@awak-app/simy-cli 0.7.36 → 0.7.38

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.
Files changed (55) hide show
  1. package/README.md +36 -36
  2. package/package.json +1 -1
  3. package/src/agent.js +1 -1
  4. package/src/auto-update.js +25 -13
  5. package/src/backend-executable.js +1 -1
  6. package/src/cli-contract.js +1 -1
  7. package/src/cli-observability.js +2 -2
  8. package/src/codex-plugins/simy/.codex-plugin/plugin.json +5 -5
  9. package/src/codex-plugins/simy/hooks/README.md +3 -3
  10. package/src/codex-plugins/simy/hooks/delivery-loop.mjs +242 -26
  11. package/src/codex-plugins/simy/hooks/purpose-workspace.mjs +4 -1
  12. package/src/codex-plugins/simy/hooks/registered-requirement-checker.mjs +201 -0
  13. package/src/codex-plugins/simy/hooks/requirement-acceptance.mjs +69 -0
  14. package/src/codex-plugins/simy/skills/continue/SKILL.md +4 -4
  15. package/src/codex-plugins/simy/skills/continue/agents/openai.yaml +3 -3
  16. package/src/codex-plugins/simy/skills/diagnose/SKILL.md +2 -2
  17. package/src/codex-plugins/simy/skills/sqm-loop/SKILL.md +3 -3
  18. package/src/codex-plugins/simy/skills/sqm-loop/agents/openai.yaml +2 -2
  19. package/src/console/app.js +7 -7
  20. package/src/delivery-hooks/claude-workflow.md +1 -1
  21. package/src/delivery-hooks/command.js +5 -5
  22. package/src/delivery-hooks/supervisor-watchdog.js +3 -2
  23. package/src/desktop-host.js +53 -23
  24. package/src/index.js +1 -1
  25. package/src/local-task.js +1 -1
  26. package/src/orchestrator/presentation.js +1 -1
  27. package/src/runner.js +2 -2
  28. package/src/session-supervision/acceptance-reads.js +11 -2
  29. package/src/session-supervision/analysis-manager.js +9 -2
  30. package/src/session-supervision/automatic-continuation.js +11 -11
  31. package/src/session-supervision/catalog-refresh.js +16 -0
  32. package/src/session-supervision/claude-native-agent-contract.js +62 -0
  33. package/src/session-supervision/claude-outcome-review.js +7 -6
  34. package/src/session-supervision/claude-sessions.js +137 -76
  35. package/src/session-supervision/codex-app-server.js +3 -3
  36. package/src/session-supervision/codex-children.js +2 -2
  37. package/src/session-supervision/continuation-analysis.js +1 -1
  38. package/src/session-supervision/continuation-command.js +3 -3
  39. package/src/session-supervision/continuation-readiness.js +1 -1
  40. package/src/session-supervision/conversation-context.js +58 -26
  41. package/src/session-supervision/conversation-presentation.js +3 -3
  42. package/src/session-supervision/decision-presentation.js +8 -8
  43. package/src/session-supervision/discovery-window.js +8 -2
  44. package/src/session-supervision/index.js +38 -25
  45. package/src/session-supervision/purpose-verification.js +12 -3
  46. package/src/session-supervision/remote-errors.js +2 -2
  47. package/src/session-supervision/work-state.js +3 -3
  48. package/src/sqm/advisory-guidance.js +1 -0
  49. package/src/sqm/agent-skill.js +3 -3
  50. package/src/sqm/checkers.js +35 -1
  51. package/src/sqm/claude-plugin.js +8 -2
  52. package/src/sqm/hypothesis-knowledge.js +3 -3
  53. package/src/sqm/knowledge-recall.js +34 -12
  54. package/src/sqm/reasoned-assessment.js +22 -9
  55. package/src/update-detachment/runtime.js +2 -1
@@ -1,5 +1,7 @@
1
1
  import { createDesktopFrameWriter } from './desktop-frame-writer.js';
2
2
  import { boundedDiscovery } from './session-supervision/progressive-discovery.js';
3
+ import { createCatalogRefreshPolicy } from './session-supervision/catalog-refresh.js';
4
+ import { supervisedSession } from './session-supervision/discovery-window.js';
3
5
  import { createClaudeDesktopArchiveController } from './session-supervision/claude-desktop-archive.js';
4
6
  import { createClaudeDesktopNativeAdapter } from './session-supervision/claude-desktop-native.js';
5
7
  import { applyControllerExecutionContext } from './session-supervision/execution-context.js';
@@ -118,9 +120,9 @@ function desktopActionError(action, error) {
118
120
  const messages = {
119
121
  session_workspace_changed: "The conversation workspace changed. Refresh and select it again.",
120
122
  session_metadata_unavailable: "Conversation details are unavailable. Refresh this provider and select the conversation again.",
121
- session_model_unavailable: "This provider did not expose the conversation model; automatic continuation is unavailable.",
123
+ session_model_unavailable: "This provider did not expose the conversation model; Delivery Loop is unavailable.",
122
124
  existing_loop: "This conversation already has a Delivery Loop. Open its existing supervision instead.",
123
- task_settled: "This task has finished. Start a new conversation before enabling automatic continuation.",
125
+ task_settled: "This task has finished. Start a new conversation before turning on Delivery Loop.",
124
126
  };
125
127
  if (messages[error?.code]) return messages[error.code];
126
128
  }
@@ -203,6 +205,7 @@ export async function startDesktopHost({ input = process.stdin, output = process
203
205
  status = { ...status, update: { ...status.update,
204
206
  auto_update_supported: updateSettings.auto_update_supported === true,
205
207
  auto_update_enabled: updateSettings.auto_update_enabled === true,
208
+ auto_update_unavailable_reason: updateSettings.auto_update_unavailable_reason || null,
206
209
  auto_update_busy: updateSettings.auto_update_busy === true } };
207
210
  if (env.SIMY_REMOTE_CHILD === "1") status = { ...status, remoteSessionProtocol: 1, remoteControl: { version: 2, actions: [...allowedActions] } };
208
211
  ownerTransport?.publish({ type: 'status', status: { ...status, remoteSessionProtocol: 1, remoteControl: { version: 2, actions: [...allowedActions] } } });
@@ -480,6 +483,13 @@ export async function startDesktopHost({ input = process.stdin, output = process
480
483
  })().finally(() => { emailSync = null; });
481
484
  await emailSync;
482
485
  };
486
+ const saveEmailDeliveryObservation = (controller, completion, observation) => controllerPreferences.change(controller.controllerScope, state => {
487
+ const identity = JSON.stringify([completion.host_id || 'local', completion.provider, completion.provider_session_id]);
488
+ const row = state.conversations[identity] ||= {};
489
+ row.emailDeliveries ||= {};
490
+ row.emailDeliveries[completion.task_id] = { ...observation, checkedAt: Date.now(), provider: completion.provider,
491
+ provider_session_id: completion.provider_session_id, host_id: completion.host_id || 'local' };
492
+ });
483
493
  const submitNativeDevelopmentReportCompletions = ({ completions, controller }) => {
484
494
  const { scope, deviceSession } = controller;
485
495
  if (!deviceSession?.token || !deviceSession?.api_base_url || !sameReportController(controller, nativeDevelopmentReportsController) || nativeDevelopmentReports.state !== "ready" || nativeDevelopmentReportsScope !== scope) return;
@@ -506,23 +516,31 @@ export async function startDesktopHost({ input = process.stdin, output = process
506
516
  if (receipt && nativeDevelopmentReportsGeneration === generation && accessScope(currentAccess) === scope) {
507
517
  if (receipt.status === 'pending') nativeCompletionRetryAt.set(key, Date.now() + 30_000);
508
518
  else { nativeCompletionSuccess.add(key); nativeCompletionRetryAt.delete(key); }
509
- const receiptScope = controller.controllerScope;
510
- await controllerPreferences.change(receiptScope, state => {
511
- const identity = JSON.stringify([hostId, completion.provider, completion.provider_session_id]);
512
- const row = state.conversations[identity] ||= {};
513
- row.emailDeliveries ||= {};
514
- row.emailDeliveries[completion.task_id] = { ...receipt, checkedAt: Date.now(), provider: completion.provider, provider_session_id: completion.provider_session_id, host_id: hostId };
515
- });
519
+ await saveEmailDeliveryObservation(controller, completion, receipt);
516
520
  await requireReportController(controller);
517
521
  if (!closed) void publish().catch(() => {});
518
522
  }
519
523
  return receipt;
520
- }).catch(error => {
524
+ }).catch(async error => {
521
525
  if (nativeDevelopmentReportsGeneration === generation && accessScope(currentAccess) === scope) {
522
526
  // These are immutable authorization/contract rejections. A new
523
527
  // enrollment revision or scope produces a fresh rejection identity.
524
- if ([400, 403, 409].includes(error?.status)) nativeCompletionRejected.add(rejectionKey);
528
+ const rejected = [400, 403, 409].includes(error?.status);
529
+ if (rejected) nativeCompletionRejected.add(rejectionKey);
525
530
  else nativeCompletionRetryAt.set(key, Date.now() + 30_000);
531
+ // ON is consent, not a delivery receipt. Surface a failed submission
532
+ // without exposing backend errors, credentials or completion proofs.
533
+ // The next confirmed server receipt replaces this local observation.
534
+ try {
535
+ await requireReportController(controller);
536
+ if (closed || nativeDevelopmentReportsGeneration !== generation || accessScope(currentAccess) !== scope) return null;
537
+ await saveEmailDeliveryObservation(controller, completion, {
538
+ task_id: completion.task_id, enrollment_revision: completion.enrollment_revision,
539
+ status: rejected ? 'rejected' : 'retrying',
540
+ });
541
+ await requireReportController(controller);
542
+ if (!closed) void publish().catch(() => {});
543
+ } catch { /* A changed controller must not receive an old failure. */ }
526
544
  }
527
545
  return null;
528
546
  }).finally(() => nativeCompletionInFlight.delete(key));
@@ -568,6 +586,7 @@ export async function startDesktopHost({ input = process.stdin, output = process
568
586
  let commands;
569
587
  let sessionStatus = { sessions: [], providers: { codex: { discovery: "loading", reason: null }, claude: { discovery: "loading", reason: null } }, refreshedAt: null, initialLoading: true };
570
588
  let sessionRefresh = null, sessionRefreshedAt = 0, sessionRefreshAgain = false, discoveryController = null, discoveryStopped = false;
589
+ const catalogRefresh = createCatalogRefreshPolicy();
571
590
  let catalogRevision = 0;
572
591
  let enrichmentController = null, sessionEnrichment = null;
573
592
  let catalogAnalysisJob = null, catalogAnalysisAgain = false;
@@ -585,7 +604,9 @@ export async function startDesktopHost({ input = process.stdin, output = process
585
604
  const records = await recordReader.list();
586
605
  if (closed || relay || discoveryStopped || manager !== analysisManager ||
587
606
  !desktopAccessReady(currentAccess) || accessScope(currentAccess) !== scope || !coordinator.acceptsNewWork()) return;
588
- await manager.update(analysisSessions(withCurrentSettings(sessionStatus.sessions, records)));
607
+ await manager.update(analysisSessions(withCurrentSettings(sessionStatus.sessions, records)), {
608
+ priorityKeys: new Set(records.filter(supervisedSession).map(record => `${record.provider}:${record.session}`)),
609
+ });
589
610
  }
590
611
  }).catch(() => {}).finally(() => {
591
612
  catalogAnalysisJob = null;
@@ -640,12 +661,14 @@ export async function startDesktopHost({ input = process.stdin, output = process
640
661
  unfinishedDesktopWork = false;
641
662
  })();
642
663
  const observePullRequests = dependencies.observePullRequests || createPullRequestObserver();
643
- const refreshSessions = (force = false) => {
664
+ const refreshSessions = (force = false, { history = false } = {}) => {
665
+ if (history) catalogRefresh.invalidate();
644
666
  if (discoveryStopped && !closed && !relay) return sessionStatus;
645
667
  if (closed || relay || !desktopAccessReady(currentAccess)) return { sessions: [], providers: {}, refreshedAt: null };
646
668
  if (force && sessionRefresh) sessionRefreshAgain = true;
647
669
  if (!sessionRefresh && (force || Date.now() - sessionRefreshedAt >= 15_000)) {
648
670
  const requestedScope = accessScope(currentAccess);
671
+ const refresh = catalogRefresh.begin(JSON.stringify([requestedScope, providerLinks]));
649
672
  const controller = discoveryController = new AbortController();
650
673
  const attempt = { id: randomUUID(), requestedAt: Date.now(), state: 'pending' };
651
674
  sessionStatus = { ...sessionStatus, discoveryAttempt: attempt };
@@ -659,7 +682,7 @@ export async function startDesktopHost({ input = process.stdin, output = process
659
682
  }
660
683
  const complete = new Set(value.completeProviders || (value.historyLoading === undefined ? PROVIDER_LINKS : []));
661
684
  const merge = (previous, incoming) => {
662
- const rows = new Map(previous.filter(item => !complete.has(item.provider)).map(item => [`${item.provider}:${item.session}`, item]));
685
+ const rows = new Map(previous.filter(item => refresh.recentOnly || !complete.has(item.provider)).map(item => [`${item.provider}:${item.session}`, item]));
663
686
  for (const item of incoming || []) rows.set(`${item.provider}:${item.session}`, item);
664
687
  return [...rows.values()];
665
688
  };
@@ -674,7 +697,7 @@ export async function startDesktopHost({ input = process.stdin, output = process
674
697
  const firstPageTimeoutMs = dependencies.discoveryTimeoutMs ?? 30_000;
675
698
  const historyTimeoutMs = dependencies.historyDiscoveryTimeoutMs ?? dependencies.discoveryTimeoutMs ?? 300_000;
676
699
  sessionRefresh = boundedDiscovery((signal, startHistory) => (dependencies.discoverSessions || discoverSessions)({
677
- root: deliveryRoot, signal, onDrain: trackDiscoveryDrain, includeArchived: true, includeSelectionCandidates: true,
700
+ root: deliveryRoot, signal, onDrain: trackDiscoveryDrain, includeArchived: true, includeSelectionCandidates: true, recentOnly: refresh.recentOnly,
678
701
  readRecords: () => recordReader.list(),
679
702
  connectedProviders: PROVIDER_LINKS.filter(provider => providerLinks[provider].connected),
680
703
  timeoutMs: firstPageTimeoutMs, historyTimeoutMs,
@@ -685,6 +708,7 @@ export async function startDesktopHost({ input = process.stdin, output = process
685
708
  await applyCatalog(value);
686
709
  if (closed || relay || controller.signal.aborted) return;
687
710
  const failed = Object.values(value.providers || {}).some(provider => provider.discovery === 'unavailable');
711
+ refresh.complete(!failed);
688
712
  sessionStatus = { ...sessionStatus, historyLoading: false,
689
713
  discoveryAttempt: { ...attempt, state: failed ? 'failed' : 'succeeded', completedAt: Date.now() } };
690
714
  // Publish before PR/network observation and queue maintenance. Neither
@@ -1507,7 +1531,7 @@ export async function startDesktopHost({ input = process.stdin, output = process
1507
1531
  if(['detach_update','update_reconnect'].includes(action)) {
1508
1532
  const access=await accessSnapshot(); requireDesktopAccess(access);
1509
1533
  updateRuntime.validate(command,coordinator.snapshot(),updater?.snapshot?.());
1510
- if(action==='detach_update')await updater.maybeApply({detach:true});
1534
+ if(action==='detach_update')await updater.maybeApply({detach:true,manual:true});
1511
1535
  else await reconnectWithHistoryRecovery({root,coordinator,eligible:()=>updateRuntime.historyRecoveryEligible()});
1512
1536
  emit({type:'result',id:command.id,ok:coordinator.snapshot().state==='connected' && (action!=='detach_update' || updater.snapshot().state==='updated')});
1513
1537
  await publish();return;
@@ -1578,7 +1602,11 @@ export async function startDesktopHost({ input = process.stdin, output = process
1578
1602
  if (['detach_update', 'update_reconnect'].includes(action)) requireDesktopAccess(access);
1579
1603
  else if (!recoveryControl && desktopAccessReady(access) && !desktopActionAllowed({ state: 'unavailable' }, command)) await authorizeWork();
1580
1604
 
1581
- if (!coordinator.acceptsNewWork() && !recoveryControl && !["refresh", "reconnect", "task_detail", "check_update", "quit", "delivery_stop", "delivery_pause", "delivery_reject"].includes(action)) {
1605
+ // Signing back into the same SIMY environment is a prerequisite for
1606
+ // recovery, not permission to resume work. Keep organization changes
1607
+ // and execution fenced until the original checkpoint is reconciled.
1608
+ const recoveryAuthorization = ["login", "authorization_open", "authorization_copy", "organization_cancel"].includes(action);
1609
+ if (!coordinator.acceptsNewWork() && !recoveryControl && !recoveryAuthorization && !["refresh", "reconnect", "task_detail", "check_update", "quit", "delivery_stop", "delivery_pause", "delivery_reject"].includes(action)) {
1582
1610
  throw new Error("The update connection requires recovery.");
1583
1611
  }
1584
1612
  if (['team_threads_refresh','team_thread_open','team_thread_share'].includes(action)) {
@@ -1890,7 +1918,7 @@ export async function startDesktopHost({ input = process.stdin, output = process
1890
1918
  else if (desktopAccessReady(access)) await agent.refreshCodexStatus?.();
1891
1919
  await agent.refreshDesktopStatus();
1892
1920
  await refreshPlugins();
1893
- refreshSessions(true);
1921
+ refreshSessions(true, { history: true });
1894
1922
  if (desktopAccessReady(currentAccess) && coordinator.acceptsNewWork()) void actionAssociations.tick().then(publish).catch(() => {});
1895
1923
  }
1896
1924
  else if (action === "organizations_refresh") await agent.refreshOrganizations();
@@ -2011,7 +2039,7 @@ export async function startDesktopHost({ input = process.stdin, output = process
2011
2039
  await authorizeWork(commandScope);
2012
2040
  await actionAssociations.correct(command, analysisManager.project(candidate));
2013
2041
  } else if (action === "session_set_auto_continue") {
2014
- if (typeof command.enabled !== 'boolean') throw new Error('Choose whether automatic continuation is enabled.');
2042
+ if (typeof command.enabled !== 'boolean') throw new Error('Choose whether Delivery Loop is on.');
2015
2043
  let admission, target, providerEpoch;
2016
2044
  if (command.enabled) {
2017
2045
  const candidate = await selectedSession(command, { preferObserved: true });
@@ -2204,7 +2232,7 @@ export async function startDesktopHost({ input = process.stdin, output = process
2204
2232
  if (fresh.cwd !== candidate.cwd) throw new Error('The conversation workspace changed. Refresh and retry.');
2205
2233
  const record = await readRecord(deliveryRoot, 'claude', command.session);
2206
2234
  if (fresh.isRunning || fresh.autoContinue || record?.autoContinue || record?.run && !TERMINAL.has(record.run.state))
2207
- throw Object.assign(new Error('Stop this conversation and turn off automatic continuation before organizing it.'), { code: 'claude_desktop_running' });
2235
+ throw Object.assign(new Error('Stop this conversation and turn off Delivery Loop before organizing it.'), { code: 'claude_desktop_running' });
2208
2236
  };
2209
2237
  await verifySelection();
2210
2238
  const operation = new AbortController();
@@ -2227,7 +2255,7 @@ export async function startDesktopHost({ input = process.stdin, output = process
2227
2255
  } finally { clearInterval(watchdog); }
2228
2256
 
2229
2257
  await recordProviderArchive(deliveryRoot, 'claude', command.session, archived, Date.now());
2230
- refreshSessions(true);
2258
+ refreshSessions(true, { history: true });
2231
2259
  emit({ type: 'result', id: command.id, ok: true, data });
2232
2260
  void publish().catch(() => {});
2233
2261
  return;
@@ -2245,7 +2273,9 @@ export async function startDesktopHost({ input = process.stdin, output = process
2245
2273
  listHidden: archived,
2246
2274
  }));
2247
2275
  }
2248
- refreshSessions(true);
2276
+ // Archive/restore can move an old conversation outside both recent
2277
+ // pages. Explicit organization must be visible immediately.
2278
+ refreshSessions(true, { history: true });
2249
2279
  } else if (action === "session_track" || action === "session_untrack") {
2250
2280
  const candidate = await selectedSession(command);
2251
2281
  if (candidate.capabilities?.canMonitor !== true) throw new Error('Tracking is unavailable for this conversation.');
@@ -2380,7 +2410,7 @@ function sessionActionError(error, action) {
2380
2410
  if (error instanceof AutoContinueUnavailableError) return error.message.replace(/[\r\n\x00-\x1f]/g, ' ').slice(0, 600);
2381
2411
  const message = String(error?.message || "");
2382
2412
  if (/already has a Delivery Loop|budgets cannot be reset/i.test(message)) return "This conversation already has a task. Open its existing supervision instead.";
2383
- if (/model.*unavailable|did not expose.*model/i.test(message)) return "This provider did not expose the conversation model; automatic continuation is unavailable.";
2413
+ if (/model.*unavailable|did not expose.*model/i.test(message)) return "This provider did not expose the conversation model; Delivery Loop is unavailable.";
2384
2414
  if (/workspace changed|metadata.*unavailable|session.*changed/i.test(message)) return "The conversation identity or workspace changed. Refresh and select it again.";
2385
2415
  if (/objective|acceptance|autonomy policy|budget/i.test(message)) return message.replace(/[\r\n\x00-\x1f]/g, " ").slice(0, 240);
2386
2416
  if (/question changed|answer is empty|session understanding|context/i.test(message)) return message.replace(/[\r\n\x00-\x1f]/g, " ").slice(0, 240);
package/src/index.js CHANGED
@@ -161,7 +161,7 @@ async function runCommand(installMode) {
161
161
  Options:
162
162
  --version, -v Print the installed CLI version and exit
163
163
  --daemon, --deamon Run detached in the background
164
- --no-tui Disable the interactive Agentic Loop console
164
+ --no-tui Disable the interactive Delivery Loop console
165
165
  --no-open Do not open the browser when authorization is required
166
166
  --port <port> Bind a specific localhost port
167
167
  --host <url> Connect to a SIMY Web origin (default: ${DEFAULT_WEB_ORIGIN})
package/src/local-task.js CHANGED
@@ -1257,7 +1257,7 @@ function localTaskProviderInstruction(task) {
1257
1257
  ? "The managed output directory is reserved for final user-requested deliverables. Keep helper scripts, temporary files, caches, and intermediate files outside it (use the task workspace or an OS temporary directory), and remove those helper files before completion."
1258
1258
  : null;
1259
1259
  return [
1260
- "You are the local executor for a SIMY Local Task, not an Agentic Loop.",
1260
+ "You are the local executor for a SIMY Local Task, not a Delivery Loop.",
1261
1261
  buildWorkflowReportingInstruction(),
1262
1262
  "Complete this bounded task once. Do not create Git branches, commits, pull requests, audits, or automatic retries.",
1263
1263
  `Permission mode: ${task.permission_mode}.`,
@@ -117,7 +117,7 @@ export function summarizeCodingLoopEvent(snapshot, event) {
117
117
  };
118
118
  case "stopped":
119
119
  return {
120
- summary: "The local human operator stopped this Agentic Loop run.",
120
+ summary: "The local human operator stopped this Delivery Loop run.",
121
121
  result: "No further executor handoff will occur for this run.",
122
122
  next: "Start a new run or explicitly continue this one when ready.",
123
123
  };
package/src/runner.js CHANGED
@@ -452,7 +452,7 @@ export async function recoverLocalCodingRun(run, dependencies = {}, now = Date.n
452
452
  }
453
453
 
454
454
  function assertRunCanContinue(run) {
455
- if (run.operation || run.child) throw new Error("The selected Agentic Loop run is still active.");
455
+ if (run.operation || run.child) throw new Error("The selected Delivery Loop run is still active.");
456
456
  if (!RESUMABLE_STATES.has(run.status)) {
457
457
  throw new Error("Human guidance is available only when the selected run is waiting.");
458
458
  }
@@ -542,7 +542,7 @@ export async function continueLocalCodingRunAfterRepositoryApproval(
542
542
  if (!approvedRepository || !approvedPath) {
543
543
  throw new Error("An approved local repository and path are required.");
544
544
  }
545
- if (run.operation || run.child) throw new Error("The selected Agentic Loop run is still active.");
545
+ if (run.operation || run.child) throw new Error("The selected Delivery Loop run is still active.");
546
546
  if (run.status !== "waiting_human") {
547
547
  throw new Error("Repository approval is available only while the run is waiting.");
548
548
  }
@@ -6,7 +6,8 @@ import { evidenceExcerpt } from '../codex-plugins/simy/hooks/purpose-evidence.mj
6
6
 
7
7
  // Read only original file predicates in the authorized task directory. These
8
8
  // are current independent controller reads, not worker claims or authored proof.
9
- export async function captureAcceptanceReads(cwd, criteria, { signal } = {}) {
9
+ export async function captureAcceptanceReads(cwd, criteria, { signal, excerptLimit = 4000, publicTaskSourcePaths = [] } = {}) {
10
+ if (!Number.isSafeInteger(excerptLimit) || excerptLimit < 1 || excerptLimit > 32768) throw new Error('Acceptance excerpt bound is invalid.');
10
11
  const root = await realpath(cwd), rows = [], seen = new Set(); let retained = 0;
11
12
  for (const criterion of criteria) {
12
13
  signal?.throwIfAborted();
@@ -22,8 +23,16 @@ export async function captureAcceptanceReads(cwd, criteria, { signal } = {}) {
22
23
  if (bytesRead > 32768) continue;
23
24
  const body = data.subarray(0, bytesRead); retained += bytesRead;
24
25
  const digest = createHash('sha256').update(body).digest('hex');
26
+ // Host-only, explicit source-read permission for a native task reviewer.
27
+ // The same original Claude already has Read access to these public-task
28
+ // source paths. Business identifiers such as Token are source, not a
29
+ // credential claim. Generic artifact reads retain their redaction policy.
30
+ const sourceRead = publicTaskSourcePaths.some(file => resolve(root, file) === path);
31
+ const sourceText = bytesRead ? body.toString('utf8') : JSON.stringify({ byteLength: 0, content: '' });
32
+ const observation = sourceRead ? { text: sourceText.slice(0, excerptLimit), truncated: sourceText.length > excerptLimit,
33
+ redacted: false, policy: 'host_authorized_public_task_source' } : evidenceExcerpt(sourceText, excerptLimit);
25
34
  rows.push({ id: `file:${target}`, source: 'controller_file_read', status: 'observed', at: Date.now(), digest,
26
- recipe: { type: 'file_read', path: target }, observation: evidenceExcerpt(bytesRead ? body.toString('utf8') : JSON.stringify({ byteLength: 0, content: '' }), 4000) });
35
+ recipe: { type: 'file_read', path: target }, observation });
27
36
  } catch { signal?.throwIfAborted(); /* Missing/large/unreadable files stay unverified; tool investigation remains available. */ }
28
37
  finally { await handle?.close(); }
29
38
  }
@@ -4,6 +4,7 @@ import { lstat, mkdir, readFile, rename, rm, writeFile } from 'node:fs/promises'
4
4
  import { join } from 'node:path';
5
5
  import { analyzeContinuation, CONTINUATION_ANALYSIS_VERSION } from './continuation-analysis.js';
6
6
  import { providerGoalEvidence } from './conversation-context.js';
7
+ import { liveSessionWindow, supervisedSession } from './discovery-window.js';
7
8
 
8
9
  const BACKFILL_LIMIT = 10;
9
10
  const MAX_CACHE_ENTRIES = 2000;
@@ -276,11 +277,17 @@ export async function createContinuationAnalysisManager({ root, analyze = analyz
276
277
  });
277
278
  startDrain();
278
279
  },
279
- async update(sessions) {
280
+ async update(sessions, { priorityKeys = new Set() } = {}) {
280
281
  await exclusive(async () => {
281
282
  if (closed || suspended) return;
283
+ // The complete catalog remains navigable. Historical metadata outside
284
+ // this window must not churn the bounded cache or schedule paid work.
285
+ // Preserve queued/manual jobs and explicitly supervised conversations;
286
+ // removal, changed authority and OFF still use the incoming snapshot.
287
+ const window = liveSessionWindow(sessions, { priority: candidate => priorityKeys.has(keyOf(candidate)) || supervisedSession(candidate) ||
288
+ jobs.has(keyOf(candidate)) || ['queued', 'analyzing'].includes(entries.get(keyOf(candidate))?.result?.state) });
282
289
  candidates.clear();
283
- for (const candidate of sessions) candidates.set(keyOf(candidate), observedCandidate(candidate));
290
+ for (const candidate of window) candidates.set(keyOf(candidate), observedCandidate(candidate));
284
291
  const sorted = [...candidates.values()].sort((a, b) => activityOf(b) - activityOf(a) || keyOf(a).localeCompare(keyOf(b)));
285
292
  const backfill = document.initialized ? new Set() : new Set(sorted.slice(0, BACKFILL_LIMIT).map(keyOf));
286
293
  let dirty = !document.initialized;
@@ -6,15 +6,15 @@ function continuationFailureReason(error) {
6
6
  // Provider exceptions can contain private paths/output. Only known codes
7
7
  // select actionable UI copy; unrecognized failures retain a safe fallback.
8
8
  const reasons = {
9
- desktop_access_required: 'Sign in to an organization with SQM and Developer access before auto-continue can run.',
10
- provider_sign_in_required: 'Sign in to the original AI provider, then retry auto-continue.',
11
- unavailable: 'The local provider is unavailable. Open or configure it, then retry auto-continue.',
12
- unsupported: 'Update the original AI provider to a compatible version, then retry auto-continue.',
9
+ desktop_access_required: 'Sign in to an organization with SQM and Developer access before Delivery Loop can run.',
10
+ provider_sign_in_required: 'Sign in to the original AI provider, then retry Delivery Loop.',
11
+ unavailable: 'The local provider is unavailable. Open or configure it, then retry Delivery Loop.',
12
+ unsupported: 'Update the original AI provider to a compatible version, then retry Delivery Loop.',
13
13
  process_exited: 'The original Claude process has exited. Return to the original conversation, then refresh.',
14
14
  session_ended: 'This Claude session has ended. Return to the original conversation, then refresh.',
15
15
  hook_missing: 'Waiting for the original Claude hook. SIMY recovers automatically when the next verified prompt or response-end hook arrives. If no notifications arrive, check the SIMY plugin in the original Claude window.',
16
16
  };
17
- return Object.hasOwn(reasons, error?.code) ? reasons[error.code] : 'Could not start auto-continue. Check the original conversation and retry.';
17
+ return Object.hasOwn(reasons, error?.code) ? reasons[error.code] : 'Could not start Delivery Loop. Check the original conversation and retry.';
18
18
  }
19
19
 
20
20
  // Shared by explicit Continue and automatic enrollment. Analysis alone never
@@ -24,7 +24,7 @@ export async function continueAnalyzedSession({ candidate, manager, root, settin
24
24
  const link = await requireProviderLink({ root, provider: candidate.provider, expectedEpoch: settings.expectedProviderEpoch });
25
25
  const recommendationOptions = { allowDiagnosis: settings.purposeDriven === true && candidate.autoContinue === true };
26
26
  const recommendation = manager.requireRecommendation(candidate, recommendationOptions);
27
- if (candidate.isRunning || candidate.capabilities?.canContinue !== true) throw new Error('Automatic continuation is unavailable for this conversation.');
27
+ if (candidate.isRunning || candidate.capabilities?.canContinue !== true) throw new Error('Delivery Loop is unavailable for this conversation.');
28
28
  const context = await readContext(candidate, { signal });
29
29
  if (context.awaitingTurn || context.incomplete || !recommendation.contextDigest || context.contextDigest !== recommendation.contextDigest) {
30
30
  await manager.invalidate(candidate);
@@ -34,7 +34,7 @@ export async function continueAnalyzedSession({ candidate, manager, root, settin
34
34
  await beforeEnroll(candidate);
35
35
  await requireProviderLink({ root, provider: candidate.provider, expectedEpoch: link.epoch });
36
36
  manager.requireRecommendation(candidate, recommendationOptions);
37
- if (signal?.aborted) throw new Error('Automatic continuation was cancelled before enrollment.');
37
+ if (signal?.aborted) throw new Error('Delivery Loop was cancelled before enrollment.');
38
38
  return enroll({ root, provider: candidate.provider, session: candidate.session, ...settings, expectedProviderEpoch: link.epoch });
39
39
  }
40
40
 
@@ -46,7 +46,7 @@ function startingPresentation(candidate, status, reason) {
46
46
  const canPresent = work && !work.needsAttention && !['running', 'stopped', 'archived', 'approval', 'question'].includes(work.state);
47
47
  return { ...candidate, autoContinueStatus: status, autoContinueReason: reason,
48
48
  ...(canPresent ? { workState: { ...work, state: status,
49
- label: status === 'starting' ? '自動継続を準備中' : '自動復旧中', reason,
49
+ label: status === 'starting' ? 'Delivery Loop を準備中' : '自動復旧中', reason,
50
50
  nextAction: null, needsAttention: false } } : {}) };
51
51
  }
52
52
 
@@ -133,7 +133,7 @@ export function createAutomaticContinuation({ root, getManager, selectCandidate,
133
133
  retries.set(key, { attempt, count, at: now() + Math.min(60_000, 1000 * 2 ** Math.min(count - 1, 6)) });
134
134
  } else retries.delete(key);
135
135
  errors.set(key, { epoch: candidate.autoContinueEpoch, recoverable,
136
- reason: recoverable ? 'The AI connection is temporarily unavailable. Automatic continuation will retry when the connection is available.' : continuationFailureReason(error) });
136
+ reason: recoverable ? 'The AI connection is temporarily unavailable. Delivery Loop will retry when the connection is available.' : continuationFailureReason(error) });
137
137
  }
138
138
  }).finally(() => {
139
139
  if (controller.signal.aborted) { attempted.delete(key); retries.delete(key); }
@@ -152,14 +152,14 @@ export function createAutomaticContinuation({ root, getManager, selectCandidate,
152
152
  if (candidate.autoContinue === true && candidate.capabilities?.canAutoContinue === true &&
153
153
  !['managed', 'pending_takeover'].includes(candidate.binding?.mode)) {
154
154
  if (pending.has(key)) return startingPresentation(candidate, 'starting',
155
- 'Auto-continue is saved. Preparing to continue this conversation in the background.');
155
+ 'Delivery Loop is on. Preparing Delivery Loop for this conversation in the background.');
156
156
  const failure = errors.get(key);
157
157
  if (failure?.recoverable && failure.epoch === candidate.autoContinueEpoch)
158
158
  return startingPresentation(candidate, 'recovering', failure.reason);
159
159
  }
160
160
  if (candidate.autoContinue === true && candidate.isRunning !== true &&
161
161
  ['unmanaged', 'monitor'].includes(candidate.binding?.mode || 'unmanaged') && candidate.analysis?.state === 'needs_user') return {
162
- ...candidate, autoContinueStatus: 'needs_user', autoContinueReason: candidate.analysis.reason || 'Your decision is needed before automatic continuation can proceed.',
162
+ ...candidate, autoContinueStatus: 'needs_user', autoContinueReason: candidate.analysis.reason || 'Your decision is needed before Delivery Loop can proceed.',
163
163
  };
164
164
  const error = errors.get(`${candidate.provider}:${candidate.session}`);
165
165
  if (error && candidate.autoContinue === true && candidate.autoContinueEpoch === error.epoch &&
@@ -0,0 +1,16 @@
1
+ // History is a navigation index, not a real-time subscription. Keep ordinary
2
+ // status/preference refreshes cheap; an explicit Refresh still reloads history.
3
+ export function createCatalogRefreshPolicy({ now = Date.now, intervalMs = 300_000 } = {}) {
4
+ let completed = null, scope = null, revision = 0;
5
+ return {
6
+ invalidate() { completed = null; revision++; },
7
+ begin(nextScope) {
8
+ if (scope !== nextScope) { scope = nextScope; completed = null; revision++; }
9
+ const generation = revision;
10
+ const recentOnly = completed !== null && now() - completed < intervalMs;
11
+ return { recentOnly, complete(success) {
12
+ if (!recentOnly && success && generation === revision) completed = now();
13
+ } };
14
+ },
15
+ };
16
+ }
@@ -0,0 +1,62 @@
1
+ import { isDeepStrictEqual } from 'node:util';
2
+
3
+ // Observed foreground-only Agent contract. Omission is never evidence of
4
+ // foreground execution when the provider exposes background or unknown options.
5
+ const FOREGROUND_SCHEMA = {
6
+ $schema: 'https://json-schema.org/draft/2020-12/schema', type: 'object',
7
+ properties: { description: { type: 'string' }, prompt: { type: 'string' },
8
+ subagent_type: { type: 'string' }, isolation: { type: 'string', enum: ['worktree', 'remote'] } },
9
+ required: ['description', 'prompt'], additionalProperties: false,
10
+ };
11
+ const withoutDescriptions = value => Array.isArray(value) ? value.map(withoutDescriptions) :
12
+ value && typeof value === 'object' ? Object.fromEntries(Object.entries(value)
13
+ .filter(([key, item]) => key !== 'description' || typeof item !== 'string')
14
+ .map(([key, item]) => [key, withoutDescriptions(item)])) : value;
15
+
16
+ export function isForegroundOnlyAgentSchema(schema) {
17
+ return isDeepStrictEqual(withoutDescriptions(schema), FOREGROUND_SCHEMA);
18
+ }
19
+
20
+ export function isForegroundNativeReviewInput(toolName, args, observedSchema = null) {
21
+ if (!['Agent', 'Task'].includes(toolName) || !args || typeof args !== 'object' || Array.isArray(args) ||
22
+ typeof args.prompt !== 'string' || typeof args.subagent_type !== 'string' ||
23
+ Object.keys(args).some(key => !['description', 'prompt', 'subagent_type', 'run_in_background'].includes(key)) ||
24
+ Object.hasOwn(args, 'description') && typeof args.description !== 'string') return false;
25
+ if (Object.hasOwn(args, 'run_in_background')) return args.run_in_background === false;
26
+ // The observed Agent schema cannot establish an omitted flag for legacy Task.
27
+ return toolName === 'Agent' && typeof args.description === 'string' && isForegroundOnlyAgentSchema(observedSchema);
28
+ }
29
+
30
+ // Callers supply rows read from the exact owned, stable parent transcript,
31
+ // never a worker-provided schema. A schema first seen after the call proves nothing.
32
+ export function foregroundAgentContract(rows, candidate, beforeIndex = rows.length) {
33
+ const collector = createForegroundAgentContractCollector(candidate);
34
+ for (const row of rows.slice(0, beforeIndex)) collector.add(row);
35
+ return collector.finish();
36
+ }
37
+
38
+ // Keep the prior schema and version agreement without retaining transcript
39
+ // bodies. A snapshot taken before a call cannot be authorized by a later shape.
40
+ export function createForegroundAgentContractCollector(candidate) {
41
+ let shape = null, shapes = 0, version, mixedVersions = false;
42
+ return {
43
+ add(row) {
44
+ if (row.type === 'api-request-shape') { shapes++; if (shapes === 1) shape = row; }
45
+ if (row.version !== undefined) {
46
+ mixedVersions ||= version !== undefined && version !== row.version;
47
+ version = row.version;
48
+ }
49
+ },
50
+ finish() {
51
+ const row = shape;
52
+ if (shapes !== 1 || mixedVersions || version !== row.version || row.sessionId !== candidate.session || row.isSidechain === true ||
53
+ row.cwd !== undefined && row.cwd !== candidate.cwd ||
54
+ !/^[a-f0-9]{64}$/.test(row.shapeHash || '') || !/^\d+\.\d+\.\d+$/.test(row.version || '') ||
55
+ !Array.isArray(row.shape?.tools)) return null;
56
+ const agents = row.shape.tools.filter(tool => tool.name === 'Agent');
57
+ if (agents.length !== 1 || !isForegroundOnlyAgentSchema(agents[0].input_schema)) return null;
58
+ return { schema: agents[0].input_schema, shapeHash: row.shapeHash,
59
+ session: candidate.session, cwd: candidate.cwd, version: row.version ?? null };
60
+ },
61
+ };
62
+ }
@@ -7,7 +7,7 @@ import { basename, relative, resolve } from 'node:path';
7
7
  import { runStructuredProviderReview } from '../sqm/review-worker.js';
8
8
  import { captureAcceptanceReads } from './acceptance-reads.js';
9
9
  import { readClaudeConversation, readClaudeNativeReview } from './claude-sessions.js';
10
- import { independentRequestContext, workspaceIdentity } from './purpose-verification.js';
10
+ import { independentReviewRequestContext, workspaceIdentity } from './purpose-verification.js';
11
11
  import { effectiveAcceptanceCriteria } from '../codex-plugins/simy/hooks/acceptance-policy.mjs';
12
12
  import { liveOutcomeReviewEligible, settleLiveOutcomeReview, readRecord, taskDirectory, hasPendingUserStop, providerLinkState, processIdentity } from '../codex-plugins/simy/hooks/delivery-loop.mjs';
13
13
 
@@ -52,7 +52,7 @@ export async function claudeReviewBudgetBoundary(record, { platform = process.pl
52
52
  // original Claude process waits for its CLI child and retains its writer lease.
53
53
  // No second controller/session resumes or writes the task workspace.
54
54
  export async function reviewClaudeCheckpoint(record, { root, review = runStructuredProviderReview,
55
- readContext = readClaudeConversation, readNativeReview = readClaudeNativeReview, onFailure = () => {}, captureReads = captureAcceptanceReads, inspectBudget = claudeReviewBudgetBoundary, signal, executionTimeoutMs = 120000 } = {}) {
55
+ readContext = readClaudeConversation, readNativeReview = readClaudeNativeReview, onFailure = () => {}, captureReads = captureAcceptanceReads, fingerprintWorkspace = workspaceIdentity, inspectBudget = claudeReviewBudgetBoundary, signal, executionTimeoutMs = 120000 } = {}) {
56
56
  if (!liveOutcomeReviewEligible(record)) return record;
57
57
  const snapshot = structuredClone(record), cwd = taskDirectory(record), criteria = effectiveAcceptanceCriteria(record.run.managed);
58
58
  const abort = new AbortController();
@@ -77,17 +77,18 @@ export async function reviewClaudeCheckpoint(record, { root, review = runStructu
77
77
  summary: 'Independent review was not dispatched: the original Claude process budget arguments could not be verified. Acceptance remains unverified; inspect the original launch configuration without changing its budget.',
78
78
  findings: [], source: 'controller_review_unavailable', checkedAt: Date.now() } });
79
79
  const inventory = [];
80
- const before = await workspaceIdentity(cwd, undefined, inventory, { signal: combined, inventoryLimit: 1000 });
80
+ const before = await fingerprintWorkspace(cwd, undefined, inventory, { signal: combined, inventoryLimit: 1000 });
81
81
  const reads = await captureReads(cwd, criteria, { signal: combined });
82
82
  // This executor cannot independently inspect unnamed artifacts or services.
83
83
  // Never turn absent/truncated input into a successful semantic judgment.
84
84
  if (!reads.length || reads.some(row => row.observation.truncated || row.observation.redacted)) return await unavailable('Current outcome observations are absent or incomplete.');
85
85
  if (reads.some(row => !inventory.some(file => file.type === 'file' && file.path === relative(cwd, resolve(cwd, row.recipe.path)) && file.digest === row.digest)))
86
86
  return await unavailable('Controller observations do not match the reviewed workspace revision; reread the actual outcome.');
87
- const context = await readContext(record, { signal: combined, userOnly: true });
87
+ const context = await readContext(record, { signal: combined, userOnly: true, reviewUserContext: true });
88
88
  if (!context?.messages?.some(row => row.role === 'user') || context.incomplete || context.truncated) return await unavailable('Original user context could not be read completely; inspect the original request before retrying review.');
89
89
  if (!await current()) return await readRecord(root, record.provider, record.session);
90
- const requests = independentRequestContext(context);
90
+ const requests = independentReviewRequestContext(context);
91
+ if (requests.incomplete || requests.truncated || requests.redacted) return await unavailable('Original user requests exceed complete review coverage or contain redacted context; acceptance remains unverified.');
91
92
  if (budget === 'worker_cap' || budget === 'native_only') {
92
93
  const contextDigest = createHash('sha256').update(JSON.stringify(requests.messages.filter(row => row.text?.trim()))).digest('hex');
93
94
  const pending = record.run.managed.nativeOutcomeReview;
@@ -111,7 +112,7 @@ export async function reviewClaudeCheckpoint(record, { root, review = runStructu
111
112
  }
112
113
  const result = await review('claude', {
113
114
  system: CLAUDE_REVIEW_SYSTEM,
114
- prompt: JSON.stringify({ objective: record.run.objective, requests: independentRequestContext(context), criteria, workspaceInventory: { paths: inventory.map(row => row.path), truncated: inventory.truncated === true }, observations: reads }),
115
+ prompt: JSON.stringify({ objective: record.run.objective, requests, criteria, workspaceInventory: { paths: inventory.map(row => row.path), truncated: inventory.truncated === true }, observations: reads }),
115
116
  output_schema: CLAUDE_REVIEW_SCHEMA
116
117
  }, { signal: combined, executionTimeoutMs, subscriptionOnly: true, beforeDispatch: async () => {
117
118
  if (!await current()) throw new Error('Review target changed');