@yemi33/minions 0.1.1910 → 0.1.1912

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.
@@ -0,0 +1,512 @@
1
+ /**
2
+ * engine/cc-worker-pool.js — Persistent per-tab Copilot worker pool (W-mp3ioyh300056f36).
3
+ *
4
+ * Sub-task B of W-mp2w003600196c51 ("CC perf: persistent per-tab Copilot
5
+ * worker"). Wraps `copilot --acp` (Agent Client Protocol — long-lived JSON-RPC
6
+ * server over stdin/stdout) so that a Command Center / doc-chat tab keeps a
7
+ * single warm Copilot process across N turns instead of paying the
8
+ * spawn + initialize + session/new (~14 s) cost on every turn.
9
+ *
10
+ * Public API:
11
+ *
12
+ * pool.getSession({ tabId, model, effort, mcpServers, systemPromptHash, cwd })
13
+ * → SessionHandle: {
14
+ * sessionId,
15
+ * stream(promptText, { onChunk, onDone, onError, signal, systemPromptText }) → Promise,
16
+ * cancel(), // session/cancel notification
17
+ * close(), // tear down this tab
18
+ * }
19
+ * pool.closeTab(tabId) // cancel inflight, kill proc, drop entry
20
+ * pool.shutdown() // close every tab and stop the idle reaper
21
+ *
22
+ * Internal model (one process per tab):
23
+ *
24
+ * tabs: Map<tabId, Worker> where Worker = {
25
+ * proc, sessionId, model, effort, mcpServers, mcpServersHash,
26
+ * systemPromptHash, cwd, lastUsedAt,
27
+ * pending: Map<id, { resolve, reject }>, inflight, readBuf, nextReqId
28
+ * }
29
+ *
30
+ * TODO (sub-task C/D follow-up): multiplex N tabs onto one ACP process
31
+ * (`session/new` per tab, single `proc`). ACP supports it; left as a
32
+ * follow-up perf win once one-process-per-tab is proven in the dashboard.
33
+ *
34
+ * Eviction rules:
35
+ * - mcpServers changes → kill proc entirely + respawn
36
+ * - systemPromptHash change → keep proc, send a fresh `session/new`
37
+ * - model / effort change → no eviction (state updated in place)
38
+ * - closeTab → `session/cancel` if inflight, then kill proc
39
+ * - Idle > 10 min → kill proc (reaper sweep every 60 s)
40
+ *
41
+ * Auth precheck: when `copilot --acp` exits before the initialize handshake
42
+ * completes, surface
43
+ * "copilot --acp failed -- ensure copilot CLI >=1.0.46 and copilot login
44
+ * is complete (...)"
45
+ * so the dashboard can surface a precise reason instead of a generic timeout.
46
+ *
47
+ * No new npm deps — Node built-ins only (child_process, events, timers, crypto).
48
+ *
49
+ * See `notes/archive/2026-05-13-ripley-W-mp3h2wq2000e64f7-2026-05-13-0301.md`
50
+ * for the ACP protocol probe (line framing, request/response correlation,
51
+ * cold-vs-warm timings, cancel semantics, configOptions surface).
52
+ */
53
+
54
+ const { spawn } = require('child_process');
55
+ const crypto = require('crypto');
56
+
57
+ // 10 minutes — matches the work-item spec.
58
+ const IDLE_REAPER_MS = 10 * 60 * 1000;
59
+ // Reaper sweep cadence. Not exposed as ENGINE_DEFAULTS to keep the pool
60
+ // dependency-free; sub-task C/D can plumb a config knob if needed.
61
+ const REAPER_INTERVAL_MS = 60 * 1000;
62
+
63
+ // Test seam — every external side effect goes through `_internals` so
64
+ // test/unit/cc-worker-pool.test.js can stub spawn/now/killImmediate.
65
+ const _internals = {
66
+ spawnAcp({ cwd } = {}) {
67
+ return spawn(
68
+ 'copilot',
69
+ ['--acp', '--allow-all', '--max-autopilot-continues', '1'],
70
+ {
71
+ stdio: ['pipe', 'pipe', 'pipe'],
72
+ cwd: cwd || process.cwd(),
73
+ windowsHide: true,
74
+ // Don't pass shell:true — Copilot is a binary on PATH and a shell
75
+ // wrapper would swallow stdin/stdout framing.
76
+ }
77
+ );
78
+ },
79
+ killImmediate(proc) {
80
+ // Lazy require so unit tests that don't load engine/shared still work.
81
+ try {
82
+ const shared = require('./shared');
83
+ shared.killImmediate(proc);
84
+ } catch {
85
+ try { proc.kill('SIGKILL'); } catch { /* already dead */ }
86
+ }
87
+ },
88
+ now() { return Date.now(); },
89
+ };
90
+
91
+ const _tabs = new Map();
92
+ let _reaperTimer = null;
93
+
94
+ function _hashMcpServers(mcpServers) {
95
+ // Stable hash via JSON.stringify; mcpServers is an array of plain objects
96
+ // in practice (name/command/env) so the natural key order is fine.
97
+ const json = mcpServers == null ? '[]' : JSON.stringify(mcpServers);
98
+ return crypto.createHash('sha256').update(json).digest('hex');
99
+ }
100
+
101
+ class Worker {
102
+ constructor({ tabId, model, effort, mcpServers, mcpServersHash, systemPromptHash, cwd }) {
103
+ this.tabId = tabId;
104
+ this.model = model;
105
+ this.effort = effort;
106
+ this.mcpServers = mcpServers || [];
107
+ this.mcpServersHash = mcpServersHash;
108
+ this.systemPromptHash = systemPromptHash;
109
+ this.cwd = cwd || process.cwd();
110
+
111
+ this.proc = null;
112
+ this.sessionId = null;
113
+ this.pending = new Map(); // id → { resolve, reject }
114
+ this.inflight = null; // current { id, sessionId, onChunk, onDone, onError, signal, signalHandler, settled }
115
+ this.readBuf = '';
116
+ this.nextReqId = 1;
117
+ this.lastUsedAt = _internals.now();
118
+ this.killed = false;
119
+ this.spawnError = null;
120
+ this.firstSystemPromptSent = false;
121
+ }
122
+
123
+ // ── Spawn + initialize handshake ────────────────────────────────────────
124
+ async _spawnAndInit() {
125
+ let proc;
126
+ try {
127
+ proc = _internals.spawnAcp({ cwd: this.cwd });
128
+ } catch (err) {
129
+ throw new Error(
130
+ `copilot --acp failed -- ensure copilot CLI >=1.0.46 and copilot login is complete (${err.message})`
131
+ );
132
+ }
133
+ this.proc = proc;
134
+
135
+ proc.stdout.on('data', (chunk) => this._onStdout(chunk));
136
+ // stderr is captured for diagnostics but we don't act on it.
137
+ if (proc.stderr) proc.stderr.on('data', () => { /* ignore */ });
138
+
139
+ // Race the init handshake against an early process exit so an auth
140
+ // failure (which kills the daemon before initialize completes) surfaces
141
+ // a clear error instead of hanging forever.
142
+ let earlyExitReject;
143
+ const earlyExitPromise = new Promise((_, reject) => {
144
+ earlyExitReject = (code) => {
145
+ this.killed = true;
146
+ const err = new Error(
147
+ `copilot --acp failed -- ensure copilot CLI >=1.0.46 and copilot login is complete (exit ${code})`
148
+ );
149
+ this.spawnError = err;
150
+ this._failAllPending(err);
151
+ reject(err);
152
+ };
153
+ });
154
+ const earlyExitHandler = (code) => earlyExitReject(code);
155
+ proc.once('exit', earlyExitHandler);
156
+
157
+ const errorHandler = (err) => {
158
+ const wrapped = new Error(
159
+ `copilot --acp failed -- ensure copilot CLI >=1.0.46 and copilot login is complete (${err.message})`
160
+ );
161
+ this.spawnError = wrapped;
162
+ this.killed = true;
163
+ this._failAllPending(wrapped);
164
+ };
165
+ proc.on('error', errorHandler);
166
+
167
+ try {
168
+ await Promise.race([
169
+ this._call('initialize', { protocolVersion: 1, clientCapabilities: {} }),
170
+ earlyExitPromise,
171
+ ]);
172
+ const result = await Promise.race([
173
+ this._call('session/new', { cwd: this.cwd, mcpServers: this.mcpServers }),
174
+ earlyExitPromise,
175
+ ]);
176
+ this.sessionId = result && result.sessionId;
177
+ if (!this.sessionId) {
178
+ throw new Error('copilot --acp failed -- session/new returned no sessionId');
179
+ }
180
+ } finally {
181
+ // Either the handshake finished (swap to a persistent exit handler that
182
+ // just marks killed) or it failed (proc is dying anyway).
183
+ proc.removeListener('exit', earlyExitHandler);
184
+ }
185
+ proc.on('exit', () => {
186
+ this.killed = true;
187
+ const err = new Error('copilot --acp process exited');
188
+ this._failAllPending(err);
189
+ // Settle inflight too if it's still hanging
190
+ if (this.inflight && !this.inflight.settled) {
191
+ const cb = this.inflight.onError;
192
+ this.inflight.settled = true;
193
+ this.inflight = null;
194
+ if (cb) try { cb(err); } catch { /* swallow */ }
195
+ }
196
+ });
197
+ }
198
+
199
+ _failAllPending(err) {
200
+ for (const { reject } of this.pending.values()) {
201
+ try { reject(err); } catch { /* swallow */ }
202
+ }
203
+ this.pending.clear();
204
+ }
205
+
206
+ // ── Newline-delimited JSON-RPC line framing ────────────────────────────
207
+ _onStdout(chunk) {
208
+ this.readBuf += chunk.toString('utf8');
209
+ let i;
210
+ while ((i = this.readBuf.indexOf('\n')) >= 0) {
211
+ const line = this.readBuf.slice(0, i).trim();
212
+ this.readBuf = this.readBuf.slice(i + 1);
213
+ if (!line) continue;
214
+ let obj;
215
+ try { obj = JSON.parse(line); } catch { continue; }
216
+ this._handleMessage(obj);
217
+ }
218
+ }
219
+
220
+ _handleMessage(obj) {
221
+ // Response (matches an outbound id)
222
+ if (obj.id != null && this.pending.has(obj.id)) {
223
+ const { resolve, reject } = this.pending.get(obj.id);
224
+ this.pending.delete(obj.id);
225
+ if (obj.error) reject(new Error(obj.error.message || 'ACP error'));
226
+ else resolve(obj.result);
227
+ return;
228
+ }
229
+ // Notification (no id) — only `session/update` matters for streaming.
230
+ if (obj.method === 'session/update' && obj.params && this.inflight) {
231
+ if (obj.params.sessionId !== this.inflight.sessionId) return;
232
+ const update = obj.params.update;
233
+ if (!update) return;
234
+ if (update.sessionUpdate === 'agent_message_chunk') {
235
+ const text = _extractChunkText(update.content);
236
+ if (text && this.inflight.onChunk) {
237
+ try { this.inflight.onChunk(text); } catch { /* swallow */ }
238
+ }
239
+ }
240
+ // Other update kinds (available_commands_update, tool_call, ...) are
241
+ // ignored in sub-task B. Sub-task C/D will surface tool_call to the
242
+ // dashboard's onToolUse callback.
243
+ }
244
+ }
245
+
246
+ _writeFrame(obj) {
247
+ if (this.killed) return;
248
+ if (!this.proc || !this.proc.stdin || this.proc.stdin.destroyed) return;
249
+ try { this.proc.stdin.write(JSON.stringify(obj) + '\n'); }
250
+ catch { /* pipe may be broken; exit handler will fire */ }
251
+ }
252
+
253
+ _call(method, params) {
254
+ return new Promise((resolve, reject) => {
255
+ const id = this.nextReqId++;
256
+ this.pending.set(id, { resolve, reject });
257
+ this._writeFrame({ jsonrpc: '2.0', id, method, params });
258
+ });
259
+ }
260
+
261
+ _notify(method, params) {
262
+ this._writeFrame({ jsonrpc: '2.0', method, params });
263
+ }
264
+
265
+ // ── Stream a single turn ───────────────────────────────────────────────
266
+ stream(promptText, opts = {}) {
267
+ const { onChunk, onDone, onError, signal, systemPromptText } = opts;
268
+ if (this.killed) {
269
+ const err = new Error('cc-worker-pool: tab is closed');
270
+ if (onError) try { onError(err); } catch { /* swallow */ }
271
+ return Promise.resolve();
272
+ }
273
+ if (this.inflight) {
274
+ const err = new Error('cc-worker-pool: a prompt is already in flight on this tab');
275
+ if (onError) try { onError(err); } catch { /* swallow */ }
276
+ return Promise.resolve();
277
+ }
278
+ this.lastUsedAt = _internals.now();
279
+
280
+ // Inject the <system> block on the first prompt of a session — Copilot
281
+ // ACP has no in-protocol "set system prompt" method (per Ripley's note),
282
+ // so the per-session adapter pattern of <system>...</system> prepended
283
+ // to the first user message remains correct.
284
+ let prompt = promptText;
285
+ if (!this.firstSystemPromptSent && systemPromptText) {
286
+ prompt = `<system>\n${systemPromptText}\n</system>\n\n${promptText}`;
287
+ }
288
+ this.firstSystemPromptSent = true;
289
+
290
+ const id = this.nextReqId++;
291
+ const inflight = {
292
+ id,
293
+ sessionId: this.sessionId,
294
+ onChunk,
295
+ onDone,
296
+ onError,
297
+ signal,
298
+ signalHandler: null,
299
+ settled: false,
300
+ };
301
+ this.inflight = inflight;
302
+
303
+ if (signal && typeof signal.addEventListener === 'function') {
304
+ inflight.signalHandler = () => this.cancel();
305
+ try { signal.addEventListener('abort', inflight.signalHandler); } catch { /* swallow */ }
306
+ }
307
+
308
+ return new Promise((resolve) => {
309
+ const finalize = (err, result) => {
310
+ if (inflight.settled) return;
311
+ inflight.settled = true;
312
+ if (this.inflight === inflight) this.inflight = null;
313
+ this.lastUsedAt = _internals.now();
314
+ if (signal && inflight.signalHandler) {
315
+ try { signal.removeEventListener('abort', inflight.signalHandler); }
316
+ catch { /* swallow */ }
317
+ }
318
+ if (err) {
319
+ if (onError) try { onError(err); } catch { /* swallow */ }
320
+ } else {
321
+ if (onDone) try { onDone(result); } catch { /* swallow */ }
322
+ }
323
+ resolve();
324
+ };
325
+
326
+ this.pending.set(id, {
327
+ resolve: (result) => finalize(null, result),
328
+ reject: (err) => finalize(err, null),
329
+ });
330
+ this._writeFrame({
331
+ jsonrpc: '2.0',
332
+ id,
333
+ method: 'session/prompt',
334
+ params: { sessionId: this.sessionId, prompt: [{ type: 'text', text: prompt }] },
335
+ });
336
+ });
337
+ }
338
+
339
+ cancel() {
340
+ if (!this.inflight || this.killed) return;
341
+ // session/cancel is a JSON-RPC notification (no id). The in-flight
342
+ // session/prompt response will arrive with stopReason=cancelled per the
343
+ // ACP spec — we don't synthesize a fake response client-side.
344
+ this._notify('session/cancel', { sessionId: this.inflight.sessionId });
345
+ }
346
+
347
+ async newSession({ mcpServers, systemPromptHash }) {
348
+ // Cancel any inflight before swapping the underlying session.
349
+ if (this.inflight) {
350
+ this.cancel();
351
+ // Wait briefly for the cancelled response to settle so we don't leak
352
+ // a stale inflight reference into the new session.
353
+ const inflight = this.inflight;
354
+ await new Promise((resolve) => {
355
+ const start = _internals.now();
356
+ const wait = () => {
357
+ if (!inflight || inflight.settled || _internals.now() - start > 1000) resolve();
358
+ else setTimeout(wait, 5);
359
+ };
360
+ wait();
361
+ });
362
+ }
363
+ const result = await this._call('session/new', {
364
+ cwd: this.cwd,
365
+ mcpServers: mcpServers || [],
366
+ });
367
+ this.sessionId = result && result.sessionId;
368
+ this.systemPromptHash = systemPromptHash;
369
+ this.firstSystemPromptSent = false;
370
+ }
371
+
372
+ close() {
373
+ if (this.killed) return;
374
+ this.killed = true;
375
+ // Best-effort cancel of an inflight prompt so the daemon doesn't keep
376
+ // generating into a torn-down session before we kill the proc.
377
+ if (this.inflight) {
378
+ try { this._notify('session/cancel', { sessionId: this.inflight.sessionId }); }
379
+ catch { /* swallow */ }
380
+ }
381
+ if (this.proc) {
382
+ try { _internals.killImmediate(this.proc); } catch { /* already dead */ }
383
+ }
384
+ this._failAllPending(new Error('cc-worker-pool: worker closed'));
385
+ if (this.inflight && !this.inflight.settled) {
386
+ const cb = this.inflight.onError;
387
+ this.inflight.settled = true;
388
+ this.inflight = null;
389
+ if (cb) try { cb(new Error('cc-worker-pool: worker closed')); } catch { /* swallow */ }
390
+ }
391
+ }
392
+ }
393
+
394
+ function _extractChunkText(content) {
395
+ if (content == null) return '';
396
+ if (typeof content === 'string') return content;
397
+ if (Array.isArray(content)) {
398
+ let out = '';
399
+ for (const block of content) {
400
+ if (block && typeof block.text === 'string') out += block.text;
401
+ }
402
+ return out;
403
+ }
404
+ if (typeof content === 'object' && typeof content.text === 'string') return content.text;
405
+ return '';
406
+ }
407
+
408
+ // ── Public API ────────────────────────────────────────────────────────────
409
+
410
+ async function getSession({ tabId, model, effort, mcpServers, systemPromptHash, cwd } = {}) {
411
+ if (!tabId) throw new Error('cc-worker-pool.getSession: tabId is required');
412
+ const mcpServersHash = _hashMcpServers(mcpServers);
413
+ let worker = _tabs.get(tabId);
414
+
415
+ if (worker) {
416
+ if (worker.killed) {
417
+ _tabs.delete(tabId);
418
+ worker = null;
419
+ } else if (worker.mcpServersHash !== mcpServersHash) {
420
+ // mcpServers shape changed → must respawn the proc; the daemon
421
+ // resolves MCP server config at process boot, not per session.
422
+ _tabs.delete(tabId);
423
+ worker.close();
424
+ worker = null;
425
+ } else if (worker.systemPromptHash !== systemPromptHash) {
426
+ // System prompt changed → keep the warm process, drop the session
427
+ // and create a fresh one. Saves the ~2.1 s initialize handshake.
428
+ await worker.newSession({ mcpServers, systemPromptHash });
429
+ worker.model = model;
430
+ worker.effort = effort;
431
+ worker.lastUsedAt = _internals.now();
432
+ } else {
433
+ // Warm reuse — only update bookkeeping. model/effort changes on a
434
+ // warm session are tracked here but not pushed via configOptions in
435
+ // sub-task B; sub-task C/D will wire that to ACP `session/configure`
436
+ // (or equivalent) once the dashboard exposes the toggle live.
437
+ worker.model = model;
438
+ worker.effort = effort;
439
+ worker.lastUsedAt = _internals.now();
440
+ }
441
+ }
442
+
443
+ if (!worker) {
444
+ worker = new Worker({
445
+ tabId, model, effort, mcpServers, mcpServersHash, systemPromptHash, cwd,
446
+ });
447
+ _tabs.set(tabId, worker);
448
+ try {
449
+ await worker._spawnAndInit();
450
+ } catch (err) {
451
+ _tabs.delete(tabId);
452
+ try { worker.close(); } catch { /* already torn down */ }
453
+ throw err;
454
+ }
455
+ }
456
+
457
+ _ensureReaper();
458
+
459
+ return {
460
+ sessionId: worker.sessionId,
461
+ stream: (promptText, opts) => worker.stream(promptText, opts),
462
+ cancel: () => worker.cancel(),
463
+ close: () => closeTab(tabId),
464
+ };
465
+ }
466
+
467
+ function closeTab(tabId) {
468
+ const worker = _tabs.get(tabId);
469
+ if (!worker) return;
470
+ _tabs.delete(tabId);
471
+ worker.close();
472
+ }
473
+
474
+ function shutdown() {
475
+ for (const worker of _tabs.values()) {
476
+ try { worker.close(); } catch { /* swallow */ }
477
+ }
478
+ _tabs.clear();
479
+ if (_reaperTimer) {
480
+ clearInterval(_reaperTimer);
481
+ _reaperTimer = null;
482
+ }
483
+ }
484
+
485
+ function _ensureReaper() {
486
+ if (_reaperTimer) return;
487
+ _reaperTimer = setInterval(_reapIdleTabs, REAPER_INTERVAL_MS);
488
+ if (typeof _reaperTimer.unref === 'function') _reaperTimer.unref();
489
+ }
490
+
491
+ function _reapIdleTabs() {
492
+ const now = _internals.now();
493
+ for (const [tabId, worker] of [..._tabs]) {
494
+ if (worker.inflight) continue;
495
+ if (now - worker.lastUsedAt > IDLE_REAPER_MS) {
496
+ _tabs.delete(tabId);
497
+ try { worker.close(); } catch { /* already torn down */ }
498
+ }
499
+ }
500
+ }
501
+
502
+ module.exports = {
503
+ getSession,
504
+ closeTab,
505
+ shutdown,
506
+ // Exposed for unit tests; engine code MUST go through the public API.
507
+ _internals,
508
+ _tabs,
509
+ _reapIdleTabs,
510
+ IDLE_REAPER_MS,
511
+ REAPER_INTERVAL_MS,
512
+ };
@@ -0,0 +1,5 @@
1
+ {
2
+ "runtime": "copilot",
3
+ "models": null,
4
+ "cachedAt": "2026-05-13T19:11:26.190Z"
5
+ }
@@ -46,6 +46,12 @@ const REPO_SLUG_RE = /^[A-Za-z0-9._-]+\/[A-Za-z0-9._-]+$/;
46
46
  const MINIONS_COMMENT_MARKER_RE =
47
47
  /^<!--\s*minions:agent=([^\s]+)\s+kind=([^\s]+)(?:\s+wi=([^\s]+))?\s*-->/m;
48
48
 
49
+ // Sample marker constant for tests / fixtures that need a body with a
50
+ // canonical-format marker but don't care about the specific agent / kind /
51
+ // wi fields. Matches MINIONS_COMMENT_MARKER_RE. Production comments build
52
+ // their marker via _buildMarker / buildMinionsCommentBody.
53
+ const MINIONS_COMMENT_MARKER = '<!-- minions:agent=minions kind=marker -->';
54
+
49
55
  // Cheaper "is this body already marked?" check that matches only at position 0
50
56
  // (for idempotency in buildMinionsCommentBody). Kept separate from the
51
57
  // exported regex so the public regex can be used by downstream classifiers
@@ -231,6 +237,7 @@ module.exports = {
231
237
  // Builders / parsers (pure functions — usable from anywhere)
232
238
  buildMinionsCommentBody,
233
239
  parseMinionsMarker,
240
+ MINIONS_COMMENT_MARKER,
234
241
  MINIONS_COMMENT_MARKER_RE,
235
242
  // Validation regexes (exported for downstream consumers)
236
243
  AGENT_ID_RE,
package/engine/github.js CHANGED
@@ -7,6 +7,7 @@
7
7
  const shared = require('./shared');
8
8
  const { exec, execAsync, getProjects, projectPrPath, projectWorkItemsPath, safeJson, safeJsonArr, safeWrite, mutateJsonFileLocked, MINIONS_DIR, getPrLinks, backfillPrPrdItems, log, ts, dateStamp, PR_STATUS, PR_POLLABLE_STATUSES, createThrottleTracker, getProjectOrg } = shared;
9
9
  const { getPrs } = require('./queries');
10
+ const { MINIONS_COMMENT_MARKER_RE } = require('./gh-comment');
10
11
  const path = require('path');
11
12
 
12
13
  // Lazy require to avoid circular dependency — only needed for engine().handlePostMerge
@@ -85,131 +86,94 @@ function _isNonActionableComment(c, config = {}) {
85
86
  if (_isAgentComment(c)) return true;
86
87
  if (_isCiReportCommentBody(c?.body)) return true;
87
88
  if (_isPreviewStatusComment(c)) return true;
88
- if (_isAgentPositiveSignalComment(c, config)) return true;
89
+ if (_isMinionsAuthoredComment(c)) return true;
89
90
  return false;
90
91
  }
91
92
 
92
- // P-a3f9b2c1Detect agent self-positive-signal comments (verification SUCCESS reports,
93
- // VERDICT:APPROVE recaps, noop:true confessions) posted via the shared `gh` PAT identity.
93
+ // W-mp3bp0ha000997ab-bStructural classifier for Minions-authored PR comments.
94
94
  //
95
- // Sibling to PR #2431's `_isAgentSelfReviewDeclinedComment` but keys off body-shape markers
96
- // + an opt-in shared-minions login allowlist (REST `/issues/:n/comments` doesn't expose
97
- // GraphQL's `viewerDidAuthor` flag, so we approximate via configured logins).
95
+ // Replaces the body-shape classifier chain (PR #2431's _isAgentSelfReviewDeclinedComment,
96
+ // PR #2442's _isAgentPositiveSignalComment + _hasAgentPositiveSignalBody) with a single
97
+ // structural check: a Minions-authored comment carries `MINIONS_COMMENT_MARKER` (defined
98
+ // in engine/gh-comment.js) on its own first line AND was authored by the viewer (the
99
+ // shared PAT identity used by all minions agents).
98
100
  //
99
- // Configuration:
100
- // engine.botCommentLogins: string[] — primary opt-in list (e.g. ['yemi33'])
101
- // engine.ignoredCommentAuthors: string[] fallback (back-compat with existing knob)
101
+ // Both gates are required:
102
+ // - `viewerDidAuthor: true` prevents humans from spoofing the marker by quoting it.
103
+ // - `MINIONS_COMMENT_MARKER_RE` enforces "own first line" quoted markers in the
104
+ // middle of a comment, or markers with trailing content on the same line, do NOT
105
+ // match. This is the structural complement to the viewer check.
102
106
  //
103
- // Reference: PR #2433 (3 consecutive noop fix-dispatches when verify agents posted
104
- // "Result: SUCCESS" comments via the shared yemi33 PAT identity).
105
- function _sharedMinionsLogins(config = {}) {
106
- const primary = Array.isArray(config?.engine?.botCommentLogins) ? config.engine.botCommentLogins : [];
107
- const fallback = Array.isArray(config?.engine?.ignoredCommentAuthors) ? config.engine.ignoredCommentAuthors : [];
108
- return new Set([...primary, ...fallback].map(s => String(s || '').toLowerCase()).filter(Boolean));
107
+ // Production note: the GitHub REST endpoints we poll (/issues/:n/comments and
108
+ // /pulls/:n/comments) do NOT populate `viewerDidAuthor` that field is GraphQL-
109
+ // only. `pollPrHumanComments` therefore backfills it via `_backfillViewerDidAuthor`
110
+ // using the viewer login resolved by `_resolveViewerLogin` (one `gh api user` call
111
+ // per process). This keeps the classifier alive in production without forcing a
112
+ // GraphQL migration on every fetch site.
113
+ //
114
+ // Sub-item -a populates the marker on every minions-posted comment via a helper.
115
+ // Sub-item -c migrates existing engine call sites + playbook examples.
116
+ function _isMinionsAuthoredComment(c) {
117
+ if (!c || c.viewerDidAuthor !== true) return false;
118
+ const body = c.body || '';
119
+ const m = MINIONS_COMMENT_MARKER_RE.exec(body);
120
+ if (!m) return false;
121
+ // Strict: marker must be on its OWN first line (no leading content,
122
+ // no trailing content on the same line). MINIONS_COMMENT_MARKER_RE uses
123
+ // /m so it would otherwise match the start of any line; we additionally
124
+ // require the match to be at index 0 AND to occupy the first line
125
+ // exactly (modulo surrounding whitespace).
126
+ if (m.index !== 0) return false;
127
+ const lineEnd = body.indexOf('\n');
128
+ const firstLine = lineEnd === -1 ? body : body.slice(0, lineEnd);
129
+ return firstLine.trim() === m[0].trim();
109
130
  }
110
131
 
111
- // Pure body-shape matcher for agent positive-signal markers in PR comments.
112
- // Triggers on:
113
- // - "Verification SUCCESS" / "Verification SUCCEEDED" / "Verification ran" (verify-dispatch reports)
114
- // - "Result: SUCCESS" (verify-dispatch report header per playbooks/verify.md)
115
- // - "VERDICT: APPROVE" (review-loop positive verdict defense-in-depth alongside _hasMinionsReviewVerdict)
116
- // - `noop: true` / `"noop": true` (completion-report-style noop confession copied to comment)
117
- function _hasAgentPositiveSignalBody(body) {
118
- const text = String(body || '');
119
- if (!text) return false;
120
- if (/(?:^|\n)\s*\*{0,2}VERDICT[:\s]+\*{0,2}APPROVE\b/i.test(text)) return true;
121
- if (/"?\bnoop\b"?\s*:\s*true\b/i.test(text)) return true;
122
- if (/\bVerification\s+(?:SUCCESS(?:FUL|FULLY)?|SUCCEEDED|ran)\b/i.test(text)) return true;
123
- if (/(?:^|\n)\s*#{0,3}\s*Result\s*:\s*SUCCESS\b/i.test(text)) return true;
124
- return false;
132
+ // ─── Viewer Login Resolution ────────────────────────────────────────────────
133
+ // `_isMinionsAuthoredComment` (above) gates on `c.viewerDidAuthor === true`.
134
+ // That field is populated by GitHub's GraphQL API but NOT by the REST
135
+ // `/issues/:n/comments` and `/pulls/:n/comments` endpoints we use in
136
+ // `pollPrHumanComments` (REST returns `viewerDidAuthor: null`). To make the
137
+ // classifier alive in production without changing every fetch site to
138
+ // GraphQL, we backfill the field by comparing the comment author's login to
139
+ // the viewer login resolved via `gh api user` once per process. The shared
140
+ // PAT identity does not change during a process lifetime, so a simple
141
+ // in-memory cache is sufficient.
142
+ let _cachedViewerLogin = null;
143
+
144
+ async function _resolveViewerLogin() {
145
+ if (_cachedViewerLogin) return _cachedViewerLogin;
146
+ try {
147
+ const result = await execAsync('gh api user', { timeout: 10000, encoding: 'utf-8', maxBuffer: GH_MAX_BUFFER });
148
+ const parsed = JSON.parse(String(result || ''));
149
+ const login = parsed?.login ? String(parsed.login).toLowerCase() : null;
150
+ if (login) _cachedViewerLogin = login;
151
+ return login;
152
+ } catch (e) {
153
+ log('warn', `GitHub viewer-login resolution failed: ${e?.message || e}`);
154
+ return null;
155
+ }
125
156
  }
126
157
 
127
- function _isAgentPositiveSignalComment(c, config = {}) {
128
- if (!c) return false;
129
- const sharedLogins = _sharedMinionsLogins(config);
130
- if (sharedLogins.size === 0) return false;
131
- const login = String(c?.user?.login || '').toLowerCase();
132
- if (!login || !sharedLogins.has(login)) return false;
133
- return _hasAgentPositiveSignalBody(c.body);
158
+ // Test hook: lets tests prime or clear the cached viewer login without
159
+ // shelling out. Pass `null` to force a re-resolve on the next call.
160
+ function _setCachedViewerLogin(login) {
161
+ _cachedViewerLogin = login ? String(login).toLowerCase() : null;
134
162
  }
135
163
 
136
- // W-mp2h696g000a7bc0 Detect agent self-review-declined no-op comments.
137
- //
138
- // Per the documented contract (memory `subject: self-review`, docs/completion-reports.md:83-105):
139
- // when an agent is dispatched to review a PR they implemented, they post a `gh pr comment`
140
- // (no `VERDICT:` line) explaining the recusal and complete with `noop:true, verdict:null,
141
- // needs_rerun:true`. Because `gh` authenticates as the shared PAT user (`yemi33`), the
142
- // classifier can't tell the noop comment from a real human comment by author alone.
143
- //
144
- // Detection requires a stronger signal than author identity. Returns true only when BOTH:
145
- // 1. Body has NO `VERDICT:` line (we never want to swallow a real verdict comment), AND
146
- // 2. EITHER the body contains the canonical "Self-review declined" phrase AND references
147
- // a dispatch id matching `<agent>-<type>-<uid>` that resolves to a same-agent
148
- // review/implement dispatch in the completed history,
149
- // OR the most recent review dispatch on this PR was assigned to the same agent as the
150
- // PR author and completed with the noop:true / verdict:null / needs_rerun:true contract.
151
- //
152
- // This is a narrow allowlist for the noop pattern — generic PAT-user comments still flow
153
- // through `_isNonActionableComment` and trigger `humanFeedback.pendingFix=true` as before.
154
- function _isAgentSelfReviewDeclinedComment(c, { pr, dispatch } = {}) {
155
- if (!c || !pr) return false;
156
- const body = String(c.body || '');
157
- if (!body) return false;
158
- if (_hasMinionsReviewVerdict(body)) return false;
159
-
160
- const prAuthorAgent = String(pr.agent || '').toLowerCase();
161
- if (!prAuthorAgent) return false;
162
-
163
- const completed = (dispatch && Array.isArray(dispatch.completed)) ? dispatch.completed : [];
164
- const isNoopContract = (sc) => {
165
- if (!sc || typeof sc !== 'object') return false;
166
- if (sc.noop !== true && sc.noop !== 'true') return false;
167
- if (sc.verdict !== null && sc.verdict !== undefined && sc.verdict !== '') return false;
168
- if (sc.needs_rerun !== true && sc.needsRerun !== true) return false;
169
- return true;
170
- };
171
-
172
- // Signal A — explicit "Self-review declined" phrase + verifiable dispatch id pointing
173
- // at a same-agent review/implement/fix dispatch in the completed history.
174
- const phraseMatch = /self[\s\-_]+review\s+declined/i.test(body);
175
- if (phraseMatch) {
176
- const dispatchIdMatches = body.match(/\b[a-z][a-z0-9]*-[a-z][a-z0-9]*-[a-z0-9]{8,}\b/g) || [];
177
- for (const id of dispatchIdMatches) {
178
- const entry = completed.find(d => d && d.id === id);
179
- if (!entry) continue;
180
- const agent = String(entry.agent || '').toLowerCase();
181
- const t = String(entry.type || '').toLowerCase();
182
- if (agent !== prAuthorAgent) continue;
183
- if (t === 'review' || t === 'implement' || t === 'implement-large' || t === 'fix') {
184
- return true;
185
- }
186
- }
164
+ // Backfill `c.viewerDidAuthor` for an array of REST-fetched comments using
165
+ // the resolved viewer login. Only sets the field when REST returned
166
+ // null/undefined never overrides a value that GraphQL or the test fixture
167
+ // already supplied. Mutates the comment objects in place.
168
+ function _backfillViewerDidAuthor(comments, viewerLogin) {
169
+ if (!Array.isArray(comments) || !viewerLogin) return;
170
+ for (const c of comments) {
171
+ if (!c || typeof c !== 'object') continue;
172
+ if (c.viewerDidAuthor === true || c.viewerDidAuthor === false) continue;
173
+ const login = c.user?.login;
174
+ if (typeof login !== 'string' || !login) continue;
175
+ c.viewerDidAuthor = login.toLowerCase() === viewerLogin;
187
176
  }
188
-
189
- // Signal B — most recent review dispatch on this PR was assigned to the PR author
190
- // and completed with the documented noop contract. Catches the case where the agent
191
- // forgot the canonical phrase but the completion-report contract still flags a noop.
192
- const prId = String(pr.id || '');
193
- const prNumberRaw = pr.prNumber;
194
- const prNumber = prNumberRaw == null ? null : Number(prNumberRaw);
195
- const reviewDispatches = completed.filter(d => {
196
- if (!d || String(d.type || '').toLowerCase() !== 'review') return false;
197
- const dpr = d.meta && d.meta.pr;
198
- if (!dpr) return false;
199
- if (prId && String(dpr.id || '') === prId) return true;
200
- if (prNumber != null && Number(dpr.prNumber) === prNumber) return true;
201
- return false;
202
- });
203
- if (reviewDispatches.length === 0) return false;
204
- reviewDispatches.sort((a, b) => {
205
- const ta = String(a.completed_at || a.created_at || '');
206
- const tb = String(b.completed_at || b.created_at || '');
207
- return tb.localeCompare(ta);
208
- });
209
- const latest = reviewDispatches[0];
210
- if (!latest) return false;
211
- if (String(latest.agent || '').toLowerCase() !== prAuthorAgent) return false;
212
- return isNoopContract(latest.structuredCompletion);
213
177
  }
214
178
 
215
179
  // ─── Per-Repo Poll Backoff ──────────────────────────────────────────────────
@@ -821,13 +785,18 @@ async function pollPrStatus(config) {
821
785
 
822
786
  async function pollPrHumanComments(config) {
823
787
  // Load dispatch state once per poll cycle (cached for ~2s by queries.getDispatch).
824
- // Used by `_isAgentSelfReviewDeclinedComment` to verify dispatch ids referenced in
825
- // candidate noop comments belong to the PR author.
788
+ // (Previously used by `_isAgentSelfReviewDeclinedComment` to verify dispatch ids;
789
+ // kept loaded in case downstream phases need dispatch context — cheap on cache hit.)
826
790
  const queries = require('./queries');
827
791
  const dispatch = (() => {
828
792
  try { return queries.getDispatch(); }
829
793
  catch { return { pending: [], active: [], completed: [] }; }
830
794
  })();
795
+ // Resolve viewer login once per poll cycle so we can backfill
796
+ // `viewerDidAuthor` on REST comments — the field GraphQL would have
797
+ // populated. `_resolveViewerLogin` caches process-wide, so repeated calls
798
+ // across PRs are free.
799
+ const viewerLogin = await _resolveViewerLogin();
831
800
  const totalUpdated = await forEachActiveGhPr(config, async (project, pr, prNum, slug) => {
832
801
  // Get issue comments (general PR comments)
833
802
  const comments = await ghApi(`/issues/${prNum}/comments`, slug);
@@ -840,6 +809,11 @@ async function pollPrHumanComments(config) {
840
809
  ...(Array.isArray(reviewComments) ? reviewComments : []).map(c => ({ ...c, _type: 'review' }))
841
810
  ];
842
811
 
812
+ // Backfill viewerDidAuthor on each comment so _isMinionsAuthoredComment
813
+ // can actually fire on production traffic. No-op when GraphQL or a test
814
+ // fixture already populated the field.
815
+ _backfillViewerDidAuthor(allComments, viewerLogin);
816
+
843
817
  const cutoffStr = pr.humanFeedback?.lastProcessedCommentDate || pr.created || '1970-01-01';
844
818
  const cutoffMs = new Date(cutoffStr).getTime() || 0;
845
819
 
@@ -853,12 +827,12 @@ async function pollPrHumanComments(config) {
853
827
  for (const c of allComments) {
854
828
  const date = c.created_at || c.updated_at || '';
855
829
  const dateMs = date ? new Date(date).getTime() : 0;
856
- // W-mp2h696g000a7bc0agent self-review-declined no-op comments are posted via
857
- // `gh pr comment` and authenticate as the shared PAT user, so author identity alone
858
- // can't distinguish them from real human feedback. The narrow allowlist check uses
859
- // the comment body + dispatch history to identify the documented noop pattern.
860
- const isNonActionable = _isNonActionableComment(c, config)
861
- || _isAgentSelfReviewDeclinedComment(c, { pr, dispatch });
830
+ // W-mp3bp0ha000997ab-bAll non-actionable filtering routes through
831
+ // _isNonActionableComment, which now composes _isMinionsAuthoredComment
832
+ // (structural marker check). The previous OR chain that appended
833
+ // _isAgentSelfReviewDeclinedComment is gone that helper was deleted
834
+ // with this PR.
835
+ const isNonActionable = _isNonActionableComment(c, config);
862
836
  if (dateMs) allCommentDates.push(date);
863
837
  if (isNonActionable) continue;
864
838
  const entry = {
@@ -883,26 +857,6 @@ async function pollPrHumanComments(config) {
883
857
  }
884
858
  if (newComments.length === 0) return false;
885
859
 
886
- // P-a3f9b2c1 — Defense-in-depth: when the PR is already approved AND every new comment
887
- // is from a configured shared-minions PAT identity AND every new comment matches the
888
- // positive-signal body shape, suppress pendingFix and advance cutoff. Belt-and-suspenders
889
- // against future regressions in `_isAgentPositiveSignalComment`. This is intentionally
890
- // narrow — a single human-shaped comment in newComments (e.g. "rename _foo to bar") will
891
- // disqualify the entire batch and let the normal human-feedback flow run.
892
- if (String(pr.reviewStatus || '').toLowerCase() === 'approved') {
893
- const sharedLogins = _sharedMinionsLogins(config);
894
- if (sharedLogins.size > 0) {
895
- const allFromSharedPat = newComments.every(nc => sharedLogins.has(String(nc.author || '').toLowerCase()));
896
- const allPositiveShape = newComments.every(nc => _hasAgentPositiveSignalBody(nc.content));
897
- if (allFromSharedPat && allPositiveShape) {
898
- const cutoff = allNewDates.sort().pop() || newComments[newComments.length - 1].date;
899
- pr.humanFeedback = { ...(pr.humanFeedback || {}), lastProcessedCommentDate: cutoff };
900
- log('info', `PR ${pr.id}: ${newComments.length} new shared-PAT positive-signal comment(s) on approved PR — defense-in-depth suppression`);
901
- return true;
902
- }
903
- }
904
- }
905
-
906
860
  // Sort all comments chronologically and build full context for the fix agent
907
861
  allCommentEntries.sort((a, b) => a.date.localeCompare(b.date));
908
862
  newComments.sort((a, b) => a.date.localeCompare(b.date));
@@ -1200,9 +1154,9 @@ module.exports = {
1200
1154
  _hasMinionsReviewVerdict, // exported for testing
1201
1155
  _isAgentComment, // exported for testing
1202
1156
  _isNonActionableComment, // exported for testing
1203
- _isAgentSelfReviewDeclinedComment, // exported for testing
1204
- _isAgentPositiveSignalComment, // exported for testing (P-a3f9b2c1)
1205
- _hasAgentPositiveSignalBody, // exported for testing (P-a3f9b2c1)
1206
- _sharedMinionsLogins, // exported for testing (P-a3f9b2c1)
1157
+ _isMinionsAuthoredComment, // exported for testing (W-mp3bp0ha000997ab-b)
1207
1158
  _isPreviewStatusComment, // exported for testing
1159
+ _resolveViewerLogin, // exported for testing (W-mp3bp0ha000997ab-b backfill)
1160
+ _setCachedViewerLogin, // exported for testing (W-mp3bp0ha000997ab-b backfill)
1161
+ _backfillViewerDidAuthor, // exported for testing (W-mp3bp0ha000997ab-b backfill)
1208
1162
  };
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@yemi33/minions",
3
- "version": "0.1.1910",
3
+ "version": "0.1.1912",
4
4
  "description": "Multi-agent AI dev team that runs from ~/.minions/ — five autonomous agents share a single engine, dashboard, and knowledge base",
5
5
  "bin": {
6
6
  "minions": "bin/minions.js"