@xiangyixie/dsh-superagent 0.11.0 → 0.12.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
@@ -57,6 +57,10 @@ const ATTENTION_FENCE = /```attention[^\n]*\n([\s\S]*?)```/i
57
57
  const FOLLOWUP_MAX = 8
58
58
  /** Characters kept from one instruction or answer. */
59
59
  const FOLLOWUP_MAX_CHARS = 1000
60
+ /** Live injection receipts kept per run; each may target several workers. */
61
+ const INJECTION_MAX = 12
62
+ /** Poll period for the measurement of one step boundary, in milliseconds. */
63
+ const INJECT_POLL_MS = 25
60
64
  /** How many expired attention batches stay readable for a late answer. */
61
65
  const ATTENTION_HISTORY_MAX = 5
62
66
 
@@ -774,6 +778,8 @@ function ensureCounters(run) {
774
778
  * @param {(childId: string) => Promise<object | null>} deps.readUsage
775
779
  * @param {(run: object) => object | void | Promise<object | void>} deps.notify - wake the owning session once.
776
780
  * @param {object | null} [deps.memory] - cross-run notebook from `createMemory`; null disables it.
781
+ * @param {object | null} [deps.guard] - worker safety guard from `createGuard`; its
782
+ * snapshot is attached to each run record and its events to the report.
777
783
  * @param {(message: string) => void} [deps.log]
778
784
  * @param {() => number} [deps.now]
779
785
  * @param {(ms: number) => Promise<void>} [deps.sleep]
@@ -789,6 +795,11 @@ export function createOrchestrator(deps) {
789
795
  notify,
790
796
  memory = null,
791
797
  workspace = null,
798
+ guard = null,
799
+ audit = null,
800
+ steerChild = null,
801
+ inspectChild = null,
802
+ watchChild = null,
792
803
  log = () => {},
793
804
  now = () => Date.now(),
794
805
  sleep = ms => new Promise(resolve => setTimeout(resolve, ms)),
@@ -997,6 +1008,10 @@ export function createOrchestrator(deps) {
997
1008
 
998
1009
  function save(run) {
999
1010
  run.updatedAt = new Date(now()).toISOString()
1011
+ // The guard is process-wide and owned by the caller; copy its counters onto
1012
+ // the record so status and the stored run show what was blocked.
1013
+ const snapshot = guard?.snapshot?.(run.id)
1014
+ if (snapshot) run.guard = snapshot
1000
1015
  const previous = saveChains.get(run.id) ?? Promise.resolve()
1001
1016
  const next = previous.catch(() => {}).then(() => store.write(run))
1002
1017
  saveChains.set(run.id, next)
@@ -1164,6 +1179,21 @@ export function createOrchestrator(deps) {
1164
1179
  await pumpQueue().catch(() => {})
1165
1180
  }
1166
1181
 
1182
+ /**
1183
+ * Write one audit event. The trail is an observer: a failure is logged and
1184
+ * dropped, never allowed to change the outcome of the run it describes.
1185
+ * `audit` is optional so a host without the audit layer behaves exactly as
1186
+ * it did before it existed.
1187
+ *
1188
+ * @param {object} event - fields for `audit.record`.
1189
+ */
1190
+ function auditRecord(event) {
1191
+ if (!audit || typeof audit.record !== 'function') return Promise.resolve()
1192
+ return Promise.resolve()
1193
+ .then(() => audit.record(event))
1194
+ .catch(error => log(`cannot write an audit event for run ${event?.runId ?? 'unknown'}: ${error?.message ?? error}`))
1195
+ }
1196
+
1167
1197
  /**
1168
1198
  * Record one phase transition on the run. Milestones are observational only:
1169
1199
  * they appear in `superagent_status` and never wake the session, which is what
@@ -1204,6 +1234,16 @@ export function createOrchestrator(deps) {
1204
1234
  options: attention.options ?? [],
1205
1235
  reason: attention.reason ?? null,
1206
1236
  })
1237
+ auditRecord({
1238
+ runId: run.id,
1239
+ subtaskId: source,
1240
+ actor: 'agent',
1241
+ kind: 'attention',
1242
+ action: 'attention.request',
1243
+ target: attention.question,
1244
+ result: 'pending',
1245
+ meta: { options: attention.options ?? [], reason: attention.reason ?? null },
1246
+ })
1207
1247
  }
1208
1248
 
1209
1249
  /**
@@ -1274,6 +1314,355 @@ export function createOrchestrator(deps) {
1274
1314
  return changed
1275
1315
  }
1276
1316
 
1317
+ /**
1318
+ * Record one live injection and its per-worker receipts.
1319
+ *
1320
+ * This is a separate trail from `run.followups` on purpose: a followup is
1321
+ * carried by the next prompt built for the run, while an injection is handed
1322
+ * to an inbox that is already executing. Mixing them would make status
1323
+ * promise a boundary that the message never uses.
1324
+ *
1325
+ * @param {object} run - live run record.
1326
+ * @param {string} text - the instruction.
1327
+ * @returns {object} the stored entry, with an empty `targets` list.
1328
+ */
1329
+ function pushInjection(run, text) {
1330
+ run.injections = Array.isArray(run.injections) ? run.injections : []
1331
+ const entry = {
1332
+ id: `i${(Number(run.injectionSeq) || 0) + 1}`,
1333
+ text: capChars(text, FOLLOWUP_MAX_CHARS),
1334
+ at: new Date(now()).toISOString(),
1335
+ delivery: 'steer',
1336
+ status: 'pending',
1337
+ followupId: null,
1338
+ targets: [],
1339
+ }
1340
+ run.injectionSeq = (Number(run.injectionSeq) || 0) + 1
1341
+ run.injections.push(entry)
1342
+ if (run.injections.length > INJECTION_MAX) run.injections = run.injections.slice(-INJECTION_MAX)
1343
+ return entry
1344
+ }
1345
+
1346
+ /** Watchers are runtime state, never persisted with the run. */
1347
+ const injectionWatches = new Map()
1348
+ const watchKey = (runId, injectionId, childId) => `${runId}:${injectionId}:${childId}`
1349
+
1350
+ function stopWatch(runId, injectionId, childId) {
1351
+ const key = watchKey(runId, injectionId, childId)
1352
+ const stop = injectionWatches.get(key)
1353
+ if (!stop) return
1354
+ injectionWatches.delete(key)
1355
+ try {
1356
+ stop()
1357
+ } catch (error) {
1358
+ log(`cannot stop the injection watcher for ${childId}: ${error?.message ?? error}`)
1359
+ }
1360
+ }
1361
+
1362
+ function stopRunWatches(runId) {
1363
+ for (const key of [...injectionWatches.keys()]) {
1364
+ if (key.split(':')[0] !== runId) continue
1365
+ stopWatch(...key.split(':'))
1366
+ }
1367
+ }
1368
+
1369
+ /**
1370
+ * Recompute the run totals kept for status and tool replies.
1371
+ * @param {object} run - live run record.
1372
+ */
1373
+ function refreshInjectionStats(run) {
1374
+ const targets = (run.injections ?? []).flatMap(entry => entry.targets ?? [])
1375
+ const applied = targets.filter(target => target.status === 'applied')
1376
+ const latencies = applied.map(target => target.latencyMs).filter(value => Number.isFinite(value))
1377
+ for (const entry of run.injections ?? []) {
1378
+ if (entry.status === 'degraded') continue
1379
+ const own = entry.targets ?? []
1380
+ if (own.length === 0) entry.status = 'failed'
1381
+ else if (own.some(target => target.status === 'pending')) entry.status = 'pending'
1382
+ else if (own.every(target => target.status === 'applied')) entry.status = 'applied'
1383
+ else if (own.some(target => target.status === 'applied')) entry.status = 'partial'
1384
+ else entry.status = 'missed'
1385
+ }
1386
+ run.injectionStats = {
1387
+ accepted: targets.filter(target => Boolean(target.messageId)).length,
1388
+ applied: applied.length,
1389
+ missed: targets.filter(target => target.status === 'missed').length,
1390
+ failed: targets.filter(target => target.status === 'failed').length,
1391
+ pending: targets.filter(target => target.status === 'pending').length,
1392
+ lastLatencyMs: latencies.length ? latencies[latencies.length - 1] : null,
1393
+ worstLatencyMs: latencies.length ? Math.max(...latencies) : null,
1394
+ samples: latencies.length,
1395
+ }
1396
+ }
1397
+
1398
+ /**
1399
+ * Move one worker receipt out of `pending`, once.
1400
+ *
1401
+ * `applied` is only recorded with the claim time the caller observed; a
1402
+ * dispatch that settles without observing a claim is not upgraded here.
1403
+ *
1404
+ * @param {object} run - live run record.
1405
+ * @param {object} target - one receipt.
1406
+ * @param {'applied'|'missed'} status - observed outcome.
1407
+ * @param {{ at?: number, reason?: string }} [detail] - claim time or miss reason.
1408
+ * @returns {boolean} whether the receipt changed.
1409
+ */
1410
+ function finishInjectionTarget(run, target, status, detail = {}) {
1411
+ if (!target || target.status !== 'pending') return false
1412
+ target.status = status
1413
+ if (status === 'applied') {
1414
+ const at = Number.isFinite(detail.at) ? detail.at : now()
1415
+ const accepted = Date.parse(target.acceptedAt ?? '')
1416
+ target.appliedAt = new Date(at).toISOString()
1417
+ target.latencyMs = Number.isFinite(accepted) ? Math.max(0, at - accepted) : null
1418
+ } else {
1419
+ target.missedAt = new Date(now()).toISOString()
1420
+ target.missedReason = detail.reason ?? 'the child never reached another step boundary'
1421
+ target.latencyMs = null
1422
+ }
1423
+ return true
1424
+ }
1425
+
1426
+ /**
1427
+ * Persist a receipt change without holding up the caller's serialized action.
1428
+ * @param {object} run - live run record.
1429
+ * @param {object} target - one receipt.
1430
+ * @param {'applied'|'missed'} status - observed outcome.
1431
+ * @param {{ at?: number, reason?: string }} [detail] - claim time or miss reason.
1432
+ */
1433
+ function recordInjectionTarget(run, target, status, detail) {
1434
+ if (!finishInjectionTarget(run, target, status, detail)) return
1435
+ refreshInjectionStats(run)
1436
+ save(run).catch(error => log(`cannot persist the injection state of run ${run.id}: ${error?.message ?? error}`))
1437
+ }
1438
+
1439
+ /**
1440
+ * Poll one accepted message until the child claims it.
1441
+ *
1442
+ * @param {object} run - live run record.
1443
+ * @param {object} entry - injection receipt.
1444
+ * @param {object} target - one worker receipt.
1445
+ */
1446
+ function watchInjection(run, entry, target) {
1447
+ if (typeof watchChild !== 'function' || !target.messageId) return
1448
+ const stop = watchChild(target.childId, target.messageId, {
1449
+ onClaimed: at => recordInjectionTarget(run, target, 'applied', { at }),
1450
+ onGone: () => recordInjectionTarget(run, target, 'missed', {
1451
+ reason: 'the child activation was disposed before it reached another step boundary',
1452
+ }),
1453
+ })
1454
+ const key = watchKey(run.id, entry.id, target.childId)
1455
+ const previous = injectionWatches.get(key)
1456
+ if (previous) {
1457
+ try {
1458
+ previous()
1459
+ } catch {
1460
+ // A stale watcher that cannot stop is harmless; the new one owns the receipt.
1461
+ }
1462
+ }
1463
+ injectionWatches.set(key, () => {
1464
+ if (stop && typeof stop.stop === 'function') stop.stop()
1465
+ else if (typeof stop === 'function') stop()
1466
+ })
1467
+ // The message can be claimed before the first poll: one immediate read keeps
1468
+ // a fast delivery from being displayed as pending for a full period.
1469
+ if (typeof stop?.check === 'function') {
1470
+ try {
1471
+ stop.check()
1472
+ } catch {
1473
+ // The poll keeps running; a failed first read is not a delivery failure.
1474
+ }
1475
+ }
1476
+ }
1477
+
1478
+ /**
1479
+ * Close every receipt aimed at one child that is leaving its attempt.
1480
+ *
1481
+ * Runs before disposal, so the inbox is still readable: a message that is
1482
+ * gone was claimed, and one that is still there never reached a prompt.
1483
+ *
1484
+ * @param {object} run - live run record.
1485
+ * @param {string | null | undefined} childId - departing child.
1486
+ * @param {string} reason - why a still-pending message will not be read.
1487
+ * @returns {boolean} whether the run needs a save.
1488
+ */
1489
+ function settleChildInjections(run, childId, reason) {
1490
+ if (!childId) return false
1491
+ const targets = (run.injections ?? [])
1492
+ .flatMap(entry => (entry.targets ?? []).map(target => ({ entry, target })))
1493
+ .filter(({ target }) => target.childId === childId && target.status === 'pending')
1494
+ if (targets.length === 0) return false
1495
+ let inspected = null
1496
+ try {
1497
+ inspected = typeof inspectChild === 'function' ? inspectChild(childId) : null
1498
+ } catch {
1499
+ inspected = null
1500
+ }
1501
+ let changed = false
1502
+ for (const { entry, target } of targets) {
1503
+ stopWatch(run.id, entry.id, target.childId)
1504
+ const live = Boolean(inspected?.live)
1505
+ const stillPending = live && inspected?.pending?.has?.(String(target.messageId))
1506
+ changed = finishInjectionTarget(run, target, live && !stillPending ? 'applied' : 'missed', {
1507
+ reason: live
1508
+ ? reason
1509
+ : 'the child was gone before the message reached another step boundary',
1510
+ }) || changed
1511
+ }
1512
+ if (changed) refreshInjectionStats(run)
1513
+ return changed
1514
+ }
1515
+
1516
+ /**
1517
+ * Deliver one instruction into the workers that are running now.
1518
+ *
1519
+ * Delivery is a steer: `agent-loop` claims it at the next pre-step, so it is
1520
+ * part of the child's next model request instead of interrupting the one in
1521
+ * flight. A child that had already decided to stop runs one more turn when
1522
+ * the inbox is non-empty, which is why a live, uncancelled child consumes it.
1523
+ * The latency is measured from acceptance to that claim and written on the
1524
+ * receipt; the theoretical bound is the remainder of the running step.
1525
+ *
1526
+ * @param {string} runId - run identity.
1527
+ * @param {string} text - the instruction.
1528
+ * @param {object} parent - live agent that owns the run.
1529
+ * @param {{ subtaskIndex?: number | string | null }} [options] - narrow to one worker.
1530
+ * @returns {Promise<{ run: object, injection: object, degraded: object | null, effect: object }>}
1531
+ */
1532
+ async function inject(runId, text, parent, options = {}) {
1533
+ const id = String(runId ?? '').trim()
1534
+ if (!id) throw new Error('run_id is required')
1535
+ const message = capChars(text, FOLLOWUP_MAX_CHARS)
1536
+ if (!message) throw new Error('message is required; pass the instruction the running workers should read')
1537
+ if (!parent || typeof parent !== 'object') throw new Error('superagent requires a live parent agent')
1538
+ const rawIndex = options.subtaskIndex
1539
+ const wanted = rawIndex === undefined || rawIndex === null || rawIndex === ''
1540
+ ? null
1541
+ : Number(rawIndex)
1542
+ if (wanted !== null && !Number.isInteger(wanted)) {
1543
+ throw new Error('subtask_index must be a whole number from superagent_status')
1544
+ }
1545
+ return serialize(async () => {
1546
+ const rt = runtime.get(id)
1547
+ const run = rt?.run ?? await loadRun(id)
1548
+ assertResumeOwner(run, parent)
1549
+ if (!ACTIVE_STATUSES.has(run.status)) {
1550
+ throw new Error(
1551
+ `run ${id} is ${run.status}, so no child is running to receive a live instruction. `
1552
+ + 'Use superagent_followup to queue it for the next prompt.',
1553
+ )
1554
+ }
1555
+ if (!rt) {
1556
+ throw new Error(
1557
+ `run ${id} is active but owned by another process; a live instruction must come from the process that owns `
1558
+ + 'its children, because only that process holds their inboxes.',
1559
+ )
1560
+ }
1561
+ const running = (run.subtasks ?? []).filter(subtask => subtask.status === 'running' && subtask.childId)
1562
+ const selected = wanted === null ? running : running.filter(subtask => subtask.index === wanted)
1563
+ // A running subtask can be between attempts, where its old child id no
1564
+ // longer resolves. Steering a dead inbox would look like delivery, so the
1565
+ // live ones are selected here and the reply names the gap otherwise.
1566
+ const live = typeof inspectChild === 'function'
1567
+ ? selected.filter(subtask => {
1568
+ try {
1569
+ return Boolean(inspectChild(subtask.childId)?.live)
1570
+ } catch {
1571
+ return false
1572
+ }
1573
+ })
1574
+ : selected
1575
+ const targets = live
1576
+ if (targets.length === 0) {
1577
+ const who = wanted === null ? 'no worker is live' : `worker ${wanted} is not live`
1578
+ const activity = (run.subtasks ?? [])
1579
+ .filter(subtask => subtask.status === 'running')
1580
+ .map(subtask => `#${subtask.index}`)
1581
+ .join(', ')
1582
+ throw new Error(
1583
+ `${who} on run ${id} (status ${run.status}${activity ? `, attempts in progress: ${activity}` : ''}), so `
1584
+ + 'there is no inbox to steer right now. superagent_followup queues the instruction for the next prompt '
1585
+ + 'instead.',
1586
+ )
1587
+ }
1588
+ const entry = pushInjection(run, message)
1589
+ let unsupported = null
1590
+ for (const subtask of targets) {
1591
+ const target = {
1592
+ subtaskIndex: subtask.index,
1593
+ childId: subtask.childId,
1594
+ messageId: null,
1595
+ via: null,
1596
+ acceptedAt: new Date(now()).toISOString(),
1597
+ status: 'pending',
1598
+ appliedAt: null,
1599
+ latencyMs: null,
1600
+ missedAt: null,
1601
+ missedReason: null,
1602
+ error: null,
1603
+ }
1604
+ entry.targets.push(target)
1605
+ if (typeof steerChild !== 'function') {
1606
+ target.status = 'failed'
1607
+ target.error = 'this host exposes no live channel into a running child'
1608
+ unsupported = target.error
1609
+ continue
1610
+ }
1611
+ try {
1612
+ const receipt = await steerChild({
1613
+ parent,
1614
+ childId: target.childId,
1615
+ text: message,
1616
+ signal: rt.controller.signal,
1617
+ })
1618
+ target.messageId = String(receipt?.messageId ?? '')
1619
+ target.via = receipt?.via ?? null
1620
+ watchInjection(run, entry, target)
1621
+ } catch (error) {
1622
+ target.status = 'failed'
1623
+ target.error = error?.message ?? String(error)
1624
+ if (error?.code === 'inject-unsupported') unsupported = target.error
1625
+ else if (error?.code === 'inject-child-gone') target.error = `${target.error}; the worker already settled`
1626
+ }
1627
+ }
1628
+ let degraded = null
1629
+ if (unsupported && entry.targets.every(target => target.status === 'failed')) {
1630
+ // No live channel at all. Do not pretend the instruction was delivered:
1631
+ // queue it exactly where superagent_followup would, and label the
1632
+ // receipt as degraded so status cannot read as a live delivery.
1633
+ const followup = pushFollowup(run, { kind: 'followup', text: message })
1634
+ entry.status = 'degraded'
1635
+ entry.followupId = followup.id
1636
+ degraded = { followup, effect: followupEffect(run), reason: unsupported }
1637
+ }
1638
+ refreshInjectionStats(run)
1639
+ await save(run)
1640
+ return { run, injection: entry, degraded, effect: followupEffect(run) }
1641
+ })
1642
+ }
1643
+
1644
+ /**
1645
+ * Wait for the receipts of one injection to leave `pending`.
1646
+ *
1647
+ * The claim is observed by a poll, so the reported latency has that
1648
+ * resolution; waiting does not change delivery, only what the reply can say.
1649
+ *
1650
+ * @param {object} run - live run record.
1651
+ * @param {object} entry - injection receipt.
1652
+ * @param {number} [seconds] - upper bound on the wait.
1653
+ * @returns {Promise<object>} the same entry, updated in place.
1654
+ */
1655
+ async function waitForInjection(run, entry, seconds) {
1656
+ const limit = Math.max(0, Number(seconds) || 0) * 1000
1657
+ if (limit <= 0) return entry
1658
+ const started = now()
1659
+ while (now() - started < limit) {
1660
+ if (!(entry.targets ?? []).some(target => target.status === 'pending')) break
1661
+ await sleep(Math.min(INJECT_POLL_MS * 2, Math.max(5, limit - (now() - started))))
1662
+ }
1663
+ return entry
1664
+ }
1665
+
1277
1666
  /**
1278
1667
  * Write the final report next to the run state, so the outcome has an address
1279
1668
  * the user can open or attach, not only a paragraph in a tool reply. The
@@ -1791,7 +2180,8 @@ export function createOrchestrator(deps) {
1791
2180
  // crash cannot lose the only evidence of a model call.
1792
2181
  ensureCounters(run).spawns += 1
1793
2182
  await save(run)
1794
- return spawnAbortable(spec, rt.controller.signal)
2183
+ // runId lets the caller bind the child to this run's guard counters.
2184
+ return spawnAbortable({ ...spec, runId: run.id }, rt.controller.signal)
1795
2185
  }
1796
2186
 
1797
2187
  /**
@@ -2077,6 +2467,13 @@ export function createOrchestrator(deps) {
2077
2467
  if (handle) {
2078
2468
  // A failed attempt still spent tokens, so read it before disposing.
2079
2469
  await recordUsage(run, handle.childId)
2470
+ // Receipts are resolved while the inbox is still readable: a message
2471
+ // still waiting here never entered a prompt for this attempt.
2472
+ if (settleChildInjections(
2473
+ run,
2474
+ handle.childId,
2475
+ 'the child finished its turn before the instruction reached another step boundary',
2476
+ )) await save(run)
2080
2477
  rt.activations.delete(handle)
2081
2478
  await disposeQuietly(handle)
2082
2479
  }
@@ -2306,6 +2703,21 @@ export function createOrchestrator(deps) {
2306
2703
  entry.missedReason = 'the run settled before another boundary ran; start a new run to apply it'
2307
2704
  }
2308
2705
 
2706
+ // A live injection has no boundary of its own: once the run is over, an
2707
+ // untouched receipt is a miss, and the queue that carried it is gone.
2708
+ let injectionsChanged = false
2709
+ for (const injection of run.injections ?? []) {
2710
+ for (const target of injection.targets ?? []) {
2711
+ if (target.status !== 'pending') continue
2712
+ stopWatch(run.id, injection.id, target.childId)
2713
+ finishInjectionTarget(run, target, 'missed', {
2714
+ reason: 'the run settled before the child reached another step boundary',
2715
+ })
2716
+ injectionsChanged = true
2717
+ }
2718
+ }
2719
+ if (injectionsChanged) refreshInjectionStats(run)
2720
+
2309
2721
  // The aggregator appends a fenced notebook block to its report. Strip it in
2310
2722
  // every case so the user never sees plugin syntax, then persist it when the
2311
2723
  // notebook is enabled. A write failure is recorded on the run but never
@@ -2359,6 +2771,12 @@ export function createOrchestrator(deps) {
2359
2771
  }
2360
2772
  }
2361
2773
  }
2774
+ // The guard's event list belongs in the user-facing report, not only in the
2775
+ // status line: a blocked command is part of what happened during the run.
2776
+ const guardSection = guard?.reportSection?.(run.id)
2777
+ if (guardSection && typeof run.report === 'string' && run.report) {
2778
+ run.report = `${run.report.trimEnd()}\n\n${guardSection}`
2779
+ }
2362
2780
  // The report gets a durable address before the run is announced, so a
2363
2781
  // notification or status line can point at a file that already exists.
2364
2782
  if (TERMINAL_STATUSES.has(nextStatus)) await writeReportArtifact(run)
@@ -2416,6 +2834,24 @@ export function createOrchestrator(deps) {
2416
2834
  if (outcome.delivered !== false) run.notified = true
2417
2835
  }
2418
2836
  await save(run)
2837
+ // Awaited: the trail must contain the outcome before the run is published
2838
+ // as terminal, or a crash right here would leave a finished run with no
2839
+ // record of how it finished. A failure is still only logged.
2840
+ await auditRecord({
2841
+ runId: run.id,
2842
+ actor: `session:${run.parentSessionId}`,
2843
+ kind: 'run.finish',
2844
+ action: 'settle',
2845
+ target: oneLine(run.goal, 200),
2846
+ result: nextStatus,
2847
+ durationMs: run.activeMs,
2848
+ meta: {
2849
+ subtasks: run.subtasks.length,
2850
+ completed: run.subtasks.filter(subtask => subtask.status === 'completed').length,
2851
+ attention: Boolean(run.attention?.items?.length),
2852
+ error: error ? oneLine(error, 300) : null,
2853
+ },
2854
+ })
2419
2855
  // The run is terminal, so a leftover marker would only make the next mount
2420
2856
  // rewrite a finished run; drop it now.
2421
2857
  await clearCancel(run.id)
@@ -2702,6 +3138,7 @@ export function createOrchestrator(deps) {
2702
3138
  .then(() => {
2703
3139
  // Finished runs are read from disk from now on; keeping them here would
2704
3140
  // grow these maps for the lifetime of the host.
3141
+ stopRunWatches(run.id)
2705
3142
  runtime.delete(run.id)
2706
3143
  saveChains.delete(run.id)
2707
3144
  // The slot this run held may let a waiting run start.
@@ -2726,6 +3163,7 @@ export function createOrchestrator(deps) {
2726
3163
  allowDuplicate = false,
2727
3164
  queue: wantsQueue = false,
2728
3165
  mode: requestedMode,
3166
+ approvePolicy = false,
2729
3167
  }) {
2730
3168
  const trimmed = typeof goal === 'string' ? goal.trim() : ''
2731
3169
  if (!trimmed) throw new Error('goal is required')
@@ -2768,12 +3206,44 @@ export function createOrchestrator(deps) {
2768
3206
  )
2769
3207
  }
2770
3208
  }
3209
+ // A permission policy decides whether this start may exist at all. The
3210
+ // verdict is recorded before the run is written, so a refused start still
3211
+ // leaves evidence of what was asked for and why it was stopped.
3212
+ const runId = newRunId()
3213
+ if (audit && typeof audit.gate === 'function') {
3214
+ const verdict = await audit.gate({
3215
+ runId,
3216
+ kind: 'run.start',
3217
+ action: 'superagent_start',
3218
+ actor: `session:${parentSessionId}`,
3219
+ target: oneLine(trimmed, 200),
3220
+ meta: { goalChars: trimmed.length },
3221
+ })
3222
+ const trail = verdict.eventId ? ` (audit event ${verdict.eventId})` : ''
3223
+ if (verdict.decision === 'deny') {
3224
+ throw new Error(`refused by the permission policy: ${verdict.reason}${trail}`)
3225
+ }
3226
+ if (verdict.decision === 'confirm') {
3227
+ if (!approvePolicy) {
3228
+ throw new Error(
3229
+ `the permission policy holds this start for confirmation: ${verdict.reason}${trail}; `
3230
+ + 'call superagent_start again with approve_policy: true to release it',
3231
+ )
3232
+ }
3233
+ await audit.approve(verdict.eventId, {
3234
+ actor: `session:${parentSessionId}`,
3235
+ note: 'approve_policy on superagent_start',
3236
+ runId,
3237
+ kind: 'run.start',
3238
+ })
3239
+ }
3240
+ }
2771
3241
  // Read once, here: the digest is part of the run record, so a status call
2772
3242
  // can show exactly what the planner was told, and a resumed run does not
2773
3243
  // silently change the context its plan was based on.
2774
3244
  const context = await memoryContext(null, trimmed)
2775
3245
  const run = {
2776
- id: newRunId(),
3246
+ id: runId,
2777
3247
  schemaVersion: 1,
2778
3248
  goal: trimmed,
2779
3249
  dedupeKey,
@@ -2866,7 +3336,7 @@ export function createOrchestrator(deps) {
2866
3336
  * `reopen` lets `superagent_answer` continue a `completed` run: its subtasks
2867
3337
  * stay done and only the aggregation runs again, now with the answer.
2868
3338
  */
2869
- async function resumeLocked(runId, parent, clientTimeZone, { reopen = false } = {}) {
3339
+ async function resumeLocked(runId, parent, clientTimeZone, { reopen = false, approvePolicy = false } = {}) {
2870
3340
  const id = String(runId ?? '').trim()
2871
3341
  if (!parent || typeof parent !== 'object') throw new Error('superagent requires a live parent agent')
2872
3342
  {
@@ -2900,6 +3370,36 @@ export function createOrchestrator(deps) {
2900
3370
  assertResumeOwner(run, parent)
2901
3371
  reopenCompleted = reopen || hasUnappliedAnswer(run)
2902
3372
  if (run.status === 'completed' && !reopenCompleted) return run
3373
+ // A resume spends the remaining budget again, so a per-run deny applies
3374
+ // to it as well; otherwise a policy added mid-run could be bypassed by
3375
+ // resuming the run instead of starting a new one.
3376
+ if (audit && typeof audit.gate === 'function') {
3377
+ const verdict = await audit.gate({
3378
+ runId: run.id,
3379
+ kind: 'run.resume',
3380
+ action: 'superagent_resume',
3381
+ actor: `session:${parent.id}`,
3382
+ target: oneLine(run.goal, 200),
3383
+ })
3384
+ const trail = verdict.eventId ? ` (audit event ${verdict.eventId})` : ''
3385
+ if (verdict.decision === 'deny') {
3386
+ throw new Error(`refused by the permission policy: ${verdict.reason}${trail}`)
3387
+ }
3388
+ if (verdict.decision === 'confirm') {
3389
+ if (!approvePolicy) {
3390
+ throw new Error(
3391
+ `the permission policy holds this resume for confirmation: ${verdict.reason}${trail}; `
3392
+ + 'call superagent_resume again with approve_policy: true to release it',
3393
+ )
3394
+ }
3395
+ await audit.approve(verdict.eventId, {
3396
+ actor: `session:${parent.id}`,
3397
+ note: 'approve_policy on superagent_resume',
3398
+ runId: run.id,
3399
+ kind: 'run.resume',
3400
+ })
3401
+ }
3402
+ }
2903
3403
  if (ACTIVE_STATUSES.has(run.status)) {
2904
3404
  const verdict = await ownership(run)
2905
3405
  if (verdict.live) {
@@ -2989,6 +3489,15 @@ export function createOrchestrator(deps) {
2989
3489
  ensureCounters(run)
2990
3490
  await save(run)
2991
3491
  schedule(run, parent)
3492
+ auditRecord({
3493
+ runId: run.id,
3494
+ actor: `session:${parent.id}`,
3495
+ kind: 'run.resume',
3496
+ action: 'superagent_resume',
3497
+ target: oneLine(run.goal, 200),
3498
+ result: 'queued',
3499
+ meta: { previousStatus: run.recovery?.previousStatus ?? null },
3500
+ })
2992
3501
  scheduled = true
2993
3502
  return run
2994
3503
  } finally {
@@ -3223,6 +3732,14 @@ export function createOrchestrator(deps) {
3223
3732
  rt.run.cancelPending = true
3224
3733
  await save(rt.run).catch(() => {})
3225
3734
  }
3735
+ auditRecord({
3736
+ runId,
3737
+ actor: 'session',
3738
+ kind: 'run.cancel',
3739
+ action: 'superagent_cancel',
3740
+ result: ACTIVE_STATUSES.has(rt.run.status) ? 'requested' : 'cancelled',
3741
+ meta: { ownedHere: true },
3742
+ })
3226
3743
  return rt.run
3227
3744
  }
3228
3745
 
@@ -3234,6 +3751,14 @@ export function createOrchestrator(deps) {
3234
3751
  // Writing `cancelled` here would be a lie: the owner would overwrite it on
3235
3752
  // its next save and keep spending.
3236
3753
  await requestCancel(runId)
3754
+ auditRecord({
3755
+ runId,
3756
+ actor: 'session',
3757
+ kind: 'run.cancel',
3758
+ action: 'superagent_cancel',
3759
+ result: 'requested',
3760
+ meta: { ownedHere: false },
3761
+ })
3237
3762
  const endsAt = now() + CANCEL_ACK_MS
3238
3763
  while (now() < endsAt) {
3239
3764
  await sleep(STATUS_POLL_MS)
@@ -3308,5 +3833,19 @@ export function createOrchestrator(deps) {
3308
3833
  await Promise.all([...runtime.values()].map(rt => rt.promise?.catch(() => {})))
3309
3834
  }
3310
3835
 
3311
- return { recover, start, resume, cancel, followup, answer, status, list, idle, activeCount, liveRun }
3836
+ return {
3837
+ recover,
3838
+ start,
3839
+ resume,
3840
+ cancel,
3841
+ followup,
3842
+ answer,
3843
+ inject,
3844
+ waitForInjection,
3845
+ status,
3846
+ list,
3847
+ idle,
3848
+ activeCount,
3849
+ liveRun,
3850
+ }
3312
3851
  }