@rulemetric/cli 0.7.3 → 0.7.4

package/dist/chunk-RAXRZYOB.js

@@ -1,299 +0,0 @@
-import {
-  startTerminalTargetHeartbeat
-} from "./chunk-GIUBARWS.js";
-import {
-  processRunCleanup
-} from "./chunk-6NBS5XFS.js";
-import {
-  processSendMessage
-} from "./chunk-W53GKIZQ.js";
-import {
-  processSessionGoal
-} from "./chunk-OCKHW72U.js";
-import {
-  cronRefreshSkills
-} from "./chunk-RQ2TMLKG.js";
-import {
-  cronSuggestInstructions
-} from "./chunk-DMX5YOTM.js";
-import {
-  processAnnouncement
-} from "./chunk-OO7JDFS4.js";
-import {
-  processConversation
-} from "./chunk-XK42ZBDE.js";
-import {
-  processEval
-} from "./chunk-RMTT4KDY.js";
-import {
-  processInsights
-} from "./chunk-VA5FFOS7.js";
-import {
-  processLaunch
-} from "./chunk-JULOLTZS.js";
-import {
-  isPermanentJobError
-} from "./chunk-DGHWRQXL.js";
-import {
-  ApiError,
-  apiGet,
-  apiPost
-} from "./chunk-ZUJBM6HE.js";
-
-// src/lib/agent-loop.ts
-var HANDLERS = {
-  process_eval: processEval,
-  process_insights: processInsights,
-  process_launch: processLaunch,
-  process_send_message: processSendMessage,
-  process_conversation: processConversation,
-  // agent_jobs kinds (Phase 4c Stage 3b, migration 00083). processSessionGoal
-  // and processAnnouncement take an optional deps object in the helpers
-  // position — every key is optional, so JobHelpers passing through is
-  // harmless and they fall back to their defaults.
-  process_session_goal: processSessionGoal,
-  process_announcement: processAnnouncement,
-  // Multi-session orchestration ("runs") worktree cleanup (Task 5.2). agent_jobs
-  // kind — no mirror table. Enqueued by POST /api/runs/:id/cleanup; prunes the
-  // worktrees of a run's ended sessions, guarding against uncommitted work.
-  process_run_cleanup: processRunCleanup,
-  cron_suggest_instructions: cronSuggestInstructions,
-  // Dropped by accident in the Phase 4c graphile→agent_jobs migration: the
-  // handler survived but lost its registration + schedule, so the skills
-  // registry froze (last refresh 2026-05-21) and the suggest cron marked
-  // every suggestion stale on each run. Re-scheduled by migration 00086.
-  cron_refresh_skills: cronRefreshSkills
-};
-var CAPABILITIES = Object.keys(HANDLERS);
-var DB_REQUIRING_KINDS = /* @__PURE__ */ new Set([
-  "process_session_goal",
-  "process_announcement",
-  "process_run_cleanup"
-]);
-function availableCapabilities(env = process.env) {
-  if (env.DATABASE_URL) return CAPABILITIES;
-  return CAPABILITIES.filter((kind) => !DB_REQUIRING_KINDS.has(kind));
-}
-var DEFAULT_POLL_WAIT_MS = 3e4;
-var DEFAULT_HEARTBEAT_MS = 3e4;
-var PRESENCE_HEARTBEAT_MS = 3e4;
-var BACKOFF_INITIAL_MS = 250;
-var BACKOFF_MAX_MS = 5e3;
-var PRESENCE_CLIENT_NAME = `thin-agent-${process.pid}`;
-function startPresenceHeartbeat(capabilities, signal, log, version) {
-  const send = (status) => {
-    void apiPost("/api/workers/heartbeat", {
-      workerType: "cli-worker",
-      clientName: PRESENCE_CLIENT_NAME,
-      clientVersion: version,
-      workerVersion: "0.1.0",
-      currentJobId: null,
-      inFlightCount: 0,
-      lastError: null,
-      status,
-      metadata: { capabilities, mode: "thin-agent" }
-    }).catch((err) => {
-      structuredLog(log, {
-        event: "presence_heartbeat_error",
-        error: err instanceof Error ? err.message : String(err)
-      });
-    });
-  };
-  send("alive");
-  const timer = setInterval(() => send("alive"), PRESENCE_HEARTBEAT_MS);
-  signal.addEventListener("abort", () => {
-    clearInterval(timer);
-    send("shutting_down");
-  }, { once: true });
-}
-function structuredLog(log, event) {
-  try {
-    log(JSON.stringify({ ts: (/* @__PURE__ */ new Date()).toISOString(), ...event }));
-  } catch {
-    log(`[agent-loop] log_error event=${event["event"]}`);
-  }
-}
-function abortableSleep(ms, signal) {
-  if (signal.aborted) return Promise.resolve(false);
-  if (ms <= 0) return Promise.resolve(true);
-  return new Promise((resolve) => {
-    const timer = setTimeout(() => {
-      signal.removeEventListener("abort", onAbort);
-      resolve(true);
-    }, ms);
-    const onAbort = () => {
-      clearTimeout(timer);
-      signal.removeEventListener("abort", onAbort);
-      resolve(false);
-    };
-    signal.addEventListener("abort", onAbort, { once: true });
-  });
-}
-async function pollForWork(capabilities, waitMs) {
-  const qs = `capabilities=${encodeURIComponent(capabilities.join(","))}&waitMs=${waitMs}`;
-  const result = await apiGet(`/api/work/next?${qs}`);
-  if (result == null || typeof result === "string") return null;
-  return result;
-}
-function makeHelpers(attempts, log) {
-  return {
-    logger: {
-      info: (msg) => log(msg),
-      warn: (msg) => log(msg),
-      error: (msg) => log(msg)
-    },
-    job: { attempts }
-  };
-}
-async function runOne(job, heartbeatIntervalMs, log) {
-  const handler = HANDLERS[job.kind];
-  const startedAt = Date.now();
-  structuredLog(log, {
-    event: "claim",
-    jobId: job.id,
-    kind: job.kind,
-    attempts: job.attempts
-  });
-  if (!handler) {
-    const msg = `unknown kind: ${job.kind}`;
-    structuredLog(log, {
-      event: "fail",
-      jobId: job.id,
-      kind: job.kind,
-      terminal: true,
-      error: msg
-    });
-    await apiPost(`/api/work/${job.id}/fail`, {
-      leaseToken: job.leaseToken,
-      error: msg,
-      terminal: true
-    }).catch(() => {
-    });
-    return;
-  }
-  const heartbeatTimer = setInterval(() => {
-    structuredLog(log, { event: "heartbeat", jobId: job.id });
-    apiPost(`/api/work/${job.id}/heartbeat`, {
-      leaseToken: job.leaseToken
-    }).catch((err) => {
-      structuredLog(log, {
-        event: "heartbeat_error",
-        jobId: job.id,
-        error: err instanceof Error ? err.message : String(err)
-      });
-    });
-  }, heartbeatIntervalMs);
-  try {
-    const helpers = makeHelpers(job.attempts, log);
-    const result = await handler(job.payload, helpers) ?? {};
-    clearInterval(heartbeatTimer);
-    structuredLog(log, {
-      event: "complete",
-      jobId: job.id,
-      kind: job.kind,
-      durationMs: Date.now() - startedAt
-    });
-    await apiPost(`/api/work/${job.id}/complete`, {
-      leaseToken: job.leaseToken,
-      result
-    });
-  } catch (err) {
-    clearInterval(heartbeatTimer);
-    const terminal = isPermanentJobError(err);
-    const message = err instanceof Error ? err.message : String(err);
-    structuredLog(log, {
-      event: "fail",
-      jobId: job.id,
-      kind: job.kind,
-      durationMs: Date.now() - startedAt,
-      terminal,
-      error: message
-    });
-    await apiPost(`/api/work/${job.id}/fail`, {
-      leaseToken: job.leaseToken,
-      error: message,
-      terminal
-    }).catch((postErr) => {
-      structuredLog(log, {
-        event: "fail_post_error",
-        jobId: job.id,
-        error: postErr instanceof Error ? postErr.message : String(postErr)
-      });
-    });
-  }
-}
-async function runAgentLoop(opts) {
-  const log = opts.log ?? (() => void 0);
-  const pollWaitMs = opts.pollWaitMs ?? DEFAULT_POLL_WAIT_MS;
-  const heartbeatIntervalMs = opts.heartbeatIntervalMs ?? DEFAULT_HEARTBEAT_MS;
-  const capabilities = opts.capabilities ?? availableCapabilities();
-  let backoffMs = BACKOFF_INITIAL_MS;
-  structuredLog(log, {
-    event: "agent_start",
-    capabilities,
-    pollWaitMs,
-    heartbeatIntervalMs
-  });
-  startPresenceHeartbeat(capabilities, opts.signal, log, opts.version);
-  void startTerminalTargetHeartbeat({ signal: opts.signal, log }).catch((err) => {
-    structuredLog(log, {
-      event: "tt_heartbeat_fatal",
-      error: err instanceof Error ? err.message : String(err)
-    });
-  });
-  const VERSION_CHECK_INTERVAL_MS = 5 * 6e4;
-  let lastVersionCheck = 0;
-  const versionDrifted = () => {
-    if (!opts.readInstalledVersion || !opts.version) return false;
-    const now = Date.now();
-    if (now - lastVersionCheck < VERSION_CHECK_INTERVAL_MS) return false;
-    lastVersionCheck = now;
-    const onDisk = opts.readInstalledVersion();
-    return Boolean(onDisk && onDisk !== opts.version);
-  };
-  while (!opts.signal.aborted) {
-    if (versionDrifted()) {
-      structuredLog(log, {
-        event: "agent_version_drift",
-        bootVersion: opts.version,
-        installedVersion: opts.readInstalledVersion?.() ?? null
-      });
-      opts.onVersionDrift?.();
-      break;
-    }
-    let job = null;
-    try {
-      job = await pollForWork(capabilities, pollWaitMs);
-    } catch (err) {
-      const status = err instanceof ApiError ? err.statusCode : void 0;
-      structuredLog(log, {
-        event: "poll_error",
-        status,
-        error: err instanceof Error ? err.message : String(err)
-      });
-      const completed = await abortableSleep(backoffMs, opts.signal);
-      if (!completed) break;
-      backoffMs = Math.min(backoffMs * 2, BACKOFF_MAX_MS);
-      continue;
-    }
-    if (!job) {
-      structuredLog(log, { event: "poll_204" });
-      const completed = await abortableSleep(backoffMs, opts.signal);
-      if (!completed) break;
-      backoffMs = Math.min(backoffMs * 2, BACKOFF_MAX_MS);
-      continue;
-    }
-    backoffMs = BACKOFF_INITIAL_MS;
-    await runOne(job, heartbeatIntervalMs, log);
-  }
-  structuredLog(log, { event: "agent_stop" });
-}
-
-export {
-  HANDLERS,
-  CAPABILITIES,
-  availableCapabilities,
-  pollForWork,
-  runAgentLoop
-};
-//# sourceMappingURL=chunk-RAXRZYOB.js.map

package/dist/chunk-RAXRZYOB.js.map

@@ -1,7 +0,0 @@
-{
-  "version": 3,
-  "sources": ["../src/lib/agent-loop.ts"],
-  "sourcesContent": ["// Phase 4b thin-agent migration: the parallel-run worker loop that replaces\n// graphile-worker's LISTEN/NOTIFY pickup with HTTP long-polling against the\n// Phase 4a /api/work endpoints. See docs/plans/2026-05-25-thin-agent-migration.md.\n//\n// This file is invoked exclusively by `rulemetric evals agent` \u2014 the only\n// worker mode since Phase 4c deleted the graphile path.\n//\n// Lifecycle for one cycle:\n//   1. Long-poll GET /api/work/next?capabilities=...&waitMs=...\n//   2. If 204 (returned as \"\" by api-client because no JSON body), increment\n//      a small backoff capped at 5s and re-poll.\n//   3. If 200, dispatch by `kind` to the matching handler from HANDLERS.\n//   4. Start a heartbeat timer that POSTs /api/work/:id/heartbeat at\n//      heartbeatIntervalMs while the handler runs.\n//   5. On success: POST /api/work/:id/complete with { result, leaseToken }.\n//      On PermanentJobError: POST /fail with terminal=true (server will not\n//      requeue regardless of attempts). On any other error: POST /fail with\n//      terminal=false (server requeues if attempts < max).\n//\n// Abort behavior: an AbortSignal short-circuits the next long-poll and any\n// backoff sleep. An in-flight handler can't be cancelled (it has already\n// spawned `claude` subprocesses), so we await it to natural completion and\n// then exit. This matches listen.ts's 10s shutdown grace pattern, except we\n// don't need an outer hard-kill timer here \u2014 the caller (the oclif command)\n// owns the process.exit.\n\nimport type { JobHelpers } from './job-helpers.js';\nimport { apiGet, apiPost, ApiError } from './api-client.js';\nimport { isPermanentJobError } from './permanent-job-error.js';\nimport { processEval } from './handlers/process-eval.js';\nimport { processInsights } from './handlers/process-insights.js';\nimport { processLaunch } from './handlers/process-launch.js';\nimport { processConversation } from './handlers/process-conversation.js';\nimport { processSendMessage } from './handlers/process-send-message.js';\nimport { processSessionGoal } from './handlers/process-session-goal.js';\nimport { processAnnouncement } from './handlers/process-announcement.js';\nimport { processRunCleanup } from './handlers/process-run-cleanup.js';\nimport { cronSuggestInstructions } from './handlers/cron-suggest-instructions.js';\nimport { cronRefreshSkills } from './handlers/cron-refresh-skills.js';\nimport { startTerminalTargetHeartbeat } from './terminal-target-heartbeat.js';\n\n// Handler dispatch table. Keys MUST match the API's CAPABILITY_TO_TABLE keys\n// in apps/api/src/routes/work.ts; mismatches would cause the API to silently\n// drop the capability and starve this kind of work.\ntype Handler = (\n  payload: Record<string, unknown>,\n  helpers: JobHelpers,\n) => Promise<Record<string, unknown> | undefined | void>;\n\nexport const HANDLERS: Record<string, Handler> = {\n  process_eval: processEval as unknown as Handler,\n  process_insights: processInsights as unknown as Handler,\n  process_launch: processLaunch as unknown as Handler,\n  process_send_message: processSendMessage as unknown as Handler,\n  process_conversation: processConversation as unknown as Handler,\n  // agent_jobs kinds (Phase 4c Stage 3b, migration 00083). processSessionGoal\n  // and processAnnouncement take an optional deps object in the helpers\n  // position \u2014 every key is optional, so JobHelpers passing through is\n  // harmless and they fall back to their defaults.\n  process_session_goal: processSessionGoal as unknown as Handler,\n  process_announcement: processAnnouncement as unknown as Handler,\n  // Multi-session orchestration (\"runs\") worktree cleanup (Task 5.2). agent_jobs\n  // kind \u2014 no mirror table. Enqueued by POST /api/runs/:id/cleanup; prunes the\n  // worktrees of a run's ended sessions, guarding against uncommitted work.\n  process_run_cleanup: processRunCleanup as unknown as Handler,\n  cron_suggest_instructions: cronSuggestInstructions as unknown as Handler,\n  // Dropped by accident in the Phase 4c graphile\u2192agent_jobs migration: the\n  // handler survived but lost its registration + schedule, so the skills\n  // registry froze (last refresh 2026-05-21) and the suggest cron marked\n  // every suggestion stale on each run. Re-scheduled by migration 00086.\n  cron_refresh_skills: cronRefreshSkills as unknown as Handler,\n};\n\nexport const CAPABILITIES: readonly string[] = Object.keys(HANDLERS);\n\n// Kinds whose handlers open a DIRECT Postgres connection (getDb \u2192 requires\n// DATABASE_URL): process-session-goal, process-announcement, process-run-cleanup.\n// A worker-only npm install (the onboarding default) has API auth but no DB\n// creds \u2014 if it advertises these kinds it claims jobs it can only throw on, and\n// process_session_goal has MAX_ATTEMPTS=1, so a single claim PERMANENTLY fails\n// the job (goalSummary dead for every npm user). Advertise them only when a\n// direct DB connection is actually configured; otherwise the rows stay pending\n// for a repo-checkout worker instead of being burned.\nconst DB_REQUIRING_KINDS: ReadonlySet<string> = new Set([\n  'process_session_goal',\n  'process_announcement',\n  'process_run_cleanup',\n]);\n\n/** The capability list this process can actually serve (see DB_REQUIRING_KINDS). */\nexport function availableCapabilities(\n  env: NodeJS.ProcessEnv = process.env,\n): readonly string[] {\n  if (env.DATABASE_URL) return CAPABILITIES;\n  return CAPABILITIES.filter((kind) => !DB_REQUIRING_KINDS.has(kind));\n}\n\ninterface NextWorkResponse {\n  id: string;\n  kind: string;\n  payload: Record<string, unknown>;\n  attempts: number;\n  leaseToken: string;\n  leaseDurationSeconds: number;\n  table: string;\n}\n\nexport interface RunAgentLoopOptions {\n  signal: AbortSignal;\n  log?: (msg: string) => void;\n  /** Long-poll wait passed to GET /api/work/next?waitMs= */\n  pollWaitMs?: number;\n  /** How often to POST /api/work/:id/heartbeat while a handler runs */\n  heartbeatIntervalMs?: number;\n  /** Capability list to advertise. Defaults to all known handlers. */\n  capabilities?: readonly string[];\n  /** Running CLI version, reported as `clientVersion` in the presence heartbeat\n   *  so the dashboard can detect an outdated CLI and nudge an upgrade. */\n  version?: string;\n  /** Reads the CURRENTLY INSTALLED CLI version from disk (vs `version`, which\n   *  is frozen at boot). When they diverge \u2014 `npm i -g` upgraded the CLI under\n   *  a running worker \u2014 the loop stops between jobs and calls `onVersionDrift`\n   *  so the supervisor can respawn onto the new code. */\n  readInstalledVersion?: () => string | null;\n  /** Invoked (idle, never mid-job) when version drift is detected. */\n  onVersionDrift?: () => void;\n}\n\nconst DEFAULT_POLL_WAIT_MS = 30_000;\nconst DEFAULT_HEARTBEAT_MS = 30_000;\nconst PRESENCE_HEARTBEAT_MS = 30_000;\nconst BACKOFF_INITIAL_MS = 250;\nconst BACKOFF_MAX_MS = 5_000;\n\n// Stable per-process identifier so worker_connections groups the same agent\n// across heartbeats (the unique key is (userId, workerType, clientName)).\nconst PRESENCE_CLIENT_NAME = `thin-agent-${process.pid}`;\n\n/**\n * Background presence heartbeat. Posts to /api/workers/heartbeat every 30s so\n * worker_connections has a fresh row \u2014 that's what the dashboard's\n * useWorkerStatus reads to decide \"Worker online\" for live-send gating.\n * Distinct from the per-job heartbeat in runOne, which only extends a\n * specific job's lease. The graphile-worker runner had this since day one;\n * the thin agent was missing it, so the dashboard showed \"Worker offline\"\n * even when the agent was actively processing jobs.\n */\nfunction startPresenceHeartbeat(\n  capabilities: readonly string[],\n  signal: AbortSignal,\n  log: (msg: string) => void,\n  version: string | undefined,\n): void {\n  const send = (status: 'alive' | 'shutting_down') => {\n    void apiPost('/api/workers/heartbeat', {\n      workerType: 'cli-worker',\n      clientName: PRESENCE_CLIENT_NAME,\n      clientVersion: version,\n      workerVersion: '0.1.0',\n      currentJobId: null,\n      inFlightCount: 0,\n      lastError: null,\n      status,\n      metadata: { capabilities, mode: 'thin-agent' },\n    }).catch((err) => {\n      structuredLog(log, {\n        event: 'presence_heartbeat_error',\n        error: err instanceof Error ? err.message : String(err),\n      });\n    });\n  };\n  send('alive');\n  const timer = setInterval(() => send('alive'), PRESENCE_HEARTBEAT_MS);\n  signal.addEventListener('abort', () => {\n    clearInterval(timer);\n    send('shutting_down');\n  }, { once: true });\n}\n\nfunction structuredLog(\n  log: (msg: string) => void,\n  event: Record<string, unknown>,\n): void {\n  try {\n    log(JSON.stringify({ ts: new Date().toISOString(), ...event }));\n  } catch {\n    // JSON.stringify can throw on a circular structure; fall back to a\n    // best-effort string so a log failure never takes down the loop.\n    log(`[agent-loop] log_error event=${event['event']}`);\n  }\n}\n\n/** Sleep that resolves early when the abort signal fires. Returns true if\n * the full duration elapsed, false if aborted. */\nfunction abortableSleep(ms: number, signal: AbortSignal): Promise<boolean> {\n  if (signal.aborted) return Promise.resolve(false);\n  if (ms <= 0) return Promise.resolve(true);\n  return new Promise<boolean>((resolve) => {\n    const timer = setTimeout(() => {\n      signal.removeEventListener('abort', onAbort);\n      resolve(true);\n    }, ms);\n    const onAbort = () => {\n      clearTimeout(timer);\n      signal.removeEventListener('abort', onAbort);\n      resolve(false);\n    };\n    signal.addEventListener('abort', onAbort, { once: true });\n  });\n}\n\n/** Poll once for work. Returns the parsed response or null on 204. api-client\n * returns \"\" for an empty response body (Content-Type is not JSON on 204), so\n * we use that as the 204 sentinel. */\nexport async function pollForWork(\n  capabilities: readonly string[],\n  waitMs: number,\n): Promise<NextWorkResponse | null> {\n  const qs = `capabilities=${encodeURIComponent(capabilities.join(','))}&waitMs=${waitMs}`;\n  const result = await apiGet<NextWorkResponse | string>(`/api/work/next?${qs}`);\n  if (result == null || typeof result === 'string') return null;\n  return result;\n}\n\n/** Construct a minimal JobHelpers shim. Handlers only touch\n * helpers.logger.{info,warn,error}(msg) and helpers.job.attempts \u2014 verified\n * by grep against apps/cli/src/lib/handlers/. We cast through `unknown` so\n * TS accepts the structurally smaller object as the full JobHelpers type;\n * pulling graphile-worker into this hot path would defeat the Phase 4c\n * cleanup that removes the dep entirely. */\nfunction makeHelpers(attempts: number, log: (msg: string) => void): JobHelpers {\n  return {\n    logger: {\n      info: (msg: string) => log(msg),\n      warn: (msg: string) => log(msg),\n      error: (msg: string) => log(msg),\n    },\n    job: { attempts },\n  };\n}\n\nasync function runOne(\n  job: NextWorkResponse,\n  heartbeatIntervalMs: number,\n  log: (msg: string) => void,\n): Promise<void> {\n  const handler = HANDLERS[job.kind];\n  const startedAt = Date.now();\n  structuredLog(log, {\n    event: 'claim',\n    jobId: job.id,\n    kind: job.kind,\n    attempts: job.attempts,\n  });\n\n  // Defensive: if the API returned a kind the agent doesn't know about\n  // (capability filter mismatch, schema drift), fail terminally so the row\n  // doesn't bounce forever. The /api/work/next filter should already prevent\n  // this, but failing closed at the dispatcher means a single new kind on\n  // the API side can't wedge an older agent.\n  if (!handler) {\n    const msg = `unknown kind: ${job.kind}`;\n    structuredLog(log, {\n      event: 'fail',\n      jobId: job.id,\n      kind: job.kind,\n      terminal: true,\n      error: msg,\n    });\n    await apiPost(`/api/work/${job.id}/fail`, {\n      leaseToken: job.leaseToken,\n      error: msg,\n      terminal: true,\n    }).catch(() => {\n      // best-effort: the lease will expire and the reconciler reclaims it\n    });\n    return;\n  }\n\n  // Heartbeat timer. Starts after the first heartbeatIntervalMs elapses so\n  // we don't spam an extra request for sub-second jobs. Errors are swallowed\n  // \u2014 a single failed heartbeat doesn't abort the run; the server's\n  // 5-minute lease gives us up to ten chances at 30s intervals before we\n  // lose the claim.\n  const heartbeatTimer = setInterval(() => {\n    structuredLog(log, { event: 'heartbeat', jobId: job.id });\n    apiPost(`/api/work/${job.id}/heartbeat`, {\n      leaseToken: job.leaseToken,\n    }).catch((err) => {\n      structuredLog(log, {\n        event: 'heartbeat_error',\n        jobId: job.id,\n        error: err instanceof Error ? err.message : String(err),\n      });\n    });\n  }, heartbeatIntervalMs);\n\n  try {\n    const helpers = makeHelpers(job.attempts, log);\n    const result = (await handler(job.payload, helpers)) ?? {};\n    clearInterval(heartbeatTimer);\n    structuredLog(log, {\n      event: 'complete',\n      jobId: job.id,\n      kind: job.kind,\n      durationMs: Date.now() - startedAt,\n    });\n    await apiPost(`/api/work/${job.id}/complete`, {\n      leaseToken: job.leaseToken,\n      result,\n    });\n  } catch (err) {\n    clearInterval(heartbeatTimer);\n    const terminal = isPermanentJobError(err);\n    const message = err instanceof Error ? err.message : String(err);\n    structuredLog(log, {\n      event: 'fail',\n      jobId: job.id,\n      kind: job.kind,\n      durationMs: Date.now() - startedAt,\n      terminal,\n      error: message,\n    });\n    await apiPost(`/api/work/${job.id}/fail`, {\n      leaseToken: job.leaseToken,\n      error: message,\n      terminal,\n    }).catch((postErr) => {\n      structuredLog(log, {\n        event: 'fail_post_error',\n        jobId: job.id,\n        error: postErr instanceof Error ? postErr.message : String(postErr),\n      });\n    });\n  }\n}\n\nexport async function runAgentLoop(opts: RunAgentLoopOptions): Promise<void> {\n  const log = opts.log ?? (() => undefined);\n  const pollWaitMs = opts.pollWaitMs ?? DEFAULT_POLL_WAIT_MS;\n  const heartbeatIntervalMs = opts.heartbeatIntervalMs ?? DEFAULT_HEARTBEAT_MS;\n  const capabilities = opts.capabilities ?? availableCapabilities();\n\n  let backoffMs = BACKOFF_INITIAL_MS;\n  structuredLog(log, {\n    event: 'agent_start',\n    capabilities,\n    pollWaitMs,\n    heartbeatIntervalMs,\n  });\n\n  // Background presence heartbeat \u2014 writes to worker_connections so the\n  // dashboard's \"Worker online\" indicator turns on. Fire-and-forget; failures\n  // are logged but don't block the work poll.\n  startPresenceHeartbeat(capabilities, opts.signal, log, opts.version);\n\n  // Fire-and-forget the terminal-target liveness heartbeat. It runs on its\n  // own interval (default 60s) and shuts down on the same abort signal.\n  // Errors are caught internally \u2014 we don't await it so a slow tmux probe\n  // can't delay the work poll.\n  void startTerminalTargetHeartbeat({ signal: opts.signal, log }).catch((err) => {\n    structuredLog(log, {\n      event: 'tt_heartbeat_fatal',\n      error: err instanceof Error ? err.message : String(err),\n    });\n  });\n\n  // Version-drift self-restart: after `npm i -g` upgrades the CLI, this\n  // long-lived process keeps executing the OLD code forever \u2014 launchd bounces\n  // on dist changes only in the repo-checkout install, and the Linux systemd\n  // unit has no WatchPaths equivalent at all. Between jobs (never mid-job),\n  // compare the on-disk installed version against the one this process booted\n  // with; on drift, hand control to the caller (which exits non-zero so BOTH\n  // supervisors respawn onto the new code: systemd is Restart=on-failure,\n  // launchd KeepAlive restarts regardless). Throttled \u2014 a disk read per idle\n  // poll would be wasteful.\n  const VERSION_CHECK_INTERVAL_MS = 5 * 60_000;\n  let lastVersionCheck = 0;\n  const versionDrifted = (): boolean => {\n    if (!opts.readInstalledVersion || !opts.version) return false;\n    const now = Date.now();\n    if (now - lastVersionCheck < VERSION_CHECK_INTERVAL_MS) return false;\n    lastVersionCheck = now;\n    const onDisk = opts.readInstalledVersion();\n    return Boolean(onDisk && onDisk !== opts.version);\n  };\n\n  while (!opts.signal.aborted) {\n    if (versionDrifted()) {\n      structuredLog(log, {\n        event: 'agent_version_drift',\n        bootVersion: opts.version,\n        installedVersion: opts.readInstalledVersion?.() ?? null,\n      });\n      opts.onVersionDrift?.();\n      break;\n    }\n    let job: NextWorkResponse | null = null;\n    try {\n      job = await pollForWork(capabilities, pollWaitMs);\n    } catch (err) {\n      // Network/API errors during poll: log + backoff. The API may be\n      // restarting, the laptop may have just woken, etc. Don't crash the\n      // loop \u2014 that's the whole point of this migration.\n      const status = err instanceof ApiError ? err.statusCode : undefined;\n      structuredLog(log, {\n        event: 'poll_error',\n        status,\n        error: err instanceof Error ? err.message : String(err),\n      });\n      const completed = await abortableSleep(backoffMs, opts.signal);\n      if (!completed) break;\n      backoffMs = Math.min(backoffMs * 2, BACKOFF_MAX_MS);\n      continue;\n    }\n\n    if (!job) {\n      structuredLog(log, { event: 'poll_204' });\n      const completed = await abortableSleep(backoffMs, opts.signal);\n      if (!completed) break;\n      backoffMs = Math.min(backoffMs * 2, BACKOFF_MAX_MS);\n      continue;\n    }\n\n    // Got work \u2014 reset the backoff and run it. An in-flight handler is\n    // intentionally not abort-cancellable; we await its natural completion\n    // even if the signal fires mid-run.\n    backoffMs = BACKOFF_INITIAL_MS;\n    await runOne(job, heartbeatIntervalMs, log);\n  }\n\n  structuredLog(log, { event: 'agent_stop' });\n}\n"],
-  "mappings": ";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAiDO,IAAM,WAAoC;AAAA,EAC/C,cAAc;AAAA,EACd,kBAAkB;AAAA,EAClB,gBAAgB;AAAA,EAChB,sBAAsB;AAAA,EACtB,sBAAsB;AAAA;AAAA;AAAA;AAAA;AAAA,EAKtB,sBAAsB;AAAA,EACtB,sBAAsB;AAAA;AAAA;AAAA;AAAA,EAItB,qBAAqB;AAAA,EACrB,2BAA2B;AAAA;AAAA;AAAA;AAAA;AAAA,EAK3B,qBAAqB;AACvB;AAEO,IAAM,eAAkC,OAAO,KAAK,QAAQ;AAUnE,IAAM,qBAA0C,oBAAI,IAAI;AAAA,EACtD;AAAA,EACA;AAAA,EACA;AACF,CAAC;AAGM,SAAS,sBACd,MAAyB,QAAQ,KACd;AACnB,MAAI,IAAI,aAAc,QAAO;AAC7B,SAAO,aAAa,OAAO,CAAC,SAAS,CAAC,mBAAmB,IAAI,IAAI,CAAC;AACpE;AAiCA,IAAM,uBAAuB;AAC7B,IAAM,uBAAuB;AAC7B,IAAM,wBAAwB;AAC9B,IAAM,qBAAqB;AAC3B,IAAM,iBAAiB;AAIvB,IAAM,uBAAuB,cAAc,QAAQ,GAAG;AAWtD,SAAS,uBACP,cACA,QACA,KACA,SACM;AACN,QAAM,OAAO,CAAC,WAAsC;AAClD,SAAK,QAAQ,0BAA0B;AAAA,MACrC,YAAY;AAAA,MACZ,YAAY;AAAA,MACZ,eAAe;AAAA,MACf,eAAe;AAAA,MACf,cAAc;AAAA,MACd,eAAe;AAAA,MACf,WAAW;AAAA,MACX;AAAA,MACA,UAAU,EAAE,cAAc,MAAM,aAAa;AAAA,IAC/C,CAAC,EAAE,MAAM,CAAC,QAAQ;AAChB,oBAAc,KAAK;AAAA,QACjB,OAAO;AAAA,QACP,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,MACxD,CAAC;AAAA,IACH,CAAC;AAAA,EACH;AACA,OAAK,OAAO;AACZ,QAAM,QAAQ,YAAY,MAAM,KAAK,OAAO,GAAG,qBAAqB;AACpE,SAAO,iBAAiB,SAAS,MAAM;AACrC,kBAAc,KAAK;AACnB,SAAK,eAAe;AAAA,EACtB,GAAG,EAAE,MAAM,KAAK,CAAC;AACnB;AAEA,SAAS,cACP,KACA,OACM;AACN,MAAI;AACF,QAAI,KAAK,UAAU,EAAE,KAAI,oBAAI,KAAK,GAAE,YAAY,GAAG,GAAG,MAAM,CAAC,CAAC;AAAA,EAChE,QAAQ;AAGN,QAAI,gCAAgC,MAAM,OAAO,CAAC,EAAE;AAAA,EACtD;AACF;AAIA,SAAS,eAAe,IAAY,QAAuC;AACzE,MAAI,OAAO,QAAS,QAAO,QAAQ,QAAQ,KAAK;AAChD,MAAI,MAAM,EAAG,QAAO,QAAQ,QAAQ,IAAI;AACxC,SAAO,IAAI,QAAiB,CAAC,YAAY;AACvC,UAAM,QAAQ,WAAW,MAAM;AAC7B,aAAO,oBAAoB,SAAS,OAAO;AAC3C,cAAQ,IAAI;AAAA,IACd,GAAG,EAAE;AACL,UAAM,UAAU,MAAM;AACpB,mBAAa,KAAK;AAClB,aAAO,oBAAoB,SAAS,OAAO;AAC3C,cAAQ,KAAK;AAAA,IACf;AACA,WAAO,iBAAiB,SAAS,SAAS,EAAE,MAAM,KAAK,CAAC;AAAA,EAC1D,CAAC;AACH;AAKA,eAAsB,YACpB,cACA,QACkC;AAClC,QAAM,KAAK,gBAAgB,mBAAmB,aAAa,KAAK,GAAG,CAAC,CAAC,WAAW,MAAM;AACtF,QAAM,SAAS,MAAM,OAAkC,kBAAkB,EAAE,EAAE;AAC7E,MAAI,UAAU,QAAQ,OAAO,WAAW,SAAU,QAAO;AACzD,SAAO;AACT;AAQA,SAAS,YAAY,UAAkB,KAAwC;AAC7E,SAAO;AAAA,IACL,QAAQ;AAAA,MACN,MAAM,CAAC,QAAgB,IAAI,GAAG;AAAA,MAC9B,MAAM,CAAC,QAAgB,IAAI,GAAG;AAAA,MAC9B,OAAO,CAAC,QAAgB,IAAI,GAAG;AAAA,IACjC;AAAA,IACA,KAAK,EAAE,SAAS;AAAA,EAClB;AACF;AAEA,eAAe,OACb,KACA,qBACA,KACe;AACf,QAAM,UAAU,SAAS,IAAI,IAAI;AACjC,QAAM,YAAY,KAAK,IAAI;AAC3B,gBAAc,KAAK;AAAA,IACjB,OAAO;AAAA,IACP,OAAO,IAAI;AAAA,IACX,MAAM,IAAI;AAAA,IACV,UAAU,IAAI;AAAA,EAChB,CAAC;AAOD,MAAI,CAAC,SAAS;AACZ,UAAM,MAAM,iBAAiB,IAAI,IAAI;AACrC,kBAAc,KAAK;AAAA,MACjB,OAAO;AAAA,MACP,OAAO,IAAI;AAAA,MACX,MAAM,IAAI;AAAA,MACV,UAAU;AAAA,MACV,OAAO;AAAA,IACT,CAAC;AACD,UAAM,QAAQ,aAAa,IAAI,EAAE,SAAS;AAAA,MACxC,YAAY,IAAI;AAAA,MAChB,OAAO;AAAA,MACP,UAAU;AAAA,IACZ,CAAC,EAAE,MAAM,MAAM;AAAA,IAEf,CAAC;AACD;AAAA,EACF;AAOA,QAAM,iBAAiB,YAAY,MAAM;AACvC,kBAAc,KAAK,EAAE,OAAO,aAAa,OAAO,IAAI,GAAG,CAAC;AACxD,YAAQ,aAAa,IAAI,EAAE,cAAc;AAAA,MACvC,YAAY,IAAI;AAAA,IAClB,CAAC,EAAE,MAAM,CAAC,QAAQ;AAChB,oBAAc,KAAK;AAAA,QACjB,OAAO;AAAA,QACP,OAAO,IAAI;AAAA,QACX,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,MACxD,CAAC;AAAA,IACH,CAAC;AAAA,EACH,GAAG,mBAAmB;AAEtB,MAAI;AACF,UAAM,UAAU,YAAY,IAAI,UAAU,GAAG;AAC7C,UAAM,SAAU,MAAM,QAAQ,IAAI,SAAS,OAAO,KAAM,CAAC;AACzD,kBAAc,cAAc;AAC5B,kBAAc,KAAK;AAAA,MACjB,OAAO;AAAA,MACP,OAAO,IAAI;AAAA,MACX,MAAM,IAAI;AAAA,MACV,YAAY,KAAK,IAAI,IAAI;AAAA,IAC3B,CAAC;AACD,UAAM,QAAQ,aAAa,IAAI,EAAE,aAAa;AAAA,MAC5C,YAAY,IAAI;AAAA,MAChB;AAAA,IACF,CAAC;AAAA,EACH,SAAS,KAAK;AACZ,kBAAc,cAAc;AAC5B,UAAM,WAAW,oBAAoB,GAAG;AACxC,UAAM,UAAU,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC/D,kBAAc,KAAK;AAAA,MACjB,OAAO;AAAA,MACP,OAAO,IAAI;AAAA,MACX,MAAM,IAAI;AAAA,MACV,YAAY,KAAK,IAAI,IAAI;AAAA,MACzB;AAAA,MACA,OAAO;AAAA,IACT,CAAC;AACD,UAAM,QAAQ,aAAa,IAAI,EAAE,SAAS;AAAA,MACxC,YAAY,IAAI;AAAA,MAChB,OAAO;AAAA,MACP;AAAA,IACF,CAAC,EAAE,MAAM,CAAC,YAAY;AACpB,oBAAc,KAAK;AAAA,QACjB,OAAO;AAAA,QACP,OAAO,IAAI;AAAA,QACX,OAAO,mBAAmB,QAAQ,QAAQ,UAAU,OAAO,OAAO;AAAA,MACpE,CAAC;AAAA,IACH,CAAC;AAAA,EACH;AACF;AAEA,eAAsB,aAAa,MAA0C;AAC3E,QAAM,MAAM,KAAK,QAAQ,MAAM;AAC/B,QAAM,aAAa,KAAK,cAAc;AACtC,QAAM,sBAAsB,KAAK,uBAAuB;AACxD,QAAM,eAAe,KAAK,gBAAgB,sBAAsB;AAEhE,MAAI,YAAY;AAChB,gBAAc,KAAK;AAAA,IACjB,OAAO;AAAA,IACP;AAAA,IACA;AAAA,IACA;AAAA,EACF,CAAC;AAKD,yBAAuB,cAAc,KAAK,QAAQ,KAAK,KAAK,OAAO;AAMnE,OAAK,6BAA6B,EAAE,QAAQ,KAAK,QAAQ,IAAI,CAAC,EAAE,MAAM,CAAC,QAAQ;AAC7E,kBAAc,KAAK;AAAA,MACjB,OAAO;AAAA,MACP,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,IACxD,CAAC;AAAA,EACH,CAAC;AAWD,QAAM,4BAA4B,IAAI;AACtC,MAAI,mBAAmB;AACvB,QAAM,iBAAiB,MAAe;AACpC,QAAI,CAAC,KAAK,wBAAwB,CAAC,KAAK,QAAS,QAAO;AACxD,UAAM,MAAM,KAAK,IAAI;AACrB,QAAI,MAAM,mBAAmB,0BAA2B,QAAO;AAC/D,uBAAmB;AACnB,UAAM,SAAS,KAAK,qBAAqB;AACzC,WAAO,QAAQ,UAAU,WAAW,KAAK,OAAO;AAAA,EAClD;AAEA,SAAO,CAAC,KAAK,OAAO,SAAS;AAC3B,QAAI,eAAe,GAAG;AACpB,oBAAc,KAAK;AAAA,QACjB,OAAO;AAAA,QACP,aAAa,KAAK;AAAA,QAClB,kBAAkB,KAAK,uBAAuB,KAAK;AAAA,MACrD,CAAC;AACD,WAAK,iBAAiB;AACtB;AAAA,IACF;AACA,QAAI,MAA+B;AACnC,QAAI;AACF,YAAM,MAAM,YAAY,cAAc,UAAU;AAAA,IAClD,SAAS,KAAK;AAIZ,YAAM,SAAS,eAAe,WAAW,IAAI,aAAa;AAC1D,oBAAc,KAAK;AAAA,QACjB,OAAO;AAAA,QACP;AAAA,QACA,OAAO,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAAA,MACxD,CAAC;AACD,YAAM,YAAY,MAAM,eAAe,WAAW,KAAK,MAAM;AAC7D,UAAI,CAAC,UAAW;AAChB,kBAAY,KAAK,IAAI,YAAY,GAAG,cAAc;AAClD;AAAA,IACF;AAEA,QAAI,CAAC,KAAK;AACR,oBAAc,KAAK,EAAE,OAAO,WAAW,CAAC;AACxC,YAAM,YAAY,MAAM,eAAe,WAAW,KAAK,MAAM;AAC7D,UAAI,CAAC,UAAW;AAChB,kBAAY,KAAK,IAAI,YAAY,GAAG,cAAc;AAClD;AAAA,IACF;AAKA,gBAAY;AACZ,UAAM,OAAO,KAAK,qBAAqB,GAAG;AAAA,EAC5C;AAEA,gBAAc,KAAK,EAAE,OAAO,aAAa,CAAC;AAC5C;",
-  "names": []
-}

package/dist/chunk-RH7NHVSK.js

@@ -1,16 +0,0 @@
-// src/dashboard/launcher/tool-labels.ts
-var TOOL_LABELS = {
-  claude_code: "Claude Code",
-  codex: "Codex",
-  aider: "Aider",
-  generic: "Generic"
-};
-function toolLabel(tool) {
-  return TOOL_LABELS[tool] ?? tool;
-}
-
-export {
-  TOOL_LABELS,
-  toolLabel
-};
-//# sourceMappingURL=chunk-RH7NHVSK.js.map

package/dist/chunk-RH7NHVSK.js.map

@@ -1,7 +0,0 @@
-{
-  "version": 3,
-  "sources": ["../src/dashboard/launcher/tool-labels.ts"],
-  "sourcesContent": ["export const TOOL_LABELS: Record<string, string> = {\n  claude_code: 'Claude Code',\n  codex: 'Codex',\n  aider: 'Aider',\n  generic: 'Generic',\n};\n\nexport function toolLabel(tool: string): string {\n  return TOOL_LABELS[tool] ?? tool;\n}\n"],
-  "mappings": ";AAAO,IAAM,cAAsC;AAAA,EACjD,aAAa;AAAA,EACb,OAAO;AAAA,EACP,OAAO;AAAA,EACP,SAAS;AACX;AAEO,SAAS,UAAU,MAAsB;AAC9C,SAAO,YAAY,IAAI,KAAK;AAC9B;",
-  "names": []
-}

package/dist/chunk-RMTT4KDY.js

@@ -1,191 +0,0 @@
-import {
-  PermanentJobError
-} from "./chunk-DGHWRQXL.js";
-import {
-  judgeEvalQuality
-} from "./chunk-J7N3DLH6.js";
-import {
-  createEvalWorktree
-} from "./chunk-52WIYRZH.js";
-import {
-  gradeEvalOutput
-} from "./chunk-YNIH3KMU.js";
-import {
-  executeEval
-} from "./chunk-QN34TGD3.js";
-import {
-  apiGet,
-  apiPatch,
-  apiPost
-} from "./chunk-ZUJBM6HE.js";
-
-// src/lib/handlers/process-eval.ts
-async function processEval(payload, helpers) {
-  const { mirrorId, evalTargetId, config } = payload;
-  const log = (msg) => helpers.logger.info(msg);
-  if (helpers.job.attempts > 1) {
-    throw new PermanentJobError(
-      `process_eval is not safe to retry (attempt ${helpers.job.attempts}). Re-trigger the eval run from the UI.`
-    );
-  }
-  const target = await apiGet(`/api/eval-targets/${evalTargetId}`);
-  const evalCases = await apiGet(`/api/eval-targets/${evalTargetId}/evals`);
-  if (evalCases.length === 0) {
-    throw new PermanentJobError("No evals defined for target");
-  }
-  const runsPerEval = config.runsPerEval ?? 1;
-  const timeout = config.timeout ?? 12e4;
-  const iteration = config.iteration ?? 1;
-  const configurations = [
-    { name: "with_target", content: target.content }
-  ];
-  if (!config.skipBaseline) {
-    configurations.push({ name: "without_target" });
-  }
-  const workQueue = [];
-  for (const eval_ of evalCases) {
-    for (const cfg of configurations) {
-      for (let run = 1; run <= runsPerEval; run++) {
-        workQueue.push({ eval_, configuration: cfg.name, targetContent: cfg.content, runNumber: run });
-      }
-    }
-  }
-  const total = workQueue.length;
-  let completed = 0;
-  const concurrency = Math.min(config.parallel ?? 1, 4);
-  await apiPatch(`/api/eval-jobs/${mirrorId}/progress`, {
-    status: "running",
-    progress: { total, completed: 0, current: null, message: `Starting ${total} eval runs (concurrency: ${concurrency})...` }
-  });
-  const results = [];
-  async function processItem(item) {
-    const label = `${item.configuration} \xB7 ${item.eval_.name} \xB7 run#${item.runNumber}`;
-    log(`[${completed + 1}/${total}] ${label}`);
-    const worktree = concurrency > 1 ? createEvalWorktree() : null;
-    try {
-      const execResult = await executeEval({
-        prompt: item.eval_.prompt,
-        targetContent: item.targetContent,
-        workDir: worktree?.workDir,
-        timeout,
-        model: config.model
-      });
-      let grading = {};
-      if (!config.skipGrading && item.eval_.expectations.length > 0 && !execResult.error) {
-        const gradingResult = await gradeEvalOutput({
-          prompt: item.eval_.prompt,
-          output: execResult.output,
-          expectations: item.eval_.expectations,
-          model: config.graderModel ?? config.model,
-          timeout
-        });
-        grading = gradingResult;
-      }
-      return {
-        evalId: item.eval_.id,
-        evalTargetId: target.id,
-        evalTargetVersion: target.version,
-        configuration: item.configuration,
-        runNumber: item.runNumber,
-        iteration,
-        status: execResult.error ? "failed" : "completed",
-        outputs: execResult.error ? { error: execResult.error } : { transcript: execResult.output },
-        timing: { duration_ms: execResult.durationMs, total_tokens: execResult.totalTokens },
-        grading,
-        worktree: worktree?.isWorktree ? true : void 0
-      };
-    } finally {
-      worktree?.cleanup();
-    }
-  }
-  if (concurrency <= 1) {
-    for (const item of workQueue) {
-      const label = `${item.configuration} \xB7 ${item.eval_.name} \xB7 run#${item.runNumber}`;
-      await apiPatch(`/api/eval-jobs/${mirrorId}/progress`, {
-        status: "running",
-        progress: { total, completed, current: label, message: `Executing ${label}...` }
-      });
-      const runResult = await processItem(item);
-      await apiPost("/api/eval-runs", runResult);
-      results.push(runResult);
-      completed++;
-      await apiPatch(`/api/eval-jobs/${mirrorId}/progress`, {
-        status: "running",
-        progress: { total, completed, current: label, message: `Completed ${label}` }
-      });
-    }
-  } else {
-    log(`Running ${total} evals with concurrency=${concurrency} (worktree isolation)`);
-    for (let i = 0; i < workQueue.length; i += concurrency) {
-      const batch = workQueue.slice(i, i + concurrency);
-      const batchLabel = `batch ${Math.floor(i / concurrency) + 1}/${Math.ceil(workQueue.length / concurrency)}`;
-      await apiPatch(`/api/eval-jobs/${mirrorId}/progress`, {
-        status: "running",
-        progress: { total, completed, current: batchLabel, message: `Executing ${batchLabel} (${batch.length} parallel)...` }
-      });
-      const batchResults = await Promise.all(batch.map(processItem));
-      for (const runResult of batchResults) {
-        await apiPost("/api/eval-runs", runResult);
-        results.push(runResult);
-        completed++;
-      }
-      await apiPatch(`/api/eval-jobs/${mirrorId}/progress`, {
-        status: "running",
-        progress: { total, completed, current: batchLabel, message: `Completed ${batchLabel}` }
-      });
-    }
-  }
-  const summary = buildSummary(results);
-  let metaJudgement;
-  if (!config.skipGrading && results.some((r) => r.grading && Object.keys(r.grading).length > 0)) {
-    try {
-      log(`Running meta-judge for job ${mirrorId}...`);
-      const evalInputs = evalCases.map((e) => {
-        const matchingResults = results.filter((r) => r.evalId === e.id && r.configuration === "with_target");
-        const firstResult = matchingResults[0];
-        return {
-          evalName: e.name,
-          prompt: e.prompt,
-          expectations: e.expectations,
-          grading: firstResult?.grading ? firstResult.grading : null,
-          output: firstResult?.outputs?.transcript?.slice(0, 500)
-        };
-      });
-      metaJudgement = await judgeEvalQuality({
-        targetName: target.name,
-        evalResults: evalInputs,
-        model: config.graderModel ?? config.model,
-        timeout: 6e4
-      });
-      log(`Meta-judge: quality=${metaJudgement.quality_score}/5, issues=${metaJudgement.issues.length}`);
-    } catch (err) {
-      log(`Meta-judge failed (non-fatal): ${err instanceof Error ? err.message : String(err)}`);
-    }
-  }
-  const result = metaJudgement ? { ...summary, meta_judgement: metaJudgement } : summary;
-  log(`Job ${mirrorId} completed: ${completed}/${total} runs`);
-  await apiPatch(`/api/eval-jobs/${mirrorId}/state`, { result }).catch((err) => {
-    log(`result persist failed: ${err instanceof Error ? err.message : String(err)}`);
-  });
-  return result;
-}
-function buildSummary(results) {
-  const completed = results.filter((r) => r.status === "completed");
-  const failed = results.filter((r) => r.status === "failed");
-  const gradedRuns = completed.filter((r) => {
-    const grading = r.grading;
-    return grading && Object.keys(grading).length > 0;
-  });
-  const avgPassRate = gradedRuns.length > 0 ? gradedRuns.reduce((sum, r) => sum + (r.grading?.summary?.pass_rate ?? 0), 0) / gradedRuns.length : null;
-  return {
-    total: results.length,
-    completed: completed.length,
-    failed: failed.length,
-    avgPassRate
-  };
-}
-
-export {
-  processEval
-};
-//# sourceMappingURL=chunk-RMTT4KDY.js.map

package/dist/chunk-RMTT4KDY.js.map

@@ -1,7 +0,0 @@
-{
-  "version": 3,
-  "sources": ["../src/lib/handlers/process-eval.ts"],
-  "sourcesContent": ["import type { JobHelpers } from '../job-helpers.js';\nimport { apiGet, apiPost, apiPatch } from '../api-client.js';\nimport { executeEval } from '../eval-executor.js';\nimport { gradeEvalOutput, type GradingResult } from '../eval-grader.js';\nimport { judgeEvalQuality, type MetaJudgement } from '../eval-meta-judge.js';\nimport { createEvalWorktree } from '../eval-worktree.js';\nimport { PermanentJobError } from '../permanent-job-error.js';\n\ninterface EvalTarget {\n  id: string;\n  name: string;\n  type: string;\n  content: string;\n  version: number;\n}\n\ninterface Eval {\n  id: string;\n  name: string;\n  prompt: string;\n  expectedOutput: string;\n  expectations: string[];\n}\n\ninterface EvalPayload {\n  mirrorId: string;\n  evalTargetId: string;\n  config: {\n    parallel?: number;\n    runsPerEval?: number;\n    iteration?: number;\n    skipBaseline?: boolean;\n    skipGrading?: boolean;\n    model?: string;\n    graderModel?: string;\n    timeout?: number;\n  };\n}\n\n/**\n * Graphile-shaped handler for the `process_eval` task.\n *\n * Lifted verbatim from the legacy `eval-worker.ts:processJob` body, with the\n * queue plumbing stripped out:\n *   - no claim \u2014 Graphile already claimed the job before invoking us\n *   - no /complete POST \u2014 we return the result object (or throw)\n *   - no `processing` flag \u2014 that lived in the polling loop's closure\n *   - no conversation branch \u2014 split into its own handler (Task 14)\n *\n * Mid-job progress patches still hit the existing\n * `PATCH /api/eval-jobs/:id/progress` endpoint, just keyed by the mirror row's\n * id instead of the legacy queue id.\n */\nexport async function processEval(\n  payload: EvalPayload,\n  helpers: JobHelpers,\n): Promise<Record<string, unknown>> {\n  const { mirrorId, evalTargetId, config } = payload;\n  const log = (msg: string) => helpers.logger.info(msg);\n\n  // Fail-fast on retry. Each eval_run is INSERT'd as it completes, but\n  // there's no idempotency key tying a row to a specific job \u2014 re-running\n  // the whole batch would duplicate every previously-completed row in\n  // eval_runs. Until proper resume-from-progress-checkpoint lands, surface\n  // the failure to the user instead of silently doubling-up rows. See\n  // docs/audits/2026-05-07-worker-audit.md \"Handlers not idempotent\".\n  if (helpers.job.attempts > 1) {\n    throw new PermanentJobError(\n      `process_eval is not safe to retry (attempt ${helpers.job.attempts}). Re-trigger the eval run from the UI.`,\n    );\n  }\n\n  const target = await apiGet<EvalTarget>(`/api/eval-targets/${evalTargetId}`);\n  const evalCases = await apiGet<Eval[]>(`/api/eval-targets/${evalTargetId}/evals`);\n\n  if (evalCases.length === 0) {\n    // No retries \u2014 eval cases don't materialize on their own; the user has\n    // to add one via the UI/CLI. Skip Graphile's retry loop.\n    throw new PermanentJobError('No evals defined for target');\n  }\n\n  const runsPerEval = config.runsPerEval ?? 1;\n  const timeout = config.timeout ?? 120_000;\n  const iteration = config.iteration ?? 1;\n\n  const configurations: Array<{ name: string; content?: string }> = [\n    { name: 'with_target', content: target.content },\n  ];\n  if (!config.skipBaseline) {\n    configurations.push({ name: 'without_target' });\n  }\n\n  // Build work queue\n  const workQueue: Array<{ eval_: Eval; configuration: string; targetContent?: string; runNumber: number }> = [];\n  for (const eval_ of evalCases) {\n    for (const cfg of configurations) {\n      for (let run = 1; run <= runsPerEval; run++) {\n        workQueue.push({ eval_: eval_, configuration: cfg.name, targetContent: cfg.content, runNumber: run });\n      }\n    }\n  }\n\n  const total = workQueue.length;\n  let completed = 0;\n  const concurrency = Math.min(config.parallel ?? 1, 4); // cap at 4 parallel\n\n  await apiPatch(`/api/eval-jobs/${mirrorId}/progress`, {\n    status: 'running',\n    progress: { total, completed: 0, current: null, message: `Starting ${total} eval runs (concurrency: ${concurrency})...` },\n  });\n\n  const results: Record<string, unknown>[] = [];\n\n  // Process work items \u2014 sequential if concurrency=1, parallel otherwise\n  async function processItem(item: typeof workQueue[0]): Promise<Record<string, unknown>> {\n    const label = `${item.configuration} \u00B7 ${item.eval_.name} \u00B7 run#${item.runNumber}`;\n    log(`[${completed + 1}/${total}] ${label}`);\n\n    // Create isolated worktree for parallel execution\n    const worktree = concurrency > 1 ? createEvalWorktree() : null;\n\n    try {\n      const execResult = await executeEval({\n        prompt: item.eval_.prompt,\n        targetContent: item.targetContent,\n        workDir: worktree?.workDir,\n        timeout,\n        model: config.model,\n      });\n\n      let grading: Record<string, unknown> = {};\n      if (!config.skipGrading && item.eval_.expectations.length > 0 && !execResult.error) {\n        const gradingResult = await gradeEvalOutput({\n          prompt: item.eval_.prompt,\n          output: execResult.output,\n          expectations: item.eval_.expectations,\n          model: config.graderModel ?? config.model,\n          timeout,\n        });\n        grading = gradingResult as unknown as Record<string, unknown>;\n      }\n\n      return {\n        evalId: item.eval_.id,\n        evalTargetId: target.id,\n        evalTargetVersion: target.version,\n        configuration: item.configuration,\n        runNumber: item.runNumber,\n        iteration,\n        status: execResult.error ? 'failed' : 'completed',\n        outputs: execResult.error ? { error: execResult.error } : { transcript: execResult.output },\n        timing: { duration_ms: execResult.durationMs, total_tokens: execResult.totalTokens },\n        grading,\n        worktree: worktree?.isWorktree ? true : undefined,\n      };\n    } finally {\n      worktree?.cleanup();\n    }\n  }\n\n  if (concurrency <= 1) {\n    // Sequential execution (existing behavior)\n    for (const item of workQueue) {\n      const label = `${item.configuration} \u00B7 ${item.eval_.name} \u00B7 run#${item.runNumber}`;\n      await apiPatch(`/api/eval-jobs/${mirrorId}/progress`, {\n        status: 'running',\n        progress: { total, completed, current: label, message: `Executing ${label}...` },\n      });\n\n      const runResult = await processItem(item);\n      await apiPost('/api/eval-runs', runResult);\n      results.push(runResult);\n      completed++;\n\n      await apiPatch(`/api/eval-jobs/${mirrorId}/progress`, {\n        status: 'running',\n        progress: { total, completed, current: label, message: `Completed ${label}` },\n      });\n    }\n  } else {\n    // Parallel execution with worktree isolation\n    log(`Running ${total} evals with concurrency=${concurrency} (worktree isolation)`);\n    for (let i = 0; i < workQueue.length; i += concurrency) {\n      const batch = workQueue.slice(i, i + concurrency);\n      const batchLabel = `batch ${Math.floor(i / concurrency) + 1}/${Math.ceil(workQueue.length / concurrency)}`;\n\n      await apiPatch(`/api/eval-jobs/${mirrorId}/progress`, {\n        status: 'running',\n        progress: { total, completed, current: batchLabel, message: `Executing ${batchLabel} (${batch.length} parallel)...` },\n      });\n\n      const batchResults = await Promise.all(batch.map(processItem));\n\n      for (const runResult of batchResults) {\n        await apiPost('/api/eval-runs', runResult);\n        results.push(runResult);\n        completed++;\n      }\n\n      await apiPatch(`/api/eval-jobs/${mirrorId}/progress`, {\n        status: 'running',\n        progress: { total, completed, current: batchLabel, message: `Completed ${batchLabel}` },\n      });\n    }\n  }\n\n  const summary = buildSummary(results);\n\n  // Meta-judge: assess eval quality if we have grading results\n  let metaJudgement: MetaJudgement | undefined;\n  if (!config.skipGrading && results.some(r => r.grading && Object.keys(r.grading as object).length > 0)) {\n    try {\n      log(`Running meta-judge for job ${mirrorId}...`);\n      const evalInputs = evalCases.map(e => {\n        const matchingResults = results.filter(r => r.evalId === e.id && r.configuration === 'with_target');\n        const firstResult = matchingResults[0];\n        return {\n          evalName: e.name,\n          prompt: e.prompt,\n          expectations: e.expectations,\n          grading: firstResult?.grading ? firstResult.grading as unknown as GradingResult : null,\n          output: (firstResult?.outputs as Record<string, string>)?.transcript?.slice(0, 500),\n        };\n      });\n      metaJudgement = await judgeEvalQuality({\n        targetName: target.name,\n        evalResults: evalInputs,\n        model: config.graderModel ?? config.model,\n        timeout: 60000,\n      });\n      log(`Meta-judge: quality=${metaJudgement.quality_score}/5, issues=${metaJudgement.issues.length}`);\n    } catch (err) {\n      log(`Meta-judge failed (non-fatal): ${err instanceof Error ? err.message : String(err)}`);\n    }\n  }\n\n  const result = metaJudgement\n    ? { ...summary, meta_judgement: metaJudgement }\n    : summary;\n  log(`Job ${mirrorId} completed: ${completed}/${total} runs`);\n\n  // Graphile discards task return values, so persist the summary onto the\n  // mirror row here. Best-effort: a failure to write the result must not\n  // mark the job as failed (it succeeded; we just lose the summary blob).\n  await apiPatch(`/api/eval-jobs/${mirrorId}/state`, { result }).catch((err: unknown) => {\n    log(`result persist failed: ${err instanceof Error ? err.message : String(err)}`);\n  });\n\n  return result;\n}\n\nfunction buildSummary(results: Record<string, unknown>[]): Record<string, unknown> {\n  const completed = results.filter((r) => r.status === 'completed');\n  const failed = results.filter((r) => r.status === 'failed');\n\n  const gradedRuns = completed.filter((r) => {\n    const grading = r.grading as Record<string, any>;\n    return grading && Object.keys(grading).length > 0;\n  });\n\n  const avgPassRate = gradedRuns.length > 0\n    ? gradedRuns.reduce((sum, r) => sum + ((r.grading as any)?.summary?.pass_rate ?? 0), 0) / gradedRuns.length\n    : null;\n\n  return {\n    total: results.length,\n    completed: completed.length,\n    failed: failed.length,\n    avgPassRate,\n  };\n}\n"],
-  "mappings": ";;;;;;;;;;;;;;;;;;;;;;AAqDA,eAAsB,YACpB,SACA,SACkC;AAClC,QAAM,EAAE,UAAU,cAAc,OAAO,IAAI;AAC3C,QAAM,MAAM,CAAC,QAAgB,QAAQ,OAAO,KAAK,GAAG;AAQpD,MAAI,QAAQ,IAAI,WAAW,GAAG;AAC5B,UAAM,IAAI;AAAA,MACR,8CAA8C,QAAQ,IAAI,QAAQ;AAAA,IACpE;AAAA,EACF;AAEA,QAAM,SAAS,MAAM,OAAmB,qBAAqB,YAAY,EAAE;AAC3E,QAAM,YAAY,MAAM,OAAe,qBAAqB,YAAY,QAAQ;AAEhF,MAAI,UAAU,WAAW,GAAG;AAG1B,UAAM,IAAI,kBAAkB,6BAA6B;AAAA,EAC3D;AAEA,QAAM,cAAc,OAAO,eAAe;AAC1C,QAAM,UAAU,OAAO,WAAW;AAClC,QAAM,YAAY,OAAO,aAAa;AAEtC,QAAM,iBAA4D;AAAA,IAChE,EAAE,MAAM,eAAe,SAAS,OAAO,QAAQ;AAAA,EACjD;AACA,MAAI,CAAC,OAAO,cAAc;AACxB,mBAAe,KAAK,EAAE,MAAM,iBAAiB,CAAC;AAAA,EAChD;AAGA,QAAM,YAAsG,CAAC;AAC7G,aAAW,SAAS,WAAW;AAC7B,eAAW,OAAO,gBAAgB;AAChC,eAAS,MAAM,GAAG,OAAO,aAAa,OAAO;AAC3C,kBAAU,KAAK,EAAE,OAAc,eAAe,IAAI,MAAM,eAAe,IAAI,SAAS,WAAW,IAAI,CAAC;AAAA,MACtG;AAAA,IACF;AAAA,EACF;AAEA,QAAM,QAAQ,UAAU;AACxB,MAAI,YAAY;AAChB,QAAM,cAAc,KAAK,IAAI,OAAO,YAAY,GAAG,CAAC;AAEpD,QAAM,SAAS,kBAAkB,QAAQ,aAAa;AAAA,IACpD,QAAQ;AAAA,IACR,UAAU,EAAE,OAAO,WAAW,GAAG,SAAS,MAAM,SAAS,YAAY,KAAK,4BAA4B,WAAW,OAAO;AAAA,EAC1H,CAAC;AAED,QAAM,UAAqC,CAAC;AAG5C,iBAAe,YAAY,MAA6D;AACtF,UAAM,QAAQ,GAAG,KAAK,aAAa,SAAM,KAAK,MAAM,IAAI,aAAU,KAAK,SAAS;AAChF,QAAI,IAAI,YAAY,CAAC,IAAI,KAAK,KAAK,KAAK,EAAE;AAG1C,UAAM,WAAW,cAAc,IAAI,mBAAmB,IAAI;AAE1D,QAAI;AACF,YAAM,aAAa,MAAM,YAAY;AAAA,QACnC,QAAQ,KAAK,MAAM;AAAA,QACnB,eAAe,KAAK;AAAA,QACpB,SAAS,UAAU;AAAA,QACnB;AAAA,QACA,OAAO,OAAO;AAAA,MAChB,CAAC;AAED,UAAI,UAAmC,CAAC;AACxC,UAAI,CAAC,OAAO,eAAe,KAAK,MAAM,aAAa,SAAS,KAAK,CAAC,WAAW,OAAO;AAClF,cAAM,gBAAgB,MAAM,gBAAgB;AAAA,UAC1C,QAAQ,KAAK,MAAM;AAAA,UACnB,QAAQ,WAAW;AAAA,UACnB,cAAc,KAAK,MAAM;AAAA,UACzB,OAAO,OAAO,eAAe,OAAO;AAAA,UACpC;AAAA,QACF,CAAC;AACD,kBAAU;AAAA,MACZ;AAEA,aAAO;AAAA,QACL,QAAQ,KAAK,MAAM;AAAA,QACnB,cAAc,OAAO;AAAA,QACrB,mBAAmB,OAAO;AAAA,QAC1B,eAAe,KAAK;AAAA,QACpB,WAAW,KAAK;AAAA,QAChB;AAAA,QACA,QAAQ,WAAW,QAAQ,WAAW;AAAA,QACtC,SAAS,WAAW,QAAQ,EAAE,OAAO,WAAW,MAAM,IAAI,EAAE,YAAY,WAAW,OAAO;AAAA,QAC1F,QAAQ,EAAE,aAAa,WAAW,YAAY,cAAc,WAAW,YAAY;AAAA,QACnF;AAAA,QACA,UAAU,UAAU,aAAa,OAAO;AAAA,MAC1C;AAAA,IACF,UAAE;AACA,gBAAU,QAAQ;AAAA,IACpB;AAAA,EACF;AAEA,MAAI,eAAe,GAAG;AAEpB,eAAW,QAAQ,WAAW;AAC5B,YAAM,QAAQ,GAAG,KAAK,aAAa,SAAM,KAAK,MAAM,IAAI,aAAU,KAAK,SAAS;AAChF,YAAM,SAAS,kBAAkB,QAAQ,aAAa;AAAA,QACpD,QAAQ;AAAA,QACR,UAAU,EAAE,OAAO,WAAW,SAAS,OAAO,SAAS,aAAa,KAAK,MAAM;AAAA,MACjF,CAAC;AAED,YAAM,YAAY,MAAM,YAAY,IAAI;AACxC,YAAM,QAAQ,kBAAkB,SAAS;AACzC,cAAQ,KAAK,SAAS;AACtB;AAEA,YAAM,SAAS,kBAAkB,QAAQ,aAAa;AAAA,QACpD,QAAQ;AAAA,QACR,UAAU,EAAE,OAAO,WAAW,SAAS,OAAO,SAAS,aAAa,KAAK,GAAG;AAAA,MAC9E,CAAC;AAAA,IACH;AAAA,EACF,OAAO;AAEL,QAAI,WAAW,KAAK,2BAA2B,WAAW,uBAAuB;AACjF,aAAS,IAAI,GAAG,IAAI,UAAU,QAAQ,KAAK,aAAa;AACtD,YAAM,QAAQ,UAAU,MAAM,GAAG,IAAI,WAAW;AAChD,YAAM,aAAa,SAAS,KAAK,MAAM,IAAI,WAAW,IAAI,CAAC,IAAI,KAAK,KAAK,UAAU,SAAS,WAAW,CAAC;AAExG,YAAM,SAAS,kBAAkB,QAAQ,aAAa;AAAA,QACpD,QAAQ;AAAA,QACR,UAAU,EAAE,OAAO,WAAW,SAAS,YAAY,SAAS,aAAa,UAAU,KAAK,MAAM,MAAM,gBAAgB;AAAA,MACtH,CAAC;AAED,YAAM,eAAe,MAAM,QAAQ,IAAI,MAAM,IAAI,WAAW,CAAC;AAE7D,iBAAW,aAAa,cAAc;AACpC,cAAM,QAAQ,kBAAkB,SAAS;AACzC,gBAAQ,KAAK,SAAS;AACtB;AAAA,MACF;AAEA,YAAM,SAAS,kBAAkB,QAAQ,aAAa;AAAA,QACpD,QAAQ;AAAA,QACR,UAAU,EAAE,OAAO,WAAW,SAAS,YAAY,SAAS,aAAa,UAAU,GAAG;AAAA,MACxF,CAAC;AAAA,IACH;AAAA,EACF;AAEA,QAAM,UAAU,aAAa,OAAO;AAGpC,MAAI;AACJ,MAAI,CAAC,OAAO,eAAe,QAAQ,KAAK,OAAK,EAAE,WAAW,OAAO,KAAK,EAAE,OAAiB,EAAE,SAAS,CAAC,GAAG;AACtG,QAAI;AACF,UAAI,8BAA8B,QAAQ,KAAK;AAC/C,YAAM,aAAa,UAAU,IAAI,OAAK;AACpC,cAAM,kBAAkB,QAAQ,OAAO,OAAK,EAAE,WAAW,EAAE,MAAM,EAAE,kBAAkB,aAAa;AAClG,cAAM,cAAc,gBAAgB,CAAC;AACrC,eAAO;AAAA,UACL,UAAU,EAAE;AAAA,UACZ,QAAQ,EAAE;AAAA,UACV,cAAc,EAAE;AAAA,UAChB,SAAS,aAAa,UAAU,YAAY,UAAsC;AAAA,UAClF,QAAS,aAAa,SAAoC,YAAY,MAAM,GAAG,GAAG;AAAA,QACpF;AAAA,MACF,CAAC;AACD,sBAAgB,MAAM,iBAAiB;AAAA,QACrC,YAAY,OAAO;AAAA,QACnB,aAAa;AAAA,QACb,OAAO,OAAO,eAAe,OAAO;AAAA,QACpC,SAAS;AAAA,MACX,CAAC;AACD,UAAI,uBAAuB,cAAc,aAAa,cAAc,cAAc,OAAO,MAAM,EAAE;AAAA,IACnG,SAAS,KAAK;AACZ,UAAI,kCAAkC,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC,EAAE;AAAA,IAC1F;AAAA,EACF;AAEA,QAAM,SAAS,gBACX,EAAE,GAAG,SAAS,gBAAgB,cAAc,IAC5C;AACJ,MAAI,OAAO,QAAQ,eAAe,SAAS,IAAI,KAAK,OAAO;AAK3D,QAAM,SAAS,kBAAkB,QAAQ,UAAU,EAAE,OAAO,CAAC,EAAE,MAAM,CAAC,QAAiB;AACrF,QAAI,0BAA0B,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG,CAAC,EAAE;AAAA,EAClF,CAAC;AAED,SAAO;AACT;AAEA,SAAS,aAAa,SAA6D;AACjF,QAAM,YAAY,QAAQ,OAAO,CAAC,MAAM,EAAE,WAAW,WAAW;AAChE,QAAM,SAAS,QAAQ,OAAO,CAAC,MAAM,EAAE,WAAW,QAAQ;AAE1D,QAAM,aAAa,UAAU,OAAO,CAAC,MAAM;AACzC,UAAM,UAAU,EAAE;AAClB,WAAO,WAAW,OAAO,KAAK,OAAO,EAAE,SAAS;AAAA,EAClD,CAAC;AAED,QAAM,cAAc,WAAW,SAAS,IACpC,WAAW,OAAO,CAAC,KAAK,MAAM,OAAQ,EAAE,SAAiB,SAAS,aAAa,IAAI,CAAC,IAAI,WAAW,SACnG;AAEJ,SAAO;AAAA,IACL,OAAO,QAAQ;AAAA,IACf,WAAW,UAAU;AAAA,IACrB,QAAQ,OAAO;AAAA,IACf;AAAA,EACF;AACF;",
-  "names": []
-}