient) clearBackoff(claimBackoff, job.id);
     } catch (e) {
       log2(`finalize failed for ${job.id}: ${e instanceof Error ? e.message : e}`);
     }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lumi.ai/runner",
-  "version": "0.6.0",
+  "version": "0.6.2",
   "type": "module",
   "description": "Lumi Crew runner daemon — claims jobs from your Ships and executes them as headless Claude sessions on your own machine.",
   "//name": "The ONLY package in this monorepo published to the public registry, so it is the one that does not follow the internal @lumi/crew-* convention: `@lumi` is not a scope we own, `@lumi.ai` is (the npm org). The workspace DIRECTORY stays packages/crew/runner — renaming the package is not renaming the folder.",