@evident-ai/cli 3.1.1-dev.5a3c6a5 → 3.1.1-dev.5d5a3d3

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.
package/dist/index.js CHANGED
@@ -38,11 +38,6 @@ function getApiUrl() {
38
38
  function getTunnelUrl() {
39
39
  return tunnelOverride ?? process.env.EVIDENT_TUNNEL_URL ?? defaults.tunnelUrl;
40
40
  }
41
- var config = new Conf({
42
- projectName: "evident",
43
- projectSuffix: "",
44
- defaults
45
- });
46
41
  var credentials = new Conf({
47
42
  projectName: "evident",
48
43
  projectSuffix: "",
@@ -267,16 +262,28 @@ async function getToken() {
267
262
  }
268
263
  return null;
269
264
  }
265
+ function toError(err) {
266
+ return err instanceof Error ? err : new Error(String(err));
267
+ }
270
268
  async function deleteToken(options = {}) {
271
269
  const keytar = await getKeytar();
270
+ const failures = [];
272
271
  if (keytar) {
273
272
  if (options.all) {
274
- const all = await keytar.findCredentials(SERVICE_NAME).catch(() => []);
273
+ let accounts = [];
274
+ try {
275
+ accounts = await keytar.findCredentials(SERVICE_NAME);
276
+ } catch (err) {
277
+ failures.push({ type: "enumerate", error: toError(err) });
278
+ }
275
279
  await Promise.all(
276
- all.map(
277
- (entry) => keytar.deletePassword(SERVICE_NAME, entry.account).catch(() => {
278
- })
279
- )
280
+ accounts.map(async (entry) => {
281
+ try {
282
+ await keytar.deletePassword(SERVICE_NAME, entry.account);
283
+ } catch (err) {
284
+ failures.push({ type: "delete", account: entry.account, error: toError(err) });
285
+ }
286
+ })
280
287
  );
281
288
  } else {
282
289
  await keytar.deletePassword(SERVICE_NAME, keychainAccount());
@@ -287,6 +294,7 @@ async function deleteToken(options = {}) {
287
294
  } else {
288
295
  clearCredentials();
289
296
  }
297
+ return { failures };
290
298
  }
291
299
 
292
300
  // src/utils/ui.ts
@@ -404,8 +412,10 @@ async function deviceFlowLogin(options) {
404
412
  }
405
413
  async function tokenLogin() {
406
414
  console.log("Token login mode.");
407
- console.log("Run `evident login` on a machine with a browser to get a token.");
408
- console.log("Manage or revoke existing tokens under Settings \u2192 CLI tokens.");
415
+ console.log("Create a token under Settings \u2192 CLI tokens in the dashboard, then paste it below.");
416
+ console.log(
417
+ "(Alternatively, run `evident login` on a machine with a browser, or set EVIDENT_TOKEN for CI.)"
418
+ );
409
419
  blank();
410
420
  process.stdout.write("Paste token: ");
411
421
  const token = await new Promise((resolve3) => {
@@ -429,13 +439,22 @@ async function tokenLogin() {
429
439
  printError("No token provided.");
430
440
  process.exit(1);
431
441
  }
442
+ await validateAndStoreToken(token);
443
+ }
444
+ async function validateAndStoreToken(token) {
432
445
  const spinner = ora("Validating token...").start();
433
446
  try {
434
- const result = await api.post("/auth/token/validate", { token });
447
+ const result = await api.get("/me", {
448
+ headers: { Authorization: `Bearer ${token}` }
449
+ });
450
+ if (!result.user) {
451
+ throw new Error(
452
+ "This token is not a user login (e.g. a runner key). Paste a CLI token instead."
453
+ );
454
+ }
435
455
  await storeToken({
436
456
  token,
437
- user: result.user,
438
- expiresAt: result.expires_at
457
+ user: { email: result.user.email }
439
458
  });
440
459
  spinner.stop();
441
460
  printSuccess(`Logged in as ${chalk2.bold(result.user.email)}`);
@@ -455,9 +474,22 @@ async function login(options) {
455
474
  }
456
475
 
457
476
  // src/commands/logout.ts
477
+ function describeFailure(failure) {
478
+ if (failure.type === "enumerate") {
479
+ return `could not list stored keychain entries (${failure.error.message})`;
480
+ }
481
+ return `${failure.account} (${failure.error.message})`;
482
+ }
458
483
  async function logout(options = {}) {
459
484
  if (options.all) {
460
- await deleteToken({ all: true });
485
+ const result = await deleteToken({ all: true });
486
+ if (result.failures.length > 0) {
487
+ printError(
488
+ `Failed to fully clear your keychain: ${result.failures.map(describeFailure).join("; ")}. Your local credentials file was cleared, but stale keychain entries may remain \u2014 run \`evident logout --all\` again, or remove them manually from your OS keychain / credential manager.`
489
+ );
490
+ process.exitCode = 1;
491
+ return;
492
+ }
461
493
  printSuccess("Logged out of all endpoints.");
462
494
  return;
463
495
  }
@@ -482,7 +514,9 @@ async function whoami() {
482
514
  blank();
483
515
  console.log(keyValue("Endpoint", apiUrl));
484
516
  console.log(keyValue("User", chalk3.bold(credentials2.user.email)));
485
- console.log(keyValue("User ID", credentials2.user.id));
517
+ if (credentials2.user.id) {
518
+ console.log(keyValue("User ID", credentials2.user.id));
519
+ }
486
520
  if (credentials2.expiresAt) {
487
521
  const expiresAt = new Date(credentials2.expiresAt);
488
522
  const now = /* @__PURE__ */ new Date();
@@ -498,9 +532,132 @@ async function whoami() {
498
532
  blank();
499
533
  }
500
534
 
535
+ // src/lib/claude-usage.ts
536
+ import { execFileSync } from "child_process";
537
+ import { readFileSync } from "fs";
538
+ import { homedir } from "os";
539
+ import { join } from "path";
540
+ var CLAUDE_USAGE_URL = "https://api.anthropic.com/api/oauth/usage";
541
+ var KEYCHAIN_SERVICE = "Claude Code-credentials";
542
+ function parseClaudeCliCredentials(raw) {
543
+ let parsed;
544
+ try {
545
+ parsed = JSON.parse(raw);
546
+ } catch {
547
+ return null;
548
+ }
549
+ const data = parsed.claudeAiOauth ?? parsed;
550
+ const creds = data;
551
+ if (typeof creds.accessToken !== "string" || typeof creds.expiresAt !== "number") {
552
+ return null;
553
+ }
554
+ return { accessToken: creds.accessToken, expiresAt: creds.expiresAt };
555
+ }
556
+ function readClaudeCliCredentials() {
557
+ if (process.platform === "darwin") {
558
+ try {
559
+ const raw = execFileSync(
560
+ "/usr/bin/security",
561
+ ["find-generic-password", "-s", KEYCHAIN_SERVICE, "-w"],
562
+ { encoding: "utf-8", timeout: 2e3, stdio: ["pipe", "pipe", "ignore"] }
563
+ );
564
+ return parseClaudeCliCredentials(raw);
565
+ } catch {
566
+ return null;
567
+ }
568
+ }
569
+ try {
570
+ const raw = readFileSync(join(homedir(), ".claude", ".credentials.json"), "utf-8");
571
+ return parseClaudeCliCredentials(raw);
572
+ } catch {
573
+ return null;
574
+ }
575
+ }
576
+ var ClaudeUsageError = class extends Error {
577
+ constructor(message, reason) {
578
+ super(message);
579
+ this.reason = reason;
580
+ }
581
+ };
582
+ function isLocalCredentialProblem(err) {
583
+ return err instanceof ClaudeUsageError && (err.reason === "no_credentials" || err.reason === "credentials_expired");
584
+ }
585
+ function normalizeResetsAt(value) {
586
+ const ms = Date.parse(value);
587
+ return Number.isNaN(ms) ? null : new Date(ms).toISOString();
588
+ }
589
+ function toWindow(value) {
590
+ if (!value || typeof value !== "object") {
591
+ return null;
592
+ }
593
+ const window = value;
594
+ if (typeof window.utilization !== "number" || typeof window.resets_at !== "string") {
595
+ return null;
596
+ }
597
+ const resetsAt = normalizeResetsAt(window.resets_at);
598
+ if (resetsAt === null) {
599
+ return null;
600
+ }
601
+ return { utilization: window.utilization, resetsAt };
602
+ }
603
+ async function getClaudeUsage() {
604
+ const credentials2 = readClaudeCliCredentials();
605
+ if (!credentials2) {
606
+ throw new ClaudeUsageError(
607
+ "No local Claude Code login found. Run `claude` once to sign in with your Claude subscription.",
608
+ "no_credentials"
609
+ );
610
+ }
611
+ if (credentials2.expiresAt < Date.now()) {
612
+ throw new ClaudeUsageError(
613
+ "Claude Code credentials have expired. Run `claude` to refresh them.",
614
+ "credentials_expired"
615
+ );
616
+ }
617
+ const res = await fetch(CLAUDE_USAGE_URL, {
618
+ headers: {
619
+ Authorization: `Bearer ${credentials2.accessToken}`,
620
+ "Content-Type": "application/json",
621
+ "anthropic-version": "2023-06-01"
622
+ }
623
+ });
624
+ if (!res.ok) {
625
+ throw new ClaudeUsageError(`Claude usage request failed: HTTP ${res.status}`, "request_failed");
626
+ }
627
+ const body = await res.json();
628
+ return {
629
+ fiveHour: toWindow(body.five_hour),
630
+ sevenDay: toWindow(body.seven_day)
631
+ };
632
+ }
633
+
634
+ // src/commands/claude-usage.ts
635
+ function formatWindow(label, window) {
636
+ if (!window) {
637
+ return keyValue(label, "not available for this plan");
638
+ }
639
+ const resetsAt = new Date(window.resetsAt);
640
+ return keyValue(label, `${window.utilization}% used, resets ${resetsAt.toLocaleString()}`);
641
+ }
642
+ async function claudeUsage() {
643
+ try {
644
+ const usage = await getClaudeUsage();
645
+ blank();
646
+ console.log(formatWindow("5-hour session", usage.fiveHour));
647
+ console.log(formatWindow("7-day", usage.sevenDay));
648
+ blank();
649
+ } catch (err) {
650
+ if (err instanceof ClaudeUsageError) {
651
+ printError(err.message);
652
+ process.exit(1);
653
+ }
654
+ throw err;
655
+ }
656
+ }
657
+
501
658
  // src/commands/run.ts
502
- import { homedir as homedir2 } from "os";
503
- import { isAbsolute as isAbsolute2, join as join2, parse, resolve as resolvePath } from "path";
659
+ import { homedir as homedir3 } from "os";
660
+ import { isAbsolute as isAbsolute2, join as join3, parse, resolve as resolvePath } from "path";
504
661
  import chalk6 from "chalk";
505
662
 
506
663
  // ../../packages/types/src/telemetry/index.ts
@@ -510,7 +667,10 @@ var TelemetryEventTypes = {
510
667
  AGENT_DISCONNECTED: "agent.disconnected",
511
668
  AGENT_MESSAGE_PROCESSING: "agent.message_processing",
512
669
  AGENT_MESSAGE_DONE: "agent.message_done",
513
- AGENT_MESSAGE_FAILED: "agent.message_failed"
670
+ AGENT_MESSAGE_FAILED: "agent.message_failed",
671
+ // A `warn`/`error` runner-side log line forwarded server-side for
672
+ // observability (issue #916) — see `apps/cli/src/lib/runner-activity-telemetry.ts`.
673
+ RUNNER_ACTIVITY: "runner.activity"
514
674
  };
515
675
 
516
676
  // ../../packages/types/src/tunnel/index.ts
@@ -565,6 +725,13 @@ var isShuttingDown = false;
565
725
  var FLUSH_INTERVAL_MS = 5e3;
566
726
  var MAX_BUFFER_SIZE = 50;
567
727
  var FLUSH_TIMEOUT_MS = 3e3;
728
+ var authProvider = null;
729
+ function setTelemetryAuthProvider(provider) {
730
+ authProvider = provider;
731
+ }
732
+ var FLUSH_FAILURE_LOG_INTERVAL_MS = 6e4;
733
+ var lastFlushFailureLoggedAt = 0;
734
+ var suppressedFlushFailureCount = 0;
568
735
  function logEvent(eventType, options = {}) {
569
736
  const event = {
570
737
  event_type: eventType,
@@ -599,9 +766,16 @@ async function flushEvents() {
599
766
  flushTimeout = null;
600
767
  }
601
768
  try {
602
- const credentials2 = await getToken();
603
- if (!credentials2) {
604
- return;
769
+ const providerContext = authProvider?.();
770
+ let authHeader;
771
+ if (providerContext?.authHeader) {
772
+ authHeader = providerContext.authHeader;
773
+ } else {
774
+ const credentials2 = await getToken();
775
+ if (!credentials2) {
776
+ return;
777
+ }
778
+ authHeader = `Bearer ${credentials2.token}`;
605
779
  }
606
780
  const apiUrl = getApiUrlConfig();
607
781
  const controller = new AbortController();
@@ -616,7 +790,7 @@ async function flushEvents() {
616
790
  method: "POST",
617
791
  headers: {
618
792
  "Content-Type": "application/json",
619
- Authorization: `Bearer ${credentials2.token}`
793
+ Authorization: authHeader
620
794
  },
621
795
  body: JSON.stringify(request),
622
796
  signal: controller.signal
@@ -628,8 +802,15 @@ async function flushEvents() {
628
802
  clearTimeout(timeout);
629
803
  }
630
804
  } catch (error2) {
631
- if (process.env.DEBUG) {
632
- console.error("Telemetry error:", error2);
805
+ const now = Date.now();
806
+ if (now - lastFlushFailureLoggedAt >= FLUSH_FAILURE_LOG_INTERVAL_MS) {
807
+ const message = error2 instanceof Error ? error2.message : String(error2);
808
+ const suffix = suppressedFlushFailureCount > 0 ? ` (${suppressedFlushFailureCount} more suppressed in the last ${FLUSH_FAILURE_LOG_INTERVAL_MS / 1e3}s)` : "";
809
+ console.error(`Telemetry flush error: ${message}${suffix}`);
810
+ lastFlushFailureLoggedAt = now;
811
+ suppressedFlushFailureCount = 0;
812
+ } else {
813
+ suppressedFlushFailureCount++;
633
814
  }
634
815
  }
635
816
  }
@@ -698,6 +879,69 @@ var EventTypes = {
698
879
  DEPRECATED_AGENT_KEY_ENV_USED: "cli.deprecated_agent_key_env_used"
699
880
  };
700
881
 
882
+ // src/lib/runner-activity-telemetry.ts
883
+ var FORWARDED_LEVELS = /* @__PURE__ */ new Set(["warn", "error"]);
884
+ var SEVERITY_BY_LEVEL = {
885
+ warn: "warning",
886
+ error: "error"
887
+ };
888
+ var MAX_MESSAGE_LENGTH = 500;
889
+ var TRUNCATION_MARKER = "\u2026";
890
+ function redact(message) {
891
+ return message.replace(/esk_[A-Za-z0-9_-]+/g, "esk_***").replace(/ct_[A-Za-z0-9_-]+/g, "ct_***").replace(/https?:\/\/\S+/g, "<url>");
892
+ }
893
+ function truncate(message) {
894
+ if (message.length <= MAX_MESSAGE_LENGTH) return message;
895
+ return message.slice(0, MAX_MESSAGE_LENGTH - TRUNCATION_MARKER.length) + TRUNCATION_MARKER;
896
+ }
897
+ var RATE_LIMIT_WINDOW_MS = 6e4;
898
+ var RATE_LIMIT_MAX_EVENTS = 30;
899
+ var windowStartedAt = 0;
900
+ var windowCount = 0;
901
+ var windowDroppedCount = 0;
902
+ function admitUnderRateLimit(now) {
903
+ if (now - windowStartedAt >= RATE_LIMIT_WINDOW_MS) {
904
+ if (windowDroppedCount > 0) {
905
+ console.error(
906
+ `[runner-activity-telemetry] rate cap reached: dropped ${windowDroppedCount} ${windowDroppedCount === 1 ? "entry" : "entries"} in the last ${RATE_LIMIT_WINDOW_MS / 1e3}s (cap ${RATE_LIMIT_MAX_EVENTS}/min)`
907
+ );
908
+ }
909
+ windowStartedAt = now;
910
+ windowCount = 0;
911
+ windowDroppedCount = 0;
912
+ }
913
+ if (windowCount >= RATE_LIMIT_MAX_EVENTS) {
914
+ windowDroppedCount++;
915
+ if (windowDroppedCount === 1) {
916
+ console.error(
917
+ `[runner-activity-telemetry] rate cap reached (${RATE_LIMIT_MAX_EVENTS} per ${RATE_LIMIT_WINDOW_MS / 1e3}s) \u2014 dropping further entries this window`
918
+ );
919
+ }
920
+ return false;
921
+ }
922
+ windowCount++;
923
+ return true;
924
+ }
925
+ function forwardRunnerActivity(entry, context) {
926
+ try {
927
+ if (!FORWARDED_LEVELS.has(entry.level)) return;
928
+ if (!context.agentId || !context.authHeader) return;
929
+ if (!admitUnderRateLimit(Date.now())) return;
930
+ const rawMessage = entry.error ?? entry.message ?? "";
931
+ const message = truncate(redact(rawMessage));
932
+ logEvent(TelemetryEventTypes.RUNNER_ACTIVITY, {
933
+ severity: SEVERITY_BY_LEVEL[entry.level],
934
+ message,
935
+ metadata: { source: "cli.run" },
936
+ agentId: context.agentId
937
+ });
938
+ } catch (err) {
939
+ console.error(
940
+ `[runner-activity-telemetry] failed to forward runner activity: ${err instanceof Error ? err.message : String(err)}`
941
+ );
942
+ }
943
+ }
944
+
701
945
  // src/lib/auth.ts
702
946
  async function getAuthCredentials() {
703
947
  const runnerKey = process.env.EVIDENT_RUNNER_KEY;
@@ -1532,6 +1776,9 @@ function isPreamblePinnedRunning(messages, userMessageId) {
1532
1776
  const reply = findLastAssistantReplyFor(messages, userMessageId);
1533
1777
  return completedOf(reply) != null && finishOf(reply) === "tool-calls";
1534
1778
  }
1779
+ function isB2AbandonmentConfirmed(params) {
1780
+ return params.pinnedForMs >= params.minPinnedMs && params.descendantOngoing === false;
1781
+ }
1535
1782
  function messageError(messages, userMessageId) {
1536
1783
  const reply = findLastAssistantReplyFor(messages, userMessageId);
1537
1784
  const error2 = errorOf(reply);
@@ -1545,6 +1792,42 @@ function messageError(messages, userMessageId) {
1545
1792
  }
1546
1793
  return "The agent run failed.";
1547
1794
  }
1795
+ function messageFailure(messages, userMessageId) {
1796
+ const reply = findLastAssistantReplyFor(messages, userMessageId);
1797
+ const error2 = errorOf(reply);
1798
+ if (error2 == null || typeof error2 !== "object") return null;
1799
+ const e = error2;
1800
+ const replyProviderId = reply?.info?.providerID ?? null;
1801
+ const replyModelId = reply?.info?.modelID ?? null;
1802
+ if (e.name === "ProviderAuthError") {
1803
+ const data = e.data;
1804
+ const providerId = typeof data?.providerID === "string" && data.providerID || replyProviderId;
1805
+ return { kind: "model_auth", providerId, modelId: replyModelId, reason: "missing" };
1806
+ }
1807
+ if (e.name === "APIError") {
1808
+ const data = e.data;
1809
+ const statusCode = data?.statusCode;
1810
+ if (statusCode === 401 || statusCode === 403) {
1811
+ return {
1812
+ kind: "model_auth",
1813
+ providerId: replyProviderId,
1814
+ modelId: replyModelId,
1815
+ reason: "rejected"
1816
+ };
1817
+ }
1818
+ }
1819
+ return null;
1820
+ }
1821
+ function applyZeroProviderFallback(classified, hasConfiguredProvider, replyProviderId, replyModelId = null) {
1822
+ if (classified != null) return classified;
1823
+ if (hasConfiguredProvider !== false) return null;
1824
+ return {
1825
+ kind: "model_auth",
1826
+ providerId: replyProviderId,
1827
+ modelId: replyModelId,
1828
+ reason: "missing"
1829
+ };
1830
+ }
1548
1831
  function hasRunningAssistantExcept(messages, exceptUserMessageId) {
1549
1832
  if (!messages || messages.length === 0) return false;
1550
1833
  return messages.some(
@@ -2073,13 +2356,52 @@ var RunnerConnection = class {
2073
2356
  }
2074
2357
  };
2075
2358
 
2359
+ // src/lib/tunnel/ready-marker.ts
2360
+ import { writeFileSync } from "fs";
2361
+ function writeTunnelReadyMarker(path, agentId) {
2362
+ try {
2363
+ writeFileSync(path, `${agentId}
2364
+ `);
2365
+ return { ok: true };
2366
+ } catch (error2) {
2367
+ return { ok: false, error: error2 instanceof Error ? error2.message : String(error2) };
2368
+ }
2369
+ }
2370
+
2371
+ // src/lib/claude-usage-reporting.ts
2372
+ var VALID_MODES = ["auto", "on", "off"];
2373
+ function resolveClaudeUsageReportingMode(flagValue, env) {
2374
+ const raw = flagValue ?? env.EVIDENT_CLAUDE_USAGE_REPORTING;
2375
+ if (raw === void 0 || raw === "") {
2376
+ return { mode: "auto", warnings: [] };
2377
+ }
2378
+ const normalized = raw.trim().toLowerCase();
2379
+ if (VALID_MODES.includes(normalized)) {
2380
+ return { mode: normalized, warnings: [] };
2381
+ }
2382
+ const source = flagValue !== void 0 ? "--claude-usage-reporting" : "EVIDENT_CLAUDE_USAGE_REPORTING";
2383
+ return {
2384
+ mode: "auto",
2385
+ warnings: [
2386
+ `Ignoring invalid ${source} "${raw}": expected one of ${VALID_MODES.join(", ")}; using auto`
2387
+ ]
2388
+ };
2389
+ }
2390
+ var BASE_REPORT_DELAY_MS = 10 * 6e4;
2391
+ var REPORT_DELAY_JITTER_FRACTION = 0.2;
2392
+ function nextReportDelayMs(random = Math.random) {
2393
+ const jitterRangeMs = BASE_REPORT_DELAY_MS * REPORT_DELAY_JITTER_FRACTION;
2394
+ return BASE_REPORT_DELAY_MS - jitterRangeMs + random() * (2 * jitterRangeMs);
2395
+ }
2396
+ var FIRST_REPORT_DELAY_MS = 5e3 + Math.random() * 1e4;
2397
+
2076
2398
  // src/lib/channels/driver.ts
2077
- import { homedir } from "os";
2399
+ import { homedir as homedir2 } from "os";
2078
2400
 
2079
2401
  // src/lib/file-push.ts
2080
2402
  import { randomUUID } from "crypto";
2081
2403
  import { chmod, mkdir, open as open2, realpath, rename, unlink } from "fs/promises";
2082
- import { basename, dirname as dirname2, isAbsolute, join, relative, resolve as resolve2, sep } from "path";
2404
+ import { basename, dirname as dirname2, isAbsolute, join as join2, relative, resolve as resolve2, sep } from "path";
2083
2405
  var FILE_MODE = 384;
2084
2406
  var DIRECTORY_MODE = 448;
2085
2407
  async function writePushedFile(request) {
@@ -2112,7 +2434,7 @@ async function writePushedFile(request) {
2112
2434
  const { existingAncestor, missingSegments } = await resolveNearestExistingAncestor(
2113
2435
  dirname2(candidate)
2114
2436
  );
2115
- const realTarget = join(existingAncestor, ...missingSegments, basename(candidate));
2437
+ const realTarget = join2(existingAncestor, ...missingSegments, basename(candidate));
2116
2438
  const allowedDirectory = await findContainingAllowedDirectory(allowedDirectories, realTarget);
2117
2439
  if (allowedDirectory === null) {
2118
2440
  return refuse("path_not_allowed", "The runner does not allow writing to that location.", {
@@ -2148,7 +2470,7 @@ function expandAndValidate(requestedPath, homeDir) {
2148
2470
  if (requestedPath.trim() === "" || requestedPath.includes("\0")) {
2149
2471
  return null;
2150
2472
  }
2151
- const expanded = requestedPath === "~" ? homeDir : requestedPath.startsWith("~/") ? join(homeDir, requestedPath.slice(2)) : requestedPath;
2473
+ const expanded = requestedPath === "~" ? homeDir : requestedPath.startsWith("~/") ? join2(homeDir, requestedPath.slice(2)) : requestedPath;
2152
2474
  if (expanded.split(/[/\\]/).includes("..")) {
2153
2475
  return null;
2154
2476
  }
@@ -2221,13 +2543,13 @@ function contains(realDirectory, realTarget) {
2221
2543
  async function createMissingDirectories(existingAncestor, missingSegments) {
2222
2544
  let current = existingAncestor;
2223
2545
  for (const segment of missingSegments) {
2224
- current = join(current, segment);
2546
+ current = join2(current, segment);
2225
2547
  await mkdir(current, { recursive: true, mode: DIRECTORY_MODE });
2226
2548
  await chmod(current, DIRECTORY_MODE);
2227
2549
  }
2228
2550
  }
2229
2551
  async function writeAtomically(realTarget, content) {
2230
- const temporaryPath = join(dirname2(realTarget), `.evident-push-${randomUUID()}.tmp`);
2552
+ const temporaryPath = join2(dirname2(realTarget), `.evident-push-${randomUUID()}.tmp`);
2231
2553
  let handle;
2232
2554
  try {
2233
2555
  handle = await open2(temporaryPath, "wx", FILE_MODE);
@@ -2499,6 +2821,8 @@ var DEFAULT_PAUSED_MAX_WAIT_MS = 10 * 60 * 1e3;
2499
2821
  var DEFAULT_STUCK_QUEUED_MS = 6e4;
2500
2822
  var HEARTBEAT_MS = 6e4;
2501
2823
  var ABSOLUTE_MAX_PROCESSING_MS = 6 * 60 * 60 * 1e3;
2824
+ var B2_ABANDONMENT_MIN_PINNED_MS = 3 * 6e4;
2825
+ var B2_ABANDONMENT_RECHECK_MS = HEARTBEAT_MS;
2502
2826
  var POLL_MISS_GRACE_MS = HEARTBEAT_MS;
2503
2827
  var MAX_SUPERSEDED_CONVERSATIONS = 256;
2504
2828
  var ChannelAuthError = class extends Error {
@@ -2724,23 +3048,23 @@ var ChannelDriver = class _ChannelDriver {
2724
3048
  * and stops opencode.
2725
3049
  */
2726
3050
  stopped = false;
2727
- constructor(config2) {
2728
- this.agentId = config2.agentId;
2729
- this.port = config2.port;
2730
- this.apiUrl = config2.apiUrl.replace(/\/$/, "");
2731
- this.getAuthHeader = config2.getAuthHeader;
2732
- this.conversationFilter = config2.conversationFilter ?? null;
2733
- this.retry = { ...DEFAULT_RETRY_POLICY, ...config2.retry };
2734
- this.log = config2.log ?? (() => {
3051
+ constructor(config) {
3052
+ this.agentId = config.agentId;
3053
+ this.port = config.port;
3054
+ this.apiUrl = config.apiUrl.replace(/\/$/, "");
3055
+ this.getAuthHeader = config.getAuthHeader;
3056
+ this.conversationFilter = config.conversationFilter ?? null;
3057
+ this.retry = { ...DEFAULT_RETRY_POLICY, ...config.retry };
3058
+ this.log = config.log ?? (() => {
2735
3059
  });
2736
- this.fetchImpl = config2.fetchImpl ?? fetch;
2737
- this.sleep = config2.sleep ?? ((ms) => new Promise((r) => setTimeout(r, ms)));
2738
- this.pausedPollIntervalMs = config2.pausedPollIntervalMs ?? DEFAULT_PAUSED_POLL_INTERVAL_MS;
2739
- this.pausedMaxWaitMs = config2.pausedMaxWaitMs ?? DEFAULT_PAUSED_MAX_WAIT_MS;
2740
- this.stuckQueuedMs = config2.stuckQueuedMs ?? DEFAULT_STUCK_QUEUED_MS;
2741
- this.now = config2.now ?? (() => Date.now());
2742
- this.fileSyncDirectories = config2.fileSyncDirectories ?? [];
2743
- this.homeDir = config2.homeDir ?? homedir();
3060
+ this.fetchImpl = config.fetchImpl ?? fetch;
3061
+ this.sleep = config.sleep ?? ((ms) => new Promise((r) => setTimeout(r, ms)));
3062
+ this.pausedPollIntervalMs = config.pausedPollIntervalMs ?? DEFAULT_PAUSED_POLL_INTERVAL_MS;
3063
+ this.pausedMaxWaitMs = config.pausedMaxWaitMs ?? DEFAULT_PAUSED_MAX_WAIT_MS;
3064
+ this.stuckQueuedMs = config.stuckQueuedMs ?? DEFAULT_STUCK_QUEUED_MS;
3065
+ this.now = config.now ?? (() => Date.now());
3066
+ this.fileSyncDirectories = config.fileSyncDirectories ?? [];
3067
+ this.homeDir = config.homeDir ?? homedir2();
2744
3068
  }
2745
3069
  /** The IPv4-loopback base URL for the local `opencode serve`. */
2746
3070
  get opencodeBase() {
@@ -3125,7 +3449,12 @@ var ChannelDriver = class _ChannelDriver {
3125
3449
  const directory = await this.resolveOpenCodeDirectory();
3126
3450
  const sessionId = await createOpenCodeSession(this.port, directory);
3127
3451
  this.sessions.set(conversationId, sessionId);
3128
- await this.persistSession(conversationId, sessionId).catch(() => {
3452
+ await this.persistSession(conversationId, sessionId).catch((err) => {
3453
+ this.log({
3454
+ level: "warn",
3455
+ message: `Persisting the OpenCode session binding ${sessionId.slice(0, 8)} for conversation ${conversationId.slice(0, 8)} failed (best-effort, not retried) \u2014 the completion PATCH also carries opencode_session_id, so the binding is repaired when the turn finishes: ${err instanceof Error ? err.message : String(err)}`,
3456
+ conversation_id: conversationId
3457
+ });
3129
3458
  });
3130
3459
  return sessionId;
3131
3460
  }
@@ -3317,12 +3646,17 @@ var ChannelDriver = class _ChannelDriver {
3317
3646
  stuckReported: false,
3318
3647
  lastAliveAt: 0,
3319
3648
  aliveInFlight: false,
3649
+ titleSynced: false,
3650
+ titleSyncInFlight: false,
3320
3651
  awaitingHumanLatched: false,
3321
3652
  pausedOnQuestion: false,
3322
3653
  pausedOnPermission: false,
3323
3654
  pausedClearConfirmed: false,
3324
3655
  pausedInFlight: false,
3325
- deliveryDeadlineAnchored: false
3656
+ deliveryDeadlineAnchored: false,
3657
+ b2PinnedSinceMs: 0,
3658
+ b2LastDescendantCheckMs: 0,
3659
+ b2AbandonedSignalled: false
3326
3660
  });
3327
3661
  }
3328
3662
  /**
@@ -3390,12 +3724,17 @@ var ChannelDriver = class _ChannelDriver {
3390
3724
  // with no extra `re_adopted` signal needed (folds old WI-6).
3391
3725
  lastAliveAt: 0,
3392
3726
  aliveInFlight: false,
3727
+ titleSynced: false,
3728
+ titleSyncInFlight: false,
3393
3729
  awaitingHumanLatched: false,
3394
3730
  pausedOnQuestion: false,
3395
3731
  pausedOnPermission: false,
3396
3732
  pausedClearConfirmed: false,
3397
3733
  pausedInFlight: false,
3398
- deliveryDeadlineAnchored: false
3734
+ deliveryDeadlineAnchored: false,
3735
+ b2PinnedSinceMs: 0,
3736
+ b2LastDescendantCheckMs: 0,
3737
+ b2AbandonedSignalled: false
3399
3738
  });
3400
3739
  }
3401
3740
  /**
@@ -3557,58 +3896,7 @@ var ChannelDriver = class _ChannelDriver {
3557
3896
  }
3558
3897
  }
3559
3898
  if (state === "done") {
3560
- this.anchorDeliveryDeadline(inFlight);
3561
- if (!inFlight.done) {
3562
- this.log({
3563
- level: "info",
3564
- message: `Message ${inFlight.evidentMessageId.slice(0, 8)} completed \u2014 marking done`,
3565
- conversation_id: conv.id,
3566
- message_id: inFlight.evidentMessageId
3567
- });
3568
- const title = await this.resolveSessionTitle(sessionId, watcher.conv.id);
3569
- const usage = messageUsage(messages, inFlight.opencodeMessageId);
3570
- try {
3571
- await this.markDone(
3572
- conv.id,
3573
- inFlight.evidentMessageId,
3574
- sessionId,
3575
- inFlight.opencodeMessageId,
3576
- title,
3577
- usage
3578
- );
3579
- } catch (err) {
3580
- if (err instanceof ChannelAuthError) throw err;
3581
- if (err instanceof ChannelTerminalError) {
3582
- this.log({
3583
- level: "warn",
3584
- message: `Failed to mark message ${inFlight.evidentMessageId.slice(0, 8)} done (terminal HTTP ${err.status}) \u2014 leaving for the cron safety net: ${err.message}`,
3585
- conversation_id: conv.id,
3586
- message_id: inFlight.evidentMessageId
3587
- });
3588
- this.removeInFlight(watcher, inFlight.evidentMessageId);
3589
- return;
3590
- }
3591
- if (this.now() >= inFlight.deadline) {
3592
- this.log({
3593
- level: "warn",
3594
- message: `Failed to mark message ${inFlight.evidentMessageId.slice(0, 8)} done within the watch window \u2014 leaving for the cron safety net: ${err instanceof Error ? err.message : String(err)}`,
3595
- conversation_id: conv.id,
3596
- message_id: inFlight.evidentMessageId
3597
- });
3598
- this.removeInFlight(watcher, inFlight.evidentMessageId);
3599
- return;
3600
- }
3601
- this.log({
3602
- level: "warn",
3603
- message: `Failed to mark message ${inFlight.evidentMessageId.slice(0, 8)} done (will retry next tick): ${err instanceof Error ? err.message : String(err)}`,
3604
- conversation_id: conv.id,
3605
- message_id: inFlight.evidentMessageId
3606
- });
3607
- return;
3608
- }
3609
- inFlight.done = true;
3610
- }
3611
- this.removeInFlight(watcher, inFlight.evidentMessageId);
3899
+ await this.settleMessageDone(sessionId, watcher, inFlight, messages);
3612
3900
  return;
3613
3901
  }
3614
3902
  if (state === "failed") {
@@ -3622,8 +3910,16 @@ var ChannelDriver = class _ChannelDriver {
3622
3910
  message_id: inFlight.evidentMessageId
3623
3911
  });
3624
3912
  const usage = messageUsage(messages, inFlight.opencodeMessageId);
3913
+ const failure = await this.classifyModelAuthFailure(messages, inFlight.opencodeMessageId);
3625
3914
  try {
3626
- await this.markFailed(conv.id, inFlight.evidentMessageId, sessionId, error2, usage);
3915
+ await this.markFailed(
3916
+ conv.id,
3917
+ inFlight.evidentMessageId,
3918
+ sessionId,
3919
+ error2,
3920
+ usage,
3921
+ failure
3922
+ );
3627
3923
  } catch (err) {
3628
3924
  if (err instanceof ChannelAuthError) throw err;
3629
3925
  if (err instanceof ChannelTerminalError) {
@@ -3668,6 +3964,44 @@ var ChannelDriver = class _ChannelDriver {
3668
3964
  });
3669
3965
  }
3670
3966
  const activelyRunning = state === "running" && !awaitingHuman;
3967
+ const pinnedNow = activelyRunning && isPreamblePinnedRunning(messages, inFlight.opencodeMessageId);
3968
+ const snapshotReadable = messages != null && messages.length > 0;
3969
+ if (!pinnedNow) {
3970
+ if (snapshotReadable) {
3971
+ inFlight.b2PinnedSinceMs = 0;
3972
+ inFlight.b2LastDescendantCheckMs = 0;
3973
+ inFlight.b2AbandonedSignalled = false;
3974
+ }
3975
+ } else {
3976
+ if (inFlight.b2AbandonedSignalled) {
3977
+ await this.settleMessageDone(sessionId, watcher, inFlight, messages);
3978
+ return;
3979
+ }
3980
+ if (inFlight.b2PinnedSinceMs === 0) inFlight.b2PinnedSinceMs = this.now();
3981
+ const pinnedForMs = this.now() - inFlight.b2PinnedSinceMs;
3982
+ if (pinnedForMs >= B2_ABANDONMENT_MIN_PINNED_MS && this.now() - inFlight.b2LastDescendantCheckMs >= B2_ABANDONMENT_RECHECK_MS) {
3983
+ inFlight.b2LastDescendantCheckMs = this.now();
3984
+ const descendantOngoing = await this.isAnyDescendantSessionOngoing(sessionId);
3985
+ if (isB2AbandonmentConfirmed({
3986
+ pinnedForMs,
3987
+ minPinnedMs: B2_ABANDONMENT_MIN_PINNED_MS,
3988
+ descendantOngoing
3989
+ })) {
3990
+ inFlight.b2AbandonedSignalled = true;
3991
+ this.log({
3992
+ level: "warn",
3993
+ message: `Message ${id.slice(0, 8)} b2-pinned for ${Math.round(pinnedForMs / 1e3)}s with no ongoing descendant sub-agent session (status-map confirmed) \u2014 treating the delegated/tool turn as abandoned, resolving done`,
3994
+ conversation_id: conv.id,
3995
+ message_id: id
3996
+ });
3997
+ void this.postSignal(conv.id, id, "b2_abandoned_resolved", {
3998
+ watched_for_ms: pinnedForMs
3999
+ });
4000
+ await this.settleMessageDone(sessionId, watcher, inFlight, messages);
4001
+ return;
4002
+ }
4003
+ }
4004
+ }
3671
4005
  if (activelyRunning && this.now() - inFlight.processingAnchorMs >= ABSOLUTE_MAX_PROCESSING_MS) {
3672
4006
  this.log({
3673
4007
  level: "warn",
@@ -3687,6 +4021,18 @@ var ChannelDriver = class _ChannelDriver {
3687
4021
  inFlight.aliveInFlight = false;
3688
4022
  if (ok) inFlight.lastAliveAt = this.now();
3689
4023
  });
4024
+ if (!inFlight.titleSynced && !inFlight.titleSyncInFlight) {
4025
+ inFlight.titleSyncInFlight = true;
4026
+ void this.resolveSessionTitle(sessionId, conv.id).then(async (title) => {
4027
+ if (!title) {
4028
+ inFlight.titleSyncInFlight = false;
4029
+ return;
4030
+ }
4031
+ const ok = await this.patchConversationTitle(conv.id, title);
4032
+ inFlight.titleSyncInFlight = false;
4033
+ if (ok) inFlight.titleSynced = true;
4034
+ });
4035
+ }
3690
4036
  }
3691
4037
  if (awaitingHuman) {
3692
4038
  if (!inFlight.awaitingHumanLatched) {
@@ -3724,6 +4070,70 @@ var ChannelDriver = class _ChannelDriver {
3724
4070
  this.removeInFlight(watcher, inFlight.evidentMessageId);
3725
4071
  }
3726
4072
  }
4073
+ /**
4074
+ * Settle a message whose run-state has resolved `'done'` — extracted verbatim
4075
+ * (pure refactor, no behavior change) from `serviceInFlightMessage`'s former
4076
+ * inline `state === 'done'` branch body, so a SECOND caller (the #721
4077
+ * b2-abandonment resolution) can reach the exact same completion behavior
4078
+ * (delivery-deadline anchoring, title resolution, usage extraction, and
4079
+ * `markDone`'s auth/terminal/transient-retry discipline) without duplicating it
4080
+ * and risking the two copies silently drifting apart.
4081
+ */
4082
+ async settleMessageDone(sessionId, watcher, inFlight, messages) {
4083
+ const conv = watcher.conv;
4084
+ this.anchorDeliveryDeadline(inFlight);
4085
+ if (!inFlight.done) {
4086
+ this.log({
4087
+ level: "info",
4088
+ message: `Message ${inFlight.evidentMessageId.slice(0, 8)} completed \u2014 marking done`,
4089
+ conversation_id: conv.id,
4090
+ message_id: inFlight.evidentMessageId
4091
+ });
4092
+ const title = await this.resolveSessionTitle(sessionId, watcher.conv.id);
4093
+ const usage = messageUsage(messages, inFlight.opencodeMessageId);
4094
+ try {
4095
+ await this.markDone(
4096
+ conv.id,
4097
+ inFlight.evidentMessageId,
4098
+ sessionId,
4099
+ inFlight.opencodeMessageId,
4100
+ title,
4101
+ usage
4102
+ );
4103
+ } catch (err) {
4104
+ if (err instanceof ChannelAuthError) throw err;
4105
+ if (err instanceof ChannelTerminalError) {
4106
+ this.log({
4107
+ level: "warn",
4108
+ message: `Failed to mark message ${inFlight.evidentMessageId.slice(0, 8)} done (terminal HTTP ${err.status}) \u2014 leaving for the cron safety net: ${err.message}`,
4109
+ conversation_id: conv.id,
4110
+ message_id: inFlight.evidentMessageId
4111
+ });
4112
+ this.removeInFlight(watcher, inFlight.evidentMessageId);
4113
+ return;
4114
+ }
4115
+ if (this.now() >= inFlight.deadline) {
4116
+ this.log({
4117
+ level: "warn",
4118
+ message: `Failed to mark message ${inFlight.evidentMessageId.slice(0, 8)} done within the watch window \u2014 leaving for the cron safety net: ${err instanceof Error ? err.message : String(err)}`,
4119
+ conversation_id: conv.id,
4120
+ message_id: inFlight.evidentMessageId
4121
+ });
4122
+ this.removeInFlight(watcher, inFlight.evidentMessageId);
4123
+ return;
4124
+ }
4125
+ this.log({
4126
+ level: "warn",
4127
+ message: `Failed to mark message ${inFlight.evidentMessageId.slice(0, 8)} done (will retry next tick): ${err instanceof Error ? err.message : String(err)}`,
4128
+ conversation_id: conv.id,
4129
+ message_id: inFlight.evidentMessageId
4130
+ });
4131
+ return;
4132
+ }
4133
+ inFlight.done = true;
4134
+ }
4135
+ this.removeInFlight(watcher, inFlight.evidentMessageId);
4136
+ }
3727
4137
  // Restart recovery: re-adopt `processing` messages (ADR-0046, WI-3/4/5)
3728
4138
  /**
3729
4139
  * Re-adopt this agent's `processing` messages on drain (ADR-0046 Decision §1).
@@ -3890,6 +4300,7 @@ var ChannelDriver = class _ChannelDriver {
3890
4300
  if (state === "failed") {
3891
4301
  const error2 = messageError(messages, ocId ?? "") ?? void 0;
3892
4302
  const usage = messageUsage(messages, ocId ?? "");
4303
+ const failure = await this.classifyModelAuthFailure(messages, ocId ?? "");
3893
4304
  this.log({
3894
4305
  level: "error",
3895
4306
  message: `Re-adopt: message ${row.id.slice(0, 8)} errored while unwatched \u2014 marking failed: ${error2 ?? "(no error text)"}`,
@@ -3897,7 +4308,7 @@ var ChannelDriver = class _ChannelDriver {
3897
4308
  message_id: row.id
3898
4309
  });
3899
4310
  try {
3900
- await this.markFailed(row.conversation_id, row.id, sessionId, error2, usage);
4311
+ await this.markFailed(row.conversation_id, row.id, sessionId, error2, usage, failure);
3901
4312
  } catch (err) {
3902
4313
  if (err instanceof ChannelAuthError) throw err;
3903
4314
  if (err instanceof ChannelTerminalError) {
@@ -4303,6 +4714,47 @@ var ChannelDriver = class _ChannelDriver {
4303
4714
  }
4304
4715
  return false;
4305
4716
  }
4717
+ /**
4718
+ * Tri-state variant of the upward parentID membership walk (#721), used ONLY
4719
+ * by `isAnyDescendantSessionOngoing`. Walks the SAME cached
4720
+ * `resolveSessionParent` chain `sessionBelongsTo` uses above, but — unlike
4721
+ * `sessionBelongsTo`, which deliberately collapses "confirmed not a
4722
+ * descendant" and "the walk's fetch failed" into the same `false` (safe for
4723
+ * its OTHER callers: interaction attribution and the recovery-path
4724
+ * `isAnyDescendantSessionAlive`, both of which just retry next tick with no
4725
+ * safety consequence either way) — this variant keeps those two outcomes
4726
+ * SEPARATE, because `isAnyDescendantSessionOngoing`'s caller
4727
+ * (`isB2AbandonmentConfirmed`) must never treat "couldn't tell" as "confirmed
4728
+ * not ongoing".
4729
+ *
4730
+ * Return contract:
4731
+ * - `true` → the walk reached `rootSessionId` — `sessionId` IS a descendant.
4732
+ * - `false` → the walk reached a definitive, parent-less root session
4733
+ * WITHOUT ever matching `rootSessionId` — `sessionId` is
4734
+ * CONFIRMED NOT a descendant of it.
4735
+ * - `null` → INDETERMINATE: a `GET /session/:id` fetch failed partway
4736
+ * through the walk (`resolveSessionParent` returned `undefined`),
4737
+ * or the depth cap (32) was hit without a definitive answer (a
4738
+ * pathological/cyclic chain proves nothing either way). NEVER
4739
+ * treat this the same as `false` — see `sessionBelongsTo`'s own
4740
+ * doc comment above for why that collapse is safe THERE but not
4741
+ * here.
4742
+ *
4743
+ * `sessionBelongsTo` itself is UNCHANGED — this is an additive helper scoped
4744
+ * to the live-path descendant check, not a modification of shared code used
4745
+ * by interaction attribution or the recovery path.
4746
+ */
4747
+ async resolveSessionMembership(sessionId, rootSessionId) {
4748
+ let current = sessionId;
4749
+ for (let depth = 0; current && depth < 32; depth++) {
4750
+ if (current === rootSessionId) return true;
4751
+ const parent = await this.resolveSessionParent(current);
4752
+ if (parent === void 0) return null;
4753
+ if (parent === null) return false;
4754
+ current = parent;
4755
+ }
4756
+ return null;
4757
+ }
4306
4758
  /**
4307
4759
  * Resolve (and cache) a session's `parentID` via `GET /session/:id`. Returns
4308
4760
  * `null` for a root session (no parent) and `undefined` when opencode is
@@ -4387,6 +4839,54 @@ var ChannelDriver = class _ChannelDriver {
4387
4839
  }
4388
4840
  return null;
4389
4841
  }
4842
+ /**
4843
+ * Best-effort mid-turn title sync (#711 follow-up): PATCH a resolved OpenCode
4844
+ * session title onto the conversation via the PLAIN conversation-update
4845
+ * endpoint (`PATCH /runners/:agentId/conversations/:conversationId`) — NOT the
4846
+ * message-status endpoint `markProcessing`/`markDone` use. Deliberately a
4847
+ * separate, lighter call: it carries no `status`, so it cannot re-trigger the
4848
+ * `processing`/`done` transition side effects (Slack notices, activity-log
4849
+ * rows, delivery jobs) those PATCHes gate on `transitioned` — this call only
4850
+ * ever touches `conversations.title`. That route (`routes/conversations.ts`)
4851
+ * skips a title write matching the stored value, so a redundant call with the
4852
+ * same title is a real no-op — it does not bump `updated_at`, which the
4853
+ * conversation list sorts and paginates on. (Note this is a DIFFERENT guard
4854
+ * from `threads.ts`'s "non-empty AND changed" one, which only covers the
4855
+ * message-status PATCH; the non-empty half is enforced here instead, by
4856
+ * `resolveSessionTitle` never returning an empty/placeholder title.)
4857
+ *
4858
+ * Telemetry-only / never blocks the caller, mirroring `postSignal`: a failure
4859
+ * is logged and the title is simply retried on the next heartbeat tick (the
4860
+ * caller only latches `titleSynced` on `true`).
4861
+ */
4862
+ async patchConversationTitle(conversationId, title) {
4863
+ try {
4864
+ const res = await this.fetchImpl(
4865
+ `${this.apiUrl}/runners/${this.agentId}/conversations/${conversationId}`,
4866
+ {
4867
+ method: "PATCH",
4868
+ headers: { Authorization: this.getAuthHeader(), "Content-Type": "application/json" },
4869
+ body: JSON.stringify({ title })
4870
+ }
4871
+ );
4872
+ if (!res.ok) {
4873
+ this.log({
4874
+ level: "debug",
4875
+ message: `Mid-turn title sync PATCH for conversation ${conversationId.slice(0, 8)} returned HTTP ${res.status} (best-effort, will retry next heartbeat)`,
4876
+ conversation_id: conversationId
4877
+ });
4878
+ return false;
4879
+ }
4880
+ return true;
4881
+ } catch (err) {
4882
+ this.log({
4883
+ level: "debug",
4884
+ message: `Best-effort mid-turn title sync PATCH failed for conversation ${conversationId.slice(0, 8)} (will retry next heartbeat): ${err instanceof Error ? err.message : String(err)}`,
4885
+ conversation_id: conversationId
4886
+ });
4887
+ return false;
4888
+ }
4889
+ }
4390
4890
  /**
4391
4891
  * DEFENSIVE cross-check for the restart-recovery path (WI-2): is any descendant
4392
4892
  * (`task` sub-agent) session under `rootSessionId` still genuinely doing work?
@@ -4445,6 +4945,84 @@ var ChannelDriver = class _ChannelDriver {
4445
4945
  }
4446
4946
  return false;
4447
4947
  }
4948
+ /**
4949
+ * LIVE-PATH descendant-liveness check (#721): is any descendant (`task`
4950
+ * sub-agent) session under `rootSessionId` currently ONGOING per OpenCode's own
4951
+ * in-memory status map (`isSessionOngoing` — `busy`/`retry`)?
4952
+ *
4953
+ * Deliberately NOT `isAnyDescendantSessionAlive` (the RECOVERY-path
4954
+ * cross-check above): that method judges liveness from the child's OWN
4955
+ * TRANSCRIPT (`isSessionActivelyGenerating`), which is the right (only) option
4956
+ * on the recovery path because a restart WIPES `SessionStatus`. On the LIVE
4957
+ * path the local opencode server IS running, so its in-memory status map is
4958
+ * live and authoritative — and per ADR-0047 §4a ("the child has its own entry
4959
+ * [in the map]"), a `task` descendant's OWN busy/retry entry reflects its
4960
+ * ENTIRE turn (including any tool call it is itself executing), not a
4961
+ * per-message transcript snapshot. This sidesteps the "child's own tool is
4962
+ * executing, between its step's completion and the next generation step"
4963
+ * transcript gap that a transcript-based check would need a second,
4964
+ * sustained-window bound to guard against — it is simply not derived from
4965
+ * message timestamps at all.
4966
+ *
4967
+ * Why not just check `isSessionOngoing(port, rootSessionId)` (the ROOT's own
4968
+ * status, as the recovery path does per §4a)? Because on the LIVE path the
4969
+ * root session can be shared: a SECOND, unrelated user message can land on the
4970
+ * SAME session (issue #721's own root cause) and keep the root `busy` for a
4971
+ * reason that has nothing to do with THIS message's delegation. A `task`
4972
+ * descendant session is spawned for exactly one delegated turn and never
4973
+ * reused, so its OWN status-map entry is unambiguous evidence about that one
4974
+ * delegation — which the root's status is not.
4975
+ *
4976
+ * Why membership is checked via `resolveSessionMembership`, NOT
4977
+ * `sessionBelongsTo`: `sessionBelongsTo` collapses a transient
4978
+ * `GET /session/:id` fetch failure into "not a descendant", which would
4979
+ * silently drop a genuinely-live candidate from consideration on the one
4980
+ * unlucky tick its membership-walk fetch hiccups (#721).
4981
+ * `resolveSessionMembership` keeps that failure mode as a distinct `null`
4982
+ * (indeterminate) so it is folded into THIS method's own `indeterminate` flag
4983
+ * instead.
4984
+ *
4985
+ * Return contract (note the DIFFERENT judge vs. `isAnyDescendantSessionAlive`):
4986
+ * - `true` → some descendant session is `busy`/`retry` (genuinely ongoing).
4987
+ * - `false` → enumeration succeeded, EVERY candidate's MEMBERSHIP was
4988
+ * confirmed either way (`resolveSessionMembership` never
4989
+ * returned `null`), and every CONFIRMED descendant's status read
4990
+ * succeeded and is not ongoing (includes "no descendant session
4991
+ * exists at all" — e.g. a plain, non-`task` tool call).
4992
+ * - `null` → INDETERMINATE: `listSessions` failed, OR at least one
4993
+ * candidate's MEMBERSHIP could not be confirmed
4994
+ * (`resolveSessionMembership` returned `null` — a fetch failure
4995
+ * or pathological chain partway through the parent walk), OR at
4996
+ * least one CONFIRMED descendant's `isSessionOngoing` read
4997
+ * failed — and no OTHER candidate was already confirmed `true`.
4998
+ * The caller MUST NOT treat `null` the same as `false` here
4999
+ * (unlike the recovery cross-check's contract) — see
5000
+ * `isB2AbandonmentConfirmed`.
5001
+ */
5002
+ async isAnyDescendantSessionOngoing(rootSessionId) {
5003
+ const sessions = await listSessions(this.port);
5004
+ if (!sessions) {
5005
+ this.log({
5006
+ level: "warn",
5007
+ message: `Could not enumerate sessions to check descendant liveness for root ${rootSessionId} (listSessions failed) \u2014 treating descendant liveness as indeterminate`
5008
+ });
5009
+ return null;
5010
+ }
5011
+ let indeterminate = false;
5012
+ for (const candidate of sessions) {
5013
+ if (!candidate?.id || candidate.id === rootSessionId) continue;
5014
+ const membership = await this.resolveSessionMembership(candidate.id, rootSessionId);
5015
+ if (membership === null) {
5016
+ indeterminate = true;
5017
+ continue;
5018
+ }
5019
+ if (membership === false) continue;
5020
+ const ongoing = await isSessionOngoing(this.port, candidate.id);
5021
+ if (ongoing === true) return true;
5022
+ if (ongoing === null) indeterminate = true;
5023
+ }
5024
+ return indeterminate ? null : false;
5025
+ }
4448
5026
  /**
4449
5027
  * Cheap decision-telemetry label for a running row's LAST correlated reply
4450
5028
  * (WI-2 Task 2.2): which running SHAPE it is, for the status-gated recovery log.
@@ -4708,7 +5286,7 @@ var ChannelDriver = class _ChannelDriver {
4708
5286
  * exists but is wedged, so the next attempt must get a fresh one
4709
5287
  * instead of reusing it — see WI-1's server-side null-clearing PATCH).
4710
5288
  */
4711
- async markFailed(conversationId, messageId, sessionId, error2, usage) {
5289
+ async markFailed(conversationId, messageId, sessionId, error2, usage, failure) {
4712
5290
  const body = { status: "failed" };
4713
5291
  if (sessionId === null) {
4714
5292
  body.opencode_session_id = null;
@@ -4717,6 +5295,12 @@ var ChannelDriver = class _ChannelDriver {
4717
5295
  }
4718
5296
  if (error2 !== void 0) body.error = error2;
4719
5297
  if (usage) Object.assign(body, usage);
5298
+ if (failure) {
5299
+ body.failure_kind = failure.kind;
5300
+ body.failure_provider_id = failure.providerId;
5301
+ body.failure_model_id = failure.modelId;
5302
+ body.failure_reason = failure.reason;
5303
+ }
4720
5304
  await this.callWithRetry(
4721
5305
  "marking message as failed",
4722
5306
  () => this.fetchImpl(
@@ -4729,6 +5313,29 @@ var ChannelDriver = class _ChannelDriver {
4729
5313
  )
4730
5314
  );
4731
5315
  }
5316
+ /**
5317
+ * Classify an errored turn as a model-auth failure (#736 P1-4), for `markFailed`.
5318
+ *
5319
+ * `messageFailure` alone (structured OpenCode error → `model_auth`) covers
5320
+ * most cases; when it returns `null` on this ALREADY-FAILED turn, fall back
5321
+ * to the P1-2b zero-provider check — one extra loopback call to
5322
+ * `hasAnyConfiguredProvider`, only reached when the structured classifier
5323
+ * couldn't place it. Fails open (never throws): a fallback probe failure
5324
+ * (`null`/indeterminate) leaves the classification `null`, which produces
5325
+ * today's byte-identical PATCH body via `markFailed`'s `if (failure)` guard.
5326
+ */
5327
+ async classifyModelAuthFailure(messages, userMessageId) {
5328
+ const classified = messageFailure(messages, userMessageId);
5329
+ if (classified != null) return classified;
5330
+ const reply = findLastAssistantReplyFor(messages, userMessageId);
5331
+ const hasProvider = await hasAnyConfiguredProvider(this.port);
5332
+ return applyZeroProviderFallback(
5333
+ classified,
5334
+ hasProvider,
5335
+ reply?.info?.providerID ?? null,
5336
+ reply?.info?.modelID ?? null
5337
+ );
5338
+ }
4732
5339
  /**
4733
5340
  * Best-effort, SINGLE-ATTEMPT runner→server telemetry ping
4734
5341
  * (queued-followup-redrive). `POST .../messages/:id/signal {signal, ...extra}`
@@ -4881,10 +5488,16 @@ var ChannelDriver = class _ChannelDriver {
4881
5488
  import chalk5 from "chalk";
4882
5489
  import ora2 from "ora";
4883
5490
  import { select as select2 } from "@inquirer/prompts";
5491
+ var INTERACTIVE_START_TIMEOUT_MS = 3e4;
4884
5492
  async function ensureOpenCodeRunning(ctx) {
4885
5493
  const healthCheck = await checkOpenCodeHealth(ctx.port);
4886
5494
  if (healthCheck.healthy) {
4887
- return { port: ctx.port, process: null, version: healthCheck.version ?? null };
5495
+ return {
5496
+ port: ctx.port,
5497
+ process: null,
5498
+ version: healthCheck.version ?? null,
5499
+ notReadyReason: null
5500
+ };
4888
5501
  }
4889
5502
  const runningInstances = await findHealthyOpenCodeInstances();
4890
5503
  if (runningInstances.length > 0) {
@@ -4905,7 +5518,7 @@ async function ensureOpenCodeRunning(ctx) {
4905
5518
  console.log(chalk5.yellow("Tip: Run with the correct port:"));
4906
5519
  console.log(
4907
5520
  chalk5.dim(
4908
- ` ${getCliName()} run --agent ${ctx.agentId} --port ${runningInstances[0].port}`
5521
+ ` ${getCliName()} run --runner ${ctx.agentId} --port ${runningInstances[0].port}`
4909
5522
  )
4910
5523
  );
4911
5524
  }
@@ -4925,14 +5538,22 @@ async function ensureOpenCodeRunning(ctx) {
4925
5538
  if (!ctx.interactive) {
4926
5539
  ctx.log(`OpenCode is not running on port ${ctx.port}. Starting it automatically...`);
4927
5540
  const proc = await startOpenCode(ctx.port);
4928
- const health = await waitForOpenCodeHealth(ctx.port, 3e4);
5541
+ const health = await waitForOpenCodeHealth(ctx.port, ctx.startTimeoutMs);
4929
5542
  if (!health.healthy) {
4930
- throw new Error(
4931
- `OpenCode failed to start on port ${ctx.port}. Install with: npm install -g opencode-ai`
4932
- );
5543
+ return {
5544
+ port: ctx.port,
5545
+ process: proc,
5546
+ version: null,
5547
+ notReadyReason: `it did not become healthy within ${Math.round(ctx.startTimeoutMs / 1e3)}s`
5548
+ };
4933
5549
  }
4934
5550
  ctx.log(`OpenCode started on port ${ctx.port}${health.version ? ` (v${health.version})` : ""}`);
4935
- return { port: ctx.port, process: proc, version: health.version ?? null };
5551
+ return {
5552
+ port: ctx.port,
5553
+ process: proc,
5554
+ version: health.version ?? null,
5555
+ notReadyReason: null
5556
+ };
4936
5557
  }
4937
5558
  let port = ctx.port;
4938
5559
  if (isPortInUse(port)) {
@@ -4985,15 +5606,15 @@ Port ${port} is already in use.`));
4985
5606
  if (action === "start") {
4986
5607
  const spinner = ora2("Starting OpenCode...").start();
4987
5608
  const proc = await startOpenCode(port);
4988
- const health = await waitForOpenCodeHealth(port, 3e4);
5609
+ const health = await waitForOpenCodeHealth(port, INTERACTIVE_START_TIMEOUT_MS);
4989
5610
  if (!health.healthy) {
4990
5611
  spinner.fail("Failed to start OpenCode");
4991
5612
  throw new Error("OpenCode failed to start");
4992
5613
  }
4993
5614
  spinner.stop();
4994
- return { port, process: proc, version: health.version ?? null };
5615
+ return { port, process: proc, version: health.version ?? null, notReadyReason: null };
4995
5616
  }
4996
- return { port, process: null, version: null };
5617
+ return { port, process: null, version: null, notReadyReason: "you chose to continue without it" };
4997
5618
  }
4998
5619
 
4999
5620
  // src/commands/agent-lookup.ts
@@ -5035,19 +5656,21 @@ async function resolveAgentIdFromKey(authHeader) {
5035
5656
  return { agent_id: data.agent_id };
5036
5657
  }
5037
5658
  return {
5038
- error: "Cannot resolve runner ID: auth type is not agent_key. Please provide --agent explicitly."
5659
+ error: "Cannot resolve runner ID: auth type is not agent_key. Please provide --runner explicitly."
5039
5660
  };
5040
5661
  } catch (error2) {
5041
5662
  const message = error2 instanceof Error ? error2.message : "Unknown error";
5042
5663
  return { error: `Failed to resolve runner from key: ${message}` };
5043
5664
  }
5044
5665
  }
5666
+ var BEST_EFFORT_NOTIFY_TIMEOUT_MS = 2e3;
5045
5667
  async function notifyAgentDisconnected(agentId, authHeader) {
5046
5668
  const apiUrl = getApiUrlConfig();
5047
5669
  try {
5048
5670
  const response = await fetch(`${apiUrl}/runners/${agentId}/disconnect`, {
5049
5671
  method: "POST",
5050
- headers: { Authorization: authHeader }
5672
+ headers: { Authorization: authHeader },
5673
+ signal: AbortSignal.timeout(BEST_EFFORT_NOTIFY_TIMEOUT_MS)
5051
5674
  });
5052
5675
  if (!response.ok) {
5053
5676
  const serverMessage = await readErrorMessage(response);
@@ -5058,7 +5681,63 @@ async function notifyAgentDisconnected(agentId, authHeader) {
5058
5681
  }
5059
5682
  return { ok: true };
5060
5683
  } catch (error2) {
5061
- return { ok: false, error: error2 instanceof Error ? error2.message : String(error2) };
5684
+ return { ok: false, error: describeBestEffortError(error2) };
5685
+ }
5686
+ }
5687
+ function describeBestEffortError(error2) {
5688
+ const name = error2?.name;
5689
+ if (name === "TimeoutError" || name === "AbortError") {
5690
+ return `timed out after ${BEST_EFFORT_NOTIFY_TIMEOUT_MS}ms`;
5691
+ }
5692
+ return error2 instanceof Error ? error2.message : String(error2);
5693
+ }
5694
+ async function reportMicrovmId(agentId, authHeader, microvmId) {
5695
+ try {
5696
+ const apiUrl = getApiUrlConfig();
5697
+ const response = await fetch(`${apiUrl}/runners/${agentId}/microvm`, {
5698
+ method: "POST",
5699
+ headers: { Authorization: authHeader, "Content-Type": "application/json" },
5700
+ body: JSON.stringify({ microvm_id: microvmId }),
5701
+ signal: AbortSignal.timeout(BEST_EFFORT_NOTIFY_TIMEOUT_MS)
5702
+ });
5703
+ if (!response.ok) {
5704
+ const serverMessage = await readErrorMessage(response);
5705
+ return {
5706
+ ok: false,
5707
+ error: `HTTP ${response.status}${serverMessage ? `: ${serverMessage}` : ""}`
5708
+ };
5709
+ }
5710
+ return { ok: true };
5711
+ } catch (error2) {
5712
+ return { ok: false, error: describeBestEffortError(error2) };
5713
+ }
5714
+ }
5715
+ function toReportedWindow(window) {
5716
+ if (!window) return null;
5717
+ return { utilization: window.utilization, resets_at: window.resetsAt };
5718
+ }
5719
+ async function reportClaudeUsage(agentId, authHeader, snapshot) {
5720
+ try {
5721
+ const apiUrl = getApiUrlConfig();
5722
+ const response = await fetch(`${apiUrl}/runners/${agentId}/claude-usage`, {
5723
+ method: "POST",
5724
+ headers: { Authorization: authHeader, "Content-Type": "application/json" },
5725
+ body: JSON.stringify({
5726
+ five_hour: toReportedWindow(snapshot.fiveHour),
5727
+ seven_day: toReportedWindow(snapshot.sevenDay)
5728
+ }),
5729
+ signal: AbortSignal.timeout(BEST_EFFORT_NOTIFY_TIMEOUT_MS)
5730
+ });
5731
+ if (!response.ok) {
5732
+ const serverMessage = await readErrorMessage(response);
5733
+ return {
5734
+ ok: false,
5735
+ error: `HTTP ${response.status}${serverMessage ? `: ${serverMessage}` : ""}`
5736
+ };
5737
+ }
5738
+ return { ok: true };
5739
+ } catch (error2) {
5740
+ return { ok: false, error: describeBestEffortError(error2) };
5062
5741
  }
5063
5742
  }
5064
5743
  async function getAgentInfo(agentId, authHeader) {
@@ -5108,6 +5787,7 @@ var MAX_ACTIVITY_LOG_ENTRIES = 10;
5108
5787
  var CHANNEL_POLL_INTERVAL_MS = Number(process.env.EVIDENT_CHANNEL_POLL_INTERVAL_MS) || 2e3;
5109
5788
  var CHANNEL_STUCK_QUEUED_MS = Number(process.env.EVIDENT_STUCK_QUEUED_MS) || void 0;
5110
5789
  var SHUTDOWN_DRAIN_TIMEOUT_MS = Number(process.env.EVIDENT_SHUTDOWN_DRAIN_MS) || 25e3;
5790
+ var TELEMETRY_SHUTDOWN_TIMEOUT_MS = 5e3;
5111
5791
  function resolveLogLevel(options) {
5112
5792
  const accepted = Object.keys(LOG_LEVELS);
5113
5793
  const validate = (value, source) => {
@@ -5138,7 +5818,7 @@ function resolveFileSyncDirectories(raw, homeDir) {
5138
5818
  if (trimmed === "") {
5139
5819
  throw new Error("--enable-file-sync-to requires a directory path (got an empty value)");
5140
5820
  }
5141
- const expanded = trimmed === "~" ? homeDir : trimmed.startsWith("~/") ? join2(homeDir, trimmed.slice(2)) : trimmed;
5821
+ const expanded = trimmed === "~" ? homeDir : trimmed.startsWith("~/") ? join3(homeDir, trimmed.slice(2)) : trimmed;
5142
5822
  if (!isAbsolute2(expanded)) {
5143
5823
  throw new Error(`--enable-file-sync-to requires an absolute directory path; got "${entry}"`);
5144
5824
  }
@@ -5159,6 +5839,35 @@ function resolveFileSyncDirectories(raw, homeDir) {
5159
5839
  }
5160
5840
  return directories;
5161
5841
  }
5842
+ var DEFAULT_OPENCODE_START_TIMEOUT_SECONDS = 180;
5843
+ var MAX_OPENCODE_START_TIMEOUT_SECONDS = 3600;
5844
+ var OPENCODE_START_TIMEOUT_ENV = "EVIDENT_OPENCODE_START_TIMEOUT";
5845
+ function resolveOpenCodeStartTimeoutMs(options, env = process.env) {
5846
+ const defaultMs = DEFAULT_OPENCODE_START_TIMEOUT_SECONDS * 1e3;
5847
+ let raw;
5848
+ let source;
5849
+ if (options.opencodeStartTimeout !== void 0) {
5850
+ raw = options.opencodeStartTimeout;
5851
+ source = "--opencode-start-timeout";
5852
+ } else if (env[OPENCODE_START_TIMEOUT_ENV] !== void 0 && env[OPENCODE_START_TIMEOUT_ENV] !== "") {
5853
+ raw = env[OPENCODE_START_TIMEOUT_ENV];
5854
+ source = OPENCODE_START_TIMEOUT_ENV;
5855
+ } else {
5856
+ return { timeoutMs: defaultMs, warnings: [] };
5857
+ }
5858
+ const trimmed = raw.trim();
5859
+ const seconds = Number(trimmed);
5860
+ const isPositiveInteger = /^\d+$/.test(trimmed) && Number.isInteger(seconds) && seconds > 0;
5861
+ if (!isPositiveInteger || seconds > MAX_OPENCODE_START_TIMEOUT_SECONDS) {
5862
+ return {
5863
+ timeoutMs: defaultMs,
5864
+ warnings: [
5865
+ `Ignoring invalid ${source} "${raw}": expected a positive integer number of seconds (at most ${MAX_OPENCODE_START_TIMEOUT_SECONDS}); using the default ${DEFAULT_OPENCODE_START_TIMEOUT_SECONDS}s`
5866
+ ]
5867
+ };
5868
+ }
5869
+ return { timeoutMs: seconds * 1e3, warnings: [] };
5870
+ }
5162
5871
  function meetsThreshold(state, level) {
5163
5872
  return LOG_LEVELS[level] >= LOG_LEVELS[state.logLevel];
5164
5873
  }
@@ -5180,6 +5889,10 @@ function log2(state, message, level = "info") {
5180
5889
  function logActivity(state, entry) {
5181
5890
  const level = entry.level ?? (entry.type === "error" ? "error" : "info");
5182
5891
  if (!meetsThreshold(state, level)) return;
5892
+ forwardRunnerActivity(
5893
+ { level, message: entry.message, error: entry.error },
5894
+ { agentId: state.agentId, authHeader: state.authHeader }
5895
+ );
5183
5896
  const fullEntry = {
5184
5897
  ...entry,
5185
5898
  level,
@@ -5279,6 +5992,7 @@ async function handleAuthError(state, error2) {
5279
5992
  }
5280
5993
  async function driveChannels(state, driver) {
5281
5994
  let idlePolls = 0;
5995
+ let consecutiveDrainFailures = 0;
5282
5996
  let lastSeenProxiedActivityAt = state.lastProxiedActivityAt;
5283
5997
  let lastSeenAppliedFiles = driver.fileSyncActivity().appliedFiles;
5284
5998
  while (state.running) {
@@ -5296,6 +6010,7 @@ async function driveChannels(state, driver) {
5296
6010
  );
5297
6011
  try {
5298
6012
  const processed = await driver.drainPending();
6013
+ consecutiveDrainFailures = 0;
5299
6014
  state.messageCount += processed;
5300
6015
  const proxiedActivity = state.lastProxiedActivityAt !== lastSeenProxiedActivityAt;
5301
6016
  lastSeenProxiedActivityAt = state.lastProxiedActivityAt;
@@ -5330,8 +6045,32 @@ async function driveChannels(state, driver) {
5330
6045
  const errorMessage = error2 instanceof Error ? error2.message : String(error2);
5331
6046
  logActivity(state, { type: "error", error: `Channel processing error: ${errorMessage}` });
5332
6047
  if (state.interactive) displayStatus(state);
6048
+ if (driver.hasInFlightWatchers()) {
6049
+ consecutiveDrainFailures = 0;
6050
+ } else if (state.idleTimeout !== null) {
6051
+ consecutiveDrainFailures++;
6052
+ if (consecutiveDrainFailures === 1) {
6053
+ logActivity(state, {
6054
+ type: "info",
6055
+ message: `Cannot reach Evident, will exit if this persists past the idle timeout (timeout: ${state.idleTimeout}s)...`
6056
+ });
6057
+ if (state.interactive) displayStatus(state);
6058
+ }
6059
+ }
5333
6060
  }
5334
6061
  await new Promise((resolve3) => setTimeout(resolve3, CHANNEL_POLL_INTERVAL_MS));
6062
+ if (state.idleTimeout !== null && consecutiveDrainFailures >= 2) {
6063
+ const unreachableMs = consecutiveDrainFailures * CHANNEL_POLL_INTERVAL_MS;
6064
+ if (unreachableMs > state.idleTimeout * 1e3) {
6065
+ logActivity(state, {
6066
+ type: "info",
6067
+ level: "warn",
6068
+ message: `Exiting: could not reach Evident for ${consecutiveDrainFailures} consecutive polls (${Math.round(unreachableMs / 1e3)}s)`
6069
+ });
6070
+ if (state.interactive) displayStatus(state);
6071
+ break;
6072
+ }
6073
+ }
5335
6074
  if (state.idleTimeout !== null && idlePolls >= 2) {
5336
6075
  const idleMs = idlePolls * CHANNEL_POLL_INTERVAL_MS;
5337
6076
  if (idleMs > state.idleTimeout * 1e3) {
@@ -5343,8 +6082,8 @@ async function driveChannels(state, driver) {
5343
6082
  }
5344
6083
  }
5345
6084
  var SESSION_CLEANUP_FIRST_SWEEP_MS = 1e4;
5346
- async function runSweep(state, driver, config2) {
5347
- const mode = `age=${config2.maxAgeMs ?? "\u2014"} count=${config2.maxCount ?? "\u2014"}`;
6085
+ async function runSweep(state, driver, config) {
6086
+ const mode = `age=${config.maxAgeMs ?? "\u2014"} count=${config.maxCount ?? "\u2014"}`;
5348
6087
  try {
5349
6088
  const sessions = await listSessions(state.port);
5350
6089
  if (sessions === null) {
@@ -5357,8 +6096,8 @@ async function runSweep(state, driver, config2) {
5357
6096
  const toDelete = selectSessionsToDelete(
5358
6097
  sessions.map((s) => ({ id: s.id, lastActivityMs: sessionLastActivityMs(s) })),
5359
6098
  {
5360
- maxAgeMs: config2.maxAgeMs,
5361
- maxCount: config2.maxCount,
6099
+ maxAgeMs: config.maxAgeMs,
6100
+ maxCount: config.maxCount,
5362
6101
  nowMs: Date.now(),
5363
6102
  protectedIds: driver.protectedSessionIds()
5364
6103
  }
@@ -5394,7 +6133,7 @@ async function runSweep(state, driver, config2) {
5394
6133
  }
5395
6134
  }
5396
6135
  function scheduleSessionCleanup(state, driver, options) {
5397
- const config2 = resolveSessionCleanupConfig(
6136
+ const config = resolveSessionCleanupConfig(
5398
6137
  {
5399
6138
  maxAge: options.sessionCleanupMaxAge,
5400
6139
  maxCount: options.sessionCleanupMaxCount,
@@ -5402,21 +6141,109 @@ function scheduleSessionCleanup(state, driver, options) {
5402
6141
  },
5403
6142
  process.env
5404
6143
  );
5405
- for (const warning2 of config2.warnings) {
6144
+ for (const warning2 of config.warnings) {
5406
6145
  logActivity(state, { type: "info", level: "warn", message: `Session cleanup: ${warning2}` });
5407
6146
  }
5408
- if (!config2.enabled) return;
6147
+ if (!config.enabled) return;
5409
6148
  logActivity(state, {
5410
6149
  type: "info",
5411
- message: `Session cleanup enabled (age=${config2.maxAgeMs ?? "\u2014"}, count=${config2.maxCount ?? "\u2014"}, interval=${config2.intervalMs}ms)`
6150
+ message: `Session cleanup enabled (age=${config.maxAgeMs ?? "\u2014"}, count=${config.maxCount ?? "\u2014"}, interval=${config.intervalMs}ms)`
5412
6151
  });
5413
- const interval = setInterval(() => void runSweep(state, driver, config2), config2.intervalMs);
6152
+ const interval = setInterval(() => void runSweep(state, driver, config), config.intervalMs);
5414
6153
  const firstSweep = setTimeout(
5415
- () => void runSweep(state, driver, config2),
6154
+ () => void runSweep(state, driver, config),
5416
6155
  SESSION_CLEANUP_FIRST_SWEEP_MS
5417
6156
  );
5418
6157
  state.sessionCleanupTimers.push(interval, firstSweep);
5419
6158
  }
6159
+ function scheduleClaudeUsageReporting(state, options) {
6160
+ const { mode, warnings } = resolveClaudeUsageReportingMode(
6161
+ options.claudeUsageReporting,
6162
+ process.env
6163
+ );
6164
+ for (const warning2 of warnings) {
6165
+ logActivity(state, {
6166
+ type: "info",
6167
+ level: "warn",
6168
+ message: `Claude usage reporting: ${warning2}`
6169
+ });
6170
+ }
6171
+ if (mode === "off") {
6172
+ logActivity(state, {
6173
+ type: "info",
6174
+ level: "debug",
6175
+ message: "Claude usage reporting is off (--claude-usage-reporting off)"
6176
+ });
6177
+ return;
6178
+ }
6179
+ let consecutiveFailures = 0;
6180
+ const scheduleNextTick = () => {
6181
+ state.claudeUsageTimer = setTimeout(() => void tick(false), nextReportDelayMs());
6182
+ };
6183
+ const tick = async (isFirst) => {
6184
+ try {
6185
+ const usage = await getClaudeUsage();
6186
+ const result = await reportClaudeUsage(state.agentId, state.authHeader, usage);
6187
+ if (result.ok) {
6188
+ if (consecutiveFailures > 0) {
6189
+ logActivity(state, {
6190
+ type: "info",
6191
+ level: "info",
6192
+ message: "Claude usage reporting recovered"
6193
+ });
6194
+ }
6195
+ consecutiveFailures = 0;
6196
+ logActivity(state, {
6197
+ type: "info",
6198
+ level: "debug",
6199
+ message: "Reported Claude usage to Evident"
6200
+ });
6201
+ } else {
6202
+ consecutiveFailures++;
6203
+ logActivity(state, {
6204
+ type: "info",
6205
+ level: consecutiveFailures === 1 ? "warn" : "debug",
6206
+ message: `Failed to report Claude usage: ${result.error}`
6207
+ });
6208
+ }
6209
+ scheduleNextTick();
6210
+ } catch (error2) {
6211
+ if (error2 instanceof ClaudeUsageError && isLocalCredentialProblem(error2)) {
6212
+ if (mode === "on") {
6213
+ logActivity(state, {
6214
+ type: "info",
6215
+ level: "warn",
6216
+ message: "Claude usage reporting is forced on but no usable Claude Code login was found \u2014 run `claude` to sign in; reporting will keep retrying"
6217
+ });
6218
+ scheduleNextTick();
6219
+ } else if (isFirst) {
6220
+ logActivity(state, {
6221
+ type: "info",
6222
+ level: "debug",
6223
+ message: `Claude usage reporting: ${error2.message}`
6224
+ });
6225
+ } else {
6226
+ logActivity(state, {
6227
+ type: "info",
6228
+ level: "debug",
6229
+ message: `Claude usage reporting: ${error2.message}`
6230
+ });
6231
+ scheduleNextTick();
6232
+ }
6233
+ } else {
6234
+ consecutiveFailures++;
6235
+ const message = error2 instanceof Error ? error2.message : String(error2);
6236
+ logActivity(state, {
6237
+ type: "info",
6238
+ level: consecutiveFailures === 1 ? "warn" : "debug",
6239
+ message: `Claude usage reporting failed: ${message}`
6240
+ });
6241
+ scheduleNextTick();
6242
+ }
6243
+ }
6244
+ };
6245
+ state.claudeUsageTimer = setTimeout(() => void tick(true), FIRST_REPORT_DELAY_MS);
6246
+ }
5420
6247
  async function notifyOffline(state) {
5421
6248
  if (!state.agentId || !state.authHeader) return;
5422
6249
  if (!state.connected) {
@@ -5434,13 +6261,28 @@ async function notifyOffline(state) {
5434
6261
  if (state.interactive) displayStatus(state);
5435
6262
  }
5436
6263
  }
6264
+ async function timeShutdownPhase(state, durations, name, run2) {
6265
+ const startedAt = Date.now();
6266
+ try {
6267
+ return await run2();
6268
+ } finally {
6269
+ const elapsedMs = Date.now() - startedAt;
6270
+ durations[name] = elapsedMs;
6271
+ log2(state, `Shutdown phase ${name}: ${elapsedMs}ms`);
6272
+ }
6273
+ }
5437
6274
  async function cleanup(state, opts = {}) {
6275
+ const durations = {};
5438
6276
  state.running = false;
5439
6277
  for (const timer of state.sessionCleanupTimers) {
5440
6278
  clearInterval(timer);
5441
6279
  clearTimeout(timer);
5442
6280
  }
5443
6281
  state.sessionCleanupTimers = [];
6282
+ if (state.claudeUsageTimer) {
6283
+ clearTimeout(state.claudeUsageTimer);
6284
+ state.claudeUsageTimer = null;
6285
+ }
5444
6286
  if (opts.graceful && state.channelDriver) {
5445
6287
  state.channelDriver.stop();
5446
6288
  log2(state, "Draining in-flight channel work before shutdown...");
@@ -5448,7 +6290,13 @@ async function cleanup(state, opts = {}) {
5448
6290
  logActivity(state, { type: "info", message: "Draining in-flight work before shutdown..." });
5449
6291
  displayStatus(state);
5450
6292
  }
5451
- const settled = await state.channelDriver.waitForInFlight(SHUTDOWN_DRAIN_TIMEOUT_MS);
6293
+ const driver = state.channelDriver;
6294
+ const settled = await timeShutdownPhase(
6295
+ state,
6296
+ durations,
6297
+ "drain",
6298
+ () => driver.waitForInFlight(SHUTDOWN_DRAIN_TIMEOUT_MS)
6299
+ );
5452
6300
  if (!settled) {
5453
6301
  logActivity(state, {
5454
6302
  type: "info",
@@ -5457,13 +6305,15 @@ async function cleanup(state, opts = {}) {
5457
6305
  if (state.interactive) displayStatus(state);
5458
6306
  }
5459
6307
  }
5460
- await notifyOffline(state);
6308
+ await timeShutdownPhase(state, durations, "offline_notify", () => notifyOffline(state));
5461
6309
  if (state.connection) {
5462
- state.connection.close();
6310
+ const connection = state.connection;
6311
+ await timeShutdownPhase(state, durations, "tunnel_close", () => connection.close());
5463
6312
  state.connection = null;
5464
6313
  }
5465
6314
  if (state.opencodeProcess) {
5466
- stopOpenCode(state.opencodeProcess);
6315
+ const opencodeProcess = state.opencodeProcess;
6316
+ await timeShutdownPhase(state, durations, "opencode_stop", () => stopOpenCode(opencodeProcess));
5467
6317
  if (state.interactive) {
5468
6318
  logActivity(state, { type: "info", message: "Stopped OpenCode process" });
5469
6319
  displayStatus(state);
@@ -5472,6 +6322,7 @@ async function cleanup(state, opts = {}) {
5472
6322
  }
5473
6323
  state.opencodeProcess = null;
5474
6324
  }
6325
+ return durations;
5475
6326
  }
5476
6327
  async function run(options) {
5477
6328
  const interactive = isInteractive(options.json);
@@ -5479,7 +6330,7 @@ async function run(options) {
5479
6330
  let fileSyncDirectories;
5480
6331
  try {
5481
6332
  logLevel = resolveLogLevel(options);
5482
- fileSyncDirectories = resolveFileSyncDirectories(options.enableFileSyncTo, homedir2());
6333
+ fileSyncDirectories = resolveFileSyncDirectories(options.enableFileSyncTo, homedir3());
5483
6334
  } catch (error2) {
5484
6335
  const message = error2 instanceof Error ? error2.message : String(error2);
5485
6336
  if (options.json) {
@@ -5512,8 +6363,10 @@ async function run(options) {
5512
6363
  messageCount: 0,
5513
6364
  lastProxiedActivityAt: null,
5514
6365
  sessionCleanupTimers: [],
6366
+ claudeUsageTimer: null,
5515
6367
  authHeader: ""
5516
6368
  };
6369
+ setTelemetryAuthProvider(() => ({ authHeader: state.authHeader, agentId: state.agentId }));
5517
6370
  if (fileSyncDirectories.length > 0) {
5518
6371
  log2(state, `File sync enabled for: ${fileSyncDirectories.join(", ")}`);
5519
6372
  } else {
@@ -5542,14 +6395,38 @@ async function run(options) {
5542
6395
  const handleSignal = async () => {
5543
6396
  if (state.shuttingDown) return;
5544
6397
  state.shuttingDown = true;
6398
+ const shutdownStartedAt = Date.now();
5545
6399
  if (state.interactive) {
5546
6400
  logActivity(state, { type: "info", message: "Shutting down..." });
5547
6401
  displayStatus(state);
5548
6402
  } else {
5549
6403
  log2(state, "Shutting down...");
5550
6404
  }
5551
- await cleanup(state, { graceful: true });
5552
- await shutdownTelemetry();
6405
+ const durations = await cleanup(state, { graceful: true });
6406
+ const telemetryBudgetMs = Number(process.env.EVIDENT_TELEMETRY_SHUTDOWN_MS) || TELEMETRY_SHUTDOWN_TIMEOUT_MS;
6407
+ await timeShutdownPhase(state, durations, "telemetry_flush", async () => {
6408
+ let timer;
6409
+ const flushed = shutdownTelemetry().then(
6410
+ () => true,
6411
+ (error2) => {
6412
+ log2(
6413
+ state,
6414
+ `Telemetry flush failed during shutdown: ${error2 instanceof Error ? error2.message : String(error2)}`,
6415
+ "warn"
6416
+ );
6417
+ return true;
6418
+ }
6419
+ );
6420
+ const timedOut = new Promise((resolve3) => {
6421
+ timer = setTimeout(() => resolve3(false), telemetryBudgetMs);
6422
+ });
6423
+ if (!await Promise.race([flushed, timedOut])) {
6424
+ log2(state, `Telemetry flush exceeded ${telemetryBudgetMs}ms \u2014 exiting anyway`, "warn");
6425
+ }
6426
+ clearTimeout(timer);
6427
+ });
6428
+ const breakdown = Object.entries(durations).map(([phase, ms]) => `${phase}=${ms}ms`).join(" ");
6429
+ log2(state, `Shutdown complete in ${Date.now() - shutdownStartedAt}ms (${breakdown})`);
5553
6430
  process.exit(0);
5554
6431
  };
5555
6432
  process.on("SIGINT", handleSignal);
@@ -5663,40 +6540,67 @@ async function run(options) {
5663
6540
  }
5664
6541
  spinner?.succeed(`Runner: ${validation.agent.name || state.agentId}`);
5665
6542
  state.agentName = validation.agent.name;
6543
+ const microvmId = process.env.MICROVM_ID?.trim();
6544
+ if (microvmId) {
6545
+ const reported = await reportMicrovmId(state.agentId, state.authHeader, microvmId);
6546
+ if (reported.ok) {
6547
+ log2(state, "Reported MicroVM identity so this runner can be resumed rather than restarted");
6548
+ } else {
6549
+ const message = `Could not report MicroVM identity (future wakes will cold-start): ${reported.error}`;
6550
+ log2(state, message, "warn");
6551
+ if (state.interactive && !state.json) {
6552
+ logActivity(state, { type: "info", level: "warn", message });
6553
+ }
6554
+ }
6555
+ } else {
6556
+ log2(state, "Not running in a MicroVM (MICROVM_ID unset) \u2014 nothing to report", "debug");
6557
+ }
6558
+ const { timeoutMs: opencodeStartTimeoutMs, warnings: opencodeStartTimeoutWarnings } = resolveOpenCodeStartTimeoutMs(options, process.env);
6559
+ for (const warning2 of opencodeStartTimeoutWarnings) {
6560
+ logActivity(state, { type: "info", level: "warn", message: warning2 });
6561
+ }
5666
6562
  const ocSpinner = interactive && !state.json ? ora3("Checking OpenCode...").start() : null;
5667
6563
  try {
5668
6564
  const oc = await ensureOpenCodeRunning({
5669
6565
  port: state.port,
5670
6566
  interactive: state.interactive,
5671
6567
  agentId: state.agentId,
5672
- log: (message) => log2(state, message)
6568
+ log: (message) => log2(state, message),
6569
+ startTimeoutMs: opencodeStartTimeoutMs
5673
6570
  });
5674
6571
  state.port = oc.port;
5675
6572
  state.opencodeProcess = oc.process;
5676
6573
  state.opencodeVersion = oc.version;
5677
- state.opencodeConnected = oc.process !== null || oc.version !== null;
6574
+ state.opencodeConnected = oc.notReadyReason === null;
5678
6575
  const version2 = state.opencodeVersion ? ` (v${state.opencodeVersion})` : "";
5679
6576
  ocSpinner?.succeed(`OpenCode running on port ${state.port}${version2}`);
5680
- const versionWarning = buildOpenCodeVersionWarning(state.opencodeVersion);
5681
- if (versionWarning) {
5682
- log2(state, versionWarning, "warn");
5683
- if (state.interactive && !state.json) {
5684
- logActivity(state, { type: "info", level: "warn", message: versionWarning });
6577
+ if (!state.interactive && oc.notReadyReason !== null) {
6578
+ const message = `OpenCode is not ready on port ${state.port}: ${oc.notReadyReason}. The runner will still come online, but messages will fail until opencode answers \u2014 raise the wait with --opencode-start-timeout <seconds> (env ${OPENCODE_START_TIMEOUT_ENV}).`;
6579
+ logActivity(state, { type: "info", level: "warn", message });
6580
+ } else {
6581
+ const versionWarning = buildOpenCodeVersionWarning(state.opencodeVersion);
6582
+ if (versionWarning) {
6583
+ log2(state, versionWarning, "warn");
6584
+ if (state.interactive && !state.json) {
6585
+ logActivity(state, { type: "info", level: "warn", message: versionWarning });
6586
+ }
5685
6587
  }
5686
- }
5687
- const noProviderWarning = buildNoProviderWarning(await hasAnyConfiguredProvider(state.port));
5688
- if (noProviderWarning) {
5689
- log2(state, noProviderWarning, "warn");
5690
- if (state.interactive && !state.json) {
5691
- logActivity(state, { type: "info", level: "warn", message: noProviderWarning });
5692
- blank();
5693
- console.log(chalk6.yellow("\u26A0 No OpenCode model provider is configured."));
5694
- console.log(
5695
- chalk6.dim(
5696
- `Run ${chalk6.cyan("opencode auth login")} to set one up \u2014 messages will fail until then.`
5697
- )
5698
- );
5699
- blank();
6588
+ const noProviderWarning = buildNoProviderWarning(
6589
+ await hasAnyConfiguredProvider(state.port)
6590
+ );
6591
+ if (noProviderWarning) {
6592
+ log2(state, noProviderWarning, "warn");
6593
+ if (state.interactive && !state.json) {
6594
+ logActivity(state, { type: "info", level: "warn", message: noProviderWarning });
6595
+ blank();
6596
+ console.log(chalk6.yellow("\u26A0 No OpenCode model provider is configured."));
6597
+ console.log(
6598
+ chalk6.dim(
6599
+ `Run ${chalk6.cyan("opencode auth login")} to set one up \u2014 messages will fail until then.`
6600
+ )
6601
+ );
6602
+ blank();
6603
+ }
5700
6604
  }
5701
6605
  }
5702
6606
  } catch (error2) {
@@ -5714,7 +6618,7 @@ async function run(options) {
5714
6618
  // #559: the `--enable-file-sync-to` allow-list. Empty ⇒ pending files are
5715
6619
  // REJECTED with `file_sync_disabled` on the ack, not silently ignored.
5716
6620
  fileSyncDirectories,
5717
- homeDir: homedir2(),
6621
+ homeDir: homedir3(),
5718
6622
  log: (entry) => (
5719
6623
  // Thread the driver's real level straight through so `debug`/`warn`
5720
6624
  // survive the sink filter (they no longer collapse to info). `type`
@@ -5741,6 +6645,18 @@ async function run(options) {
5741
6645
  type: "info",
5742
6646
  message: `Tunnel ${isReconnect ? "reconnected" : "connected"} (runner: ${agentId})`
5743
6647
  });
6648
+ if (options.tunnelReadyFile) {
6649
+ const marker = writeTunnelReadyMarker(options.tunnelReadyFile, agentId);
6650
+ if (marker.ok) {
6651
+ log2(state, `Wrote tunnel readiness marker to ${options.tunnelReadyFile}`, "debug");
6652
+ } else {
6653
+ log2(
6654
+ state,
6655
+ `Failed to write tunnel readiness marker to ${options.tunnelReadyFile}: ${marker.error}`,
6656
+ "error"
6657
+ );
6658
+ }
6659
+ }
5744
6660
  emitAgentConnected(state.agentId, {
5745
6661
  port: state.port,
5746
6662
  cli_version: getCliVersion(),
@@ -5831,6 +6747,7 @@ async function run(options) {
5831
6747
  throw error2;
5832
6748
  }
5833
6749
  scheduleSessionCleanup(state, channelDriver, options);
6750
+ scheduleClaudeUsageReporting(state, options);
5834
6751
  if (!interactive || state.json) {
5835
6752
  log2(state, "Driving channel messages...");
5836
6753
  }
@@ -5885,13 +6802,17 @@ program.name("evident").description("Run OpenCode locally and connect it to Evid
5885
6802
  program.command("login").description("Authenticate with Evident").option("--token", "Use token-based authentication (for CI/CD)").option("--no-browser", "Do not open the browser automatically").action(login);
5886
6803
  program.command("logout").description("Remove stored credentials for the current endpoint").option("--all", "Remove stored credentials for all endpoints").action((options) => logout({ all: options.all }));
5887
6804
  program.command("whoami").description("Show the currently logged in user").action(whoami);
6805
+ program.command("claude-usage").description("[spike] Show Claude subscription usage (requires a local `claude login`)").action(claudeUsage);
5888
6806
  program.command("run").description("Connect to Evident and process messages").option("--runner [id]", "Runner ID to connect to (optional when EVIDENT_RUNNER_KEY is set)").option(
5889
6807
  "-a, --agent [id]",
5890
6808
  "Deprecated alias for --runner (still supported; --runner wins if both are given)"
5891
6809
  ).option("-p, --port <port>", "OpenCode port (default: 4096)", "4096").option(
5892
6810
  "--log-level <level>",
5893
6811
  "Log verbosity: debug | info | warn | error (default: info). Env: EVIDENT_LOG_LEVEL"
5894
- ).option("-v, --verbose", "Alias for --log-level debug (ignored if --log-level is set)").option("-c, --conversation <id>", "Process only this specific conversation").option("--idle-timeout <seconds>", "Exit after N seconds idle").option("--json", "Output in JSON format").option(
6812
+ ).option("-v, --verbose", "Alias for --log-level debug (ignored if --log-level is set)").option("-c, --conversation <id>", "Process only this specific conversation").option("--idle-timeout <seconds>", "Exit after N seconds idle").option(
6813
+ "--opencode-start-timeout <seconds>",
6814
+ "Seconds to wait for OpenCode to become healthy when the runner starts it (default: 180). Env: EVIDENT_OPENCODE_START_TIMEOUT"
6815
+ ).option("--json", "Output in JSON format").option(
5895
6816
  "--session-cleanup-max-age <duration>",
5896
6817
  "Delete OpenCode sessions idle longer than this (e.g. 7d, 24h). Enables cleanup. Env: EVIDENT_SESSION_CLEANUP_MAX_AGE"
5897
6818
  ).option(
@@ -5900,11 +6821,17 @@ program.command("run").description("Connect to Evident and process messages").op
5900
6821
  ).option(
5901
6822
  "--session-cleanup-interval <duration>",
5902
6823
  "How often the cleanup sweep runs (default: 1h). Env: EVIDENT_SESSION_CLEANUP_INTERVAL"
6824
+ ).option(
6825
+ "--claude-usage-reporting <mode>",
6826
+ "Report Claude subscription usage to Evident: auto | on | off (default: auto). Env: EVIDENT_CLAUDE_USAGE_REPORTING"
5903
6827
  ).option(
5904
6828
  "--enable-file-sync-to <dir>",
5905
6829
  "Allow Evident to write files into this directory (repeatable). Omit to disable file sync entirely.",
5906
6830
  (value, previous) => previous.concat([value]),
5907
6831
  []
6832
+ ).option(
6833
+ "--tunnel-ready-file <path>",
6834
+ "Path to write once the tunnel is connected (set by the MicroVM hooks; unused on a developer machine)"
5908
6835
  ).action(
5909
6836
  (options) => {
5910
6837
  run({
@@ -5917,14 +6844,21 @@ program.command("run").description("Connect to Evident and process messages").op
5917
6844
  verbose: options.verbose,
5918
6845
  conversation: options.conversation,
5919
6846
  idleTimeout: options.idleTimeout ? parseInt(options.idleTimeout, 10) : void 0,
6847
+ // Raw string — validation/precedence is single-sourced in run.ts's
6848
+ // resolveOpenCodeStartTimeoutMs (flag > EVIDENT_OPENCODE_START_TIMEOUT > 180s default).
6849
+ opencodeStartTimeout: options.opencodeStartTimeout,
5920
6850
  json: options.json,
5921
6851
  // Raw strings — the resolver in run.ts single-sources parsing (M1).
5922
6852
  sessionCleanupMaxAge: options.sessionCleanupMaxAge,
5923
6853
  sessionCleanupMaxCount: options.sessionCleanupMaxCount,
5924
6854
  sessionCleanupInterval: options.sessionCleanupInterval,
6855
+ // Raw string — the resolver in run.ts single-sources parsing
6856
+ // (resolveClaudeUsageReportingMode).
6857
+ claudeUsageReporting: options.claudeUsageReporting,
5925
6858
  // Raw values — expansion/validation is single-sourced in run.ts's
5926
6859
  // resolveFileSyncDirectories.
5927
- enableFileSyncTo: options.enableFileSyncTo
6860
+ enableFileSyncTo: options.enableFileSyncTo,
6861
+ tunnelReadyFile: options.tunnelReadyFile
5928
6862
  });
5929
6863
  }
5930
6864
  );