@lumi.ai/runner 0.15.18 → 0.15.20

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 (2) hide show
  1. package/dist/cli.js +148 -17
  2. package/package.json +1 -1
package/dist/cli.js CHANGED
@@ -63,6 +63,9 @@ var ENGINES = {
63
63
  claude: CLAUDE_ENGINE
64
64
  };
65
65
  var DEFAULT_ENGINE_ID = "claude";
66
+ function listEngines() {
67
+ return Object.values(ENGINES);
68
+ }
66
69
  function isKnownEngine(id) {
67
70
  return !!id && Object.hasOwn(ENGINES, id);
68
71
  }
@@ -82,6 +85,8 @@ function estimateCostUsd(engineId, modelId, tokens) {
82
85
  const fresh = Math.max(0, tokens.inputTokens - tokens.cachedInputTokens);
83
86
  return (fresh * model.rates.input + tokens.cachedInputTokens * model.rates.cachedInput + tokens.outputTokens * model.rates.output) / 1e6;
84
87
  }
88
+ var AGENT_ENGINE_IDS = Object.freeze(listEngines().map((e) => e.id));
89
+ var AGENT_MODEL_IDS = Object.freeze(listEngines().flatMap((e) => e.models.map((m) => m.id)));
85
90
 
86
91
  // ../shared/dist/agent.js
87
92
  var DEFAULT_AGENT_TOOLS = {
@@ -144,6 +149,41 @@ var ASSISTANT_PROFILE = Object.freeze({
144
149
  })
145
150
  });
146
151
 
152
+ // ../shared/dist/assistantAuthority.js
153
+ var ASSISTANT_TOOL_AUTHORITY = Object.freeze({
154
+ // Reads. Everything the asker could open in the app.
155
+ ship_info: "member",
156
+ agent_list: "member",
157
+ task_list: "member",
158
+ task_get: "member",
159
+ knowledge_get: "member",
160
+ chat_get: "member",
161
+ // Writes a member can already make by clicking.
162
+ task_create: "member",
163
+ task_comment: "member",
164
+ task_update_status: "member",
165
+ task_relate: "member",
166
+ task_assign: "member",
167
+ knowledge_write: "member",
168
+ chat_send: "member",
169
+ // Captain-only, because the rules say so for a direct write — and `agent_write` also bills.
170
+ agent_write: "captain",
171
+ workflow_write: "captain",
172
+ knowledge_folder: "captain"
173
+ });
174
+ var ASSISTANT_NEVER_REGISTER = Object.freeze([
175
+ "approval_decide",
176
+ "approval_request",
177
+ "memory_write",
178
+ "media_attach"
179
+ ]);
180
+ var ASSISTANT_FROZEN_ARGS = Object.freeze({
181
+ agent_write: Object.freeze(["tools"])
182
+ });
183
+
184
+ // ../shared/dist/toolStats.js
185
+ var TOOL_STATS_HALF_LIFE_MS = 14 * 24 * 60 * 60 * 1e3;
186
+
147
187
  // ../shared/dist/browser.js
148
188
  var BROWSER_ONLINE_WINDOW_MS = 9e4;
149
189
  var DEFAULT_BROWSER_CONSENT_MS = 60 * 60 * 1e3;
@@ -332,6 +372,8 @@ var COLLECTIONS = {
332
372
  engineLimits: "engine_limits",
333
373
  /** `ships/{shipId}/events/{id}` — append-only Ship-level audit log. */
334
374
  events: "events",
375
+ /** §15.70: per-actor tool usage, so the assistant declares what it actually reaches for. */
376
+ toolStats: "toolStats",
335
377
  /** `ships/{shipId}/secrets/{docId}` — runner creds + MCP token hash (see secrets.ts). */
336
378
  secrets: "secrets",
337
379
  /**
@@ -1023,7 +1065,7 @@ function mcpUrl(config2) {
1023
1065
  }
1024
1066
 
1025
1067
  // src/version.ts
1026
- var RUNNER_VERSION = true ? "0.15.18" : "0.0.0-dev";
1068
+ var RUNNER_VERSION = true ? "0.15.20" : "0.0.0-dev";
1027
1069
 
1028
1070
  // src/auth.ts
1029
1071
  import { signInWithCustomToken } from "firebase/auth";
@@ -2085,16 +2127,40 @@ function detectClaudeLimit(text, now, fallbackMs) {
2085
2127
  detail: text.trim().slice(0, 300)
2086
2128
  };
2087
2129
  }
2088
- function workspaceMcpProblem(initEvent, required) {
2089
- if (!required) return null;
2130
+ function expectedMcpServers(agent, extraServers = []) {
2131
+ const out = [];
2132
+ if (effectiveAgentTools(agent).workspaceMcp) {
2133
+ out.push({ key: WORKSPACE_MCP_KEY, label: "the Workspace MCP" });
2134
+ }
2135
+ for (const s of extraServers) out.push({ key: s.key, label: s.name || s.key });
2136
+ return out;
2137
+ }
2138
+ function mcpConnectionFailures(initEvent, expected) {
2139
+ if (expected.length === 0) return [];
2140
+ const byKey = new Map(expected.map((e) => [e.key, e]));
2141
+ const found = /* @__PURE__ */ new Map();
2090
2142
  const errors = initEvent.mcp_server_errors;
2091
- if (!Array.isArray(errors)) return null;
2092
- const mine = errors.find(
2093
- (e) => e && typeof e === "object" && e.name === WORKSPACE_MCP_KEY
2094
- );
2095
- if (!mine) return null;
2096
- const detail = mine.message || mine.type || "no reason given";
2097
- return `The Claude CLI refused to load the Workspace MCP ("${WORKSPACE_MCP_KEY}"): ${detail}. The session would have had none of this Ship's tools \u2014 no task_comment, no run_report, no chat_send \u2014 so it was stopped instead of run blind. This is a runner/CLI mismatch on this machine: update the daemon (\`npm i -g @lumi.ai/runner\`) and the Claude CLI, then retry.`;
2143
+ if (Array.isArray(errors)) {
2144
+ for (const e of errors) {
2145
+ if (!e || typeof e !== "object" || typeof e.name !== "string") continue;
2146
+ const want = byKey.get(e.name);
2147
+ if (want) found.set(want.key, { ...want, detail: e.message || e.type || "no reason given" });
2148
+ }
2149
+ }
2150
+ const servers = initEvent.mcp_servers;
2151
+ if (Array.isArray(servers)) {
2152
+ for (const s of servers) {
2153
+ if (!s || typeof s !== "object" || typeof s.name !== "string") continue;
2154
+ const want = byKey.get(s.name);
2155
+ if (want && s.status !== "connected" && !found.has(want.key)) {
2156
+ found.set(want.key, { ...want, detail: `status "${s.status ?? "unknown"}"` });
2157
+ }
2158
+ }
2159
+ }
2160
+ return [...found.values()];
2161
+ }
2162
+ function workspaceRefusal(failure) {
2163
+ return `The Claude CLI did not load ${failure.label} ("${failure.key}"): ${failure.detail}. The session would have had none of this Ship's tools \u2014 no task_comment, no run_report, no chat_send \u2014 so it was stopped instead of run blind. If this keeps happening, update the daemon (\`npm i -g @lumi.ai/runner\`) and the Claude CLI on this machine.`;
2098
2164
  }
2099
2165
  function allowedTools(agent, extraServers = []) {
2100
2166
  const granted = effectiveAgentTools(agent);
@@ -2202,6 +2268,21 @@ async function runSession(input, bin, dirs) {
2202
2268
  const claudeToken = input.secrets?.claudeToken;
2203
2269
  const env = {
2204
2270
  ...process.env,
2271
+ // THE INIT EVENT MUST BE TRUTHFUL, and one inherited variable is enough to make it lie.
2272
+ //
2273
+ // `MCP_CONNECTION_NONBLOCKING` lets the CLI start before its MCP servers have settled. The init
2274
+ // event then reports `mcp_servers: []` with no status and no error — indistinguishable from a
2275
+ // server that failed — while the connection completes moments later. `mcpConnectionFailures`
2276
+ // reads that event and can only be as honest as it is, so a daemon started from a shell that
2277
+ // happens to set this would either miss every blind session or refuse every healthy one.
2278
+ //
2279
+ // It is DELETED rather than set to a value, because "off" for these is absence: the CLI reads
2280
+ // the variable's presence. A daemon is a long-lived process started from somebody's terminal,
2281
+ // an npm script, or a launchd unit, and it must not inherit a decision about its own
2282
+ // observability from any of them. Found the hard way — every local session for a day reported
2283
+ // an empty server list because the shell that launched the daemon set this.
2284
+ MCP_CONNECTION_NONBLOCKING: void 0,
2285
+ MCP_SERVER_CONNECTION_BATCH_SIZE: void 0,
2205
2286
  ...claudeToken ? { CLAUDE_CODE_OAUTH_TOKEN: claudeToken } : {},
2206
2287
  ...input.githubToken ? {
2207
2288
  GH_TOKEN: input.githubToken,
@@ -2222,6 +2303,8 @@ async function runSession(input, bin, dirs) {
2222
2303
  let resultEvent = null;
2223
2304
  let mcpProblem = null;
2224
2305
  const workspaceRequired = effectiveAgentTools(input.agent).workspaceMcp;
2306
+ const expectedServers = expectedMcpServers(input.agent, input.extraMcpServers ?? []);
2307
+ const mcpFailed = [];
2225
2308
  const mcpWatch = createWorkspaceMcpWatch();
2226
2309
  const exitCode = await new Promise((resolve) => {
2227
2310
  const child = spawn3(bin, args, { cwd: dirs.workdir, env, stdio: ["ignore", "pipe", "pipe"] });
@@ -2259,7 +2342,14 @@ async function runSession(input, bin, dirs) {
2259
2342
  if (event.type === "result") resultEvent = event;
2260
2343
  if (event.type === "assistant") input.log("claude: assistant turn");
2261
2344
  if (event.type === "system" && event.subtype === "init" && !mcpProblem) {
2262
- stopBlindSession(workspaceMcpProblem(event, workspaceRequired));
2345
+ for (const failure of mcpConnectionFailures(event, expectedServers)) {
2346
+ if (failure.key === WORKSPACE_MCP_KEY) {
2347
+ stopBlindSession(workspaceRefusal(failure));
2348
+ continue;
2349
+ }
2350
+ input.log(`MCP connection "${failure.key}" did not load: ${failure.detail}.`);
2351
+ if (!mcpFailed.includes(failure.key)) mcpFailed.push(failure.key);
2352
+ }
2263
2353
  }
2264
2354
  if (workspaceRequired && !mcpProblem) {
2265
2355
  stopBlindSession(mcpWatch.observe(event));
@@ -2307,7 +2397,8 @@ ${stderrLines.slice(-20).join("\n")}`,
2307
2397
  `,
2308
2398
  usage,
2309
2399
  resultText: resultText2,
2310
- ...limit3 ? { limit: limit3 } : {}
2400
+ ...limit3 ? { limit: limit3 } : {},
2401
+ ...mcpFailed.length ? { mcpFailed } : {}
2311
2402
  };
2312
2403
  }
2313
2404
  var CLAUDE_DRIVER_ID = "claude";
@@ -3106,7 +3197,7 @@ async function markChatFailed(db, shipId, job, error) {
3106
3197
  ${error.slice(0, 500)}
3107
3198
  \`\`\`
3108
3199
 
3109
- Write again to start a fresh run.`;
3200
+ Use Try again above to run it once more.`;
3110
3201
  await addDoc(collection6(chatRef, COLLECTIONS.chatMessages), {
3111
3202
  author: { type: "agent", id: job.agentId },
3112
3203
  content,
@@ -3116,8 +3207,45 @@ Write again to start a fresh run.`;
3116
3207
  });
3117
3208
  }
3118
3209
  var REPLY_SCAN_LIMIT = 20;
3210
+ var NARRATION_MARKERS = [
3211
+ // Our own heading and the harness's own word for the block above it — see `chatStandingRules`.
3212
+ /\bsystem prompt\b/i,
3213
+ /\bstanding rules\b/i,
3214
+ // "Based on my available tools, …"
3215
+ /\bmy (?:available |current )?tools?\b/i,
3216
+ // "Let me check what tools I do have access to."
3217
+ /\btools?\b[^.?!]{0,40}\bI (?:do )?have access to\b/i,
3218
+ // Both orders of "<the toolbox> is <missing>", bounded to one sentence so a paragraph that
3219
+ // merely mentions tools and, later, something being unavailable does not match.
3220
+ /\b(?:mcp|tools?)\b[^.?!]{0,60}\b(?:not available|unavailable|(?:are|is)n'?t available|not connected|not loaded|no access)\b/i,
3221
+ /\b(?:not available|unavailable|no access|(?:do not|don'?t|does not|doesn'?t) have access)\b[^.?!]{0,60}\b(?:mcp|tools?)\b/i
3222
+ ];
3223
+ function isSessionNarration(paragraph) {
3224
+ if (paragraph.includes("```")) return false;
3225
+ if (!/\b(?:i|i'm|i'll|i've|my|me)\b/i.test(paragraph)) return false;
3226
+ return NARRATION_MARKERS.some((re) => re.test(paragraph));
3227
+ }
3228
+ function paragraphsWithOffsets(text) {
3229
+ const out = [];
3230
+ const breaks = /\n{2,}/g;
3231
+ let start = 0;
3232
+ let m;
3233
+ while ((m = breaks.exec(text)) !== null) {
3234
+ out.push({ text: text.slice(start, m.index), start });
3235
+ start = m.index + m[0].length;
3236
+ }
3237
+ out.push({ text: text.slice(start), start });
3238
+ return out;
3239
+ }
3240
+ function stripSessionNarration(text) {
3241
+ const paragraphs = paragraphsWithOffsets(text);
3242
+ let dropped = 0;
3243
+ while (dropped < paragraphs.length && isSessionNarration(paragraphs[dropped].text)) dropped += 1;
3244
+ if (dropped === 0 || dropped === paragraphs.length) return text;
3245
+ return text.slice(paragraphs[dropped].start).trim();
3246
+ }
3119
3247
  function backstopReplyContent(resultText2) {
3120
- const text = resultText2.trim();
3248
+ const text = stripSessionNarration(resultText2.trim());
3121
3249
  if (!text) {
3122
3250
  return "I finished that run without writing a reply. Write again to start a fresh one.";
3123
3251
  }
@@ -4402,6 +4530,7 @@ async function startDaemon() {
4402
4530
  let engineId = DEFAULT_ENGINE_ID;
4403
4531
  let transcript = "";
4404
4532
  let resultText2 = "";
4533
+ let mcpConnected = [];
4405
4534
  let usage = {
4406
4535
  engine: DEFAULT_ENGINE_ID,
4407
4536
  inputTokens: 0,
@@ -4574,6 +4703,8 @@ async function startDaemon() {
4574
4703
  armLimitTimer();
4575
4704
  }
4576
4705
  resultText2 = session.resultText;
4706
+ const failed = new Set(session.mcpFailed ?? []);
4707
+ mcpConnected = extraMcpServers.map((s) => s.key).filter((k) => !failed.has(k));
4577
4708
  if (!session.ok) failure = session.resultText || "Session failed.";
4578
4709
  } catch (e) {
4579
4710
  failure = e instanceof Error ? e.message : String(e);
@@ -4597,7 +4728,7 @@ async function startDaemon() {
4597
4728
  // mid-thought, so it almost certainly never reached `run_report` — and whatever it had
4598
4729
  // got to is what the next run on this task would otherwise have to rediscover.
4599
4730
  resultText: resultText2,
4600
- mcpServers: extraMcpServers.map((s) => s.key)
4731
+ mcpServers: mcpConnected
4601
4732
  });
4602
4733
  const by = slot.stop.by;
4603
4734
  if (target.kind === "chat") {
@@ -4647,7 +4778,7 @@ async function startDaemon() {
4647
4778
  // §15.41. Only used when the session never called `run_report` — the ordinary path is
4648
4779
  // that it did, and a real report always wins inside the transaction.
4649
4780
  resultText: resultText2,
4650
- mcpServers: extraMcpServers.map((s) => s.key)
4781
+ mcpServers: mcpConnected
4651
4782
  });
4652
4783
  log2(`Job ${job.id} done (${usage.inputTokens}in/${usage.outputTokens}out tokens).`);
4653
4784
  let delivery = "agent-replied";
@@ -4704,7 +4835,7 @@ async function startDaemon() {
4704
4835
  // is spent — starts from the context pack alone. `error` is the tail for a human; this
4705
4836
  // is the continuity for the next session, and they are read by different readers.
4706
4837
  resultText: resultText2,
4707
- mcpServers: extraMcpServers.map((s) => s.key)
4838
+ mcpServers: mcpConnected
4708
4839
  });
4709
4840
  if (target.kind === "chat") {
4710
4841
  await markChatFailed(sess(shipId).fb.db, shipId, { ...job, chatId: target.chatId }, failure);
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@lumi.ai/runner",
3
- "version": "0.15.18",
3
+ "version": "0.15.20",
4
4
  "type": "module",
5
5
  "description": "Lumi Crew runner daemon — claims jobs from your Ships and executes them as headless Claude sessions on your own machine.",
6
6
  "//name": "The ONLY package in this monorepo published to the public registry, so it is the one that does not follow the internal @lumi/crew-* convention: `@lumi` is not a scope we own, `@lumi.ai` is (the npm org). The workspace DIRECTORY stays packages/crew/runner — renaming the package is not renaming the folder.",