@hegemonart/get-design-done 1.27.1 → 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.
- package/.claude-plugin/marketplace.json +2 -2
- package/.claude-plugin/plugin.json +1 -1
- package/CHANGELOG.md +45 -0
- package/SKILL.md +1 -0
- package/agents/design-reflector.md +52 -0
- package/hooks/budget-enforcer.ts +249 -5
- package/package.json +1 -1
- package/reference/bandit-integration.md +163 -0
- package/reference/registry.json +7 -0
- package/scripts/lib/bandit-arbitrage.cjs +423 -0
- package/scripts/lib/bandit-router/integration.cjs +309 -0
- package/scripts/lib/session-runner/index.ts +206 -0
- package/skills/bandit-status/SKILL.md +129 -0
- package/skills/peers/SKILL.md +27 -8
|
@@ -0,0 +1,309 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* scripts/lib/bandit-router/integration.cjs — Plan 27.5-01
|
|
3
|
+
*
|
|
4
|
+
* Production-integration shim for the Phase 23.5 bandit posterior +
|
|
5
|
+
* Phase 27-07 delegate dimension. Hides the `pull` vs `pullWithDelegate`
|
|
6
|
+
* + `update` vs `updateWithDelegate` choice from callers.
|
|
7
|
+
*
|
|
8
|
+
* Two functions:
|
|
9
|
+
* consultBandit({agent, bin, delegate, agentFrontmatter, adaptiveMode, baseDir?, posteriorPath?})
|
|
10
|
+
* → {tier, decision_log}
|
|
11
|
+
* recordOutcome({agent, bin, delegate, tier, status, costUsd, adaptiveMode, baseDir?, posteriorPath?})
|
|
12
|
+
* → void (best-effort write per D-04)
|
|
13
|
+
*
|
|
14
|
+
* Routing rules (D-05 + D-07):
|
|
15
|
+
* 1. agentFrontmatter.tier_override is set → bypass bandit, return tier_override
|
|
16
|
+
* 2. adaptiveMode !== 'full' → bandit silent, return frontmatter.default_tier
|
|
17
|
+
* (covers 'static' and 'hedge' per D-07)
|
|
18
|
+
* 3. adaptiveMode === 'full' && delegate is 'none' / undefined
|
|
19
|
+
* → call pull()
|
|
20
|
+
* 4. adaptiveMode === 'full' && delegate is a peer name
|
|
21
|
+
* → call pullWithDelegate({delegates:[delegate]})
|
|
22
|
+
*
|
|
23
|
+
* recordOutcome is symmetric on the adaptive_mode gate:
|
|
24
|
+
* - non-'full' → no-op
|
|
25
|
+
* - 'full' && delegate 'none'/undefined → update()
|
|
26
|
+
* - 'full' && delegate is a peer → updateWithDelegate()
|
|
27
|
+
*
|
|
28
|
+
* Reward function = Phase 23.5's computeReward unchanged (D-08).
|
|
29
|
+
*
|
|
30
|
+
* Posterior writes are best-effort — all throws are swallowed. The
|
|
31
|
+
* shim's job is to plumb the call; telemetry resilience is downstream.
|
|
32
|
+
*/
|
|
33
|
+
|
|
34
|
+
'use strict';
|
|
35
|
+
|
|
36
|
+
const banditRouter = require('../bandit-router.cjs');
|
|
37
|
+
const adaptiveModeLib = require('../adaptive-mode.cjs');
|
|
38
|
+
|
|
39
|
+
const DELEGATE_NONE = banditRouter.DELEGATE_NONE; // 'none'
|
|
40
|
+
const VALID_DELEGATES = banditRouter.DEFAULT_DELEGATES; // ['none','gemini','codex','cursor','copilot','qwen']
|
|
41
|
+
|
|
42
|
+
/**
|
|
43
|
+
* Validate that `delegate` is either undefined, DELEGATE_NONE, or a
|
|
44
|
+
* member of VALID_DELEGATES. Returns the canonical delegate string
|
|
45
|
+
* (undefined → 'none').
|
|
46
|
+
*
|
|
47
|
+
* @param {string|undefined} delegate
|
|
48
|
+
* @param {string} fnName — for error message context
|
|
49
|
+
* @returns {string}
|
|
50
|
+
*/
|
|
51
|
+
function resolveDelegate(delegate, fnName) {
|
|
52
|
+
if (delegate === undefined || delegate === null) return DELEGATE_NONE;
|
|
53
|
+
if (typeof delegate !== 'string') {
|
|
54
|
+
throw new TypeError(
|
|
55
|
+
`integration.${fnName}: delegate must be a string when provided, got ${typeof delegate}`,
|
|
56
|
+
);
|
|
57
|
+
}
|
|
58
|
+
if (!VALID_DELEGATES.includes(delegate)) {
|
|
59
|
+
throw new RangeError(
|
|
60
|
+
`integration.${fnName}: unknown delegate '${delegate}'; expected one of ${VALID_DELEGATES.join(',')}`,
|
|
61
|
+
);
|
|
62
|
+
}
|
|
63
|
+
return delegate;
|
|
64
|
+
}
|
|
65
|
+
|
|
66
|
+
/**
|
|
67
|
+
* Resolve the adaptive_mode for a call. If the caller passed it
|
|
68
|
+
* explicitly we use that; otherwise we read it from disk via
|
|
69
|
+
* adaptive-mode.getMode (D-07: single gating surface).
|
|
70
|
+
*
|
|
71
|
+
* @param {string|undefined} adaptiveMode
|
|
72
|
+
* @param {{baseDir?: string}} opts
|
|
73
|
+
* @returns {'static'|'hedge'|'full'}
|
|
74
|
+
*/
|
|
75
|
+
function resolveAdaptiveMode(adaptiveMode, opts) {
|
|
76
|
+
if (typeof adaptiveMode === 'string' && adaptiveMode.length > 0) {
|
|
77
|
+
return /** @type {'static'|'hedge'|'full'} */ (adaptiveMode);
|
|
78
|
+
}
|
|
79
|
+
return adaptiveModeLib.getMode({ baseDir: opts && opts.baseDir, quiet: true });
|
|
80
|
+
}
|
|
81
|
+
|
|
82
|
+
/**
|
|
83
|
+
* consultBandit — single canonical lookup that returns a tier + a
|
|
84
|
+
* decision_log explaining how the tier was chosen. Five paths:
|
|
85
|
+
*
|
|
86
|
+
* Path 1 — static mode → frontmatter.default_tier (or 'sonnet' fallback)
|
|
87
|
+
* Path 2 — tier_override set on frontmatter → bypass bandit
|
|
88
|
+
* Path 3 — full mode + delegate='none' (or undefined) → pull()
|
|
89
|
+
* Path 4 — full mode + delegate=<peer> → pullWithDelegate()
|
|
90
|
+
* Path 5 — hedge mode → frontmatter.default_tier (bandit silent)
|
|
91
|
+
*
|
|
92
|
+
* Path 2 takes precedence over Path 1 / 3 / 4 / 5 (tier_override is the
|
|
93
|
+
* explicit operator override per D-05).
|
|
94
|
+
*
|
|
95
|
+
* @param {{
|
|
96
|
+
* agent: string,
|
|
97
|
+
* bin: string,
|
|
98
|
+
* delegate?: string,
|
|
99
|
+
* agentFrontmatter?: {tier_override?: string, default_tier?: string},
|
|
100
|
+
* adaptiveMode?: 'static'|'hedge'|'full',
|
|
101
|
+
* baseDir?: string,
|
|
102
|
+
* posteriorPath?: string,
|
|
103
|
+
* }} input
|
|
104
|
+
* @returns {{
|
|
105
|
+
* tier: string,
|
|
106
|
+
* decision_log: {
|
|
107
|
+
* source: 'frontmatter'|'tier_override_bypass'|'bandit_pull'|'bandit_pull_with_delegate',
|
|
108
|
+
* samples?: object,
|
|
109
|
+
* delegate?: string,
|
|
110
|
+
* adaptive_mode: 'static'|'hedge'|'full',
|
|
111
|
+
* reason?: string,
|
|
112
|
+
* }
|
|
113
|
+
* }}
|
|
114
|
+
*/
|
|
115
|
+
function consultBandit(input) {
|
|
116
|
+
if (!input || typeof input !== 'object') {
|
|
117
|
+
throw new TypeError('integration.consultBandit: input object required');
|
|
118
|
+
}
|
|
119
|
+
if (typeof input.agent !== 'string' || input.agent.length === 0) {
|
|
120
|
+
throw new TypeError('integration.consultBandit: agent (string) required');
|
|
121
|
+
}
|
|
122
|
+
if (typeof input.bin !== 'string' || input.bin.length === 0) {
|
|
123
|
+
throw new TypeError('integration.consultBandit: bin (string) required');
|
|
124
|
+
}
|
|
125
|
+
|
|
126
|
+
const agentFrontmatter = input.agentFrontmatter && typeof input.agentFrontmatter === 'object'
|
|
127
|
+
? input.agentFrontmatter
|
|
128
|
+
: {};
|
|
129
|
+
const adaptiveMode = resolveAdaptiveMode(input.adaptiveMode, input);
|
|
130
|
+
const delegate = resolveDelegate(input.delegate, 'consultBandit');
|
|
131
|
+
|
|
132
|
+
// Step 1 — tier_override bypass (D-05). Highest priority; beats both
|
|
133
|
+
// bandit consultation and static/hedge frontmatter.default_tier.
|
|
134
|
+
if (typeof agentFrontmatter.tier_override === 'string' && agentFrontmatter.tier_override.length > 0) {
|
|
135
|
+
return {
|
|
136
|
+
tier: agentFrontmatter.tier_override,
|
|
137
|
+
decision_log: {
|
|
138
|
+
source: 'tier_override_bypass',
|
|
139
|
+
adaptive_mode: adaptiveMode,
|
|
140
|
+
reason: 'frontmatter_tier_override_set',
|
|
141
|
+
},
|
|
142
|
+
};
|
|
143
|
+
}
|
|
144
|
+
|
|
145
|
+
// Step 2 — non-full short-circuit (D-07). Static and hedge are both
|
|
146
|
+
// "bandit silent"; frontmatter.default_tier (or 'sonnet' fallback)
|
|
147
|
+
// is authoritative. No posterior read or write.
|
|
148
|
+
if (adaptiveMode !== 'full') {
|
|
149
|
+
const fallbackTier = (typeof agentFrontmatter.default_tier === 'string' && agentFrontmatter.default_tier.length > 0)
|
|
150
|
+
? agentFrontmatter.default_tier
|
|
151
|
+
: 'sonnet';
|
|
152
|
+
return {
|
|
153
|
+
tier: fallbackTier,
|
|
154
|
+
decision_log: {
|
|
155
|
+
source: 'frontmatter',
|
|
156
|
+
adaptive_mode: adaptiveMode,
|
|
157
|
+
reason: adaptiveMode === 'hedge' ? 'hedge_mode_skips_bandit' : 'static_mode_authoritative',
|
|
158
|
+
},
|
|
159
|
+
};
|
|
160
|
+
}
|
|
161
|
+
|
|
162
|
+
// Step 3/4 — full mode → consult the bandit. Choice of pull vs
|
|
163
|
+
// pullWithDelegate is driven by `delegate`:
|
|
164
|
+
// delegate === 'none' (or undefined → 'none') → pull()
|
|
165
|
+
// delegate ∈ {gemini,codex,cursor,copilot,qwen} → pullWithDelegate
|
|
166
|
+
if (delegate === DELEGATE_NONE) {
|
|
167
|
+
const result = banditRouter.pull({
|
|
168
|
+
agent: input.agent,
|
|
169
|
+
bin: input.bin,
|
|
170
|
+
baseDir: input.baseDir,
|
|
171
|
+
posteriorPath: input.posteriorPath,
|
|
172
|
+
});
|
|
173
|
+
return {
|
|
174
|
+
tier: result.tier,
|
|
175
|
+
decision_log: {
|
|
176
|
+
source: 'bandit_pull',
|
|
177
|
+
samples: result.samples,
|
|
178
|
+
delegate: DELEGATE_NONE,
|
|
179
|
+
adaptive_mode: 'full',
|
|
180
|
+
},
|
|
181
|
+
};
|
|
182
|
+
}
|
|
183
|
+
|
|
184
|
+
// Path 4 — peer delegate. Constrain the delegate axis to the single
|
|
185
|
+
// requested peer so the bandit samples the (tier × delegate) joint
|
|
186
|
+
// restricted to {delegate}. Same posterior file, same arm shape; the
|
|
187
|
+
// arm's `delegate` field is set so the slice is distinct from local.
|
|
188
|
+
const result = banditRouter.pullWithDelegate({
|
|
189
|
+
agent: input.agent,
|
|
190
|
+
bin: input.bin,
|
|
191
|
+
delegates: [delegate],
|
|
192
|
+
baseDir: input.baseDir,
|
|
193
|
+
posteriorPath: input.posteriorPath,
|
|
194
|
+
});
|
|
195
|
+
return {
|
|
196
|
+
tier: result.tier,
|
|
197
|
+
decision_log: {
|
|
198
|
+
source: 'bandit_pull_with_delegate',
|
|
199
|
+
samples: result.samples,
|
|
200
|
+
delegate: result.delegate,
|
|
201
|
+
adaptive_mode: 'full',
|
|
202
|
+
},
|
|
203
|
+
};
|
|
204
|
+
}
|
|
205
|
+
|
|
206
|
+
/**
|
|
207
|
+
* recordOutcome — post-spawn telemetry update. Computes a reward via
|
|
208
|
+
* computeReward (Phase 23.5 D-08, unchanged) and writes the posterior
|
|
209
|
+
* arm. Best-effort: all errors swallowed so telemetry can never break
|
|
210
|
+
* a session (D-04).
|
|
211
|
+
*
|
|
212
|
+
* No-op when adaptive_mode is not 'full' (D-07).
|
|
213
|
+
*
|
|
214
|
+
* @param {{
|
|
215
|
+
* agent: string,
|
|
216
|
+
* bin: string,
|
|
217
|
+
* delegate?: string,
|
|
218
|
+
* tier: string,
|
|
219
|
+
* status: string,
|
|
220
|
+
* costUsd?: number,
|
|
221
|
+
* adaptiveMode?: 'static'|'hedge'|'full',
|
|
222
|
+
* baseDir?: string,
|
|
223
|
+
* posteriorPath?: string,
|
|
224
|
+
* }} input
|
|
225
|
+
* @returns {void}
|
|
226
|
+
*/
|
|
227
|
+
function recordOutcome(input) {
|
|
228
|
+
if (!input || typeof input !== 'object') {
|
|
229
|
+
throw new TypeError('integration.recordOutcome: input object required');
|
|
230
|
+
}
|
|
231
|
+
if (typeof input.agent !== 'string' || input.agent.length === 0) {
|
|
232
|
+
throw new TypeError('integration.recordOutcome: agent (string) required');
|
|
233
|
+
}
|
|
234
|
+
if (typeof input.bin !== 'string' || input.bin.length === 0) {
|
|
235
|
+
throw new TypeError('integration.recordOutcome: bin (string) required');
|
|
236
|
+
}
|
|
237
|
+
if (typeof input.tier !== 'string' || input.tier.length === 0) {
|
|
238
|
+
throw new TypeError('integration.recordOutcome: tier (string) required');
|
|
239
|
+
}
|
|
240
|
+
if (typeof input.status !== 'string') {
|
|
241
|
+
throw new TypeError('integration.recordOutcome: status (string) required');
|
|
242
|
+
}
|
|
243
|
+
|
|
244
|
+
const adaptiveMode = resolveAdaptiveMode(input.adaptiveMode, input);
|
|
245
|
+
|
|
246
|
+
// D-07 + D-04: posterior is silent in static/hedge. No-op early.
|
|
247
|
+
if (adaptiveMode !== 'full') {
|
|
248
|
+
return undefined;
|
|
249
|
+
}
|
|
250
|
+
|
|
251
|
+
const delegate = resolveDelegate(input.delegate, 'recordOutcome');
|
|
252
|
+
|
|
253
|
+
// D-08: reward function unchanged. wall_time_ms always 0 per
|
|
254
|
+
// Phase 23.5 / 27.5 — the wall-time tiebreaker is not used at the
|
|
255
|
+
// recordOutcome boundary; correctness + cost are the only signals.
|
|
256
|
+
const reward = banditRouter.computeReward({
|
|
257
|
+
solidify_pass: input.status === 'completed',
|
|
258
|
+
cost_usd: typeof input.costUsd === 'number' ? input.costUsd : 0,
|
|
259
|
+
wall_time_ms: 0,
|
|
260
|
+
});
|
|
261
|
+
|
|
262
|
+
// D-04: best-effort write. Swallow ALL exceptions so a broken
|
|
263
|
+
// posterior file never breaks a session.
|
|
264
|
+
try {
|
|
265
|
+
if (delegate === DELEGATE_NONE) {
|
|
266
|
+
banditRouter.update({
|
|
267
|
+
agent: input.agent,
|
|
268
|
+
bin: input.bin,
|
|
269
|
+
tier: input.tier,
|
|
270
|
+
reward,
|
|
271
|
+
baseDir: input.baseDir,
|
|
272
|
+
posteriorPath: input.posteriorPath,
|
|
273
|
+
});
|
|
274
|
+
} else {
|
|
275
|
+
banditRouter.updateWithDelegate({
|
|
276
|
+
agent: input.agent,
|
|
277
|
+
bin: input.bin,
|
|
278
|
+
tier: input.tier,
|
|
279
|
+
delegate,
|
|
280
|
+
reward,
|
|
281
|
+
baseDir: input.baseDir,
|
|
282
|
+
posteriorPath: input.posteriorPath,
|
|
283
|
+
});
|
|
284
|
+
}
|
|
285
|
+
} catch (err) {
|
|
286
|
+
// Live-tail breadcrumb opt-in via env var. Inner try/catch around
|
|
287
|
+
// the stderr write itself keeps the swallow guarantee even when
|
|
288
|
+
// stderr is closed/unavailable.
|
|
289
|
+
if (process.env.GDD_BANDIT_DEBUG === '1') {
|
|
290
|
+
try {
|
|
291
|
+
process.stderr.write(
|
|
292
|
+
'[bandit-integration] recordOutcome swallowed: ' +
|
|
293
|
+
(err && err.message ? err.message : String(err)) +
|
|
294
|
+
'\n',
|
|
295
|
+
);
|
|
296
|
+
} catch {
|
|
297
|
+
/* swallow */
|
|
298
|
+
}
|
|
299
|
+
}
|
|
300
|
+
}
|
|
301
|
+
|
|
302
|
+
return undefined;
|
|
303
|
+
}
|
|
304
|
+
|
|
305
|
+
module.exports = {
|
|
306
|
+
consultBandit,
|
|
307
|
+
recordOutcome,
|
|
308
|
+
DELEGATE_NONE,
|
|
309
|
+
};
|
|
@@ -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
|
|
@@ -680,6 +798,22 @@ export async function run(opts: SessionRunnerOptions): Promise<SessionResult> {
|
|
|
680
798
|
const toolCalls: SessionResult['tool_calls'] = [];
|
|
681
799
|
const usage = { input: 0, output: 0, model: null as string | null };
|
|
682
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
|
+
}
|
|
683
817
|
let finalText: string | undefined;
|
|
684
818
|
|
|
685
819
|
// -- 4. Emit session.started. -------------------------------------------
|
|
@@ -719,6 +853,20 @@ export async function run(opts: SessionRunnerOptions): Promise<SessionResult> {
|
|
|
719
853
|
transcript_path: transcriptPath,
|
|
720
854
|
sanitizer: { applied: [...result.sanitizer.applied], removedSections: [...result.sanitizer.removedSections] },
|
|
721
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
|
+
});
|
|
722
870
|
transcript.close();
|
|
723
871
|
return result;
|
|
724
872
|
}
|
|
@@ -765,6 +913,37 @@ export async function run(opts: SessionRunnerOptions): Promise<SessionResult> {
|
|
|
765
913
|
transcript_path: transcriptPath,
|
|
766
914
|
sanitizer: { applied: [...peerResult.sanitizer.applied], removedSections: [...peerResult.sanitizer.removedSections] },
|
|
767
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
|
+
});
|
|
768
947
|
transcript.close();
|
|
769
948
|
if (opts.signal !== undefined) opts.signal.removeEventListener('abort', onExternalAbort);
|
|
770
949
|
return peerResult;
|
|
@@ -796,6 +975,18 @@ export async function run(opts: SessionRunnerOptions): Promise<SessionResult> {
|
|
|
796
975
|
transcript_path: transcriptPath,
|
|
797
976
|
sanitizer: { applied: [...sanResult.applied], removedSections: [...sanResult.removedSections] },
|
|
798
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
|
+
});
|
|
799
990
|
transcript.close();
|
|
800
991
|
if (opts.signal !== undefined) opts.signal.removeEventListener('abort', onExternalAbort);
|
|
801
992
|
return result;
|
|
@@ -890,6 +1081,21 @@ export async function run(opts: SessionRunnerOptions): Promise<SessionResult> {
|
|
|
890
1081
|
transcript_path: transcriptPath,
|
|
891
1082
|
sanitizer: { applied: [...sanResult.applied], removedSections: [...sanResult.removedSections] },
|
|
892
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
|
+
});
|
|
893
1099
|
|
|
894
1100
|
return result;
|
|
895
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.
|