@lumi.ai/runner 0.6.3 → 0.6.4

package/dist/cli.js

@@ -230,6 +230,8 @@ function jobTarget(job) {
     return { kind: "chat", chatId: job.chatId };
   return null;
 }
+var MAX_RUN_REPORT_CHARS = 6e3;
+var MAX_RUN_REPORT_SUMMARY_CHARS = 300;
 
 // ../shared/dist/jobProgress.js
 var MAX_JOB_STEPS = 10;
@@ -772,7 +774,7 @@ function mcpUrl(config2) {
 }
 
 // src/version.ts
-var RUNNER_VERSION = true ? "0.6.3" : "0.0.0-dev";
+var RUNNER_VERSION = true ? "0.6.4" : "0.0.0-dev";
 
 // src/auth.ts
 import { signInWithCustomToken } from "firebase/auth";
@@ -1196,6 +1198,13 @@ async function loadSettledBlockers(shipRef, task) {
     return [];
   }
 }
+function reportSummaryLine(r) {
+  if (r.summary?.trim()) return r.summary.trim();
+  const body = r.report?.trim();
+  if (!body) return "No report.";
+  const firstLine2 = body.split("\n").find((l) => l.trim())?.trim() ?? body;
+  return firstLine2.length <= MAX_RUN_REPORT_SUMMARY_CHARS ? firstLine2 : `${firstLine2.slice(0, MAX_RUN_REPORT_SUMMARY_CHARS - 1)}\u2026`;
+}
 async function loadJobContext(db, shipId, job) {
   const shipRef = doc(db, COLLECTIONS.ships, shipId);
   const taskRef = doc(shipRef, COLLECTIONS.tasks, job.taskId);
@@ -1276,7 +1285,13 @@ async function loadJobContext(db, shipId, job) {
     activity: activityDocs.map((d) => ({ id: d.id, ...d.data() })).reverse(),
     // read newest-first, rendered chronologically
     activityTruncated: activitySnap.docs.length > MAX_ACTIVITY_IN_PROMPT,
-    previousReports: jobsSnap.docs.filter((d) => d.id !== job.id).slice(0, MAX_PREVIOUS_JOBS_READ).map((d) => d.data().report).filter((r) => !!r).reverse()
+    // §15.41. No longer filtered down to jobs that HAVE a report: the catalog carries a line per
+    // previous run, and "that run left no report" is itself worth knowing — it is the difference
+    // between a task nobody has worked and one where three runs died without saying why.
+    previousReports: jobsSnap.docs.filter((d) => d.id !== job.id).slice(0, MAX_PREVIOUS_JOBS_READ).map((d) => {
+      const data = d.data();
+      return { jobId: d.id, summary: data.reportSummary ?? null, report: data.report ?? null };
+    }).reverse()
     // chronological
   };
 }
@@ -1301,7 +1316,11 @@ function standingRules(ship2, statuses, task, agent) {
     // Amended with the pointer clause, which is the only thing keeping the three tiers from
     // duplicating each other: without it a diligent agent writes the same fact into its report,
     // its notebook and the org brain, and the report grows into a transcript of the other two.
-    "End every session by calling run_report: what you did, what remains, what the next run must know. Do not repeat what you already put in your memory or in Ship knowledge \u2014 point at it instead (the entry id, or the slug)."
+    // §15.41 added the summary clause. It is second, not an afterthought at the end, because that
+    // one string is what every LATER run on this task reads instead of this report — and what a
+    // human sees in their notification. An agent that treats it as a label writes "Run report",
+    // and four runs later the catalog says nothing at all.
+    "End every session by calling run_report: what you did, what remains, what the next run must know. Give it a summary too \u2014 one or two sentences that stand in for the whole report in later runs, so write the sentence you would want the next run to read first. Do not repeat what you already put in your memory or in Ship knowledge \u2014 point at it instead (the entry id, or the slug)."
   ].join("\n- ");
 }
 function formatTaskDate(millis, now = Date.now()) {
@@ -1341,26 +1360,17 @@ function buildPrompt(ctx, reason, mcpServers = [], tokenRepos, githubUnavailable
   const parts = [];
   const statuses = shipTaskStatuses(ctx.ship);
   const playbook = usableWorkflow(ctx.workflow);
+  let preamble = null;
   if (reason === "activity") {
-    parts.push(
-      "# Why this session started\n\nSomeone posted on your task after your last run. Read the newest activity below" + (playbook ? ", then follow your playbook below." : ", respond via task_comment, do any follow-up work it asks for, and set the task status honestly if it needs to change.")
-    );
+    preamble = "# Why this session started\n\nSomeone posted on your task after your last run. Read the newest activity below" + (playbook ? ", then follow your playbook above." : ", respond via task_comment, do any follow-up work it asks for, and set the task status honestly if it needs to change.");
   } else if (reason === "watch") {
-    parts.push(
-      "# Why this session started\n\nOne of your playbooks watches the board, and this task matched it. Nobody assigned it to you and nobody else is working it right now \u2014 you are picking it up.\n\n**Leave it in a status that says what happened.** The board is how the watch decides what still needs attention, so a task you leave in the same status will come back to you on the next scan" + (playbook ? ", which your playbook below is written for." : ".")
-    );
+    preamble = "# Why this session started\n\nOne of your playbooks watches the board, and this task matched it. Nobody assigned it to you and nobody else is working it right now \u2014 you are picking it up.\n\n**Leave it in a status that says what happened.** The board is how the watch decides what still needs attention, so a task you leave in the same status will come back to you on the next scan" + (playbook ? ", which your playbook above is written for." : ".");
   } else if (reason === "resume_after_approval") {
-    parts.push(
-      "# Why this session started\n\nYou asked a captain for permission and stopped. They have now answered \u2014 their decision is the newest entry in the task activity below. Read it first.\n\n**If they approved it, do that thing now** \u2014 the permission is granted for this task and may be single-use, so do not ask again for the same thing. **If they refused, do not retry and do not look for a way around it**: say what you will do instead, or hand the task back with task_assign."
-    );
+    preamble = "# Why this session started\n\nYou asked a captain for permission and stopped. They have now answered \u2014 their decision is the newest entry in the task activity below. Read it first.\n\n**If they approved it, do that thing now** \u2014 the permission is granted for this task and may be single-use, so do not ask again for the same thing. **If they refused, do not retry and do not look for a way around it**: say what you will do instead, or hand the task back with task_assign.";
   } else if (reason === "unblocked") {
-    parts.push(
-      '# Why this session started\n\nThis task was blocked and is not any more: everything it was waiting on is now done. Nobody has just assigned it to you \u2014 you have had it all along, and the work has become startable.\n\n**Read "What you were waiting on" below before you do anything else.** It carries what those tasks produced, which is the input this work was held up for; starting without it means redoing or contradicting somebody else' + (playbook ? ", then follow your playbook below." : "'s work.")
-    );
+    preamble = '# Why this session started\n\nThis task was blocked and is not any more: everything it was waiting on is now done. Nobody has just assigned it to you \u2014 you have had it all along, and the work has become startable.\n\n**Read "What you were waiting on" below before you do anything else.** It carries what those tasks produced, which is the input this work was held up for; starting without it means redoing or contradicting somebody else' + (playbook ? ", then follow your playbook above." : "'s work.");
   } else if (reason === "schedule") {
-    parts.push(
-      "# Why this session started\n\nThis is a scheduled run: one of your own playbooks created this task on its cron and assigned it to you. It is routine work, not a request from a person, so nobody is waiting on a reply" + (playbook ? " \u2014 the playbook below is the work, and the task's description is only a record of why it exists." : ".")
-    );
+    preamble = "# Why this session started\n\nThis is a scheduled run: one of your own playbooks created this task on its cron and assigned it to you. It is routine work, not a request from a person, so nobody is waiting on a reply" + (playbook ? " \u2014 the playbook above is the work, and the task's description is only a record of why it exists." : ".");
   }
   parts.push(`# Who you are
 
@@ -1385,6 +1395,7 @@ This is the playbook for this kind of work \u2014 follow it. Your contract and t
 ${playbook.instructions}`
     );
   }
+  if (preamble) parts.push(preamble);
   const t = ctx.task;
   const now = Date.now();
   const statusLabel = statusById(statuses, t.status)?.label ?? t.status;
@@ -1436,12 +1447,28 @@ ${p.description || "(no description)"}`
 ${note2}${feed}`);
   }
   if (ctx.previousReports.length > 0) {
-    const reports = ctx.previousReports.map((r, i) => `## Run ${i + 1}
+    const runs = ctx.previousReports;
+    const latest = runs[runs.length - 1];
+    const older = runs.slice(0, -1);
+    const sections = [];
+    if (older.length > 0) {
+      const lines = older.map((r) => `- (\`${r.jobId}\`) ${reportSummaryLine(r)}`).join("\n");
+      sections.push(
+        `## Earlier runs
 
-${r}`).join("\n\n");
+Oldest first, one line each. To read one in full, call \`task_get\` with this task's id and \`reportJobId\` set to the id in brackets.
+
+${lines}`
+      );
+    }
+    sections.push(
+      `## Most recent run (\`${latest.jobId}\`)
+
+${latest.report ?? "That run left no report."}`
+    );
     parts.push(`# Reports from previous runs on this task
 
-${reports}`);
+${sections.join("\n\n")}`);
   }
   const githubRepos = tokenRepos ?? ctx.agent.tools.github.repos;
   if (ctx.agent.tools.github.enabled && githubUnavailable) {
@@ -1517,7 +1544,10 @@ async function loadChatContext(db, shipId, job) {
           limit2(MAX_CHAT_REPORTS + 1)
         )
       );
-      previousReports = jobsSnap.docs.filter((d) => d.id !== job.id).slice(0, MAX_CHAT_REPORTS).map((d) => d.data().report).filter((r) => !!r).reverse();
+      previousReports = jobsSnap.docs.filter((d) => d.id !== job.id).slice(0, MAX_CHAT_REPORTS).map((d) => {
+        const data = d.data();
+        return { jobId: d.id, summary: data.reportSummary ?? null, report: data.report ?? null };
+      }).reverse();
     } catch {
       previousReports = [];
     }
@@ -1553,7 +1583,7 @@ function chatStandingRules(ship2, agent) {
     scopeRule,
     registerRule,
     routingRule,
-    "End the session by calling run_report: what was discussed and anything the next run must know. Do not repeat what you already put in your memory or in Ship knowledge \u2014 point at it instead."
+    "End the session by calling run_report: what was discussed and anything the next run must know. Give it a summary too \u2014 one or two sentences that stand in for the whole report once this conversation is long enough to scroll out of the window. Do not repeat what you already put in your memory or in Ship knowledge \u2014 point at it instead."
   ].join("\n- ");
 }
 function buildChatPrompt(ctx, mcpServers = [], tokenRepos, githubUnavailable) {
@@ -1576,15 +1606,22 @@ ${ctx.agent.contract}`);
 
 - ${chatStandingRules(ctx.ship, ctx.agent)}`);
   if (ctx.previousReports.length > 0) {
-    const reports = ctx.previousReports.map((r, i) => `## Earlier run ${i + 1}
+    const runs = ctx.previousReports;
+    const latest = runs[runs.length - 1];
+    const older = runs.slice(0, -1);
+    const sections = [];
+    if (older.length > 0) {
+      sections.push(older.map((r) => `- ${reportSummaryLine(r)}`).join("\n"));
+    }
+    sections.push(`## Most recent run
 
-${r}`).join("\n\n");
+${latest.report ?? "That run left no report."}`);
     parts.push(
       `# Earlier in this chat
 
 The conversation is longer than the window below. These are your own notes from previous runs on this chat, oldest first.
 
-${reports}`
+${sections.join("\n\n")}`
     );
   }
   const title = ctx.chat.title?.trim();
@@ -2445,6 +2482,19 @@ async function uploadTranscript(storage, shipId, jobId, redacted) {
 function utcDay(millis) {
   return new Date(millis).toISOString().slice(0, 10);
 }
+function backstopReportContent(resultText) {
+  const text = resultText.trim();
+  if (!text) {
+    const none = "This run ended without writing a report, and left no final text to fall back on.";
+    return { report: none, summary: none };
+  }
+  const marker = "\n\n\u2026(truncated)";
+  const report4 = text.length <= MAX_RUN_REPORT_CHARS ? text : `${text.slice(0, MAX_RUN_REPORT_CHARS - marker.length)}${marker}`;
+  const firstLine2 = text.split("\n").find((l) => l.trim())?.trim() ?? text;
+  const ellipsis = "\u2026";
+  const summary = firstLine2.length <= MAX_RUN_REPORT_SUMMARY_CHARS ? firstLine2 : `${firstLine2.slice(0, MAX_RUN_REPORT_SUMMARY_CHARS - ellipsis.length)}${ellipsis}`;
+  return { report: report4, summary };
+}
 async function finalizeJob(db, shipId, job, input) {
   const shipRef = doc6(db, COLLECTIONS.ships, shipId);
   const jobRef = doc6(shipRef, COLLECTIONS.jobs, job.id);
@@ -2452,6 +2502,7 @@ async function finalizeJob(db, shipId, job, input) {
   const usageRef = doc6(shipRef, COLLECTIONS.usageDaily, utcDay(now));
   await runTransaction(db, async (tx) => {
     const usageSnap = await tx.get(usageRef);
+    const jobSnap = await tx.get(jobRef);
     const totals = { ...EMPTY_USAGE_TOTALS, ...usageSnap.data()?.totals ?? {} };
     const byAgent = usageSnap.data()?.byAgent ?? {};
     const agentAgg = {
@@ -2459,7 +2510,10 @@ async function finalizeJob(db, shipId, job, input) {
       ...byAgent[job.agentId] ?? {}
     };
     const u = input.usage;
+    const wroteReport = !!jobSnap.data()?.report;
+    const backstop = !wroteReport && input.resultText !== void 0 ? backstopReportContent(input.resultText) : null;
     tx.update(jobRef, {
+      ...backstop ? { report: backstop.report, reportSummary: backstop.summary } : {},
       status: input.status,
       endedAt: now,
       usage: u,
@@ -4036,6 +4090,10 @@ async function startDaemon() {
           status: "stopped",
           usage,
           transcriptPath,
+          // §15.41. A stopped run is the case that needs the backstop MOST: it was cut off
+          // mid-thought, so it almost certainly never reached `run_report` — and whatever it had
+          // got to is what the next run on this task would otherwise have to rediscover.
+          resultText,
           mcpServers: extraMcpServers.map((s) => s.key)
         });
         const by = slot.stop.by;
@@ -4077,6 +4135,9 @@ async function startDaemon() {
           status: "done",
           usage,
           transcriptPath,
+          // §15.41. Only used when the session never called `run_report` — the ordinary path is
+          // that it did, and a real report always wins inside the transaction.
+          resultText,
           mcpServers: extraMcpServers.map((s) => s.key)
         });
         log2(`Job ${job.id} done (${usage.inputTokens}in/${usage.outputTokens}out tokens).`);
@@ -4130,6 +4191,10 @@ async function startDaemon() {
           usage,
           transcriptPath,
           error: failure,
+          // §15.41. A failed run still did work, and the retry — or the next run after the retry
+          // is spent — starts from the context pack alone. `error` is the tail for a human; this
+          // is the continuity for the next session, and they are read by different readers.
+          resultText,
           mcpServers: extraMcpServers.map((s) => s.key)
         });
         if (target.kind === "chat") {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lumi.ai/runner",
-  "version": "0.6.3",
+  "version": "0.6.4",
   "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.",