openmausbot 0.1.69 → 0.1.71

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 (66) hide show
  1. package/dist/assets/{index-BmHupgVJ.js → index-CJtyBonr.js} +1 -1
  2. package/dist/assets/index-CYLdYtYp.css +1 -0
  3. package/dist/assets/index-Di4ovwu5.js +289 -0
  4. package/dist/index.html +2 -2
  5. package/dist-server/container-mcp.js +6 -2
  6. package/dist-server/drivers/agents-proxy.js +305 -16
  7. package/dist-server/drivers/pi-mcp-extension.ts +1 -1
  8. package/dist-server/index.js +3141 -1124
  9. package/dist-server/local-computer-proxy.js +2 -0
  10. package/dist-server/local-computer.js +2 -1
  11. package/dist-server/mcp-gate.js +543 -0
  12. package/dist-server/mcp-server.js +38 -0
  13. package/dist-server/openmausbot.js +772 -260
  14. package/dist-server/pair-cli.js +772 -260
  15. package/dist-server/prepare-cloudflared.js +20 -3
  16. package/dist-server/proxy-paths.js +1 -0
  17. package/dist-server/server/agent-tool-policy.js +29 -0
  18. package/dist-server/server/bot-package.js +55 -2
  19. package/dist-server/server/browser-live.js +12 -7
  20. package/dist-server/server/config.js +40 -4
  21. package/dist-server/server/delegations.js +132 -74
  22. package/dist-server/server/drivers/acp/core.js +2 -0
  23. package/dist-server/server/drivers/acp/cursor.js +1 -1
  24. package/dist-server/server/drivers/acp/grok.js +5 -1
  25. package/dist-server/server/drivers/agents-proxy.js +324 -16
  26. package/dist-server/server/drivers/boxagent.js +1 -1
  27. package/dist-server/server/drivers/claude-login-auth.js +102 -9
  28. package/dist-server/server/drivers/claude.js +298 -21
  29. package/dist-server/server/drivers/codex.js +13 -4
  30. package/dist-server/server/drivers/pi.js +1 -1
  31. package/dist-server/server/engine-install.js +137 -0
  32. package/dist-server/server/env-path.js +12 -0
  33. package/dist-server/server/harness/bus.js +1 -1
  34. package/dist-server/server/harness/registry.js +31 -6
  35. package/dist-server/server/index.js +860 -215
  36. package/dist-server/server/local-computer.js +1 -1
  37. package/dist-server/server/mcp-gate-config.js +56 -0
  38. package/dist-server/server/mcp-gate.js +227 -0
  39. package/dist-server/server/mcp-trim.js +182 -0
  40. package/dist-server/server/memory-journal.js +440 -0
  41. package/dist-server/server/memory-store.js +308 -0
  42. package/dist-server/server/message-db.js +90 -3
  43. package/dist-server/server/package-export.js +25 -0
  44. package/dist-server/server/peer-approval.js +1 -1
  45. package/dist-server/server/peer-provenance.js +6 -2
  46. package/dist-server/server/proxy-paths.js +1 -0
  47. package/dist-server/server/redact.js +15 -0
  48. package/dist-server/server/resume-recovery.js +46 -0
  49. package/dist-server/server/routines.js +116 -55
  50. package/dist-server/server/screen-frame-gate.js +9 -6
  51. package/dist-server/server/skill-library.js +5 -5
  52. package/dist-server/server/steer-queue.js +20 -1
  53. package/dist-server/server/store.js +29 -7
  54. package/dist-server/server/system-prompt.js +14 -1
  55. package/dist-server/server/team-backup.js +53 -4
  56. package/dist-server/server/tool-summary.js +39 -0
  57. package/dist-server/server/tts/speech-text.js +3 -1
  58. package/dist-server/server/turn-context.js +18 -0
  59. package/dist-server/server/workspace.js +321 -22
  60. package/dist-server/shared/approval-mode.js +1 -1
  61. package/dist-server/shared/team-backup.js +14 -2
  62. package/dist-server/vps-container-mcp.js +6 -2
  63. package/package.json +1 -1
  64. package/skills/create-verification-skill/SKILL.md +2 -3
  65. package/dist/assets/index-COp5lMnv.js +0 -279
  66. package/dist/assets/index-Jk7TIeyW.css +0 -1
@@ -16,9 +16,12 @@ import { join, dirname, isAbsolute, normalize } from "node:path";
16
16
  import { DATA_DIR, stripWorkspaceCredentialEnv } from "../config.js";
17
17
  import { augmentedPath } from "../env-path.js";
18
18
  import { brokerSocketPath, describeSpawnFailure, execCli, killCliTree, spawnCli } from "../procs.js";
19
+ import { classifyResumeFailure, mayReplay, recoveryPromptFor } from "../resume-recovery.js";
19
20
  import { ClaudeLoginController } from "./claude-login-auth.js";
20
21
  import { computerProxyEnv } from "../container-computer.js";
22
+ import { gateServer, resultBudget } from "../mcp-gate-config.js";
21
23
  import { newEventId, newId } from "../contracts.js";
24
+ import { askInputSummary, commandSummary } from "../tool-summary.js";
22
25
  import { classifyError, computeBackoff, interruptibleDelay, RETRY_MAX_ATTEMPTS } from "./retry.js";
23
26
  import { applyClaudeInject, decodeInjectId, mergeLocalInject, probeLocalInjects, resolveInjectId, } from "./local-inject.js";
24
27
  import { appendNative } from "./native.js";
@@ -131,6 +134,133 @@ function claudeEnvironment(model, source = process.env, configDir, instanceEnvir
131
134
  delete env.ANTHROPIC_API_KEY;
132
135
  return env;
133
136
  }
137
+ /** Escape hatch back to the pre-isolation launch, where a bot inherited this
138
+ * machine's Claude Code setup: its MCP servers and connectors, skills,
139
+ * agents, hooks and personal CLAUDE.md. Set it only to recover a bot that
140
+ * genuinely depended on a user- or local-scope MCP server; the supported way
141
+ * to give a bot a server is the app's own `mcpServers` config or the bot
142
+ * project's `.mcp.json`. */
143
+ function inheritsUserConfig(env) {
144
+ return env.OMB_CLAUDE_INHERIT_USER_CONFIG === "1";
145
+ }
146
+ /** MCP servers the bot's own project declares in `<cwd>/.mcp.json`.
147
+ *
148
+ * The CLI would find this file itself, but the harness launches it with
149
+ * --strict-mcp-config, which makes the harness's config the only source.
150
+ * The project file IS part of the bot's definition (its cwd is chosen per
151
+ * bot), so it is forwarded verbatim — including `type: "http"`/`"sse"`
152
+ * entries the harness never mounts itself, because the CLI, not this code,
153
+ * is what has to understand them. A malformed file is ignored rather than
154
+ * failing the turn: an unreadable project config must not brick a bot. */
155
+ function projectMcpServers(cwd) {
156
+ let parsed;
157
+ try {
158
+ parsed = JSON.parse(readFileSync(join(cwd, ".mcp.json"), "utf8"));
159
+ }
160
+ catch {
161
+ return {};
162
+ }
163
+ const servers = parsed?.mcpServers;
164
+ if (!servers || typeof servers !== "object" || Array.isArray(servers))
165
+ return {};
166
+ const out = {};
167
+ for (const [name, server] of Object.entries(servers)) {
168
+ if (server && typeof server === "object" && !Array.isArray(server))
169
+ out[name] = server;
170
+ }
171
+ return out;
172
+ }
173
+ /** The CLI compacts its own session when it approaches a window. Left alone
174
+ * that window is the model's, so a Sonnet 5 session runs to something near a
175
+ * million tokens before anything happens — and every model call until then
176
+ * re-reads the whole thing. The measured food-ordering thread sat at 330k
177
+ * tokens per call and looked perfectly healthy to the CLI.
178
+ *
179
+ * So the harness picks the window instead. This delegates the actual
180
+ * compaction to the CLI, which owns the session and already has a summarizer
181
+ * for it; the harness only decides when it is worth paying for.
182
+ *
183
+ * OMB_CLAUDE_AUTOCOMPACT takes a token count, "auto" to hand the decision
184
+ * back to the CLI, or "off" to pass nothing at all. The CLI rejects a window
185
+ * outside 100k-1M as a hard argument error, so a configured value is clamped
186
+ * rather than passed through: a mistyped setting must not fail every turn. */
187
+ export function autoCompactWindow(env) {
188
+ const raw = (env.OMB_CLAUDE_AUTOCOMPACT ?? "").trim().toLowerCase();
189
+ if (raw === "off")
190
+ return null;
191
+ if (raw === "auto")
192
+ return "auto";
193
+ const parsed = raw ? Number(raw) : DEFAULT_AUTOCOMPACT_TOKENS;
194
+ if (!Number.isFinite(parsed) || parsed <= 0)
195
+ return String(DEFAULT_AUTOCOMPACT_TOKENS);
196
+ return String(Math.min(1_000_000, Math.max(100_000, Math.floor(parsed))));
197
+ }
198
+ /** Generous for real work, and still a third of where a 1M-window session
199
+ * would otherwise get to. With bot tool results gated (mcp-gate.ts) most
200
+ * threads never reach it; this is the backstop for the ones that do. */
201
+ const DEFAULT_AUTOCOMPACT_TOKENS = 200_000;
202
+ /** The Claude CLI version that first accepted each flag the harness passes
203
+ * for context control. An unknown flag is a hard argument error, so passing
204
+ * one to an older CLI would fail every turn rather than degrade; each flag
205
+ * is therefore only passed to a CLI known to accept it.
206
+ *
207
+ * Verified against the published binaries, not the changelog (which never
208
+ * records `--autocompact`): `--strict-mcp-config` is present in 1.0.60 and
209
+ * absent from 1.0.0; `--setting-sources` first appears in 1.0.122 (1.0.120
210
+ * lacks it); `--autocompact` first appears in 2.1.122 (2.1.121 lacks it). */
211
+ export const CLAUDE_FLAG_FLOORS = {
212
+ "--strict-mcp-config": [1, 0, 60],
213
+ "--setting-sources": [1, 0, 122],
214
+ "--autocompact": [2, 1, 122],
215
+ };
216
+ /** The newest floor above: a CLI at or past it accepts everything the
217
+ * harness sends. Below it the engine still works, minus the flags the CLI
218
+ * predates, and the Engines page suggests an update. */
219
+ export const CLAUDE_CONTEXT_CONTROL_MIN_VERSION = CLAUDE_FLAG_FLOORS["--autocompact"];
220
+ /** `claude --version` prints "2.1.232 (Claude Code)"; the first dotted triple
221
+ * is the version. Null when nothing parses, e.g. a wrapper that prints its
222
+ * own banner first — see claudeCliSupports for how that is treated. */
223
+ export function parseClaudeCliVersion(stdout) {
224
+ const match = /(\d+)\.(\d+)\.(\d+)/.exec(stdout ?? "");
225
+ if (!match)
226
+ return null;
227
+ return [Number(match[1]), Number(match[2]), Number(match[3])];
228
+ }
229
+ function versionAtLeast(installed, floor) {
230
+ for (let i = 0; i < 3; i += 1) {
231
+ if (installed[i] !== floor[i])
232
+ return installed[i] > floor[i];
233
+ }
234
+ return true;
235
+ }
236
+ /** Whether a CLI reporting `version` accepts `flag`. A version that could
237
+ * not be parsed counts as current: every CLI that predates a floor prints a
238
+ * plain "x.y.z (Claude Code)", so an unreadable version is far more likely
239
+ * a newer wrapper than an old build, and withholding the flags from a modern
240
+ * CLI would silently re-open the context leak this file exists to close. */
241
+ export function claudeCliSupports(version, flag) {
242
+ return version === null || versionAtLeast(version, CLAUDE_FLAG_FLOORS[flag]);
243
+ }
244
+ /** The Engines-page notice for a CLI older than the newest floor. The engine
245
+ * keeps working: turns run without the flags the CLI predates, which means
246
+ * no harness-picked compaction window and, on a very old CLI, no isolation
247
+ * from this machine's own Claude Code setup. */
248
+ export function claudeCliUpdate(version, cli) {
249
+ const parsed = parseClaudeCliVersion(version);
250
+ if (!parsed || versionAtLeast(parsed, CLAUDE_CONTEXT_CONTROL_MIN_VERSION))
251
+ return undefined;
252
+ const floor = CLAUDE_CONTEXT_CONTROL_MIN_VERSION.join(".");
253
+ const missing = Object.keys(CLAUDE_FLAG_FLOORS)
254
+ .filter((flag) => !claudeCliSupports(parsed, flag));
255
+ return {
256
+ title: "Update Claude Code for context controls",
257
+ message: `Claude Code ${parsed.join(".")} predates ${floor}, so bots run without ${missing.join(", ")}: ` +
258
+ "no compaction window picked by OpenMausBot" +
259
+ (missing.includes("--setting-sources") ? ", and bots still see this machine's own Claude Code setup" : "") +
260
+ ". Update it, then refresh Engines.",
261
+ command: cli === "claude" ? "claude update" : `${cli} update`,
262
+ };
263
+ }
134
264
  const DRIVER_KIND = "claudeAgent";
135
265
  // model catalog ported from upstream packages/contracts/src/model.ts
136
266
  export const STATIC_CLAUDE_MODELS = {
@@ -235,15 +365,7 @@ function systemEndedReply(kind) {
235
365
  }
236
366
  /** One human-readable line for an ask — what the card subtitle shows. */
237
367
  function askSummary(ask) {
238
- const input = ask.input ?? {};
239
- if (typeof input.question === "string")
240
- return input.question.slice(0, 300);
241
- if (typeof input.command === "string")
242
- return input.command.slice(0, 200);
243
- if (typeof input.url === "string")
244
- return input.url.slice(0, 200);
245
- const text = JSON.stringify(input);
246
- return text === "{}" ? (ask.tool ?? "tool") : text.slice(0, 200);
368
+ return askInputSummary(ask.input) ?? ask.tool ?? "tool";
247
369
  }
248
370
  export function permissionSocketPath(threadId) {
249
371
  // A readable prefix alone is not unique: ids that agree on their first
@@ -435,7 +557,7 @@ export async function createPermissionBroker(opts) {
435
557
  if (!server)
436
558
  throw new Error("claude: permission broker could not bind a local socket");
437
559
  const drain = () => {
438
- for (const p of [...pending.values()]) {
560
+ for (const p of Array.from(pending.values())) {
439
561
  const { behavior, message } = systemEndedReply(p.ask.kind);
440
562
  p.finish(behavior, message, "system");
441
563
  }
@@ -524,6 +646,17 @@ function firstText(content) {
524
646
  /** Claude's stream-json input accepts the same image source blocks as the
525
647
  * Anthropic Messages API. Keep the old string form for text-only turns so a
526
648
  * CLI update cannot disturb the overwhelmingly common path. */
649
+ /** How a mid-session change to the volatile half of the system prompt
650
+ * reaches a model whose process was launched with the old copy. The CLI's
651
+ * own out-of-band convention inside a user turn, and it costs one short
652
+ * append rather than a relaunch that re-uploads the whole prompt cache. */
653
+ function withVolatileNote(text, volatile) {
654
+ const body = volatile.trim()
655
+ ? `This part of your instructions changed since this session started. It replaces the earlier copy:\n\n${volatile.trim()}`
656
+ : "The notes that were in your instructions when this session started have been cleared.";
657
+ const note = `<system-reminder>\n${body}\n</system-reminder>`;
658
+ return text ? `${note}\n\n${text}` : note;
659
+ }
527
660
  function claudeUserMessage(text, images) {
528
661
  if (!images?.length)
529
662
  return { type: "user", message: { role: "user", content: text } };
@@ -594,6 +727,15 @@ export const ClaudeDriver = {
594
727
  }
595
728
  };
596
729
  await refreshModels();
730
+ // The installed CLI's version as snapshot() last read it, so a flag the
731
+ // CLI does not know is never passed to it (CLAUDE_FLAG_FLOORS). The
732
+ // harness snapshots every instance whenever it describes them — app
733
+ // load, the Engines page, and right after `claude update`, which is
734
+ // exactly when the answer changes — so a turn normally finds it filled.
735
+ // A turn before any snapshot assumes a current CLI rather than paying a
736
+ // CLI start-up of its own: the flags are the default, the exception is
737
+ // the older install, and the next snapshot corrects it.
738
+ let cliVersion = null;
597
739
  const listeners = new Set();
598
740
  // one active turn per thread; a second send while busy is a caller bug
599
741
  const active = new Map();
@@ -659,7 +801,7 @@ export const ClaudeDriver = {
659
801
  });
660
802
  };
661
803
  const emit = (event) => {
662
- for (const l of [...listeners])
804
+ for (const l of Array.from(listeners))
663
805
  l(event);
664
806
  };
665
807
  const base = (threadId, turnId) => ({
@@ -718,6 +860,26 @@ export const ClaudeDriver = {
718
860
  args.push("--disallowedTools", config.disallowedTools.join(","));
719
861
  }
720
862
  const turnEnvironment = environment();
863
+ const isolated = !inheritsUserConfig(turnEnvironment);
864
+ if (isolated) {
865
+ // A bot gets the tools and instructions its owner gave it, not
866
+ // whatever this machine's Claude Code happens to be set up with.
867
+ // Without these the CLI silently adds, to EVERY turn of every bot:
868
+ // the desktop's own MCP servers and claude.ai connectors (one
869
+ // measured desktop mounted 407 extra tools, ~10k tokens), its skill
870
+ // and agent listings, its hooks, and its personal CLAUDE.md. Every
871
+ // model call in the session then re-reads all of it.
872
+ // Each flag only on a CLI that accepts it: an unknown flag is an
873
+ // argument error that would fail every turn (CLAUDE_FLAG_FLOORS).
874
+ if (claudeCliSupports(cliVersion, "--strict-mcp-config"))
875
+ args.push("--strict-mcp-config");
876
+ if (claudeCliSupports(cliVersion, "--setting-sources"))
877
+ args.push("--setting-sources", "project");
878
+ }
879
+ const compactWindow = autoCompactWindow(turnEnvironment);
880
+ if (compactWindow && claudeCliSupports(cliVersion, "--autocompact")) {
881
+ args.push("--autocompact", compactWindow);
882
+ }
721
883
  const turnModel = await resolveClaudeTurnModel(turn.model, turnEnvironment);
722
884
  const injected = applyClaudeInject({ ...turnEnvironment }, turnModel);
723
885
  if (injected.model)
@@ -794,10 +956,39 @@ export const ClaudeDriver = {
794
956
  // routes every custom tool call through the ogb permission broker
795
957
  // into an Allow/Deny card. Reserved names were filtered upstream;
796
958
  // skip any residual collision instead of clobbering a built-in.
959
+ // Bot-owned servers, gated below: they are the ones that answer for a
960
+ // machine rather than for a context window.
961
+ const botOwned = new Set();
797
962
  for (const [name, server] of Object.entries(turn.integrations?.custom ?? {})) {
798
963
  if (name in mcpServers)
799
964
  continue;
800
965
  mcpServers[name] = { ...server };
966
+ botOwned.add(name);
967
+ }
968
+ // --strict-mcp-config (above) makes this config the CLI's only source
969
+ // of MCP servers, so a server the bot's OWN project declares would
970
+ // otherwise vanish with the machine's. Merge it last: a project file
971
+ // can add servers but never shadow a harness-owned mount.
972
+ if (isolated && turn.cwd) {
973
+ for (const [name, server] of Object.entries(projectMcpServers(turn.cwd))) {
974
+ if (name in mcpServers)
975
+ continue;
976
+ mcpServers[name] = server;
977
+ botOwned.add(name);
978
+ }
979
+ }
980
+ // One tool call can put more into the conversation than the whole rest
981
+ // of the session: a single product search measured 60-140 KB of JSON,
982
+ // and the CLI re-reads it on every later model call. The harness never
983
+ // sees these calls — the CLI runs the server itself — so the only place
984
+ // to stand is between the two processes. Harness-owned mounts (the
985
+ // permission broker, computer, browser, agents, dweb) are already
986
+ // bounded and are deliberately left alone.
987
+ const budget = resultBudget(turnEnvironment);
988
+ for (const name of botOwned) {
989
+ const gated = gateServer({ name, server: mcpServers[name], threadId, budget, nodeEnv: NODE_ENV_FLAG });
990
+ if (gated)
991
+ mcpServers[name] = gated;
801
992
  }
802
993
  // Keep ask_user available even in Full access. Native bypass skips
803
994
  // permission prompts, not questions requiring a person's answer.
@@ -832,7 +1023,9 @@ export const ClaudeDriver = {
832
1023
  const keyArgs = args.filter((a, i) => !privateFileFlags.has(a) && !privateFileFlags.has(args[i - 1] ?? ""));
833
1024
  const argsKey = JSON.stringify({
834
1025
  args: keyArgs,
835
- system: turn.system ?? null,
1026
+ // the volatile half is deliberately absent: it must not respawn a
1027
+ // healthy session (see Session.volatile)
1028
+ system: turn.systemStable ?? turn.system ?? null,
836
1029
  mcpServers,
837
1030
  cwd,
838
1031
  model: injected.model ?? null,
@@ -852,7 +1045,12 @@ export const ClaudeDriver = {
852
1045
  killCliTree(live.child);
853
1046
  }, turnId, broker: live.broker });
854
1047
  emit({ ...base(threadId, turnId), type: "turn.started" });
855
- const written = await writeUser(live, threadId, promptMsg);
1048
+ const volatile = turn.systemVolatile ?? "";
1049
+ const message = volatile === live.volatile
1050
+ ? promptMsg
1051
+ : claudeUserMessage(withVolatileNote(turn.text, volatile), turn.images);
1052
+ live.volatile = volatile;
1053
+ const written = await writeUser(live, threadId, message);
856
1054
  if (!written) {
857
1055
  active.delete(threadId);
858
1056
  live.turn = null;
@@ -982,7 +1180,9 @@ export const ClaudeDriver = {
982
1180
  mcpConfigPath,
983
1181
  systemPromptPath,
984
1182
  argsKey,
1183
+ volatile: turn.systemVolatile ?? "",
985
1184
  sessionId: sessionId ?? newSessionId,
1185
+ sawInit: false,
986
1186
  turn: { turnId, settled: false, sawStreamDelta: false },
987
1187
  idleTimer: null,
988
1188
  closing: false,
@@ -1035,6 +1235,7 @@ export const ClaudeDriver = {
1035
1235
  switch (o.type) {
1036
1236
  case "system":
1037
1237
  if (o.subtype === "init") {
1238
+ session.sawInit = true;
1038
1239
  if (typeof o.session_id === "string")
1039
1240
  session.sessionId = o.session_id;
1040
1241
  emit({ ...base(threadId, currentTurnId()), type: "session.started", sessionId: o.session_id, model: o.model });
@@ -1087,7 +1288,14 @@ export const ClaudeDriver = {
1087
1288
  }
1088
1289
  for (const b of Array.isArray(msg.content) ? msg.content : []) {
1089
1290
  if (b.type === "tool_use") {
1090
- emit({ ...base(threadId, currentTurnId()), type: "item.started", itemType: "tool", itemId: b.id, title: b.name });
1291
+ emit({
1292
+ ...base(threadId, currentTurnId()),
1293
+ type: "item.started",
1294
+ itemType: "tool",
1295
+ itemId: b.id,
1296
+ title: b.name,
1297
+ summary: commandSummary(b.input),
1298
+ });
1091
1299
  }
1092
1300
  }
1093
1301
  if (msg.usage) {
@@ -1240,6 +1448,67 @@ export const ClaudeDriver = {
1240
1448
  })();
1241
1449
  return;
1242
1450
  }
1451
+ // A --resume that the CLI never acknowledged: it exited without
1452
+ // an `init` frame, so it never read the prompt and this turn has
1453
+ // caused nothing. Without this the thread is BRICKED — the dead
1454
+ // cursor is never cleared, so every later turn resumes the same
1455
+ // missing session and fails identically, and the user has no way
1456
+ // back except switching engines. One fresh session, carrying the
1457
+ // harness's rebuild of the conversation. Exactly one: the relaunch
1458
+ // offers no cursor, so `attempted` is false there and a second
1459
+ // failure is reported like any other.
1460
+ const resumeFailure = classifyResumeFailure({
1461
+ attempted: Boolean(sessionId),
1462
+ rejected: !session.sawInit,
1463
+ promptSubmitted: session.sawInit,
1464
+ producedOutput: session.turn.sawStreamDelta,
1465
+ });
1466
+ if (mayReplay(resumeFailure) && !retry.cancelled) {
1467
+ const recovery = recoveryPromptFor({
1468
+ recoveryText: turn.recoveryText,
1469
+ currentText: turn.text,
1470
+ failure: resumeFailure,
1471
+ });
1472
+ session.broker?.close();
1473
+ session.broker = undefined;
1474
+ if (session.mcpConfigPath) {
1475
+ try {
1476
+ rmSync(dirname(session.mcpConfigPath), { recursive: true, force: true });
1477
+ }
1478
+ catch { }
1479
+ session.mcpConfigPath = null;
1480
+ }
1481
+ if (session.systemPromptPath) {
1482
+ removePrivateTempDir(session.systemPromptPath);
1483
+ session.systemPromptPath = null;
1484
+ }
1485
+ sessions.delete(threadId);
1486
+ session.turn = null;
1487
+ active.delete(threadId);
1488
+ emit({
1489
+ ...base(threadId, turnId),
1490
+ type: "turn.retrying",
1491
+ attempt: retry.attempt + 1,
1492
+ delayMs: 0,
1493
+ reason: "resume_rejected",
1494
+ });
1495
+ void (async () => {
1496
+ try {
1497
+ // no cursor: a fresh session, carrying the rebuild
1498
+ await sendTurn({ ...turn, resumeCursor: undefined, recoveryText: undefined, text: recovery.text });
1499
+ }
1500
+ catch (e) {
1501
+ retryState.delete(threadId);
1502
+ emit({
1503
+ ...base(threadId, turnId),
1504
+ type: "runtime.error",
1505
+ message: e instanceof Error ? e.message : String(e),
1506
+ });
1507
+ emit({ ...base(threadId, turnId), type: "turn.completed", ok: false, stopReason: "exit_before_result", cost: null });
1508
+ }
1509
+ })();
1510
+ return;
1511
+ }
1243
1512
  retryState.delete(threadId);
1244
1513
  emit({
1245
1514
  ...base(threadId, currentTurnId()),
@@ -1298,11 +1567,13 @@ export const ClaudeDriver = {
1298
1567
  });
1299
1568
  if (!version)
1300
1569
  return { state: "unavailable", reason: `\`${config.cli}\` CLI not found` };
1570
+ cliVersion = parseClaudeCliVersion(version);
1301
1571
  const auth = await claudeAuthStatus(config.cli, env);
1302
1572
  // claudeEnvironment strips ANTHROPIC_API_KEY, so turns run on the
1303
1573
  // CLI's own login (Pro/Max): the cost it reports is what the call
1304
1574
  // WOULD bill, not a charge
1305
- return { state: "available", version, ...auth, billing: "subscription" };
1575
+ const update = claudeCliUpdate(version, config.cli);
1576
+ return { state: "available", version, ...auth, ...(update ? { update } : {}), billing: "subscription" };
1306
1577
  };
1307
1578
  /** One-shot Claude call with the prompt on stdin, never argv. Approval
1308
1579
  * summaries can contain paths, commands, or secrets, so the generic
@@ -1376,6 +1647,7 @@ export const ClaudeDriver = {
1376
1647
  getAuthentication: (flowId) => login.get(flowId),
1377
1648
  completeAuthentication: (flowId, code) => login.complete(flowId, code),
1378
1649
  cancelAuthentication: () => login.cancel(),
1650
+ signOut: () => login.signOut(),
1379
1651
  adapter: {
1380
1652
  provider: DRIVER_KIND,
1381
1653
  capabilities: {
@@ -1412,7 +1684,7 @@ export const ClaudeDriver = {
1412
1684
  stopAll: async () => {
1413
1685
  for (const { stop } of active.values())
1414
1686
  stop();
1415
- for (const threadId of [...sessions.keys()])
1687
+ for (const threadId of Array.from(sessions.keys()))
1416
1688
  closeSession(threadId, "stopAll");
1417
1689
  },
1418
1690
  onEvent: (listener) => {
@@ -1423,11 +1695,16 @@ export const ClaudeDriver = {
1423
1695
  generateText: (prompt) => generateReview(prompt),
1424
1696
  reviewPermission: generateReview,
1425
1697
  dispose: async () => {
1426
- for (const { stop } of active.values())
1427
- stop();
1428
- for (const threadId of [...sessions.keys()])
1429
- closeSession(threadId, "dispose");
1430
- listeners.clear();
1698
+ try {
1699
+ await login.dispose();
1700
+ }
1701
+ finally {
1702
+ for (const { stop } of active.values())
1703
+ stop();
1704
+ for (const threadId of Array.from(sessions.keys()))
1705
+ closeSession(threadId, "dispose");
1706
+ listeners.clear();
1707
+ }
1431
1708
  },
1432
1709
  };
1433
1710
  },
@@ -22,6 +22,7 @@ import { codexLocalProviderArgs } from "./local-inject.js";
22
22
  import { augmentedPath, splitCliString } from "../env-path.js";
23
23
  import { classifyError, computeBackoff, RETRY_MAX_ATTEMPTS } from "./retry.js";
24
24
  import { appendNative } from "./native.js";
25
+ import { commandSummary } from "../tool-summary.js";
25
26
  import { codexDeveloperInstructions, syncCodexInstructions } from "./codex-instructions.js";
26
27
  import { CodexDeviceAuthController } from "./codex-device-auth.js";
27
28
  import { codexAccountEmail } from "./codex-identity.js";
@@ -434,7 +435,7 @@ export const CodexDriver = {
434
435
  const listeners = new Set();
435
436
  const active = new Map();
436
437
  const emit = (event) => {
437
- for (const l of [...listeners])
438
+ for (const l of Array.from(listeners))
438
439
  l(event);
439
440
  };
440
441
  const base = (threadId, turnId) => ({
@@ -571,7 +572,7 @@ export const CodexDriver = {
571
572
  if (state.settled)
572
573
  return;
573
574
  state.settled = true;
574
- for (const finish of [...asks.values()])
575
+ for (const finish of Array.from(asks.values()))
575
576
  finish("deny", "OpenMausBot: the turn ended", "system");
576
577
  for (const p of rpcPending.values())
577
578
  p.reject(new Error("turn settled"));
@@ -738,8 +739,16 @@ export const CodexDriver = {
738
739
  : item.type === "webSearch"
739
740
  ? "web_search"
740
741
  : null;
741
- if (title)
742
- emit({ ...base(threadId, turnId), type: "item.started", itemType: "tool", itemId: item.id, title });
742
+ if (title) {
743
+ emit({
744
+ ...base(threadId, turnId),
745
+ type: "item.started",
746
+ itemType: "tool",
747
+ itemId: item.id,
748
+ title,
749
+ summary: item.type === "commandExecution" ? commandSummary({ command: item.command }) : undefined,
750
+ });
751
+ }
743
752
  break;
744
753
  }
745
754
  case "item/completed": {
@@ -389,7 +389,7 @@ export const PiDriver = {
389
389
  // one active turn per thread
390
390
  const active = new Map();
391
391
  const emit = (event) => {
392
- for (const l of [...listeners])
392
+ for (const l of Array.from(listeners))
393
393
  l(event);
394
394
  };
395
395
  const base = (threadId, turnId) => ({
@@ -0,0 +1,137 @@
1
+ // Install an engine's command-line app from Settings, on the machine that
2
+ // runs this server, as this process's own user, into a directory the app
3
+ // owns. No sudo, no shell, no terminal: the package name comes from the
4
+ // driver's own install descriptor and never from a request, npm runs with
5
+ // a fixed argument list, and the binary is found on the engines' PATH
6
+ // afterwards because that directory is registered ahead of everything else.
7
+ import { existsSync, mkdirSync } from "node:fs";
8
+ import { join } from "node:path";
9
+ import { stripVTControlCharacters } from "node:util";
10
+ import { DATA_DIR, stripWorkspaceCredentialEnv } from "./config.js";
11
+ import { augmentedPath, findCliCandidates, registerPathDir, resetPathCache } from "./env-path.js";
12
+ import { killCliTree, spawnCli } from "./procs.js";
13
+ const MAX_OUTPUT = 64 * 1024;
14
+ const DEFAULT_TIMEOUT_MS = 10 * 60_000;
15
+ /** npm's global prefix for engines the app installs itself. */
16
+ export function enginesPrefix(baseDir = DATA_DIR) {
17
+ return join(baseDir, "tools", "npm");
18
+ }
19
+ /** Where that prefix puts executables: `bin/` on POSIX, the prefix itself on Windows. */
20
+ export function enginesBinDir(baseDir = DATA_DIR, platform = process.platform) {
21
+ return platform === "win32" ? enginesPrefix(baseDir) : join(enginesPrefix(baseDir), "bin");
22
+ }
23
+ /** Called once at boot: engines installed here win over any other copy on PATH. */
24
+ export function registerEnginesBinDir(baseDir = DATA_DIR) {
25
+ registerPathDir(enginesBinDir(baseDir));
26
+ }
27
+ /** The npm package a driver's own install one-liner names, when it is one. */
28
+ export function npmPackageOf(install) {
29
+ const line = install?.command?.linux ?? install?.command?.darwin ?? install?.command?.win32;
30
+ const match = line ? /^npm install -g ((?:@[\w.-]+\/)?[\w.-]+)$/.exec(line.trim()) : null;
31
+ return match ? match[1] : null;
32
+ }
33
+ export function npmAvailable() {
34
+ return findCliCandidates("npm").length > 0;
35
+ }
36
+ /** What Settings may install for this engine on this machine, or null. A
37
+ * managed engine keeps its own verified download; anything else needs a
38
+ * plain npm package and npm on PATH. */
39
+ export function serverInstallFor(install, npmPresent = npmAvailable()) {
40
+ if (!install || install.managed)
41
+ return null;
42
+ const pkg = npmPackageOf(install);
43
+ return pkg && npmPresent ? { package: pkg } : null;
44
+ }
45
+ const running = new Map();
46
+ /** Install or update one npm engine. Concurrent clicks share one npm run. */
47
+ export function installNpmEngine(pkg, options = {}) {
48
+ const key = `${options.baseDir ?? DATA_DIR} ${pkg}`;
49
+ const existing = running.get(key);
50
+ if (existing)
51
+ return existing;
52
+ const run = installOnce(pkg, options).finally(() => running.delete(key));
53
+ running.set(key, run);
54
+ return run;
55
+ }
56
+ async function installOnce(pkg, options) {
57
+ const prefix = enginesPrefix(options.baseDir);
58
+ mkdirSync(prefix, { recursive: true });
59
+ const env = {
60
+ ...(options.env ?? process.env),
61
+ PATH: options.path ?? augmentedPath(),
62
+ NO_COLOR: "1",
63
+ npm_config_update_notifier: "false",
64
+ npm_config_fund: "false",
65
+ npm_config_audit: "false",
66
+ };
67
+ // Workspace credentials (xai/box/voice keys) are not npm's to see.
68
+ stripWorkspaceCredentialEnv(env);
69
+ // npm 11 skips a dependency's install script unless the package is named
70
+ // here; the engines that need one (Claude Code) are exactly these.
71
+ const args = ["install", "-g", "--prefix", prefix, "--loglevel=error", `--allow-scripts=${pkg}`, `${pkg}@latest`];
72
+ const result = await runNpm(args, env, prefix, options.timeoutMs ?? DEFAULT_TIMEOUT_MS);
73
+ if (result.code !== 0) {
74
+ throw new Error(`npm could not install ${pkg} on this server.${tail(result.output)}`);
75
+ }
76
+ // Whatever the boot registration did, the directory we just filled must
77
+ // be on the engines' PATH from here on.
78
+ registerPathDir(enginesBinDir(options.baseDir));
79
+ resetPathCache();
80
+ if (options.cli) {
81
+ const binDir = enginesBinDir(options.baseDir);
82
+ if (!findCliCandidates(options.cli).some((path) => path.startsWith(binDir))) {
83
+ throw new Error(`${pkg} installed, but it did not provide a \`${options.cli}\` command. Check the package name in this engine's install descriptor.`);
84
+ }
85
+ }
86
+ }
87
+ function tail(output) {
88
+ const clean = stripVTControlCharacters(output).trim();
89
+ if (!clean)
90
+ return "";
91
+ const lines = clean.split(/\r?\n/).filter((line) => line.trim()).slice(-6);
92
+ return `\n${lines.join("\n").slice(-600)}`;
93
+ }
94
+ function runNpm(args, env, cwd, timeoutMs) {
95
+ return new Promise((resolveRun, rejectRun) => {
96
+ let child;
97
+ try {
98
+ child = spawnCli("npm", args, { env, cwd, stdio: ["pipe", "pipe", "pipe"] });
99
+ }
100
+ catch {
101
+ rejectRun(new Error("npm could not start on this server. Install Node.js with npm for the user running OpenMausBot, then try again."));
102
+ return;
103
+ }
104
+ child.stdin.end();
105
+ let output = "";
106
+ const receive = (chunk) => {
107
+ if (output.length < MAX_OUTPUT)
108
+ output += chunk.toString("utf8").slice(0, MAX_OUTPUT - output.length);
109
+ };
110
+ child.stdout.on("data", receive);
111
+ child.stderr.on("data", receive);
112
+ let timedOut = false;
113
+ const timer = setTimeout(() => {
114
+ timedOut = true;
115
+ void killCliTree(child);
116
+ }, timeoutMs);
117
+ timer.unref();
118
+ child.once("error", (error) => {
119
+ clearTimeout(timer);
120
+ rejectRun(new Error(error.code === "ENOENT"
121
+ ? "npm is not installed on this server. Install Node.js with npm for the user running OpenMausBot, then try again."
122
+ : "npm could not start on this server. Check that Node.js is installed for the user running OpenMausBot."));
123
+ });
124
+ child.once("close", (code) => {
125
+ clearTimeout(timer);
126
+ if (timedOut) {
127
+ rejectRun(new Error("The install took too long and was stopped. Check the server's network connection and try again."));
128
+ return;
129
+ }
130
+ resolveRun({ code, output });
131
+ });
132
+ });
133
+ }
134
+ /** True once something has been installed here, for status pages. */
135
+ export function enginesInstalled(baseDir = DATA_DIR) {
136
+ return existsSync(enginesBinDir(baseDir));
137
+ }