@lumi.ai/runner 0.5.0 → 0.5.2

package/dist/cli.js

@@ -552,7 +552,7 @@ function mcpUrl(config2) {
 }
 
 // src/version.ts
-var RUNNER_VERSION = true ? "0.5.0" : "0.0.0-dev";
+var RUNNER_VERSION = true ? "0.5.2" : "0.0.0-dev";
 
 // src/auth.ts
 import { signInWithCustomToken } from "firebase/auth";
@@ -909,52 +909,80 @@ import {
   query,
   where
 } from "firebase/firestore";
+
+// src/jobs/retry.ts
+var FIRESTORE_RETRY_DELAYS_MS = [250, 750, 2e3];
+function isTransientFirestoreError(error) {
+  const raw = error?.code;
+  const code = typeof raw === "string" ? raw.replace(/^firestore\//, "") : "";
+  if (code === "unavailable" || code === "deadline-exceeded" || code === "internal" || code === "cancelled" || code === "aborted" || code === "resource-exhausted") {
+    return true;
+  }
+  const message = error?.message;
+  return typeof message === "string" && message.toLowerCase().includes("client is offline");
+}
+async function withFirestoreRetry(read, sleep2 = (ms) => new Promise((r) => setTimeout(r, ms)), delays = FIRESTORE_RETRY_DELAYS_MS) {
+  for (let i = 0; ; i++) {
+    try {
+      return await read();
+    } catch (error) {
+      if (i >= delays.length || !isTransientFirestoreError(error)) throw error;
+      await sleep2(delays[i]);
+    }
+  }
+}
+
+// src/jobs/contextPack.ts
 var MAX_ACTIVITY_IN_PROMPT = 40;
 var MAX_PREVIOUS_JOBS_READ = 5;
 async function loadJobContext(db, shipId, job) {
   const shipRef = doc(db, COLLECTIONS.ships, shipId);
   const taskRef = doc(shipRef, COLLECTIONS.tasks, job.taskId);
-  const [shipSnap, agentSnap, taskSnap, activitySnap, jobsSnap, indexSnap] = await Promise.all([
-    getDoc(shipRef),
-    getDoc(doc(shipRef, COLLECTIONS.agents, job.agentId)),
-    getDoc(taskRef),
-    // DESC + reverse, not ASC + limit: a limit takes the FIRST rows the order produces, so an
-    // ascending query with a limit would hand the agent the OLDEST 40 events and hide everything
-    // that has happened since — the exact opposite of what continuity needs. The extra row is
-    // how truncation is detected without a second count query.
-    getDocs(
-      query(
-        collection(taskRef, COLLECTIONS.activity),
-        orderBy("createdAt", "desc"),
-        limit(MAX_ACTIVITY_IN_PROMPT + 1)
-      )
-    ),
-    // The extra row here is for a different reason: this job itself is usually the newest match
-    // and is filtered out below, so without it a full window yields one report short.
-    getDocs(
-      query(
-        collection(shipRef, COLLECTIONS.jobs),
-        where("taskId", "==", job.taskId),
-        orderBy("createdAt", "desc"),
-        limit(MAX_PREVIOUS_JOBS_READ + 1)
-      )
-    ),
-    // The knowledge CATALOG — one document, so this is +1 read regardless of how much the Ship
-    // knows, and it rides the existing Promise.all so it costs no extra latency either. The
-    // agent's memory is free: it is a field on the agent doc already being fetched above.
-    //
-    // ENRICHMENT, not identity, exactly like the playbook read below: a Ship whose catalog is
-    // missing, unreadable or not yet deployed still has a perfectly valid session. Caught rather
-    // than thrown, and deliberately NOT placed where a rejection would propagate.
-    getDoc(doc(shipRef, COLLECTIONS.indexes, INDEX_DOCS.knowledge)).catch(() => null)
-  ]);
+  const [shipSnap, agentSnap, taskSnap, activitySnap, jobsSnap, indexSnap] = await withFirestoreRetry(
+    () => Promise.all([
+      getDoc(shipRef),
+      getDoc(doc(shipRef, COLLECTIONS.agents, job.agentId)),
+      getDoc(taskRef),
+      // DESC + reverse, not ASC + limit: a limit takes the FIRST rows the order produces, so an
+      // ascending query with a limit would hand the agent the OLDEST 40 events and hide everything
+      // that has happened since — the exact opposite of what continuity needs. The extra row is
+      // how truncation is detected without a second count query.
+      getDocs(
+        query(
+          collection(taskRef, COLLECTIONS.activity),
+          orderBy("createdAt", "desc"),
+          limit(MAX_ACTIVITY_IN_PROMPT + 1)
+        )
+      ),
+      // The extra row here is for a different reason: this job itself is usually the newest match
+      // and is filtered out below, so without it a full window yields one report short.
+      getDocs(
+        query(
+          collection(shipRef, COLLECTIONS.jobs),
+          where("taskId", "==", job.taskId),
+          orderBy("createdAt", "desc"),
+          limit(MAX_PREVIOUS_JOBS_READ + 1)
+        )
+      ),
+      // The knowledge CATALOG — one document, so this is +1 read regardless of how much the Ship
+      // knows, and it rides the existing Promise.all so it costs no extra latency either. The
+      // agent's memory is free: it is a field on the agent doc already being fetched above.
+      //
+      // ENRICHMENT, not identity, exactly like the playbook read below: a Ship whose catalog is
+      // missing, unreadable or not yet deployed still has a perfectly valid session. Caught rather
+      // than thrown, and deliberately NOT placed where a rejection would propagate.
+      getDoc(doc(shipRef, COLLECTIONS.indexes, INDEX_DOCS.knowledge)).catch(() => null)
+    ])
+  );
   if (!shipSnap.exists() || !agentSnap.exists() || !taskSnap.exists()) {
     throw new Error("Job context incomplete: ship, agent or task missing.");
   }
   const task = { id: taskSnap.id, ...taskSnap.data() };
   let parentTask = null;
   if (task.parentTaskId) {
-    const parentSnap = await getDoc(doc(shipRef, COLLECTIONS.tasks, task.parentTaskId));
+    const parentSnap = await withFirestoreRetry(
+      () => getDoc(doc(shipRef, COLLECTIONS.tasks, task.parentTaskId))
+    );
     if (parentSnap.exists()) {
       parentTask = { id: parentSnap.id, ...parentSnap.data() };
     }
@@ -1151,7 +1179,7 @@ async function loadChatContext(db, shipId, job) {
   const shipRef = doc2(db, COLLECTIONS.ships, shipId);
   const chatRef = doc2(shipRef, COLLECTIONS.chats, job.chatId);
   const messagesCol = collection2(chatRef, COLLECTIONS.chatMessages);
-  const [shipSnap, agentSnap, chatSnap, messagesSnap, countSnap, membersSnap, agentsSnap, indexSnap] = await Promise.all([
+  const [shipSnap, agentSnap, chatSnap, messagesSnap, countSnap, membersSnap, agentsSnap, indexSnap] = await withFirestoreRetry(() => Promise.all([
     getDoc2(shipRef),
     getDoc2(doc2(shipRef, COLLECTIONS.agents, job.agentId)),
     getDoc2(chatRef),
@@ -1165,7 +1193,7 @@ async function loadChatContext(db, shipId, job) {
     // ENRICHMENT, not identity — same posture as the task pack: a Ship whose catalog is
     // missing or unreadable still has a valid session.
     getDoc2(doc2(shipRef, COLLECTIONS.indexes, INDEX_DOCS.knowledge)).catch(() => null)
-  ]);
+  ]));
   if (!shipSnap.exists() || !agentSnap.exists() || !chatSnap.exists()) {
     throw new Error("Chat job context incomplete: ship, agent or chat missing.");
   }
@@ -1575,8 +1603,8 @@ import {
 var TerminalJobError = class extends Error {
 };
 async function readIntegration(db, shipId) {
-  const snap = await getDoc3(
-    doc3(db, COLLECTIONS.ships, shipId, COLLECTIONS.integrations, INTEGRATION_DOCS.github)
+  const snap = await withFirestoreRetry(
+    () => getDoc3(doc3(db, COLLECTIONS.ships, shipId, COLLECTIONS.integrations, INTEGRATION_DOCS.github))
   );
   return snap.exists() ? snap.data() : null;
 }
@@ -1606,8 +1634,8 @@ async function resolveGithubToken(input) {
 // src/jobs/secrets.ts
 import { doc as doc4, getDoc as getDoc4 } from "firebase/firestore";
 async function loadRunnerSecrets(db, shipId) {
-  const snap = await getDoc4(
-    doc4(db, COLLECTIONS.ships, shipId, COLLECTIONS.secrets, SECRET_DOCS.runner)
+  const snap = await withFirestoreRetry(
+    () => getDoc4(doc4(db, COLLECTIONS.ships, shipId, COLLECTIONS.secrets, SECRET_DOCS.runner))
   );
   return snap.exists() ? snap.data() : null;
 }
@@ -2227,6 +2255,7 @@ async function startDaemon() {
     );
     let failure = null;
     let terminal = false;
+    let transient = false;
     let sessionLimit = null;
     let engineId = DEFAULT_ENGINE_ID;
     let transcript = "";
@@ -2315,6 +2344,7 @@ async function startDaemon() {
       if (!session.ok) failure = session.resultText || "Session failed.";
     } catch (e) {
       failure = e instanceof Error ? e.message : String(e);
+      transient = isTransientFirestoreError(e);
     }
     wake?.release();
     try {
@@ -2355,6 +2385,16 @@ async function startDaemon() {
           "Crew job finished",
           target.kind === "task" ? `Task ${target.taskId} is done.` : "Agent replied in a chat."
         );
+      } else if (transient) {
+        await releaseJob(
+          sess(shipId).fb.db,
+          shipId,
+          job,
+          `Lost the connection to Firestore \u2014 released without consuming a retry: ${failure.slice(0, 120)}`
+        );
+        log2(
+          `Job ${job.id} released: lost the connection to Firestore. Attempt ${job.attempt} preserved \u2014 ${failure.slice(0, 120)}`
+        );
       } else if (!terminal && job.attempt < MAX_ATTEMPTS) {
         log2(`Job ${job.id} failed (attempt ${job.attempt}) \u2014 re-queueing: ${failure.slice(0, 120)}`);
         await requeueForRetry(sess(shipId).fb.db, shipId, job, failure);

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lumi.ai/runner",
-  "version": "0.5.0",
+  "version": "0.5.2",
   "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.",