pi-smart-router 1.2.0 → 1.2.2
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/.pi/extensions/smart-router/delegate-stream.ts +7 -1
- package/.pi/extensions/smart-router/delegation-runtime.ts +12 -2
- package/.pi/extensions/smart-router/fleet-bootstrap.ts +2 -0
- package/.pi/extensions/smart-router/pi-model-scope.ts +18 -3
- package/.pi/extensions/smart-router/routing-context.ts +95 -11
- package/dist/config/pi-model-mapper.d.ts +12 -1
- package/dist/config/pi-model-mapper.d.ts.map +1 -1
- package/dist/config/pi-model-mapper.js +33 -8
- package/dist/config/pi-model-mapper.js.map +1 -1
- package/dist/domain/delegation/delegation-context.d.ts +49 -0
- package/dist/domain/delegation/delegation-context.d.ts.map +1 -1
- package/dist/domain/delegation/delegation-context.js +92 -0
- package/dist/domain/delegation/delegation-context.js.map +1 -1
- package/dist/domain/matching/hydra-input.d.ts.map +1 -1
- package/dist/domain/matching/hydra-input.js +4 -17
- package/dist/domain/matching/hydra-input.js.map +1 -1
- package/dist/domain/pinning/loop-escalation.d.ts.map +1 -1
- package/dist/domain/pinning/loop-escalation.js +1 -49
- package/dist/domain/pinning/loop-escalation.js.map +1 -1
- package/dist/domain/pinning/tool-failure.d.ts +20 -0
- package/dist/domain/pinning/tool-failure.d.ts.map +1 -0
- package/dist/domain/pinning/tool-failure.js +62 -0
- package/dist/domain/pinning/tool-failure.js.map +1 -0
- package/dist/index.d.ts +1 -1
- package/dist/index.d.ts.map +1 -1
- package/dist/index.js +1 -1
- package/dist/index.js.map +1 -1
- package/dist/infrastructure/persistence/sqlite-store.d.ts.map +1 -1
- package/dist/infrastructure/persistence/sqlite-store.js +37 -9
- package/dist/infrastructure/persistence/sqlite-store.js.map +1 -1
- package/dist/infrastructure/pricing/price-broker.d.ts +6 -1
- package/dist/infrastructure/pricing/price-broker.d.ts.map +1 -1
- package/dist/infrastructure/pricing/price-broker.js +16 -1
- package/dist/infrastructure/pricing/price-broker.js.map +1 -1
- package/package.json +3 -3
- package/src/config/pi-model-mapper.ts +48 -9
- package/src/domain/delegation/delegation-context.ts +123 -0
- package/src/domain/matching/hydra-input.ts +4 -19
- package/src/domain/pinning/loop-escalation.ts +1 -52
- package/src/domain/pinning/tool-failure.ts +66 -0
- package/src/index.ts +6 -0
- package/src/infrastructure/persistence/sqlite-store.ts +45 -9
- package/src/infrastructure/pricing/price-broker.ts +18 -2
|
@@ -1 +1 @@
|
|
|
1
|
-
{"version":3,"file":"price-broker.js","sourceRoot":"","sources":["../../../src/infrastructure/pricing/price-broker.ts"],"names":[],"mappings":"AAAA
|
|
1
|
+
{"version":3,"file":"price-broker.js","sourceRoot":"","sources":["../../../src/infrastructure/pricing/price-broker.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAGH,OAAO,EAAE,uBAAuB,EAA2B,MAAM,iCAAiC,CAAC;AAEnG,OAAO,EACL,YAAY,EACZ,yBAAyB,GAE1B,MAAM,0CAA0C,CAAC;AAClD,OAAO,EACL,YAAY,GAEb,MAAM,0CAA0C,CAAC;AAOlD,SAAS,oBAAoB,CAC3B,KAAmB,EACnB,OAAqB;IAErB,MAAM,QAAQ,GAAG,OAAO,CAAC,wBAAwB,CAAC;IAClD,IAAI,CAAC,QAAQ,EAAE,CAAC;QACd,OAAO,SAAS,CAAC;IACnB,CAAC;IAED,MAAM,WAAW,GAAG,KAAK,CAAC,OAAO,CAAC,YAAY,IAAI,KAAK,CAAC,EAAE,CAAC;IAC3D,OAAO,QAAQ,CAAC,WAAW,CAAC,IAAI,QAAQ,CAAC,KAAK,CAAC,EAAE,CAAC,CAAC;AACrD,CAAC;AAED;;;GAGG;AACH,SAAS,sBAAsB,CAAC,KAAmB;IACjD,OAAQ,KAA4B,CAAC,kBAAkB,CAAC;AAC1D,CAAC;AAED;;;;;GAKG;AACH,MAAM,UAAU,aAAa,CAC3B,KAAmB,EACnB,OAA4B;IAE5B,MAAM,UAAU,GAAG,KAAK,CAAC,MAAM,CAAC;IAChC,MAAM,cAAc,GAAG,OAAO,CAAC,CAAC,CAAC,oBAAoB,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC,CAAC,CAAC,SAAS,CAAC;IAClF,MAAM,gBAAgB,GAAG,sBAAsB,CAAC,KAAK,CAAC,CAAC;IACvD,MAAM,QAAQ,GAAG,uBAAuB,CAAC,KAAK,CAAC,IAAI,CAAC,CAAC;IAErD,OAAO;QACL,QAAQ,EAAE,KAAK,CAAC,EAAE;QAClB,MAAM,EAAE;YACN,gBAAgB,EACd,UAAU,EAAE,gBAAgB;gBAC5B,cAAc,EAAE,gBAAgB;gBAChC,gBAAgB,EAAE,gBAAgB;gBAClC,QAAQ,CAAC,gBAAgB;YAC3B,iBAAiB,EACf,UAAU,EAAE,iBAAiB;gBAC7B,cAAc,EAAE,iBAAiB;gBACjC,gBAAgB,EAAE,iBAAiB;gBACnC,QAAQ,CAAC,iBAAiB;SAC7B;KACF,CAAC;AACJ,CAAC;AAED;;GAEG;AACH,MAAM,UAAU,yBAAyB,CACvC,KAA8B,EAC9B,OAA4B;IAE5B,OAAO,KAAK,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE;QAC3B,MAAM,QAAQ,GAAG,aAAa,CAAC,OAAO,EAAE,OAAO,CAAC,CAAC;QACjD,OAAO;YACL,GAAG,OAAO;YACV,MAAM,EAAE,QAAQ,CAAC,MAAM;SACxB,CAAC;IACJ,CAAC,CAAC,CAAC;AACL,CAAC;AAED;;;;GAIG;AACH,MAAM,UAAU,kBAAkB,CAChC,KAA8B,EAC9B,OAA4B;IAE5B,MAAM,MAAM,GAAG,IAAI,GAAG,EAAyB,CAAC;IAChD,KAAK,MAAM,KAAK,IAAI,KAAK,EAAE,CAAC;QAC1B,MAAM,CAAC,GAAG,CAAC,KAAK,CAAC,EAAE,EAAE,YAAY,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC,CAAC;IACrD,CAAC;IACD,OAAO,MAAM,CAAC;AAChB,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,yBAAyB,CACvC,KAA8B,EAC9B,OAA4B;IAE5B,IAAI,CAAC,OAAO,EAAE,CAAC;QACb,OAAO,yBAAyB,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC;IACnD,CAAC;IAED,MAAM,MAAM,GAAG,kBAAkB,CAAC,KAAK,EAAE,OAAO,CAAC,CAAC;IAElD,MAAM,MAAM,GAAG,KAAK,CAAC,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE;QACnC,MAAM,QAAQ,GAAG,MAAM,CAAC,GAAG,CAAC,OAAO,CAAC,EAAE,CAAC,CAAC;QACxC,IAAI,CAAC,QAAQ,IAAI,QAAQ,CAAC,MAAM,KAAK,eAAe,EAAE,CAAC;YACrD,OAAO,OAAO,CAAC;QACjB,CAAC;QAED,OAAO;YACL,GAAG,OAAO;YACV,OAAO,EAAE;gBACP,GAAG,OAAO,CAAC,OAAO;gBAClB,oBAAoB,EAAE,QAAQ,CAAC,kBAAkB;aAClD;SACF,CAAC;IACJ,CAAC,CAAC,CAAC;IAEH,OAAO,yBAAyB,CAAC,MAAM,EAAE,OAAO,CAAC,CAAC;AACpD,CAAC"}
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "pi-smart-router",
|
|
3
|
-
"version": "1.2.
|
|
3
|
+
"version": "1.2.2",
|
|
4
4
|
"description": "Auto-model router middleware for the pi.dev coding agent",
|
|
5
5
|
"type": "module",
|
|
6
6
|
"license": "MIT",
|
|
@@ -110,7 +110,8 @@
|
|
|
110
110
|
}
|
|
111
111
|
},
|
|
112
112
|
"devDependencies": {
|
|
113
|
-
"@earendil-works/pi-
|
|
113
|
+
"@earendil-works/pi-ai": "^1.3.0",
|
|
114
|
+
"@earendil-works/pi-coding-agent": "^1.3.0",
|
|
114
115
|
"@types/better-sqlite3": "^7.6.13",
|
|
115
116
|
"@types/node": "^22.15.21",
|
|
116
117
|
"@typescript-eslint/eslint-plugin": "^8.62.1",
|
|
@@ -124,7 +125,6 @@
|
|
|
124
125
|
"vitest": "^3.2.3"
|
|
125
126
|
},
|
|
126
127
|
"dependencies": {
|
|
127
|
-
"@earendil-works/pi-ai": "^0.85.1",
|
|
128
128
|
"@huggingface/transformers": "^4.2.0",
|
|
129
129
|
"better-sqlite3": "^12.11.1",
|
|
130
130
|
"yaml": "^2.9.0",
|
|
@@ -33,6 +33,13 @@ export type CapabilitySource = 'benchmark' | 'pattern_default';
|
|
|
33
33
|
/** ModelProfile plus operator-visible capability provenance (SP-174). */
|
|
34
34
|
export interface MappedModelProfile extends ModelProfile {
|
|
35
35
|
readonly capability_source: CapabilitySource;
|
|
36
|
+
/**
|
|
37
|
+
* Context limits carried from the pi model registry (`Model.contextWindow` /
|
|
38
|
+
* `Model.maxTokens`) — SP-299 / #177. Kept separate from `limits` (YAML
|
|
39
|
+
* overrides) so `resolveLimits` can rank them below LiteLLM catalog data
|
|
40
|
+
* and above tier defaults.
|
|
41
|
+
*/
|
|
42
|
+
readonly pi_registry_limits?: ModelLimits;
|
|
36
43
|
}
|
|
37
44
|
|
|
38
45
|
const benchmarkCapabilitiesSchema = z.object({
|
|
@@ -181,7 +188,7 @@ function withBenchmarkCapabilities(
|
|
|
181
188
|
};
|
|
182
189
|
}
|
|
183
190
|
|
|
184
|
-
/** Pi registry `Model.cost` shape — per-token USD rates. */
|
|
191
|
+
/** Pi registry `Model.cost` shape — per-1M-token USD rates (SP-298). */
|
|
185
192
|
export interface PiRegistryCost {
|
|
186
193
|
readonly input: number;
|
|
187
194
|
readonly output: number;
|
|
@@ -194,6 +201,10 @@ export interface PiModelInput {
|
|
|
194
201
|
readonly id: string;
|
|
195
202
|
readonly name?: string;
|
|
196
203
|
readonly cost?: PiRegistryCost;
|
|
204
|
+
/** Pi registry `Model.contextWindow` — total input token window (SP-299 / #177). */
|
|
205
|
+
readonly contextWindow?: number;
|
|
206
|
+
/** Pi registry `Model.maxTokens` — max output tokens (SP-299 / #177). */
|
|
207
|
+
readonly maxTokens?: number;
|
|
197
208
|
}
|
|
198
209
|
|
|
199
210
|
interface ModelPatternRule {
|
|
@@ -355,19 +366,24 @@ function matchPatternRules(id: string): ModelFamilyDefaults | undefined {
|
|
|
355
366
|
}
|
|
356
367
|
|
|
357
368
|
/**
|
|
358
|
-
* Registry cost → USD per 1M tokens: average of input and output rates
|
|
359
|
-
*
|
|
360
|
-
*
|
|
369
|
+
* Registry cost → USD per 1M tokens: average of input and output rates.
|
|
370
|
+
* Pi registry `Model.cost` rates are already denominated per 1M tokens
|
|
371
|
+
* (SP-298), so no scaling is applied — unlike LiteLLM per-token rates handled
|
|
372
|
+
* by `computeCostPer1MTokens` in litellm-fetch (SP-045/SP-046).
|
|
373
|
+
* Returns undefined when both input and output rates are zero (use pattern
|
|
374
|
+
* default), or when either rate is negative — negative values are vendor
|
|
375
|
+
* unknown/sentinel rates, not valid pricing, so cost stays unknown (SP-298).
|
|
361
376
|
*/
|
|
362
377
|
function deriveFallbackCostPer1M(cost: PiRegistryCost): number | undefined {
|
|
363
378
|
const { input, output } = cost;
|
|
364
379
|
if (input === 0 && output === 0) {
|
|
365
380
|
return undefined;
|
|
366
381
|
}
|
|
382
|
+
if (input < 0 || output < 0) {
|
|
383
|
+
return undefined;
|
|
384
|
+
}
|
|
367
385
|
|
|
368
|
-
|
|
369
|
-
const outputPer1M = output * 1_000_000;
|
|
370
|
-
return (inputPer1M + outputPer1M) / 2;
|
|
386
|
+
return (input + output) / 2;
|
|
371
387
|
}
|
|
372
388
|
|
|
373
389
|
function resolveFallbackCost(
|
|
@@ -382,6 +398,23 @@ function resolveFallbackCost(
|
|
|
382
398
|
return fromRegistry ?? defaults.pricing.fallback_cost_per_1m;
|
|
383
399
|
}
|
|
384
400
|
|
|
401
|
+
/**
|
|
402
|
+
* Pi registry contextWindow/maxTokens → ModelLimits (SP-299 / #177).
|
|
403
|
+
* Non-positive or non-finite registry values are vendor sentinels, not real
|
|
404
|
+
* limits — ignored so tier defaults still apply.
|
|
405
|
+
*/
|
|
406
|
+
function derivePiRegistryLimits(input: PiModelInput): ModelLimits | undefined {
|
|
407
|
+
const limits: { max_input_tokens?: number; max_output_tokens?: number } = {};
|
|
408
|
+
const { contextWindow, maxTokens } = input;
|
|
409
|
+
if (contextWindow !== undefined && Number.isFinite(contextWindow) && contextWindow > 0) {
|
|
410
|
+
limits.max_input_tokens = Math.floor(contextWindow);
|
|
411
|
+
}
|
|
412
|
+
if (maxTokens !== undefined && Number.isFinite(maxTokens) && maxTokens > 0) {
|
|
413
|
+
limits.max_output_tokens = Math.floor(maxTokens);
|
|
414
|
+
}
|
|
415
|
+
return Object.keys(limits).length > 0 ? limits : undefined;
|
|
416
|
+
}
|
|
417
|
+
|
|
385
418
|
function buildProfile(
|
|
386
419
|
input: PiModelInput,
|
|
387
420
|
defaults: ModelFamilyDefaults,
|
|
@@ -410,11 +443,17 @@ function buildProfile(
|
|
|
410
443
|
? { ...profile, performance: { ...defaults.performance } }
|
|
411
444
|
: profile;
|
|
412
445
|
|
|
446
|
+
const piRegistryLimits = derivePiRegistryLimits(input);
|
|
447
|
+
const withLimits =
|
|
448
|
+
piRegistryLimits !== undefined
|
|
449
|
+
? { ...withPerformance, pi_registry_limits: piRegistryLimits }
|
|
450
|
+
: withPerformance;
|
|
451
|
+
|
|
413
452
|
if (defaults.endpoint !== undefined) {
|
|
414
|
-
return { ...
|
|
453
|
+
return { ...withLimits, endpoint: defaults.endpoint };
|
|
415
454
|
}
|
|
416
455
|
|
|
417
|
-
return
|
|
456
|
+
return withLimits;
|
|
418
457
|
}
|
|
419
458
|
|
|
420
459
|
/**
|
|
@@ -85,6 +85,30 @@ export function isGoogleOriginAssistantMessage(message: AssistantMessage): boole
|
|
|
85
85
|
return provider === 'cursor' && /gemini/i.test(message.model);
|
|
86
86
|
}
|
|
87
87
|
|
|
88
|
+
const CURSOR_PROVIDER = 'cursor' as const;
|
|
89
|
+
|
|
90
|
+
/**
|
|
91
|
+
* Env gate for the Cursor lifecycle strip (SP-300, #180). Default on;
|
|
92
|
+
* `SMART_ROUTER_STRIP_CURSOR_LIFECYCLE=0` disables the sanitizer.
|
|
93
|
+
*/
|
|
94
|
+
export const STRIP_CURSOR_LIFECYCLE_ENV = 'SMART_ROUTER_STRIP_CURSOR_LIFECYCLE' as const;
|
|
95
|
+
|
|
96
|
+
/** Whether the delegation target is a Cursor model (cursor/auto, composer-*). */
|
|
97
|
+
export function isCursorDelegationTarget<TApi extends Api>(model: Model<TApi>): boolean {
|
|
98
|
+
return model.provider.trim().toLowerCase() === CURSOR_PROVIDER;
|
|
99
|
+
}
|
|
100
|
+
|
|
101
|
+
/**
|
|
102
|
+
* Whether an assistant message was produced by the Cursor provider (e.g.
|
|
103
|
+
* `cursor/auto`). Cursor-origin thinking blocks carry pi-cursor-sdk lifecycle
|
|
104
|
+
* progress lines (`Cursor shell: …`, `Cursor MCP: …`) rather than model
|
|
105
|
+
* reasoning — replaying them cross-provider teaches the target to imitate
|
|
106
|
+
* shell commands as prose (issue #180).
|
|
107
|
+
*/
|
|
108
|
+
export function isCursorOriginAssistantMessage(message: AssistantMessage): boolean {
|
|
109
|
+
return message.provider.trim().toLowerCase() === CURSOR_PROVIDER;
|
|
110
|
+
}
|
|
111
|
+
|
|
88
112
|
export function hasReplaySensitiveState(messages: readonly Message[]): boolean {
|
|
89
113
|
for (const message of messages) {
|
|
90
114
|
if (message.role !== 'assistant') {
|
|
@@ -236,3 +260,102 @@ export function repairGeminiReplayContext<TApi extends Api>(
|
|
|
236
260
|
|
|
237
261
|
return { ...context, messages };
|
|
238
262
|
}
|
|
263
|
+
|
|
264
|
+
export interface StripCursorLifecycleOptions {
|
|
265
|
+
/** Env override for tests; defaults to `process.env`. */
|
|
266
|
+
readonly env?: Record<string, string | undefined>;
|
|
267
|
+
}
|
|
268
|
+
|
|
269
|
+
export interface CursorLifecycleStripResult {
|
|
270
|
+
readonly context: Context;
|
|
271
|
+
/** Thinking blocks removed from Cursor-origin assistant messages. */
|
|
272
|
+
readonly strippedThinkingBlocks: number;
|
|
273
|
+
/** Total characters of removed thinking text (observability for #180). */
|
|
274
|
+
readonly strippedThinkingChars: number;
|
|
275
|
+
}
|
|
276
|
+
|
|
277
|
+
/**
|
|
278
|
+
* SP-300 (#180): drop Cursor lifecycle thinking from Cursor-origin assistant
|
|
279
|
+
* messages before delegating to a non-Cursor target.
|
|
280
|
+
*
|
|
281
|
+
* The third-party pi-cursor-sdk provider emits tool progress as thinking lines
|
|
282
|
+
* (`Cursor shell: …`). pi-ai's transformMessages converts foreign thinking
|
|
283
|
+
* blocks to plain text when the provider/model differs, so a Cursor-origin
|
|
284
|
+
* history handed to GLM/Gemini replays those progress lines as assistant
|
|
285
|
+
* prose — the target then imitates shell commands as text instead of emitting
|
|
286
|
+
* real toolCalls (2026-10-01 glm-5.3-flash runaway).
|
|
287
|
+
*
|
|
288
|
+
* Scope guardrails:
|
|
289
|
+
* - Cursor targets are returned unchanged (Cursor→Cursor replay is native).
|
|
290
|
+
* - Non-Cursor thinking (Gemini, Anthropic, …) is never touched.
|
|
291
|
+
* - Text, toolCall, and other blocks are preserved; messages are never
|
|
292
|
+
* dropped — a message left with an empty content array is kept so
|
|
293
|
+
* tool-call/result pairing stays intact.
|
|
294
|
+
* - Disabled via `SMART_ROUTER_STRIP_CURSOR_LIFECYCLE=0`.
|
|
295
|
+
*
|
|
296
|
+
* Call after {@link normalizeDelegationContext} so virtual-tagged messages
|
|
297
|
+
* already resolved to a Cursor session execution are caught as Cursor-origin.
|
|
298
|
+
*/
|
|
299
|
+
export function stripCursorLifecycleThinking<TApi extends Api>(
|
|
300
|
+
context: Context,
|
|
301
|
+
targetModel: Model<TApi>,
|
|
302
|
+
options?: StripCursorLifecycleOptions,
|
|
303
|
+
): CursorLifecycleStripResult {
|
|
304
|
+
const unchanged: CursorLifecycleStripResult = {
|
|
305
|
+
context,
|
|
306
|
+
strippedThinkingBlocks: 0,
|
|
307
|
+
strippedThinkingChars: 0,
|
|
308
|
+
};
|
|
309
|
+
|
|
310
|
+
const env = options?.env ?? process.env;
|
|
311
|
+
if (env[STRIP_CURSOR_LIFECYCLE_ENV] === '0') {
|
|
312
|
+
return unchanged;
|
|
313
|
+
}
|
|
314
|
+
|
|
315
|
+
if (isCursorDelegationTarget(targetModel)) {
|
|
316
|
+
return unchanged;
|
|
317
|
+
}
|
|
318
|
+
|
|
319
|
+
let strippedThinkingBlocks = 0;
|
|
320
|
+
let strippedThinkingChars = 0;
|
|
321
|
+
let changed = false;
|
|
322
|
+
|
|
323
|
+
const messages: Message[] = [];
|
|
324
|
+
for (const message of context.messages) {
|
|
325
|
+
if (message.role !== 'assistant' || !isCursorOriginAssistantMessage(message)) {
|
|
326
|
+
messages.push(message);
|
|
327
|
+
continue;
|
|
328
|
+
}
|
|
329
|
+
|
|
330
|
+
const kept: typeof message.content = [];
|
|
331
|
+
for (const block of message.content) {
|
|
332
|
+
if (block.type === 'thinking') {
|
|
333
|
+
strippedThinkingBlocks += 1;
|
|
334
|
+
strippedThinkingChars += block.thinking.length;
|
|
335
|
+
changed = true;
|
|
336
|
+
continue;
|
|
337
|
+
}
|
|
338
|
+
kept.push(block);
|
|
339
|
+
}
|
|
340
|
+
|
|
341
|
+
if (kept.length === message.content.length) {
|
|
342
|
+
messages.push(message);
|
|
343
|
+
continue;
|
|
344
|
+
}
|
|
345
|
+
|
|
346
|
+
// Never drop messages: keep the message with its remaining blocks (an
|
|
347
|
+
// empty content array is fine) so user/toolResult turn pairings stay
|
|
348
|
+
// intact in the delegated history.
|
|
349
|
+
messages.push({ ...message, content: kept });
|
|
350
|
+
}
|
|
351
|
+
|
|
352
|
+
if (!changed) {
|
|
353
|
+
return unchanged;
|
|
354
|
+
}
|
|
355
|
+
|
|
356
|
+
return {
|
|
357
|
+
context: { ...context, messages },
|
|
358
|
+
strippedThinkingBlocks,
|
|
359
|
+
strippedThinkingChars,
|
|
360
|
+
};
|
|
361
|
+
}
|
|
@@ -8,18 +8,7 @@
|
|
|
8
8
|
|
|
9
9
|
import type { Message, RoutingRequest } from '../types/index.js';
|
|
10
10
|
import type { TriageResult } from '../triage/triage-engine.js';
|
|
11
|
-
|
|
12
|
-
const FAILURE_PATTERNS = [
|
|
13
|
-
'error',
|
|
14
|
-
'fail',
|
|
15
|
-
'exception',
|
|
16
|
-
'timed out',
|
|
17
|
-
'timeout',
|
|
18
|
-
'econnrefused',
|
|
19
|
-
'enotfound',
|
|
20
|
-
'econnreset',
|
|
21
|
-
'epipe',
|
|
22
|
-
] as const;
|
|
11
|
+
import { isToolFailure } from '../pinning/tool-failure.js';
|
|
23
12
|
|
|
24
13
|
function resolveMessageCount(request: RoutingRequest): number {
|
|
25
14
|
return request.messages?.length ?? 0;
|
|
@@ -45,14 +34,10 @@ function resolveCompactionFlag(request: RoutingRequest): 0 | 1 {
|
|
|
45
34
|
return request.compaction_flag ? 1 : 0;
|
|
46
35
|
}
|
|
47
36
|
|
|
48
|
-
function looksLikeToolFailure(content: string): boolean {
|
|
49
|
-
const lower = content.toLowerCase();
|
|
50
|
-
return FAILURE_PATTERNS.some((pattern) => lower.includes(pattern));
|
|
51
|
-
}
|
|
52
|
-
|
|
53
37
|
/**
|
|
54
38
|
* Observational loop-pressure flag: 1 when the latest tool message looks
|
|
55
|
-
* like a failure.
|
|
39
|
+
* like a failure. Uses the shared loop-escalation failure detection
|
|
40
|
+
* (`is_error` → `status` → tightened body heuristic; SP-303 / #186).
|
|
56
41
|
*/
|
|
57
42
|
function resolveLoopPressure(request: RoutingRequest): 0 | 1 {
|
|
58
43
|
const messages = request.messages;
|
|
@@ -63,7 +48,7 @@ function resolveLoopPressure(request: RoutingRequest): 0 | 1 {
|
|
|
63
48
|
for (let i = messages.length - 1; i >= 0; i--) {
|
|
64
49
|
const message = messages[i]!;
|
|
65
50
|
if (message.role === 'tool') {
|
|
66
|
-
return
|
|
51
|
+
return isToolFailure(message) ? 1 : 0;
|
|
67
52
|
}
|
|
68
53
|
}
|
|
69
54
|
|
|
@@ -22,6 +22,7 @@
|
|
|
22
22
|
*/
|
|
23
23
|
|
|
24
24
|
import type { Message, ModelProfile, RoutingRequest, SessionPin } from '../types/index.js';
|
|
25
|
+
import { isToolFailure } from './tool-failure.js';
|
|
25
26
|
|
|
26
27
|
// ─── Types ────────────────────────────────────────────────────────────────────
|
|
27
28
|
|
|
@@ -46,32 +47,6 @@ export const ZERO_TIER_TOOL_CHURN_SIGNATURE = 'zt:tool_churn' as const;
|
|
|
46
47
|
|
|
47
48
|
// ─── Failure signature extraction ─────────────────────────────────────────────
|
|
48
49
|
|
|
49
|
-
const FAILURE_PATTERNS = [
|
|
50
|
-
'error',
|
|
51
|
-
'fail',
|
|
52
|
-
'exception',
|
|
53
|
-
'timed out',
|
|
54
|
-
'timeout',
|
|
55
|
-
'econnrefused',
|
|
56
|
-
'enotfound',
|
|
57
|
-
'econnreset',
|
|
58
|
-
'epipe',
|
|
59
|
-
] as const;
|
|
60
|
-
|
|
61
|
-
/** Rate-limit and quota failures whose body text may omit generic error keywords. */
|
|
62
|
-
const RATE_LIMIT_PATTERNS = [
|
|
63
|
-
'rate limit',
|
|
64
|
-
'429',
|
|
65
|
-
'too many requests',
|
|
66
|
-
'quota exceeded',
|
|
67
|
-
] as const;
|
|
68
|
-
|
|
69
|
-
/** Auth/access failures that are not rate limits but still indicate tool failure. */
|
|
70
|
-
const AUTH_DENIED_PATTERNS = [
|
|
71
|
-
'permission denied',
|
|
72
|
-
'access denied',
|
|
73
|
-
] as const;
|
|
74
|
-
|
|
75
50
|
/** Host/agent signals that the model invoked a tool the runtime does not expose. */
|
|
76
51
|
const UNSUPPORTED_TOOL_PATTERNS = [
|
|
77
52
|
'unknown tool',
|
|
@@ -125,32 +100,6 @@ function lastToolMessage(request: RoutingRequest): Message | null {
|
|
|
125
100
|
return null;
|
|
126
101
|
}
|
|
127
102
|
|
|
128
|
-
/**
|
|
129
|
-
* Decide whether a tool-result message represents a failure.
|
|
130
|
-
* Prefers structured host signals (`is_error` / `status`); falls back to a
|
|
131
|
-
* tightened body heuristic only when no structured signal is present.
|
|
132
|
-
*/
|
|
133
|
-
function isToolFailure(msg: Message): boolean {
|
|
134
|
-
if (msg.is_error === true) return true;
|
|
135
|
-
if (msg.is_error === false) return false;
|
|
136
|
-
if (msg.status !== undefined) {
|
|
137
|
-
return msg.status >= 400;
|
|
138
|
-
}
|
|
139
|
-
return looksLikeFailure(msg.content ?? '');
|
|
140
|
-
}
|
|
141
|
-
|
|
142
|
-
function looksLikeFailure(content: string): boolean {
|
|
143
|
-
const lower = content.toLowerCase();
|
|
144
|
-
// Benign negations must not be counted as failures.
|
|
145
|
-
if (/\bno (error|errors|failure|failures)\b/.test(lower)) return false;
|
|
146
|
-
if (/\bwithout (an? )?(error|failure)\b/.test(lower)) return false;
|
|
147
|
-
return (
|
|
148
|
-
FAILURE_PATTERNS.some((p) => lower.includes(p)) ||
|
|
149
|
-
RATE_LIMIT_PATTERNS.some((p) => lower.includes(p)) ||
|
|
150
|
-
AUTH_DENIED_PATTERNS.some((p) => lower.includes(p))
|
|
151
|
-
);
|
|
152
|
-
}
|
|
153
|
-
|
|
154
103
|
/** Deterministic djb2 hash of normalised error content. */
|
|
155
104
|
function computeSignature(content: string): string {
|
|
156
105
|
const normalized = content.trim().slice(0, 256).toLowerCase();
|
|
@@ -0,0 +1,66 @@
|
|
|
1
|
+
/**
|
|
2
|
+
* Shared tool-failure detection — SP-303, GitHub #186.
|
|
3
|
+
*
|
|
4
|
+
* Single source of truth for deciding whether a tool-result message
|
|
5
|
+
* represents a failure. Used by loop escalation (pin break) and the
|
|
6
|
+
* HyDRA `loop:` metadata flag.
|
|
7
|
+
*
|
|
8
|
+
* Precedence: structured host signals (`is_error`, then `status`) win;
|
|
9
|
+
* a tightened body heuristic applies only when no structured signal
|
|
10
|
+
* is present. Benign negations ("no errors", "without failure") never
|
|
11
|
+
* count as failures.
|
|
12
|
+
*/
|
|
13
|
+
|
|
14
|
+
import type { Message } from '../types/index.js';
|
|
15
|
+
|
|
16
|
+
const FAILURE_PATTERNS = [
|
|
17
|
+
'error',
|
|
18
|
+
'fail',
|
|
19
|
+
'exception',
|
|
20
|
+
'timed out',
|
|
21
|
+
'timeout',
|
|
22
|
+
'econnrefused',
|
|
23
|
+
'enotfound',
|
|
24
|
+
'econnreset',
|
|
25
|
+
'epipe',
|
|
26
|
+
] as const;
|
|
27
|
+
|
|
28
|
+
/** Rate-limit and quota failures whose body text may omit generic error keywords. */
|
|
29
|
+
const RATE_LIMIT_PATTERNS = [
|
|
30
|
+
'rate limit',
|
|
31
|
+
'429',
|
|
32
|
+
'too many requests',
|
|
33
|
+
'quota exceeded',
|
|
34
|
+
] as const;
|
|
35
|
+
|
|
36
|
+
/** Auth/access failures that are not rate limits but still indicate tool failure. */
|
|
37
|
+
const AUTH_DENIED_PATTERNS = [
|
|
38
|
+
'permission denied',
|
|
39
|
+
'access denied',
|
|
40
|
+
] as const;
|
|
41
|
+
|
|
42
|
+
/**
|
|
43
|
+
* Decide whether a tool-result message represents a failure.
|
|
44
|
+
* Prefers structured host signals (`is_error` / `status`); falls back to a
|
|
45
|
+
* tightened body heuristic only when no structured signal is present.
|
|
46
|
+
*/
|
|
47
|
+
export function isToolFailure(msg: Message): boolean {
|
|
48
|
+
if (msg.is_error === true) return true;
|
|
49
|
+
if (msg.is_error === false) return false;
|
|
50
|
+
if (msg.status !== undefined) {
|
|
51
|
+
return msg.status >= 400;
|
|
52
|
+
}
|
|
53
|
+
return looksLikeFailure(msg.content ?? '');
|
|
54
|
+
}
|
|
55
|
+
|
|
56
|
+
function looksLikeFailure(content: string): boolean {
|
|
57
|
+
const lower = content.toLowerCase();
|
|
58
|
+
// Benign negations must not be counted as failures.
|
|
59
|
+
if (/\bno (error|errors|failure|failures)\b/.test(lower)) return false;
|
|
60
|
+
if (/\bwithout (an? )?(error|failure)\b/.test(lower)) return false;
|
|
61
|
+
return (
|
|
62
|
+
FAILURE_PATTERNS.some((p) => lower.includes(p)) ||
|
|
63
|
+
RATE_LIMIT_PATTERNS.some((p) => lower.includes(p)) ||
|
|
64
|
+
AUTH_DENIED_PATTERNS.some((p) => lower.includes(p))
|
|
65
|
+
);
|
|
66
|
+
}
|
package/src/index.ts
CHANGED
|
@@ -190,8 +190,14 @@ export {
|
|
|
190
190
|
} from './domain/delegation/adaptive-reasoning.js';
|
|
191
191
|
export {
|
|
192
192
|
isGoogleDelegationTarget,
|
|
193
|
+
isCursorDelegationTarget,
|
|
194
|
+
isCursorOriginAssistantMessage,
|
|
193
195
|
normalizeDelegationContext,
|
|
194
196
|
repairGeminiReplayContext,
|
|
197
|
+
stripCursorLifecycleThinking,
|
|
198
|
+
STRIP_CURSOR_LIFECYCLE_ENV,
|
|
199
|
+
type CursorLifecycleStripResult,
|
|
200
|
+
type StripCursorLifecycleOptions,
|
|
195
201
|
} from './domain/delegation/delegation-context.js';
|
|
196
202
|
export { ExecutionLedger } from './domain/delegation/execution-ledger.js';
|
|
197
203
|
export {
|
|
@@ -48,7 +48,7 @@ import { EMBEDDING_DIM } from '../../domain/matching/embedding-provider.js';
|
|
|
48
48
|
|
|
49
49
|
// ─── Schema version & migrations ────────────────────────────────────────────
|
|
50
50
|
|
|
51
|
-
const CURRENT_SCHEMA_VERSION =
|
|
51
|
+
const CURRENT_SCHEMA_VERSION = 9;
|
|
52
52
|
|
|
53
53
|
const MIGRATION_V1 = `
|
|
54
54
|
CREATE TABLE IF NOT EXISTS pins (
|
|
@@ -203,6 +203,18 @@ const MIGRATION_V8 = `
|
|
|
203
203
|
ALTER TABLE dataset ADD COLUMN embedding_json TEXT;
|
|
204
204
|
`;
|
|
205
205
|
|
|
206
|
+
// SP-301 / #171: persist context-fit gate telemetry so router-explain and
|
|
207
|
+
// exports can surface why a request overflowed after restart. Previously the
|
|
208
|
+
// fields were dropped on insert and read back as null literals.
|
|
209
|
+
const MIGRATION_V9 = `
|
|
210
|
+
ALTER TABLE telemetry ADD COLUMN estimated_input_tokens INTEGER;
|
|
211
|
+
ALTER TABLE telemetry ADD COLUMN context_fit_viable_count INTEGER;
|
|
212
|
+
ALTER TABLE telemetry ADD COLUMN context_fit_rejected_json TEXT;
|
|
213
|
+
ALTER TABLE telemetry ADD COLUMN context_overflow_pin_break INTEGER NOT NULL DEFAULT 0;
|
|
214
|
+
ALTER TABLE telemetry ADD COLUMN selected_model_max_input_tokens INTEGER;
|
|
215
|
+
ALTER TABLE telemetry ADD COLUMN context_fit_reason_code TEXT;
|
|
216
|
+
`;
|
|
217
|
+
|
|
206
218
|
// ─── Token bucket result ────────────────────────────────────────────────────
|
|
207
219
|
|
|
208
220
|
export interface TokenBucketResult {
|
|
@@ -708,14 +720,20 @@ export class SqliteStore implements StorePort {
|
|
|
708
720
|
estimated_cost_usd, routing_latency_ms, pin_reason,
|
|
709
721
|
actual_cost_usd, actual_input_tokens, actual_output_tokens,
|
|
710
722
|
actual_cache_read_tokens, actual_cache_write_tokens,
|
|
711
|
-
reasoning_level_requested, reasoning_level_applied, reasoning_reason_code
|
|
723
|
+
reasoning_level_requested, reasoning_level_applied, reasoning_reason_code,
|
|
724
|
+
estimated_input_tokens, context_fit_viable_count,
|
|
725
|
+
context_fit_rejected_json, context_overflow_pin_break,
|
|
726
|
+
selected_model_max_input_tokens, context_fit_reason_code
|
|
712
727
|
) VALUES (
|
|
713
728
|
@timestamp, @session_id, @request_id, @turn_type,
|
|
714
729
|
@stage, @reason_code, @selected_model_id,
|
|
715
730
|
@estimated_cost_usd, @routing_latency_ms, @pin_reason,
|
|
716
731
|
@actual_cost_usd, @actual_input_tokens, @actual_output_tokens,
|
|
717
732
|
@actual_cache_read_tokens, @actual_cache_write_tokens,
|
|
718
|
-
@reasoning_level_requested, @reasoning_level_applied, @reasoning_reason_code
|
|
733
|
+
@reasoning_level_requested, @reasoning_level_applied, @reasoning_reason_code,
|
|
734
|
+
@estimated_input_tokens, @context_fit_viable_count,
|
|
735
|
+
@context_fit_rejected_json, @context_overflow_pin_break,
|
|
736
|
+
@selected_model_max_input_tokens, @context_fit_reason_code
|
|
719
737
|
)`,
|
|
720
738
|
)
|
|
721
739
|
.run({
|
|
@@ -737,6 +755,12 @@ export class SqliteStore implements StorePort {
|
|
|
737
755
|
reasoning_level_requested: entry.reasoning_level_requested ?? null,
|
|
738
756
|
reasoning_level_applied: entry.reasoning_level_applied ?? null,
|
|
739
757
|
reasoning_reason_code: entry.reasoning_reason_code ?? null,
|
|
758
|
+
estimated_input_tokens: entry.estimated_input_tokens ?? null,
|
|
759
|
+
context_fit_viable_count: entry.context_fit_viable_count ?? null,
|
|
760
|
+
context_fit_rejected_json: entry.context_fit_rejected_json ?? null,
|
|
761
|
+
context_overflow_pin_break: entry.context_overflow_pin_break ? 1 : 0,
|
|
762
|
+
selected_model_max_input_tokens: entry.selected_model_max_input_tokens ?? null,
|
|
763
|
+
context_fit_reason_code: entry.context_fit_reason_code ?? null,
|
|
740
764
|
});
|
|
741
765
|
}
|
|
742
766
|
|
|
@@ -935,6 +959,12 @@ export class SqliteStore implements StorePort {
|
|
|
935
959
|
if (version < 8) {
|
|
936
960
|
this.db.exec(MIGRATION_V8);
|
|
937
961
|
version = 8;
|
|
962
|
+
this.db.pragma(`user_version = ${version}`);
|
|
963
|
+
}
|
|
964
|
+
|
|
965
|
+
if (version < 9) {
|
|
966
|
+
this.db.exec(MIGRATION_V9);
|
|
967
|
+
version = 9;
|
|
938
968
|
this.db.pragma(`user_version = ${CURRENT_SCHEMA_VERSION}`);
|
|
939
969
|
}
|
|
940
970
|
}
|
|
@@ -1033,6 +1063,12 @@ interface TelemetryRow {
|
|
|
1033
1063
|
reasoning_level_requested: string | null;
|
|
1034
1064
|
reasoning_level_applied: string | null;
|
|
1035
1065
|
reasoning_reason_code: string | null;
|
|
1066
|
+
estimated_input_tokens: number | null;
|
|
1067
|
+
context_fit_viable_count: number | null;
|
|
1068
|
+
context_fit_rejected_json: string | null;
|
|
1069
|
+
context_overflow_pin_break: number;
|
|
1070
|
+
selected_model_max_input_tokens: number | null;
|
|
1071
|
+
context_fit_reason_code: string | null;
|
|
1036
1072
|
}
|
|
1037
1073
|
|
|
1038
1074
|
interface DatasetRow {
|
|
@@ -1142,12 +1178,12 @@ function telemetryRowToEntity(row: TelemetryRow): RoutingTelemetry {
|
|
|
1142
1178
|
estimated_cost_usd: row.estimated_cost_usd,
|
|
1143
1179
|
routing_latency_ms: row.routing_latency_ms,
|
|
1144
1180
|
pin_reason: row.pin_reason,
|
|
1145
|
-
estimated_input_tokens:
|
|
1146
|
-
context_fit_viable_count:
|
|
1147
|
-
context_fit_rejected_json:
|
|
1148
|
-
context_overflow_pin_break:
|
|
1149
|
-
selected_model_max_input_tokens:
|
|
1150
|
-
context_fit_reason_code:
|
|
1181
|
+
estimated_input_tokens: row.estimated_input_tokens,
|
|
1182
|
+
context_fit_viable_count: row.context_fit_viable_count,
|
|
1183
|
+
context_fit_rejected_json: row.context_fit_rejected_json,
|
|
1184
|
+
context_overflow_pin_break: row.context_overflow_pin_break === 1,
|
|
1185
|
+
selected_model_max_input_tokens: row.selected_model_max_input_tokens,
|
|
1186
|
+
context_fit_reason_code: row.context_fit_reason_code,
|
|
1151
1187
|
cluster_id: null,
|
|
1152
1188
|
cluster_similarity: null,
|
|
1153
1189
|
cluster_margin: null,
|
|
@@ -11,10 +11,14 @@
|
|
|
11
11
|
* to finish the domain→infra inversion; this module re-exports them for
|
|
12
12
|
* import-path stability. What remains here is limit resolution and
|
|
13
13
|
* fleet-mapping, which depend on `config/pi-model-mapper` tier defaults.
|
|
14
|
+
*
|
|
15
|
+
* SP-299 (#177): limit resolution is a four-level cascade — YAML profile
|
|
16
|
+
* override → LiteLLM registry snapshot → pi registry (`contextWindow` /
|
|
17
|
+
* `maxTokens` carried on the mapped profile) → tier default.
|
|
14
18
|
*/
|
|
15
19
|
|
|
16
20
|
import type { ModelProfile, ModelLimits, PriceCatalog } from '../../domain/types/index.js';
|
|
17
|
-
import { getDefaultLimitsForTier } from '../../config/pi-model-mapper.js';
|
|
21
|
+
import { getDefaultLimitsForTier, type MappedModelProfile } from '../../config/pi-model-mapper.js';
|
|
18
22
|
|
|
19
23
|
export {
|
|
20
24
|
resolvePrice,
|
|
@@ -44,10 +48,19 @@ function lookupRegistryLimits(
|
|
|
44
48
|
return snapshot[registryKey] ?? snapshot[model.id];
|
|
45
49
|
}
|
|
46
50
|
|
|
51
|
+
/**
|
|
52
|
+
* Pi registry limits attached by `mapPiModelToProfile` (SP-299 / #177).
|
|
53
|
+
* Profiles built from fleet YAML never carry this field.
|
|
54
|
+
*/
|
|
55
|
+
function lookupPiRegistryLimits(model: ModelProfile): ModelLimits | undefined {
|
|
56
|
+
return (model as MappedModelProfile).pi_registry_limits;
|
|
57
|
+
}
|
|
58
|
+
|
|
47
59
|
/**
|
|
48
60
|
* Resolve effective context limits for a model.
|
|
49
61
|
*
|
|
50
|
-
* Priority: YAML/profile override → LiteLLM registry snapshot →
|
|
62
|
+
* Priority: YAML/profile override → LiteLLM registry snapshot → pi registry
|
|
63
|
+
* (`contextWindow`/`maxTokens`) → tier default (SP-299 / #177).
|
|
51
64
|
*/
|
|
52
65
|
export function resolveLimits(
|
|
53
66
|
model: ModelProfile,
|
|
@@ -55,6 +68,7 @@ export function resolveLimits(
|
|
|
55
68
|
): ResolvedLimits {
|
|
56
69
|
const yamlLimits = model.limits;
|
|
57
70
|
const registryLimits = catalog ? lookupRegistryLimits(model, catalog) : undefined;
|
|
71
|
+
const piRegistryLimits = lookupPiRegistryLimits(model);
|
|
58
72
|
const defaults = getDefaultLimitsForTier(model.tier);
|
|
59
73
|
|
|
60
74
|
return {
|
|
@@ -63,10 +77,12 @@ export function resolveLimits(
|
|
|
63
77
|
max_input_tokens:
|
|
64
78
|
yamlLimits?.max_input_tokens ??
|
|
65
79
|
registryLimits?.max_input_tokens ??
|
|
80
|
+
piRegistryLimits?.max_input_tokens ??
|
|
66
81
|
defaults.max_input_tokens,
|
|
67
82
|
max_output_tokens:
|
|
68
83
|
yamlLimits?.max_output_tokens ??
|
|
69
84
|
registryLimits?.max_output_tokens ??
|
|
85
|
+
piRegistryLimits?.max_output_tokens ??
|
|
70
86
|
defaults.max_output_tokens,
|
|
71
87
|
},
|
|
72
88
|
};
|