@hegemonart/get-design-done 1.27.0 → 1.27.5

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.
@@ -80,12 +80,130 @@ const rateGuard = _nodeRequire(
80
80
  ingestHeaders: (provider: string, headers: unknown) => Promise<unknown>;
81
81
  };
82
82
 
83
+ // ── Plan 27.5-03 — Bandit posterior feedback loop ────────────────────────────
84
+ //
85
+ // `integration.cjs` is the Phase 27.5-01 production-integration shim for the
86
+ // Phase 23.5 bandit posterior. It exposes `recordOutcome({agent, bin, delegate,
87
+ // tier, status, costUsd, adaptiveMode, baseDir?, posteriorPath?})` which writes
88
+ // the (status + cost) reward back to the posterior arm for the (agent × bin ×
89
+ // tier × delegate) joint. Per CONTEXT D-04, the call fires AFTER every
90
+ // `emit('session.completed', …)` site so the posterior reflects the measured
91
+ // signal — correctness + cost.
92
+ //
93
+ // The shim is no-throw (best-effort write). The session-runner wraps each
94
+ // `recordOutcome` call in its own try/catch as a defensive belt-and-braces
95
+ // guard against future shim changes.
96
+ const banditIntegration = _nodeRequire(
97
+ _resolve(_REPO_ROOT, 'scripts/lib/bandit-router/integration.cjs'),
98
+ ) as {
99
+ recordOutcome: (input: {
100
+ agent: string;
101
+ bin: string;
102
+ delegate?: string;
103
+ tier: string;
104
+ status: string;
105
+ costUsd?: number;
106
+ adaptiveMode?: 'static' | 'hedge' | 'full';
107
+ baseDir?: string;
108
+ posteriorPath?: string;
109
+ }) => void;
110
+ DELEGATE_NONE: string;
111
+ };
112
+
113
+ // ── Plan 27.5-03 — adaptive-mode read once per run ──────────────────────────
114
+ //
115
+ // `adaptive-mode.cjs.getMode({quiet: true})` reads `.design/budget.json` and
116
+ // returns `'static' | 'hedge' | 'full'`. We cache the resolved mode locally
117
+ // on each `run()` invocation so the recordOutcome calls at the 4 terminal
118
+ // emit sites all see the same value (consistent gating per session).
119
+ const adaptiveModeLib = _nodeRequire(
120
+ _resolve(_REPO_ROOT, 'scripts/lib/adaptive-mode.cjs'),
121
+ ) as {
122
+ getMode: (opts?: { baseDir?: string; budgetPath?: string; quiet?: boolean }) => 'static' | 'hedge' | 'full';
123
+ };
124
+
83
125
  /** Rate-guard provider key for the Anthropic Agent SDK. */
84
126
  const RATE_GUARD_PROVIDER = 'anthropic';
85
127
 
86
128
  /** Default retries (first attempt + 1 retry). */
87
129
  const DEFAULT_MAX_RETRIES = 2;
88
130
 
131
+ /**
132
+ * Default bin marker for bandit posterior writes from session-runner.
133
+ *
134
+ * Per CONTEXT D-12, session-runner uses a deterministic placeholder bin
135
+ * (`'medium'`) for now; real complexity-class-based bin selection is
136
+ * deferred to a later plan. This matches the 27.5-02 budget-enforcer
137
+ * convention so the (agent × bin) posterior slices stay consistent
138
+ * across both write paths.
139
+ */
140
+ const SESSION_RUNNER_DEFAULT_BIN = 'medium';
141
+
142
+ /**
143
+ * Infer a tier ('opus' | 'sonnet' | 'haiku') from a model identifier.
144
+ *
145
+ * Used at the 4 terminal-emit sites where the final tier isn't already
146
+ * carried on `opts` — we fall back to inspecting `usage.model` (folded
147
+ * during the run loop from SDK chunks). Unknown / empty model names
148
+ * default to 'sonnet' (matches the DEFAULT_MODEL_RATE choice and is
149
+ * the safest middle tier for posterior arms).
150
+ */
151
+ function tierFromModel(modelName: string | null | undefined): 'opus' | 'sonnet' | 'haiku' {
152
+ if (typeof modelName !== 'string' || modelName.length === 0) return 'sonnet';
153
+ const lower = modelName.toLowerCase();
154
+ if (lower.includes('opus')) return 'opus';
155
+ if (lower.includes('haiku')) return 'haiku';
156
+ return 'sonnet';
157
+ }
158
+
159
+ /**
160
+ * Best-effort bandit posterior write following `emit('session.completed', …)`.
161
+ *
162
+ * Per CONTEXT D-04: posterior updates happen AT the terminal emit site so the
163
+ * recorded reward reflects the same (status + cost) the rest of the system
164
+ * just observed. The shim (`integration.cjs`) is no-throw and short-circuits
165
+ * silently in static/hedge mode; the outer try/catch here is a defensive
166
+ * belt-and-braces guard for any future shim change.
167
+ *
168
+ * Failures NEVER bubble out — the session-runner contract is that `run()`
169
+ * never throws, and that contract MUST hold even when telemetry is broken.
170
+ */
171
+ function _recordBanditOutcome(input: {
172
+ agent: string;
173
+ bin: string;
174
+ delegate: string;
175
+ tier: string;
176
+ status: string;
177
+ costUsd: number;
178
+ adaptiveMode: 'static' | 'hedge' | 'full';
179
+ }): void {
180
+ try {
181
+ banditIntegration.recordOutcome({
182
+ agent: input.agent,
183
+ bin: input.bin,
184
+ delegate: input.delegate,
185
+ tier: input.tier,
186
+ status: input.status,
187
+ costUsd: input.costUsd,
188
+ adaptiveMode: input.adaptiveMode,
189
+ });
190
+ } catch (err) {
191
+ // Defensive: shim is no-throw, but a future change could regress.
192
+ // Telemetry failure must never break a session — swallow.
193
+ if (process.env['GDD_BANDIT_DEBUG'] === '1') {
194
+ try {
195
+ process.stderr.write(
196
+ '[session-runner] _recordBanditOutcome swallowed: ' +
197
+ (err instanceof Error ? err.message : String(err)) +
198
+ '\n',
199
+ );
200
+ } catch {
201
+ /* swallow */
202
+ }
203
+ }
204
+ }
205
+ }
206
+
89
207
  // ── Plan 27-06 — Peer-CLI delegation primitives ─────────────────────────────
90
208
  //
91
209
  // Lazy registry loader: the registry is a .cjs module under scripts/lib/peer-cli
@@ -200,15 +318,22 @@ async function tryDelegate(args: {
200
318
  return reg !== null ? reg.dispatch : null;
201
319
  })();
202
320
  if (dispatcher === null) {
203
- // No registry available at all — fall through to local. Phase 22
204
- // event emission (Plan 27-08) hooks here as `peer_call_failed`
205
- // with reason="registry_missing". For now, a placeholder stderr
206
- // breadcrumb so operators can grep for delegation drops without
207
- // CI-failing on stdout pollution.
208
- _logPeerCallFailed({ peer: parsed.peer, role, errorClass: 'registry_missing' });
321
+ // No registry available at all — fall through to local.
322
+ _logPeerCallFailed({
323
+ peer: parsed.peer, role, errorClass: 'registry_missing',
324
+ sessionId: args.sessionId, stage: opts.stage,
325
+ });
209
326
  return null;
210
327
  }
211
328
 
329
+ // v1.27.1 — emit peer_call_started before dispatcher invocation so the
330
+ // events.jsonl trail captures the attempt even if the dispatcher hangs.
331
+ _logPeerCallStarted({
332
+ peer: parsed.peer, role,
333
+ sessionId: args.sessionId, stage: opts.stage,
334
+ });
335
+ const dispatchStartedAt = Date.now();
336
+
212
337
  let dispatchResult: { result: unknown; peer: string; protocol: 'acp' | 'asp' } | null = null;
213
338
  try {
214
339
  dispatchResult = await dispatcher(role, tier, sanitizedPrompt, { cwd: process.cwd() });
@@ -218,6 +343,8 @@ async function tryDelegate(args: {
218
343
  role,
219
344
  errorClass: 'dispatch_threw',
220
345
  message: err instanceof Error ? err.message : String(err),
346
+ sessionId: args.sessionId,
347
+ stage: opts.stage,
221
348
  });
222
349
  return null; // transparent fallback
223
350
  }
@@ -225,10 +352,28 @@ async function tryDelegate(args: {
225
352
  if (dispatchResult === null) {
226
353
  // Registry returned null — peer absent, capability mismatch, or
227
354
  // adapter-side error. Per D-07 we fall back silently.
228
- _logPeerCallFailed({ peer: parsed.peer, role, errorClass: 'registry_returned_null' });
355
+ _logPeerCallFailed({
356
+ peer: parsed.peer, role, errorClass: 'registry_returned_null',
357
+ sessionId: args.sessionId, stage: opts.stage,
358
+ });
229
359
  return null;
230
360
  }
231
361
 
362
+ // v1.27.1 — peer round-trip succeeded. Emit peer_call_complete with the
363
+ // measured latency. Token counts + cost are 0 / null because adapters
364
+ // don't surface usage in v1.27 (Plan 27-04 spec deferred it); reflector
365
+ // tolerates null cost (Plan 26-06 cost-arbitrage analysis).
366
+ _logPeerCallComplete({
367
+ peer: dispatchResult.peer,
368
+ role,
369
+ latencyMs: Date.now() - dispatchStartedAt,
370
+ tokensIn: 0,
371
+ tokensOut: 0,
372
+ costUsd: null,
373
+ sessionId: args.sessionId,
374
+ stage: opts.stage,
375
+ });
376
+
232
377
  // Peer succeeded. Build a SessionResult that mirrors the local path's
233
378
  // shape so downstream consumers (stage-handlers, transcript readers,
234
379
  // tests) treat both paths uniformly. We do NOT write a transcript file
@@ -271,34 +416,122 @@ function _coerceFinalText(result: unknown): string | undefined {
271
416
  }
272
417
 
273
418
  /**
274
- * Placeholder for Plan 27-08's `peer_call_failed` event. Until 27-08
275
- * wires real `appendEvent('peer_call_failed', ...)`, we write a single
276
- * stderr line so operators can grep for silent delegation drops. We
277
- * deliberately don't go through `appendEvent` here because the Phase 22
278
- * event-stream hasn't gained a `peer_call_failed` type yet (that's
279
- * 27-08's job) and pushing an unknown event type today would create a
280
- * migration mess for the reflector.
419
+ * v1.27.1 wires Plan 27-08's `peer_call_failed` event for real.
420
+ * Phase 22 `appendEvent` accepts the new event type (registered in
421
+ * KNOWN_EVENT_TYPES via Plan 27-08), so the reflector and downstream
422
+ * telemetry consumers see delegation drops as a measurement signal.
423
+ *
424
+ * Errors from `appendEvent` (e.g., events.jsonl unwritable) are
425
+ * swallowed peer-call telemetry is observability, not critical
426
+ * path. STATE.md remains the durable record of session outcomes.
427
+ *
428
+ * Operators can additionally set `GDD_PEER_DEBUG=1` to emit a
429
+ * one-line stderr breadcrumb mirroring the event for live tailing.
281
430
  */
282
431
  function _logPeerCallFailed(args: {
283
432
  peer: string;
284
433
  role: string;
285
434
  errorClass: string;
286
435
  message?: string;
436
+ sessionId?: string;
437
+ stage?: SessionRunnerOptions['stage'];
287
438
  }): void {
288
- // One-line, machine-greppable. Quiet by default in test runs (NODE_ENV)
289
- // so the test output stays clean. Operators set GDD_PEER_DEBUG=1 to see
290
- // the breadcrumb in production logs.
291
- if (process.env['GDD_PEER_DEBUG'] !== '1') return;
292
- const payload = JSON.stringify({
293
- type: 'peer_call_failed',
294
- peer_id: args.peer,
295
- role: args.role,
296
- error_class: args.errorClass,
297
- ...(args.message !== undefined ? { message: args.message } : {}),
298
- ts: new Date().toISOString(),
299
- });
300
- // eslint-disable-next-line no-console
301
- console.error(`[peer-cli] ${payload}`);
439
+ try {
440
+ appendEvent({
441
+ type: 'peer_call_failed',
442
+ timestamp: new Date().toISOString(),
443
+ sessionId: args.sessionId ?? 'unknown',
444
+ ...(args.stage !== undefined && args.stage !== 'init' && args.stage !== 'custom' ? { stage: args.stage } : {}),
445
+ payload: {
446
+ runtime_role: 'peer',
447
+ peer_id: args.peer,
448
+ role: args.role,
449
+ error_class: args.errorClass,
450
+ ...(args.message !== undefined ? { message: args.message } : {}),
451
+ },
452
+ });
453
+ } catch {
454
+ // Telemetry is best-effort — never let an event-stream failure
455
+ // break the actual session flow.
456
+ }
457
+ if (process.env['GDD_PEER_DEBUG'] === '1') {
458
+ const payload = JSON.stringify({
459
+ type: 'peer_call_failed',
460
+ peer_id: args.peer,
461
+ role: args.role,
462
+ error_class: args.errorClass,
463
+ ...(args.message !== undefined ? { message: args.message } : {}),
464
+ ts: new Date().toISOString(),
465
+ });
466
+ // eslint-disable-next-line no-console
467
+ console.error(`[peer-cli] ${payload}`);
468
+ }
469
+ }
470
+
471
+ /**
472
+ * v1.27.1 — emit `peer_call_started` event. Fired once at the beginning
473
+ * of a delegation attempt, before the dispatcher is invoked. Pairs with
474
+ * `peer_call_complete` (success path) or `peer_call_failed` (any failure
475
+ * path, transparent to caller per D-07).
476
+ */
477
+ function _logPeerCallStarted(args: {
478
+ peer: string;
479
+ role: string;
480
+ sessionId?: string;
481
+ stage?: SessionRunnerOptions['stage'];
482
+ }): void {
483
+ try {
484
+ appendEvent({
485
+ type: 'peer_call_started',
486
+ timestamp: new Date().toISOString(),
487
+ sessionId: args.sessionId ?? 'unknown',
488
+ ...(args.stage !== undefined && args.stage !== 'init' && args.stage !== 'custom' ? { stage: args.stage } : {}),
489
+ payload: {
490
+ runtime_role: 'peer',
491
+ peer_id: args.peer,
492
+ role: args.role,
493
+ },
494
+ });
495
+ } catch {
496
+ // best-effort
497
+ }
498
+ }
499
+
500
+ /**
501
+ * v1.27.1 — emit `peer_call_complete` event. Fired after a successful
502
+ * dispatcher round-trip. Cost is null when the adapter doesn't return
503
+ * usage data (some peers don't surface token counts); the reflector
504
+ * tolerates null cost for arbitrage analysis (Plan 26-06).
505
+ */
506
+ function _logPeerCallComplete(args: {
507
+ peer: string;
508
+ role: string;
509
+ latencyMs: number;
510
+ tokensIn: number;
511
+ tokensOut: number;
512
+ costUsd: number | null;
513
+ sessionId?: string;
514
+ stage?: SessionRunnerOptions['stage'];
515
+ }): void {
516
+ try {
517
+ appendEvent({
518
+ type: 'peer_call_complete',
519
+ timestamp: new Date().toISOString(),
520
+ sessionId: args.sessionId ?? 'unknown',
521
+ ...(args.stage !== undefined && args.stage !== 'init' && args.stage !== 'custom' ? { stage: args.stage } : {}),
522
+ payload: {
523
+ runtime_role: 'peer',
524
+ peer_id: args.peer,
525
+ role: args.role,
526
+ latency_ms: args.latencyMs,
527
+ tokens_in: args.tokensIn,
528
+ tokens_out: args.tokensOut,
529
+ cost_usd: args.costUsd,
530
+ },
531
+ });
532
+ } catch {
533
+ // best-effort
534
+ }
302
535
  }
303
536
 
304
537
  /** Baseline retry backoff parameters (matches jittered-backoff defaults for
@@ -565,6 +798,22 @@ export async function run(opts: SessionRunnerOptions): Promise<SessionResult> {
565
798
  const toolCalls: SessionResult['tool_calls'] = [];
566
799
  const usage = { input: 0, output: 0, model: null as string | null };
567
800
  let turns = 0;
801
+
802
+ // -- 3a. Resolve adaptive-mode once for the entire session (Plan 27.5-03). --
803
+ // Cached locally so all four `recordOutcome()` call sites below see the
804
+ // same gating decision (consistent posterior-write semantics across
805
+ // rate-limit, peer, turnCap=0, and terminal-completion paths).
806
+ //
807
+ // Wrapped in try/catch because adaptive-mode.getMode reads
808
+ // `.design/budget.json`; a broken fs.readFile / JSON.parse must not
809
+ // crash the session before it even starts. Fallback = 'static' which
810
+ // short-circuits the recordOutcome shim (no-op).
811
+ let adaptiveMode: 'static' | 'hedge' | 'full' = 'static';
812
+ try {
813
+ adaptiveMode = adaptiveModeLib.getMode({ quiet: true });
814
+ } catch {
815
+ // swallow — fallback to 'static' means no posterior writes
816
+ }
568
817
  let finalText: string | undefined;
569
818
 
570
819
  // -- 4. Emit session.started. -------------------------------------------
@@ -604,6 +853,20 @@ export async function run(opts: SessionRunnerOptions): Promise<SessionResult> {
604
853
  transcript_path: transcriptPath,
605
854
  sanitizer: { applied: [...result.sanitizer.applied], removedSections: [...result.sanitizer.removedSections] },
606
855
  });
856
+ // Plan 27.5-03 — feedback loop. Posterior records the
857
+ // measured outcome (status + cost) for the (agent × bin × tier × delegate)
858
+ // slice. The rate-limit preflight failure path has no peer dispatch and no
859
+ // usage data (zero cost), so delegate=DELEGATE_NONE and tier falls back to
860
+ // 'sonnet' via tierFromModel(null). Shim no-ops in static/hedge mode.
861
+ _recordBanditOutcome({
862
+ agent: opts.stage,
863
+ bin: SESSION_RUNNER_DEFAULT_BIN,
864
+ delegate: banditIntegration.DELEGATE_NONE,
865
+ tier: tierFromModel(usage.model),
866
+ status: result.status,
867
+ costUsd: result.usage.usd_cost,
868
+ adaptiveMode,
869
+ });
607
870
  transcript.close();
608
871
  return result;
609
872
  }
@@ -623,6 +886,69 @@ export async function run(opts: SessionRunnerOptions): Promise<SessionResult> {
623
886
  }
624
887
  }
625
888
 
889
+ // -- 6.5. Peer-CLI delegation try (Plan 27-06 wiring, v1.27.1). ---------
890
+ // If the agent's frontmatter declares `delegate_to: <peer>-<role>` AND the
891
+ // peer is allowlisted AND the registry can route, run the prompt on the
892
+ // peer-CLI and return early. On peer-absent / peer-error / null result,
893
+ // fall through transparently to the local SDK loop (D-07).
894
+ //
895
+ // tryDelegate is a no-op when opts.delegateTo is undefined / 'none', when
896
+ // the registry can't load, when the peer isn't allowlisted, when the
897
+ // dispatcher returns null, or when the dispatcher throws. In all those
898
+ // cases tryDelegate returns null and we proceed to the local SDK path.
899
+ const peerResult = await tryDelegate({
900
+ opts,
901
+ sanitizedPrompt,
902
+ transcriptPath,
903
+ sessionId,
904
+ sanitizer: sanResult,
905
+ });
906
+ if (peerResult !== null) {
907
+ emit('session.completed', opts.stage, sessionId, {
908
+ stage: opts.stage,
909
+ sessionId,
910
+ status: peerResult.status,
911
+ turns: peerResult.turns,
912
+ usage: peerResult.usage,
913
+ transcript_path: transcriptPath,
914
+ sanitizer: { applied: [...peerResult.sanitizer.applied], removedSections: [...peerResult.sanitizer.removedSections] },
915
+ });
916
+ // Plan 27.5-03 — feedback loop, peer path. The
917
+ // delegate dimension is the peer name parsed from opts.delegateTo (e.g.
918
+ // 'gemini-research' → 'gemini'). Per CONTEXT D-04 we use the peer name
919
+ // for the delegate slice of the posterior so peer-success arms get the
920
+ // reward signal separately from local arms. Tier is 'sonnet' by default
921
+ // since the peer adapter doesn't surface a model identifier in v1.27.
922
+ // Re-parse opts.delegateTo here — tryDelegate already validated it but
923
+ // didn't expose the peer name on the returned SessionResult.
924
+ const _peerParsed = parseDelegateTo(opts.delegateTo);
925
+ const _delegate = _peerParsed !== null
926
+ ? _peerParsed.peer
927
+ : banditIntegration.DELEGATE_NONE;
928
+ // Tier resolution priority for the peer path:
929
+ // 1. opts.delegateTier when it's a bare tier name (opus/sonnet/haiku)
930
+ // 2. tierFromModel(opts.delegateTier) when it's a model id string
931
+ // 3. tierFromModel(usage.model) fallback
932
+ // tierFromModel() is safe for any string and returns 'sonnet' on miss,
933
+ // so the second branch covers both bare-tier and model-id inputs.
934
+ const _peerTier: 'opus' | 'sonnet' | 'haiku' =
935
+ typeof opts.delegateTier === 'string' && opts.delegateTier.length > 0
936
+ ? tierFromModel(opts.delegateTier)
937
+ : tierFromModel(usage.model);
938
+ _recordBanditOutcome({
939
+ agent: opts.stage,
940
+ bin: SESSION_RUNNER_DEFAULT_BIN,
941
+ delegate: _delegate,
942
+ tier: _peerTier,
943
+ status: peerResult.status,
944
+ costUsd: peerResult.usage.usd_cost,
945
+ adaptiveMode,
946
+ });
947
+ transcript.close();
948
+ if (opts.signal !== undefined) opts.signal.removeEventListener('abort', onExternalAbort);
949
+ return peerResult;
950
+ }
951
+
626
952
  // -- 7. Retry-once loop. ------------------------------------------------
627
953
  const maxAttempts = opts.maxRetries !== undefined && opts.maxRetries > 0
628
954
  ? opts.maxRetries
@@ -649,6 +975,18 @@ export async function run(opts: SessionRunnerOptions): Promise<SessionResult> {
649
975
  transcript_path: transcriptPath,
650
976
  sanitizer: { applied: [...sanResult.applied], removedSections: [...sanResult.removedSections] },
651
977
  });
978
+ // Plan 27.5-03 — feedback loop, turnCap=0 path. No
979
+ // SDK call was ever made, so no peer involvement and no model id was
980
+ // ever observed. Reward will be 0 (status !== 'completed') with cost 0.
981
+ _recordBanditOutcome({
982
+ agent: opts.stage,
983
+ bin: SESSION_RUNNER_DEFAULT_BIN,
984
+ delegate: banditIntegration.DELEGATE_NONE,
985
+ tier: tierFromModel(usage.model),
986
+ status,
987
+ costUsd: result.usage.usd_cost,
988
+ adaptiveMode,
989
+ });
652
990
  transcript.close();
653
991
  if (opts.signal !== undefined) opts.signal.removeEventListener('abort', onExternalAbort);
654
992
  return result;
@@ -743,6 +1081,21 @@ export async function run(opts: SessionRunnerOptions): Promise<SessionResult> {
743
1081
  transcript_path: transcriptPath,
744
1082
  sanitizer: { applied: [...sanResult.applied], removedSections: [...sanResult.removedSections] },
745
1083
  });
1084
+ // Plan 27.5-03 — feedback loop, terminal main-loop path.
1085
+ // This is the dominant write site: covers natural completion, budget cap,
1086
+ // turn cap (after first turn), abort, and error (post-retry-exhaustion).
1087
+ // Tier is inferred from the model actually observed during the run
1088
+ // (usage.model). Delegate=DELEGATE_NONE because tryDelegate either returned
1089
+ // null (we wouldn't be here otherwise) or wasn't invoked at all.
1090
+ _recordBanditOutcome({
1091
+ agent: opts.stage,
1092
+ bin: SESSION_RUNNER_DEFAULT_BIN,
1093
+ delegate: banditIntegration.DELEGATE_NONE,
1094
+ tier: tierFromModel(usage.model),
1095
+ status: result.status,
1096
+ costUsd: result.usage.usd_cost,
1097
+ adaptiveMode,
1098
+ });
746
1099
 
747
1100
  return result;
748
1101
  }
@@ -0,0 +1,129 @@
1
+ ---
2
+ name: gdd-bandit-status
3
+ description: "Surface read-only per-(agent, bin, delegate) bandit posterior snapshot — alpha/beta/mean/stddev/count/last-used per arm. Phase 27.5 (v1.27.5) diagnostic. Use when investigating 'why did the bandit pick tier X for agent Y?' or when verifying posterior convergence after enabling adaptive_mode: full."
4
+ argument-hint: ""
5
+ tools: Read, Bash
6
+ ---
7
+
8
+ # gdd-bandit-status
9
+
10
+ ## Role
11
+
12
+ You are a deterministic, read-only diagnostic skill. You do not spawn agents and do not modify the bandit posterior. You read `.design/telemetry/posterior.json` (the path declared by `scripts/lib/bandit-router.cjs`'s `DEFAULT_POSTERIOR_PATH` constant), aggregate per-`(agent, bin, delegate, tier)` arm state, and emit a single Markdown table summarizing the posterior. The user runs this when they want to inspect bandit decisions without touching the posterior.
13
+
14
+ Strictly read-only per Phase 27.5 D-11. To reset the posterior, use `/gdd:bandit-reset` from Phase 23.5.
15
+
16
+ ## Invocation Contract
17
+
18
+ - **Input**: none. The skill takes no arguments.
19
+ - **Output**: a Markdown bandit-status table to stdout. No JSON wrapper. The table is the entire output.
20
+
21
+ ## Procedure
22
+
23
+ ### 1. Locate the posterior file
24
+
25
+ Read `.design/telemetry/posterior.json`. If the file does not exist:
26
+
27
+ - Emit the empty-state message:
28
+
29
+ ```
30
+ ## Bandit Posterior Snapshot
31
+
32
+ No posterior data yet — run a few pipeline cycles with `adaptive_mode: full` first.
33
+
34
+ No posterior data found at `.design/telemetry/posterior.json`.
35
+
36
+ Possible reasons:
37
+ - `adaptive_mode` is `static` or `hedge` (bandit is silent — see `.design/budget.json` `adaptive_mode` setting).
38
+ - No spawns have fired since Phase 27.5 wiring landed.
39
+ - Posterior was cleared via `/gdd:bandit-reset`.
40
+
41
+ See `reference/bandit-integration.md` for setup guidance.
42
+ ```
43
+
44
+ - Skip to Section 4 (Record).
45
+
46
+ ### 2. Parse the posterior
47
+
48
+ Parse the file as JSON. If parsing fails (truncated/corrupted file), emit:
49
+
50
+ ```
51
+ ## Bandit Posterior Snapshot
52
+
53
+ Posterior file at `.design/telemetry/posterior.json` exists but is unparseable (truncated or corrupted).
54
+
55
+ Run `/gdd:bandit-reset` to start fresh, or restore from a backup.
56
+ ```
57
+
58
+ The posterior schema is:
59
+
60
+ ```json
61
+ {
62
+ "schema_version": "1.0.0",
63
+ "generated_at": "<ISO timestamp>",
64
+ "arms": [
65
+ { "agent": "...", "bin": "...", "tier": "...", "delegate": "...", "alpha": N, "beta": N, "last_used": "...", "count": N }
66
+ ]
67
+ }
68
+ ```
69
+
70
+ The `delegate` field is optional — when absent, the arm is the Phase 23.5 legacy slice (equivalent to `delegate: 'none'`). The status output renders `delegate: '-'` for legacy arms to distinguish them visually from explicit `'none'` arms.
71
+
72
+ ### 3. Render the table
73
+
74
+ Compute per arm:
75
+
76
+ - `mean = alpha / (alpha + beta)` (rounded to 3 decimals)
77
+ - `stddev = sqrt(alpha * beta / ((alpha + beta)^2 * (alpha + beta + 1)))` (rounded to 3 decimals)
78
+
79
+ Sort arms by `(agent ascending, bin ascending, delegate ascending where '-' sorts first, tier ascending where opus < sonnet < haiku is the canonical tier ordering, last_used descending tiebreaker)`. Group rows by agent for readability.
80
+
81
+ Emit:
82
+
83
+ ```
84
+ ## Bandit Posterior Snapshot
85
+
86
+ Per-(agent, bin, delegate, tier) posterior state. Read-only — to reset the posterior, use `/gdd:bandit-reset` (Phase 23.5).
87
+
88
+ Posterior file: `.design/telemetry/posterior.json` (last updated: <generated_at>)
89
+ Total arms: <count>
90
+
91
+ | Agent | Bin | Delegate | Tier | Alpha | Beta | Mean | Stddev | Count | Last Used |
92
+ |-----------------|--------|----------|--------|-------|-------|-------|--------|-------|----------------------|
93
+ | <agent> | <bin> | <deleg> | <tier> | <a> | <b> | <m> | <s> | <c> | <last_used or '-'> |
94
+
95
+ > Mean = alpha / (alpha + beta). Stddev = sqrt(alpha*beta / ((alpha+beta)^2 * (alpha+beta+1))).
96
+ > Delegate '-' = Phase 23.5 legacy slice (equivalent to 'none').
97
+ > See `reference/bandit-integration.md` for interpretation.
98
+ > Read-only — use `/gdd:bandit-reset` to clear posterior state.
99
+ ```
100
+
101
+ Format numbers to fixed precision: alpha/beta to 2 decimals, mean/stddev to 3 decimals, count as integer, last_used truncated to the minute precision (`YYYY-MM-DDTHH:MM`).
102
+
103
+ When `last_used` is null (arm exists but never selected — possible if the arm was created by `ensureArm` without a subsequent `pull`), render `-` in the Last Used column.
104
+
105
+ After the table, surface a brief best-arm summary per `(agent, bin)` slice — for each unique `(agent, bin)` pair, identify the arm with the highest `mean` (tie-broken by `count` descending) and display it as the "best-arm" recommendation. This helps the operator answer "why did the bandit pick tier X?" at a glance.
106
+
107
+ ### 4. Record
108
+
109
+ After execution, append one JSONL line to `.design/skill-records.jsonl`:
110
+
111
+ ```json
112
+ {"skill": "gdd-bandit-status", "ts": "<ISO timestamp>", "arms_seen": <count>, "posterior_present": <bool>}
113
+ ```
114
+
115
+ The skill writes ONLY to `.design/skill-records.jsonl` for telemetry purposes. It never touches `.design/telemetry/posterior.json`.
116
+
117
+ ## Cross-references
118
+
119
+ - `scripts/lib/bandit-router.cjs` (Phase 23.5) — posterior shape, `DEFAULT_POSTERIOR_PATH` constant, `loadPosterior()` helper.
120
+ - `scripts/lib/bandit-router/integration.cjs` (Phase 27.5-01) — production-integration shim.
121
+ - `hooks/budget-enforcer.ts` (Phase 27.5-02) — bandit consultation site.
122
+ - `scripts/lib/session-runner/index.ts` (Phase 27.5-03) — outcome recording site.
123
+ - `scripts/lib/bandit-arbitrage.cjs` (Phase 27.5-04) — automated stale-frontmatter analysis.
124
+ - `reference/bandit-integration.md` (Phase 27.5-06) — operator guide.
125
+ - `/gdd:bandit-reset` (Phase 23.5) — the ONLY surface that mutates the posterior.
126
+
127
+ ## Record
128
+
129
+ See Section 4 above.