@warmdrift/kgauto-compiler 2.0.0-alpha.88 → 2.0.0-alpha.89
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/dist/{chunk-DSHWJFM2.mjs → chunk-D7RX3UD3.mjs} +1 -1
- package/dist/{chunk-GCOVYPJA.mjs → chunk-MKOOPD6S.mjs} +1 -1
- package/dist/{chunk-UNTUOZAG.mjs → chunk-OK2TMFRR.mjs} +11 -11
- package/dist/glassbox/index.d.mts +3 -3
- package/dist/glassbox/index.d.ts +3 -3
- package/dist/glassbox-routes/format.d.mts +2 -2
- package/dist/glassbox-routes/format.d.ts +2 -2
- package/dist/glassbox-routes/index.d.mts +4 -4
- package/dist/glassbox-routes/index.d.ts +4 -4
- package/dist/glassbox-routes/index.js +11 -11
- package/dist/glassbox-routes/index.mjs +2 -2
- package/dist/glassbox-routes/react/index.d.mts +2 -2
- package/dist/glassbox-routes/react/index.d.ts +2 -2
- package/dist/index.d.mts +25 -5
- package/dist/index.d.ts +25 -5
- package/dist/index.js +23 -13
- package/dist/index.mjs +13 -4
- package/dist/{ir-ZgN2KkuF.d.ts → ir-CTx026t0.d.ts} +16 -0
- package/dist/{ir-P3NEU8z4.d.mts → ir-DeYMLWge.d.mts} +16 -0
- package/dist/key-health.js +1 -1
- package/dist/key-health.mjs +1 -1
- package/dist/profiles.d.mts +1 -1
- package/dist/profiles.d.ts +1 -1
- package/dist/profiles.js +11 -11
- package/dist/profiles.mjs +1 -1
- package/dist/{types-B_MjEB0h.d.ts → types-BKbRtmUb.d.ts} +1 -1
- package/dist/{types-BgfuVIVI.d.ts → types-Cp9ot1HV.d.ts} +1 -1
- package/dist/{types-CK0zUlVE.d.mts → types-DD36cCbZ.d.mts} +1 -1
- package/dist/{types-7sF3adLO.d.mts → types-cBzinzUR.d.mts} +1 -1
- package/package.json +1 -1
|
@@ -211,7 +211,7 @@ var PROFILES_RAW = [
|
|
|
211
211
|
},
|
|
212
212
|
{
|
|
213
213
|
id: "claude-opus-4-7",
|
|
214
|
-
verifiedAgainstDocs: "2026-08-
|
|
214
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
215
215
|
provider: "anthropic",
|
|
216
216
|
status: "legacy",
|
|
217
217
|
// s76 2026-07-25: docs list as LEGACY (Opus 5 is current). Bundle said current while the brain already said legacy — the two disagreed.
|
|
@@ -261,7 +261,7 @@ var PROFILES_RAW = [
|
|
|
261
261
|
},
|
|
262
262
|
{
|
|
263
263
|
id: "claude-opus-4-6",
|
|
264
|
-
verifiedAgainstDocs: "2026-08-
|
|
264
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
265
265
|
provider: "anthropic",
|
|
266
266
|
status: "legacy",
|
|
267
267
|
maxContextTokens: 1e6,
|
|
@@ -306,7 +306,7 @@ var PROFILES_RAW = [
|
|
|
306
306
|
},
|
|
307
307
|
{
|
|
308
308
|
id: "claude-sonnet-4-6",
|
|
309
|
-
verifiedAgainstDocs: "2026-08-
|
|
309
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
310
310
|
provider: "anthropic",
|
|
311
311
|
status: "legacy",
|
|
312
312
|
// s76 2026-07-25: docs list as LEGACY (Sonnet 5 is current).
|
|
@@ -347,7 +347,7 @@ var PROFILES_RAW = [
|
|
|
347
347
|
},
|
|
348
348
|
{
|
|
349
349
|
id: "claude-haiku-4-5",
|
|
350
|
-
verifiedAgainstDocs: "2026-08-
|
|
350
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
351
351
|
provider: "anthropic",
|
|
352
352
|
status: "current",
|
|
353
353
|
maxContextTokens: 2e5,
|
|
@@ -410,7 +410,7 @@ var PROFILES_RAW = [
|
|
|
410
410
|
// ── Google ──
|
|
411
411
|
{
|
|
412
412
|
id: "gemini-2.5-flash",
|
|
413
|
-
verifiedAgainstDocs: "2026-08-
|
|
413
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
414
414
|
provider: "google",
|
|
415
415
|
status: "current",
|
|
416
416
|
maxContextTokens: 1048576,
|
|
@@ -516,7 +516,7 @@ var PROFILES_RAW = [
|
|
|
516
516
|
// thresholds. The brain will validate/relax these as evidence accumulates
|
|
517
517
|
// per (archetype, model) tuple. Currently ZERO brain rows for this model.
|
|
518
518
|
id: "gemini-2.5-flash-lite",
|
|
519
|
-
verifiedAgainstDocs: "2026-08-
|
|
519
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
520
520
|
provider: "google",
|
|
521
521
|
status: "current",
|
|
522
522
|
maxContextTokens: 1048576,
|
|
@@ -673,7 +673,7 @@ var PROFILES_RAW = [
|
|
|
673
673
|
},
|
|
674
674
|
{
|
|
675
675
|
id: "gemini-3.1-pro-preview",
|
|
676
|
-
verifiedAgainstDocs: "2026-08-
|
|
676
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
677
677
|
provider: "google",
|
|
678
678
|
status: "preview",
|
|
679
679
|
maxContextTokens: 1048576,
|
|
@@ -771,7 +771,7 @@ var PROFILES_RAW = [
|
|
|
771
771
|
// baseQuality (it carries 'reasoning') with no latency counterweight,
|
|
772
772
|
// leapfrogged gemini-2.5-flash as PB's summarize leader once reachable.
|
|
773
773
|
latencyTier: "slow",
|
|
774
|
-
notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87
|
|
774
|
+
notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87 was built + verified 2026-08-14 (prices then live), held at the publish gate 10 days, and published 2026-08-24 \u2014 by which time the 08-16 transition had killed them ($0.14/$0.28 flash, $0.435/$0.87 pro); corrected here (alpha.88). Build-then-hold is the mechanism: re-verify prices at publish when publish is not same-session as build. A single price field cannot represent a two-rate schedule; catalog-sync extractions oscillate between the rates day-to-day (peak read 08-21, off-peak read 08-22) \u2014 time-of-day pricing remains unmodeled, roadmap.",
|
|
775
775
|
// Master plan §6.2 anchor. Brain-validated tier 1 cross-provider for
|
|
776
776
|
// classify (169 rows, 0% empty). Tier 0 for summarize-with-no-tools.
|
|
777
777
|
// Falls off on hunt (sequential tools — L-040) and reasoning depth.
|
|
@@ -847,7 +847,7 @@ var PROFILES_RAW = [
|
|
|
847
847
|
// of the served set. Tag derivation would say 'medium'; the measurement says
|
|
848
848
|
// otherwise.
|
|
849
849
|
latencyTier: "slow",
|
|
850
|
-
notes: "Pro tier. 1M context, 384k max output. $0.435/$0.87 per 1M (cache-hit in $0.003625) \u2014 verified live 2026-06-22; the launch promo became the standing price (no revert). Default mode = thinking. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87
|
|
850
|
+
notes: "Pro tier. 1M context, 384k max output. $0.435/$0.87 per 1M (cache-hit in $0.003625) \u2014 verified live 2026-06-22; the launch promo became the standing price (no revert). Default mode = thinking. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87 was built + verified 2026-08-14 (prices then live), held at the publish gate 10 days, and published 2026-08-24 \u2014 by which time the 08-16 transition had killed them ($0.14/$0.28 flash, $0.435/$0.87 pro); corrected here (alpha.88). Build-then-hold is the mechanism: re-verify prices at publish when publish is not same-session as build. A single price field cannot represent a two-rate schedule; catalog-sync extractions oscillate between the rates day-to-day (peak read 08-21, off-peak read 08-22) \u2014 time-of-day pricing remains unmodeled, roadmap.",
|
|
851
851
|
// Master plan §3.3: tier 3 cross-provider for plan chain. Reasoning
|
|
852
852
|
// bumped one notch over V4-Flash; same parallel-tool ceiling.
|
|
853
853
|
archetypePerf: {
|
|
@@ -1538,7 +1538,7 @@ var PROFILES_RAW = [
|
|
|
1538
1538
|
],
|
|
1539
1539
|
strengths: ["agentic_loops", "coding", "1m_context", "parallel_tools", "thinking_mode"],
|
|
1540
1540
|
weaknesses: ["no_brain_evidence_yet"],
|
|
1541
|
-
notes: "Hand-onboarded s86 (2026-08-
|
|
1541
|
+
notes: "Hand-onboarded s86 (2026-08-14), step-zero verified: 1,048,576 ctx / 65,536 out (Google model page). PRICING re-verified 2026-08-24 at publish: $0.75/$3.75 promo through 2026-12-31, $1.50/$7.50 from 2027-01-01 \u2014 the alpha.87 profile carried the post-promo rate as if current; corrected alpha.88. Docs status Stable; enters kgauto as preview per standing rule (availability != lifecycle, promotion via brain evidence). Function calling + structured outputs + thinking per docs; parallel tools assumed per Flash family (docs silent). First flagged by the release watcher s73 (2026-07-24); ~21 days in the onboarding queue (onboarded 2026-08-14).",
|
|
1542
1542
|
archetypePerf: {
|
|
1543
1543
|
hunt: 9,
|
|
1544
1544
|
// Flash-family parallel-tool tier (L-040); inherited from 3.5-flash
|
|
@@ -1572,7 +1572,7 @@ var PROFILES_RAW = [
|
|
|
1572
1572
|
// 'preview' per the Fable precedent — no brain evidence yet; promotion to
|
|
1573
1573
|
// 'current' is an explicit call. (L-049/L-081.)
|
|
1574
1574
|
id: "claude-sonnet-5",
|
|
1575
|
-
verifiedAgainstDocs: "2026-08-
|
|
1575
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
1576
1576
|
provider: "anthropic",
|
|
1577
1577
|
status: "current",
|
|
1578
1578
|
// s76 2026-07-25: GA per docs; was preview.
|
|
@@ -1,6 +1,6 @@
|
|
|
1
|
-
import { G as GlassboxEvent } from '../types-
|
|
2
|
-
export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-
|
|
3
|
-
import '../ir-
|
|
1
|
+
import { G as GlassboxEvent } from '../types-cBzinzUR.mjs';
|
|
2
|
+
export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-cBzinzUR.mjs';
|
|
3
|
+
import '../ir-DeYMLWge.mjs';
|
|
4
4
|
import '../dialect.mjs';
|
|
5
5
|
|
|
6
6
|
/**
|
package/dist/glassbox/index.d.ts
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
|
-
import { G as GlassboxEvent } from '../types-
|
|
2
|
-
export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-
|
|
3
|
-
import '../ir-
|
|
1
|
+
import { G as GlassboxEvent } from '../types-BKbRtmUb.js';
|
|
2
|
+
export { A as AdvisoryFiredData, C as CompileDoneData, a as CompileStartData, E as ExecuteAttemptData, b as ExecuteSuccessData, F as FallbackWalkedData, c as GLASSBOX_STREAM_TTL_MS, d as GlassboxEventKind, e as GlassboxPubSub } from '../types-BKbRtmUb.js';
|
|
3
|
+
import '../ir-CTx026t0.js';
|
|
4
4
|
import '../dialect.js';
|
|
5
5
|
|
|
6
6
|
/**
|
|
@@ -1,7 +1,7 @@
|
|
|
1
|
-
import { G as GlassboxEvent } from '../types-
|
|
2
|
-
import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-
|
|
3
|
-
export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-
|
|
4
|
-
import '../ir-
|
|
1
|
+
import { G as GlassboxEvent } from '../types-cBzinzUR.mjs';
|
|
2
|
+
import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-DD36cCbZ.mjs';
|
|
3
|
+
export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-DD36cCbZ.mjs';
|
|
4
|
+
import '../ir-DeYMLWge.mjs';
|
|
5
5
|
import '../dialect.mjs';
|
|
6
6
|
|
|
7
7
|
/**
|
|
@@ -1,7 +1,7 @@
|
|
|
1
|
-
import { G as GlassboxEvent } from '../types-
|
|
2
|
-
import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-
|
|
3
|
-
export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-
|
|
4
|
-
import '../ir-
|
|
1
|
+
import { G as GlassboxEvent } from '../types-BKbRtmUb.js';
|
|
2
|
+
import { a as TraceDetail, b as TraceSummary, c as TraceCounterfactual } from '../types-Cp9ot1HV.js';
|
|
3
|
+
export { A as AdvisoryRecord, T as TraceHealth, d as TraceSectionRewrite } from '../types-Cp9ot1HV.js';
|
|
4
|
+
import '../ir-CTx026t0.js';
|
|
5
5
|
import '../dialect.js';
|
|
6
6
|
|
|
7
7
|
/**
|
|
@@ -265,7 +265,7 @@ var PROFILES_RAW = [
|
|
|
265
265
|
},
|
|
266
266
|
{
|
|
267
267
|
id: "claude-opus-4-7",
|
|
268
|
-
verifiedAgainstDocs: "2026-08-
|
|
268
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
269
269
|
provider: "anthropic",
|
|
270
270
|
status: "legacy",
|
|
271
271
|
// s76 2026-07-25: docs list as LEGACY (Opus 5 is current). Bundle said current while the brain already said legacy — the two disagreed.
|
|
@@ -315,7 +315,7 @@ var PROFILES_RAW = [
|
|
|
315
315
|
},
|
|
316
316
|
{
|
|
317
317
|
id: "claude-opus-4-6",
|
|
318
|
-
verifiedAgainstDocs: "2026-08-
|
|
318
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
319
319
|
provider: "anthropic",
|
|
320
320
|
status: "legacy",
|
|
321
321
|
maxContextTokens: 1e6,
|
|
@@ -360,7 +360,7 @@ var PROFILES_RAW = [
|
|
|
360
360
|
},
|
|
361
361
|
{
|
|
362
362
|
id: "claude-sonnet-4-6",
|
|
363
|
-
verifiedAgainstDocs: "2026-08-
|
|
363
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
364
364
|
provider: "anthropic",
|
|
365
365
|
status: "legacy",
|
|
366
366
|
// s76 2026-07-25: docs list as LEGACY (Sonnet 5 is current).
|
|
@@ -401,7 +401,7 @@ var PROFILES_RAW = [
|
|
|
401
401
|
},
|
|
402
402
|
{
|
|
403
403
|
id: "claude-haiku-4-5",
|
|
404
|
-
verifiedAgainstDocs: "2026-08-
|
|
404
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
405
405
|
provider: "anthropic",
|
|
406
406
|
status: "current",
|
|
407
407
|
maxContextTokens: 2e5,
|
|
@@ -464,7 +464,7 @@ var PROFILES_RAW = [
|
|
|
464
464
|
// ── Google ──
|
|
465
465
|
{
|
|
466
466
|
id: "gemini-2.5-flash",
|
|
467
|
-
verifiedAgainstDocs: "2026-08-
|
|
467
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
468
468
|
provider: "google",
|
|
469
469
|
status: "current",
|
|
470
470
|
maxContextTokens: 1048576,
|
|
@@ -570,7 +570,7 @@ var PROFILES_RAW = [
|
|
|
570
570
|
// thresholds. The brain will validate/relax these as evidence accumulates
|
|
571
571
|
// per (archetype, model) tuple. Currently ZERO brain rows for this model.
|
|
572
572
|
id: "gemini-2.5-flash-lite",
|
|
573
|
-
verifiedAgainstDocs: "2026-08-
|
|
573
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
574
574
|
provider: "google",
|
|
575
575
|
status: "current",
|
|
576
576
|
maxContextTokens: 1048576,
|
|
@@ -727,7 +727,7 @@ var PROFILES_RAW = [
|
|
|
727
727
|
},
|
|
728
728
|
{
|
|
729
729
|
id: "gemini-3.1-pro-preview",
|
|
730
|
-
verifiedAgainstDocs: "2026-08-
|
|
730
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
731
731
|
provider: "google",
|
|
732
732
|
status: "preview",
|
|
733
733
|
maxContextTokens: 1048576,
|
|
@@ -825,7 +825,7 @@ var PROFILES_RAW = [
|
|
|
825
825
|
// baseQuality (it carries 'reasoning') with no latency counterweight,
|
|
826
826
|
// leapfrogged gemini-2.5-flash as PB's summarize leader once reachable.
|
|
827
827
|
latencyTier: "slow",
|
|
828
|
-
notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87
|
|
828
|
+
notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87 was built + verified 2026-08-14 (prices then live), held at the publish gate 10 days, and published 2026-08-24 \u2014 by which time the 08-16 transition had killed them ($0.14/$0.28 flash, $0.435/$0.87 pro); corrected here (alpha.88). Build-then-hold is the mechanism: re-verify prices at publish when publish is not same-session as build. A single price field cannot represent a two-rate schedule; catalog-sync extractions oscillate between the rates day-to-day (peak read 08-21, off-peak read 08-22) \u2014 time-of-day pricing remains unmodeled, roadmap.",
|
|
829
829
|
// Master plan §6.2 anchor. Brain-validated tier 1 cross-provider for
|
|
830
830
|
// classify (169 rows, 0% empty). Tier 0 for summarize-with-no-tools.
|
|
831
831
|
// Falls off on hunt (sequential tools — L-040) and reasoning depth.
|
|
@@ -901,7 +901,7 @@ var PROFILES_RAW = [
|
|
|
901
901
|
// of the served set. Tag derivation would say 'medium'; the measurement says
|
|
902
902
|
// otherwise.
|
|
903
903
|
latencyTier: "slow",
|
|
904
|
-
notes: "Pro tier. 1M context, 384k max output. $0.435/$0.87 per 1M (cache-hit in $0.003625) \u2014 verified live 2026-06-22; the launch promo became the standing price (no revert). Default mode = thinking. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87
|
|
904
|
+
notes: "Pro tier. 1M context, 384k max output. $0.435/$0.87 per 1M (cache-hit in $0.003625) \u2014 verified live 2026-06-22; the launch promo became the standing price (no revert). Default mode = thinking. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87 was built + verified 2026-08-14 (prices then live), held at the publish gate 10 days, and published 2026-08-24 \u2014 by which time the 08-16 transition had killed them ($0.14/$0.28 flash, $0.435/$0.87 pro); corrected here (alpha.88). Build-then-hold is the mechanism: re-verify prices at publish when publish is not same-session as build. A single price field cannot represent a two-rate schedule; catalog-sync extractions oscillate between the rates day-to-day (peak read 08-21, off-peak read 08-22) \u2014 time-of-day pricing remains unmodeled, roadmap.",
|
|
905
905
|
// Master plan §3.3: tier 3 cross-provider for plan chain. Reasoning
|
|
906
906
|
// bumped one notch over V4-Flash; same parallel-tool ceiling.
|
|
907
907
|
archetypePerf: {
|
|
@@ -1592,7 +1592,7 @@ var PROFILES_RAW = [
|
|
|
1592
1592
|
],
|
|
1593
1593
|
strengths: ["agentic_loops", "coding", "1m_context", "parallel_tools", "thinking_mode"],
|
|
1594
1594
|
weaknesses: ["no_brain_evidence_yet"],
|
|
1595
|
-
notes: "Hand-onboarded s86 (2026-08-
|
|
1595
|
+
notes: "Hand-onboarded s86 (2026-08-14), step-zero verified: 1,048,576 ctx / 65,536 out (Google model page). PRICING re-verified 2026-08-24 at publish: $0.75/$3.75 promo through 2026-12-31, $1.50/$7.50 from 2027-01-01 \u2014 the alpha.87 profile carried the post-promo rate as if current; corrected alpha.88. Docs status Stable; enters kgauto as preview per standing rule (availability != lifecycle, promotion via brain evidence). Function calling + structured outputs + thinking per docs; parallel tools assumed per Flash family (docs silent). First flagged by the release watcher s73 (2026-07-24); ~21 days in the onboarding queue (onboarded 2026-08-14).",
|
|
1596
1596
|
archetypePerf: {
|
|
1597
1597
|
hunt: 9,
|
|
1598
1598
|
// Flash-family parallel-tool tier (L-040); inherited from 3.5-flash
|
|
@@ -1626,7 +1626,7 @@ var PROFILES_RAW = [
|
|
|
1626
1626
|
// 'preview' per the Fable precedent — no brain evidence yet; promotion to
|
|
1627
1627
|
// 'current' is an explicit call. (L-049/L-081.)
|
|
1628
1628
|
id: "claude-sonnet-5",
|
|
1629
|
-
verifiedAgainstDocs: "2026-08-
|
|
1629
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
1630
1630
|
provider: "anthropic",
|
|
1631
1631
|
status: "current",
|
|
1632
1632
|
// s76 2026-07-25: GA per docs; was preview.
|
|
@@ -1,10 +1,10 @@
|
|
|
1
1
|
import {
|
|
2
2
|
ARCHETYPE_FLOOR_DEFAULT,
|
|
3
3
|
getDefaultFallbackChain
|
|
4
|
-
} from "../chunk-
|
|
4
|
+
} from "../chunk-D7RX3UD3.mjs";
|
|
5
5
|
import {
|
|
6
6
|
tryGetProfile
|
|
7
|
-
} from "../chunk-
|
|
7
|
+
} from "../chunk-OK2TMFRR.mjs";
|
|
8
8
|
import {
|
|
9
9
|
subscribe,
|
|
10
10
|
subscribeApp
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
import * as react_jsx_runtime from 'react/jsx-runtime';
|
|
2
|
-
import { a as TraceDetail } from '../../types-
|
|
3
|
-
import '../../ir-
|
|
2
|
+
import { a as TraceDetail } from '../../types-DD36cCbZ.mjs';
|
|
3
|
+
import '../../ir-DeYMLWge.mjs';
|
|
4
4
|
import '../../dialect.mjs';
|
|
5
5
|
|
|
6
6
|
/**
|
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
import * as react_jsx_runtime from 'react/jsx-runtime';
|
|
2
|
-
import { a as TraceDetail } from '../../types-
|
|
3
|
-
import '../../ir-
|
|
2
|
+
import { a as TraceDetail } from '../../types-Cp9ot1HV.js';
|
|
3
|
+
import '../../ir-CTx026t0.js';
|
|
4
4
|
import '../../dialect.js';
|
|
5
5
|
|
|
6
6
|
/**
|
package/dist/index.d.mts
CHANGED
|
@@ -1,5 +1,5 @@
|
|
|
1
|
-
import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, F as FallbackReason, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OutcomeSource, i as OracleScore, j as Adapter, k as PerAxisMetrics, l as Provider, m as ChainEntry, G as Grounding } from './ir-
|
|
2
|
-
export { n as CallAttempt, o as CallError, p as ChainModelEntry, q as ChainWithGrounding, r as Constraints, E as EffortLevel, s as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, t as MutationApplied, u as NormalizedTokens, v as OutcomeKind, w as PerAxisMetricsByModel, x as PromptSection, y as SectionKind, z as ShadowProbeConfig, T as ToolCall, D as ToolDefinition, J as captureGoldenIr, K as hasMutation, L as mutationId, Q as parseGoldenCaptureRate, U as resolveGoldenCaptureRate, V as shouldCaptureGolden } from './ir-
|
|
1
|
+
import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, F as FallbackReason, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OutcomeSource, i as OracleScore, j as Adapter, k as PerAxisMetrics, l as Provider, m as ChainEntry, G as Grounding } from './ir-DeYMLWge.mjs';
|
|
2
|
+
export { n as CallAttempt, o as CallError, p as ChainModelEntry, q as ChainWithGrounding, r as Constraints, E as EffortLevel, s as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, t as MutationApplied, u as NormalizedTokens, v as OutcomeKind, w as PerAxisMetricsByModel, x as PromptSection, y as SectionKind, z as ShadowProbeConfig, T as ToolCall, D as ToolDefinition, J as captureGoldenIr, K as hasMutation, L as mutationId, Q as parseGoldenCaptureRate, U as resolveGoldenCaptureRate, V as shouldCaptureGolden } from './ir-DeYMLWge.mjs';
|
|
3
3
|
import { ModelProfile, ArchetypeConvention } from './profiles.mjs';
|
|
4
4
|
export { ALIASES, CacheStrategy, CliffRule, LATENCY_TIER_MS, LatencyTier, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, _testClearConsumerProfiles, allProfiles, bestEffortProfile, getProfile, inferProviderFromId, latencyTierOf, profilesByProvider, registerProfiles, resolveModelAlias, tryGetProfile } from './profiles.mjs';
|
|
5
5
|
export { BrainForwardConfig, BrainForwardRoutes, createBrainForwardRoutes } from './brain-proxy.mjs';
|
|
@@ -679,6 +679,11 @@ interface OutcomePayload {
|
|
|
679
679
|
* organic; liveness rules compute over `source IS NULL` rows.
|
|
680
680
|
*/
|
|
681
681
|
source?: string;
|
|
682
|
+
/**
|
|
683
|
+
* alpha.89 (migration 065) — per-end-user metering: the consumer's opaque
|
|
684
|
+
* customer id, truncated to 128 chars at write. NULL = no end user.
|
|
685
|
+
*/
|
|
686
|
+
end_user_id?: string;
|
|
682
687
|
error_type?: string;
|
|
683
688
|
tools_called?: string[];
|
|
684
689
|
oracle_score?: number;
|
|
@@ -950,6 +955,21 @@ declare function recordGoldenIr(input: GoldenIrRecordInput): Promise<void>;
|
|
|
950
955
|
* pollute compile_outcomes.
|
|
951
956
|
*/
|
|
952
957
|
|
|
958
|
+
/**
|
|
959
|
+
* Default judge model (judge chain tier 0). Exported so operator scripts
|
|
960
|
+
* (golden-eval-trigger's judge pre-flight) probe the SAME model the engine
|
|
961
|
+
* will call — two inline literals of one concept is the s75 drift shape.
|
|
962
|
+
* 2026-08-25: the judge had no pre-flight at all, so an Anthropic billing
|
|
963
|
+
* outage would have burned incumbent+candidate replay spend on non-Anthropic
|
|
964
|
+
* surfaces before dying verdictless at judge time.
|
|
965
|
+
*
|
|
966
|
+
* 2026-08-25 (same session): moved 'claude-opus-4-8' → 'claude-opus-5'. The
|
|
967
|
+
* guard test caught the old default sitting on a status:'legacy' roster
|
|
968
|
+
* entry (4-8 went legacy 2026-07-25 with the Opus 5 generation; nobody moved
|
|
969
|
+
* the judge). Verdict comparability across the switch is preserved by
|
|
970
|
+
* `judge_model` on every run + case row (migrations 024/036).
|
|
971
|
+
*/
|
|
972
|
+
declare const DEFAULT_JUDGE_MODEL = "claude-opus-5";
|
|
953
973
|
/**
|
|
954
974
|
* What "better" means per archetype. Fed verbatim into the pairwise judge
|
|
955
975
|
* prompt. Kept deliberately short — the judge sees the full request, so the
|
|
@@ -1020,7 +1040,7 @@ interface GoldenEvalOptions {
|
|
|
1020
1040
|
* across the loaded golden set (the surface's de-facto leader).
|
|
1021
1041
|
*/
|
|
1022
1042
|
incumbentModel?: string;
|
|
1023
|
-
/** Judge model id. Default
|
|
1043
|
+
/** Judge model id. Default `DEFAULT_JUDGE_MODEL` (judge chain tier 0). */
|
|
1024
1044
|
judgeModel?: string;
|
|
1025
1045
|
/** Max golden cases to load (newest first). Default 50. */
|
|
1026
1046
|
limit?: number;
|
|
@@ -1250,7 +1270,7 @@ declare function runStrategyEvalWithAttribution(opts: Omit<GoldenEvalOptions, 'a
|
|
|
1250
1270
|
* guard in `tests/version.test.ts` fails the suite (and therefore
|
|
1251
1271
|
* `prepublishOnly`) when they diverge — a stale constant cannot reach npm.
|
|
1252
1272
|
*/
|
|
1253
|
-
declare const LIBRARY_VERSION = "2.0.0-alpha.
|
|
1273
|
+
declare const LIBRARY_VERSION = "2.0.0-alpha.89";
|
|
1254
1274
|
|
|
1255
1275
|
/**
|
|
1256
1276
|
* Oracle contract — how an app tells the brain whether a response was good.
|
|
@@ -4070,4 +4090,4 @@ declare function planDecomposition(args: PlanDecompositionArgs): DecompositionPl
|
|
|
4070
4090
|
*/
|
|
4071
4091
|
declare function compile(ir: PromptIR, opts?: CompileOptions): CompileResult;
|
|
4072
4092
|
|
|
4073
|
-
export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ALT_BLIND_TOKEN_BUDGET_BREACH, ALT_STRATEGY_IDS, ARCHETYPE_FAMILY_FITS, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, type AdvisoryEvidenceWindow, type AdvisoryResolutionSource, type AdvisorySeverity, type AdvisoryStatus, type AdvisorySuggestedFix, ApiKeys, type AppOracle, type ApplySectionRewritesArgs, type ApplySectionRewritesResult, ArchetypeConvention, type ArchetypeFamilyFit, type ArchetypePerfMap, type ArchetypePerfNMap, type ArchetypePerfScoreResult, type AttachCacheControlResult, BLOCKED_MODEL_FAMILY_SIBLING_SERVED_CODE, BLOCKED_MODEL_NOT_IN_ROSTER_CODE, BRAIN_READ_ENV_NAMES, BURST_SPAN_MS, BestPracticeAdvisory, type BrainConfig, type BrainDeadLetterEntry, type BrainHealthSnapshot, type BrainQueryConfig, type BrainReadEnv, COACH_CFG, COST_RANKING_REFERENCE_SHAPE, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileForAISDKv6Result, type CompileOptions, CompilePolicy, CompileResult, CompiledRequest, type CreateDelegateOpts, DECOMPOSITION_TEMPLATES, DECOMPOSITION_TEMPLATES_VERSION, DEFAULT_FINDINGS_ENDPOINT, DEFAULT_MEASURED_FAILURE_ENDPOINT, DEFAULT_PROMOTIONS_ENDPOINT, DELEGATE_TOOL_DEFINITION, DISCIPLINE_GATES_V1_ALT_BLIND_HEADER, DISCIPLINE_GATES_V1_ALT_HEADER, type DecompositionPlan, type DecompositionStep, type DecompositionTemplate, type DelegateHandle, type DelegateRefusalReason, type DelegateResult, type DelegateToolArgs, type ExclusionFindingRow, type ExclusionResolutionSource, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type ExecutorCandidate, type FallbackPosture, FallbackReason, FamilyResolutionError, type GetActionableAdvisoriesOptions, type GetApplicablePromotionOpts, type GetDefaultFallbackChainOpts, type GetMeasuredFailureOpts, type GetPerAxisMetricsOpts, type GetRecommendedPrimaryOptions, type GoldenEvalAxis, type GoldenEvalCase, type GoldenEvalOptions, type GoldenEvalRunResult, type GoldenEvalStrategyId, type GoldenIrRecordInput, Grounding, IntentArchetypeName, JUDGE_RUBRICS, LIBRARY_VERSION, type LLMJudgeOptions, MEASURED_FAILURE_CFG, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type MarkExclusionFindingHandledOptions, type MarkPromoteReadyHandledOptions, type MeasuredFailureRuntime, type MeasuredFailureVerdict, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, OutputMode, PRODUCER_OWNED_RULE_CODES, PROVIDER_ENV_KEYS, PerAxisMetrics, type PlannedStep, type PricingRow, type ProbeShadowOptions, type ProbeShadowServed, type ProfileToRowOptions, type PromoteReadyFindingRow, type PromoteReadyResolution, type PromotionRow, type PromotionsRuntime, PromptIR, Provider, ProviderOverrides, type ProviderReachability, ROLLBACK_SUPPRESSION_WINDOW_DAYS, RULE_DISCIPLINE_GATES_V1, RULE_DISCIPLINE_GATES_V1_STRUCTURED, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, STRATEGY_AUTHORSHIP_INDEPENDENT, STRATEGY_AUTHORSHIP_LIMITATION, SectionRewrite, type ShadowProbeRecordInput, type StrategyAttribution, type StrategyAttributionResult, type StrategyOutcome, type SupportedProvider, type SurfaceFailureRow, type SurfaceStats, SystemModelMessage, TRANSLATOR_FLOOR, _testResetMeasuredFailure, _testResetOutcomeSourceWarning, _testResetPromotions, _testWaitForMeasuredFailureRefresh, _testWaitForPromotionsRefresh, altBlindGatesBlockFor, altGatesBlockFor, applyArchetypeConvention, applySectionRewrites, attachCacheControlToStreamTextInput, awaitMeasuredFailureReady, awaitPromotionsReady, brainHealth, buildGoldenIrRow, buildLLMJudge, buildPairwiseJudgePrompt, buildShadowProbeRow, burstCaveat, call, chainProviderSpread, classifyEvidenceWindow, classifyStrategyOutcome, clearBrain, combineOrderSwappedVerdicts, compile, compileForAISDKv6, configureBrain, configureMeasuredFailureBrain, configurePromotionsBrain, countTokens, createDelegate, deriveFamilyFromModelId, deriveOwnership, estimateChainCostUsd, estimateModelCostUsd, execute, findBetterFit, flushBrainDeadLetter, formatEvidenceSpan, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getApplicablePromotion, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getMeasuredFailureVerdict, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getRecentRollback, getRecommendedPrimary, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStaleExclusionFindings, getStarterChain, getStarterChainWithGrounding, isAltStrategy, isAutoPromoteEnabledFromEnv, isBrainQueryActiveFor, isBrainSync, isDelegateEnabledFromEnv, isExclusionFindingsBrainActive, isMeasuredFailureBrainActive, isMeasuredFailureGateEnabledFromEnv, isModelReachable, isPromotionsBrainActive, isProviderReachable, judgeMeasuredFailure, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, mapMeasuredFailureRows, markAdvisoryResolved, markExclusionFindingHandled, markPromoteReadyHandled, parseJudgeVerdict, peekBrainDeadLetter, planDecomposition, prefetchMeasuredFailure, prefetchPromotions, probeShadow, profileToRow, readBrainReadEnv, record, recordGoldenIr, recordOutcome, recordShadowProbe, renderIrForJudge, resetTokenizer, resolveConventionsForProfile, resolveOutcomeSource, resolvePricingAt, resolveProviderKey, rowToAdvisory, rubricFor, runAdvisor, runGoldenEval, runStrategyEvalWithAttribution, setTokenizer, wilsonLowerBound, withAltBlindDisciplineContract, withAltDisciplineContract, withDisciplineContract };
|
|
4093
|
+
export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ALT_BLIND_TOKEN_BUDGET_BREACH, ALT_STRATEGY_IDS, ARCHETYPE_FAMILY_FITS, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, type AdvisoryEvidenceWindow, type AdvisoryResolutionSource, type AdvisorySeverity, type AdvisoryStatus, type AdvisorySuggestedFix, ApiKeys, type AppOracle, type ApplySectionRewritesArgs, type ApplySectionRewritesResult, ArchetypeConvention, type ArchetypeFamilyFit, type ArchetypePerfMap, type ArchetypePerfNMap, type ArchetypePerfScoreResult, type AttachCacheControlResult, BLOCKED_MODEL_FAMILY_SIBLING_SERVED_CODE, BLOCKED_MODEL_NOT_IN_ROSTER_CODE, BRAIN_READ_ENV_NAMES, BURST_SPAN_MS, BestPracticeAdvisory, type BrainConfig, type BrainDeadLetterEntry, type BrainHealthSnapshot, type BrainQueryConfig, type BrainReadEnv, COACH_CFG, COST_RANKING_REFERENCE_SHAPE, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileForAISDKv6Result, type CompileOptions, CompilePolicy, CompileResult, CompiledRequest, type CreateDelegateOpts, DECOMPOSITION_TEMPLATES, DECOMPOSITION_TEMPLATES_VERSION, DEFAULT_FINDINGS_ENDPOINT, DEFAULT_JUDGE_MODEL, DEFAULT_MEASURED_FAILURE_ENDPOINT, DEFAULT_PROMOTIONS_ENDPOINT, DELEGATE_TOOL_DEFINITION, DISCIPLINE_GATES_V1_ALT_BLIND_HEADER, DISCIPLINE_GATES_V1_ALT_HEADER, type DecompositionPlan, type DecompositionStep, type DecompositionTemplate, type DelegateHandle, type DelegateRefusalReason, type DelegateResult, type DelegateToolArgs, type ExclusionFindingRow, type ExclusionResolutionSource, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type ExecutorCandidate, type FallbackPosture, FallbackReason, FamilyResolutionError, type GetActionableAdvisoriesOptions, type GetApplicablePromotionOpts, type GetDefaultFallbackChainOpts, type GetMeasuredFailureOpts, type GetPerAxisMetricsOpts, type GetRecommendedPrimaryOptions, type GoldenEvalAxis, type GoldenEvalCase, type GoldenEvalOptions, type GoldenEvalRunResult, type GoldenEvalStrategyId, type GoldenIrRecordInput, Grounding, IntentArchetypeName, JUDGE_RUBRICS, LIBRARY_VERSION, type LLMJudgeOptions, MEASURED_FAILURE_CFG, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type MarkExclusionFindingHandledOptions, type MarkPromoteReadyHandledOptions, type MeasuredFailureRuntime, type MeasuredFailureVerdict, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, OutputMode, PRODUCER_OWNED_RULE_CODES, PROVIDER_ENV_KEYS, PerAxisMetrics, type PlannedStep, type PricingRow, type ProbeShadowOptions, type ProbeShadowServed, type ProfileToRowOptions, type PromoteReadyFindingRow, type PromoteReadyResolution, type PromotionRow, type PromotionsRuntime, PromptIR, Provider, ProviderOverrides, type ProviderReachability, ROLLBACK_SUPPRESSION_WINDOW_DAYS, RULE_DISCIPLINE_GATES_V1, RULE_DISCIPLINE_GATES_V1_STRUCTURED, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, STRATEGY_AUTHORSHIP_INDEPENDENT, STRATEGY_AUTHORSHIP_LIMITATION, SectionRewrite, type ShadowProbeRecordInput, type StrategyAttribution, type StrategyAttributionResult, type StrategyOutcome, type SupportedProvider, type SurfaceFailureRow, type SurfaceStats, SystemModelMessage, TRANSLATOR_FLOOR, _testResetMeasuredFailure, _testResetOutcomeSourceWarning, _testResetPromotions, _testWaitForMeasuredFailureRefresh, _testWaitForPromotionsRefresh, altBlindGatesBlockFor, altGatesBlockFor, applyArchetypeConvention, applySectionRewrites, attachCacheControlToStreamTextInput, awaitMeasuredFailureReady, awaitPromotionsReady, brainHealth, buildGoldenIrRow, buildLLMJudge, buildPairwiseJudgePrompt, buildShadowProbeRow, burstCaveat, call, chainProviderSpread, classifyEvidenceWindow, classifyStrategyOutcome, clearBrain, combineOrderSwappedVerdicts, compile, compileForAISDKv6, configureBrain, configureMeasuredFailureBrain, configurePromotionsBrain, countTokens, createDelegate, deriveFamilyFromModelId, deriveOwnership, estimateChainCostUsd, estimateModelCostUsd, execute, findBetterFit, flushBrainDeadLetter, formatEvidenceSpan, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getApplicablePromotion, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getMeasuredFailureVerdict, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getRecentRollback, getRecommendedPrimary, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStaleExclusionFindings, getStarterChain, getStarterChainWithGrounding, isAltStrategy, isAutoPromoteEnabledFromEnv, isBrainQueryActiveFor, isBrainSync, isDelegateEnabledFromEnv, isExclusionFindingsBrainActive, isMeasuredFailureBrainActive, isMeasuredFailureGateEnabledFromEnv, isModelReachable, isPromotionsBrainActive, isProviderReachable, judgeMeasuredFailure, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, mapMeasuredFailureRows, markAdvisoryResolved, markExclusionFindingHandled, markPromoteReadyHandled, parseJudgeVerdict, peekBrainDeadLetter, planDecomposition, prefetchMeasuredFailure, prefetchPromotions, probeShadow, profileToRow, readBrainReadEnv, record, recordGoldenIr, recordOutcome, recordShadowProbe, renderIrForJudge, resetTokenizer, resolveConventionsForProfile, resolveOutcomeSource, resolvePricingAt, resolveProviderKey, rowToAdvisory, rubricFor, runAdvisor, runGoldenEval, runStrategyEvalWithAttribution, setTokenizer, wilsonLowerBound, withAltBlindDisciplineContract, withAltDisciplineContract, withDisciplineContract };
|
package/dist/index.d.ts
CHANGED
|
@@ -1,5 +1,5 @@
|
|
|
1
|
-
import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, F as FallbackReason, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OutcomeSource, i as OracleScore, j as Adapter, k as PerAxisMetrics, l as Provider, m as ChainEntry, G as Grounding } from './ir-
|
|
2
|
-
export { n as CallAttempt, o as CallError, p as ChainModelEntry, q as ChainWithGrounding, r as Constraints, E as EffortLevel, s as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, t as MutationApplied, u as NormalizedTokens, v as OutcomeKind, w as PerAxisMetricsByModel, x as PromptSection, y as SectionKind, z as ShadowProbeConfig, T as ToolCall, D as ToolDefinition, J as captureGoldenIr, K as hasMutation, L as mutationId, Q as parseGoldenCaptureRate, U as resolveGoldenCaptureRate, V as shouldCaptureGolden } from './ir-
|
|
1
|
+
import { C as CompilePolicy, N as NormalizedResponse, A as ApiKeys, P as ProviderOverrides, a as CompiledRequest, b as PromptIR, c as CallOptions, d as CallResult, S as SystemModelMessage, e as CompileResult, B as BestPracticeAdvisory, F as FallbackReason, f as SectionRewrite, R as RecordInput, g as RecordOutcomeInput, O as OutcomeResult, h as OutcomeSource, i as OracleScore, j as Adapter, k as PerAxisMetrics, l as Provider, m as ChainEntry, G as Grounding } from './ir-CTx026t0.js';
|
|
2
|
+
export { n as CallAttempt, o as CallError, p as ChainModelEntry, q as ChainWithGrounding, r as Constraints, E as EffortLevel, s as GoldenCaptureOptions, H as HistoryCachePolicy, I as IntentDeclaration, M as Message, t as MutationApplied, u as NormalizedTokens, v as OutcomeKind, w as PerAxisMetricsByModel, x as PromptSection, y as SectionKind, z as ShadowProbeConfig, T as ToolCall, D as ToolDefinition, J as captureGoldenIr, K as hasMutation, L as mutationId, Q as parseGoldenCaptureRate, U as resolveGoldenCaptureRate, V as shouldCaptureGolden } from './ir-CTx026t0.js';
|
|
3
3
|
import { ModelProfile, ArchetypeConvention } from './profiles.js';
|
|
4
4
|
export { ALIASES, CacheStrategy, CliffRule, LATENCY_TIER_MS, LatencyTier, LoweringSpec, RecoveryRule, StructuredOutputCapability, SystemPromptMode, _testClearConsumerProfiles, allProfiles, bestEffortProfile, getProfile, inferProviderFromId, latencyTierOf, profilesByProvider, registerProfiles, resolveModelAlias, tryGetProfile } from './profiles.js';
|
|
5
5
|
export { BrainForwardConfig, BrainForwardRoutes, createBrainForwardRoutes } from './brain-proxy.js';
|
|
@@ -679,6 +679,11 @@ interface OutcomePayload {
|
|
|
679
679
|
* organic; liveness rules compute over `source IS NULL` rows.
|
|
680
680
|
*/
|
|
681
681
|
source?: string;
|
|
682
|
+
/**
|
|
683
|
+
* alpha.89 (migration 065) — per-end-user metering: the consumer's opaque
|
|
684
|
+
* customer id, truncated to 128 chars at write. NULL = no end user.
|
|
685
|
+
*/
|
|
686
|
+
end_user_id?: string;
|
|
682
687
|
error_type?: string;
|
|
683
688
|
tools_called?: string[];
|
|
684
689
|
oracle_score?: number;
|
|
@@ -950,6 +955,21 @@ declare function recordGoldenIr(input: GoldenIrRecordInput): Promise<void>;
|
|
|
950
955
|
* pollute compile_outcomes.
|
|
951
956
|
*/
|
|
952
957
|
|
|
958
|
+
/**
|
|
959
|
+
* Default judge model (judge chain tier 0). Exported so operator scripts
|
|
960
|
+
* (golden-eval-trigger's judge pre-flight) probe the SAME model the engine
|
|
961
|
+
* will call — two inline literals of one concept is the s75 drift shape.
|
|
962
|
+
* 2026-08-25: the judge had no pre-flight at all, so an Anthropic billing
|
|
963
|
+
* outage would have burned incumbent+candidate replay spend on non-Anthropic
|
|
964
|
+
* surfaces before dying verdictless at judge time.
|
|
965
|
+
*
|
|
966
|
+
* 2026-08-25 (same session): moved 'claude-opus-4-8' → 'claude-opus-5'. The
|
|
967
|
+
* guard test caught the old default sitting on a status:'legacy' roster
|
|
968
|
+
* entry (4-8 went legacy 2026-07-25 with the Opus 5 generation; nobody moved
|
|
969
|
+
* the judge). Verdict comparability across the switch is preserved by
|
|
970
|
+
* `judge_model` on every run + case row (migrations 024/036).
|
|
971
|
+
*/
|
|
972
|
+
declare const DEFAULT_JUDGE_MODEL = "claude-opus-5";
|
|
953
973
|
/**
|
|
954
974
|
* What "better" means per archetype. Fed verbatim into the pairwise judge
|
|
955
975
|
* prompt. Kept deliberately short — the judge sees the full request, so the
|
|
@@ -1020,7 +1040,7 @@ interface GoldenEvalOptions {
|
|
|
1020
1040
|
* across the loaded golden set (the surface's de-facto leader).
|
|
1021
1041
|
*/
|
|
1022
1042
|
incumbentModel?: string;
|
|
1023
|
-
/** Judge model id. Default
|
|
1043
|
+
/** Judge model id. Default `DEFAULT_JUDGE_MODEL` (judge chain tier 0). */
|
|
1024
1044
|
judgeModel?: string;
|
|
1025
1045
|
/** Max golden cases to load (newest first). Default 50. */
|
|
1026
1046
|
limit?: number;
|
|
@@ -1250,7 +1270,7 @@ declare function runStrategyEvalWithAttribution(opts: Omit<GoldenEvalOptions, 'a
|
|
|
1250
1270
|
* guard in `tests/version.test.ts` fails the suite (and therefore
|
|
1251
1271
|
* `prepublishOnly`) when they diverge — a stale constant cannot reach npm.
|
|
1252
1272
|
*/
|
|
1253
|
-
declare const LIBRARY_VERSION = "2.0.0-alpha.
|
|
1273
|
+
declare const LIBRARY_VERSION = "2.0.0-alpha.89";
|
|
1254
1274
|
|
|
1255
1275
|
/**
|
|
1256
1276
|
* Oracle contract — how an app tells the brain whether a response was good.
|
|
@@ -4070,4 +4090,4 @@ declare function planDecomposition(args: PlanDecompositionArgs): DecompositionPl
|
|
|
4070
4090
|
*/
|
|
4071
4091
|
declare function compile(ir: PromptIR, opts?: CompileOptions): CompileResult;
|
|
4072
4092
|
|
|
4073
|
-
export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ALT_BLIND_TOKEN_BUDGET_BREACH, ALT_STRATEGY_IDS, ARCHETYPE_FAMILY_FITS, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, type AdvisoryEvidenceWindow, type AdvisoryResolutionSource, type AdvisorySeverity, type AdvisoryStatus, type AdvisorySuggestedFix, ApiKeys, type AppOracle, type ApplySectionRewritesArgs, type ApplySectionRewritesResult, ArchetypeConvention, type ArchetypeFamilyFit, type ArchetypePerfMap, type ArchetypePerfNMap, type ArchetypePerfScoreResult, type AttachCacheControlResult, BLOCKED_MODEL_FAMILY_SIBLING_SERVED_CODE, BLOCKED_MODEL_NOT_IN_ROSTER_CODE, BRAIN_READ_ENV_NAMES, BURST_SPAN_MS, BestPracticeAdvisory, type BrainConfig, type BrainDeadLetterEntry, type BrainHealthSnapshot, type BrainQueryConfig, type BrainReadEnv, COACH_CFG, COST_RANKING_REFERENCE_SHAPE, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileForAISDKv6Result, type CompileOptions, CompilePolicy, CompileResult, CompiledRequest, type CreateDelegateOpts, DECOMPOSITION_TEMPLATES, DECOMPOSITION_TEMPLATES_VERSION, DEFAULT_FINDINGS_ENDPOINT, DEFAULT_MEASURED_FAILURE_ENDPOINT, DEFAULT_PROMOTIONS_ENDPOINT, DELEGATE_TOOL_DEFINITION, DISCIPLINE_GATES_V1_ALT_BLIND_HEADER, DISCIPLINE_GATES_V1_ALT_HEADER, type DecompositionPlan, type DecompositionStep, type DecompositionTemplate, type DelegateHandle, type DelegateRefusalReason, type DelegateResult, type DelegateToolArgs, type ExclusionFindingRow, type ExclusionResolutionSource, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type ExecutorCandidate, type FallbackPosture, FallbackReason, FamilyResolutionError, type GetActionableAdvisoriesOptions, type GetApplicablePromotionOpts, type GetDefaultFallbackChainOpts, type GetMeasuredFailureOpts, type GetPerAxisMetricsOpts, type GetRecommendedPrimaryOptions, type GoldenEvalAxis, type GoldenEvalCase, type GoldenEvalOptions, type GoldenEvalRunResult, type GoldenEvalStrategyId, type GoldenIrRecordInput, Grounding, IntentArchetypeName, JUDGE_RUBRICS, LIBRARY_VERSION, type LLMJudgeOptions, MEASURED_FAILURE_CFG, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type MarkExclusionFindingHandledOptions, type MarkPromoteReadyHandledOptions, type MeasuredFailureRuntime, type MeasuredFailureVerdict, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, OutputMode, PRODUCER_OWNED_RULE_CODES, PROVIDER_ENV_KEYS, PerAxisMetrics, type PlannedStep, type PricingRow, type ProbeShadowOptions, type ProbeShadowServed, type ProfileToRowOptions, type PromoteReadyFindingRow, type PromoteReadyResolution, type PromotionRow, type PromotionsRuntime, PromptIR, Provider, ProviderOverrides, type ProviderReachability, ROLLBACK_SUPPRESSION_WINDOW_DAYS, RULE_DISCIPLINE_GATES_V1, RULE_DISCIPLINE_GATES_V1_STRUCTURED, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, STRATEGY_AUTHORSHIP_INDEPENDENT, STRATEGY_AUTHORSHIP_LIMITATION, SectionRewrite, type ShadowProbeRecordInput, type StrategyAttribution, type StrategyAttributionResult, type StrategyOutcome, type SupportedProvider, type SurfaceFailureRow, type SurfaceStats, SystemModelMessage, TRANSLATOR_FLOOR, _testResetMeasuredFailure, _testResetOutcomeSourceWarning, _testResetPromotions, _testWaitForMeasuredFailureRefresh, _testWaitForPromotionsRefresh, altBlindGatesBlockFor, altGatesBlockFor, applyArchetypeConvention, applySectionRewrites, attachCacheControlToStreamTextInput, awaitMeasuredFailureReady, awaitPromotionsReady, brainHealth, buildGoldenIrRow, buildLLMJudge, buildPairwiseJudgePrompt, buildShadowProbeRow, burstCaveat, call, chainProviderSpread, classifyEvidenceWindow, classifyStrategyOutcome, clearBrain, combineOrderSwappedVerdicts, compile, compileForAISDKv6, configureBrain, configureMeasuredFailureBrain, configurePromotionsBrain, countTokens, createDelegate, deriveFamilyFromModelId, deriveOwnership, estimateChainCostUsd, estimateModelCostUsd, execute, findBetterFit, flushBrainDeadLetter, formatEvidenceSpan, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getApplicablePromotion, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getMeasuredFailureVerdict, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getRecentRollback, getRecommendedPrimary, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStaleExclusionFindings, getStarterChain, getStarterChainWithGrounding, isAltStrategy, isAutoPromoteEnabledFromEnv, isBrainQueryActiveFor, isBrainSync, isDelegateEnabledFromEnv, isExclusionFindingsBrainActive, isMeasuredFailureBrainActive, isMeasuredFailureGateEnabledFromEnv, isModelReachable, isPromotionsBrainActive, isProviderReachable, judgeMeasuredFailure, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, mapMeasuredFailureRows, markAdvisoryResolved, markExclusionFindingHandled, markPromoteReadyHandled, parseJudgeVerdict, peekBrainDeadLetter, planDecomposition, prefetchMeasuredFailure, prefetchPromotions, probeShadow, profileToRow, readBrainReadEnv, record, recordGoldenIr, recordOutcome, recordShadowProbe, renderIrForJudge, resetTokenizer, resolveConventionsForProfile, resolveOutcomeSource, resolvePricingAt, resolveProviderKey, rowToAdvisory, rubricFor, runAdvisor, runGoldenEval, runStrategyEvalWithAttribution, setTokenizer, wilsonLowerBound, withAltBlindDisciplineContract, withAltDisciplineContract, withDisciplineContract };
|
|
4093
|
+
export { ABSOLUTE_FLOOR, type AISDKConvertedMessage, ALT_BLIND_TOKEN_BUDGET_BREACH, ALT_STRATEGY_IDS, ARCHETYPE_FAMILY_FITS, ARCHETYPE_FLOOR_DEFAULT, type ActionableAdvisory, Adapter, type AdvisoryEvidenceWindow, type AdvisoryResolutionSource, type AdvisorySeverity, type AdvisoryStatus, type AdvisorySuggestedFix, ApiKeys, type AppOracle, type ApplySectionRewritesArgs, type ApplySectionRewritesResult, ArchetypeConvention, type ArchetypeFamilyFit, type ArchetypePerfMap, type ArchetypePerfNMap, type ArchetypePerfScoreResult, type AttachCacheControlResult, BLOCKED_MODEL_FAMILY_SIBLING_SERVED_CODE, BLOCKED_MODEL_NOT_IN_ROSTER_CODE, BRAIN_READ_ENV_NAMES, BURST_SPAN_MS, BestPracticeAdvisory, type BrainConfig, type BrainDeadLetterEntry, type BrainHealthSnapshot, type BrainQueryConfig, type BrainReadEnv, COACH_CFG, COST_RANKING_REFERENCE_SHAPE, CallOptions, CallResult, ChainEntry, type CompatibilityIntent, type CompileForAISDKv6Result, type CompileOptions, CompilePolicy, CompileResult, CompiledRequest, type CreateDelegateOpts, DECOMPOSITION_TEMPLATES, DECOMPOSITION_TEMPLATES_VERSION, DEFAULT_FINDINGS_ENDPOINT, DEFAULT_JUDGE_MODEL, DEFAULT_MEASURED_FAILURE_ENDPOINT, DEFAULT_PROMOTIONS_ENDPOINT, DELEGATE_TOOL_DEFINITION, DISCIPLINE_GATES_V1_ALT_BLIND_HEADER, DISCIPLINE_GATES_V1_ALT_HEADER, type DecompositionPlan, type DecompositionStep, type DecompositionTemplate, type DelegateHandle, type DelegateRefusalReason, type DelegateResult, type DelegateToolArgs, type ExclusionFindingRow, type ExclusionResolutionSource, type ExecuteErr, type ExecuteOk, type ExecuteOptions, type ExecuteResult, type ExecutorCandidate, type FallbackPosture, FallbackReason, FamilyResolutionError, type GetActionableAdvisoriesOptions, type GetApplicablePromotionOpts, type GetDefaultFallbackChainOpts, type GetMeasuredFailureOpts, type GetPerAxisMetricsOpts, type GetRecommendedPrimaryOptions, type GoldenEvalAxis, type GoldenEvalCase, type GoldenEvalOptions, type GoldenEvalRunResult, type GoldenEvalStrategyId, type GoldenIrRecordInput, Grounding, IntentArchetypeName, JUDGE_RUBRICS, LIBRARY_VERSION, type LLMJudgeOptions, MEASURED_FAILURE_CFG, MEASURED_GROUNDING_MIN_N, type MarkAdvisoryResolvedOptions, type MarkExclusionFindingHandledOptions, type MarkPromoteReadyHandledOptions, type MeasuredFailureRuntime, type MeasuredFailureVerdict, type ModelBrainRow, type ModelCompatibility, ModelProfile, NormalizedResponse, type OracleContext, OracleScore, type OutcomePayload, OutcomeResult, OutputMode, PRODUCER_OWNED_RULE_CODES, PROVIDER_ENV_KEYS, PerAxisMetrics, type PlannedStep, type PricingRow, type ProbeShadowOptions, type ProbeShadowServed, type ProfileToRowOptions, type PromoteReadyFindingRow, type PromoteReadyResolution, type PromotionRow, type PromotionsRuntime, PromptIR, Provider, ProviderOverrides, type ProviderReachability, ROLLBACK_SUPPRESSION_WINDOW_DAYS, RULE_DISCIPLINE_GATES_V1, RULE_DISCIPLINE_GATES_V1_STRUCTURED, RULE_SEQUENTIAL_TOOL_CLIFF, type ReachabilityOpts, RecordInput, RecordOutcomeInput, type RunAdvisorPhase2Context, STRATEGY_AUTHORSHIP_INDEPENDENT, STRATEGY_AUTHORSHIP_LIMITATION, SectionRewrite, type ShadowProbeRecordInput, type StrategyAttribution, type StrategyAttributionResult, type StrategyOutcome, type SupportedProvider, type SurfaceFailureRow, type SurfaceStats, SystemModelMessage, TRANSLATOR_FLOOR, _testResetMeasuredFailure, _testResetOutcomeSourceWarning, _testResetPromotions, _testWaitForMeasuredFailureRefresh, _testWaitForPromotionsRefresh, altBlindGatesBlockFor, altGatesBlockFor, applyArchetypeConvention, applySectionRewrites, attachCacheControlToStreamTextInput, awaitMeasuredFailureReady, awaitPromotionsReady, brainHealth, buildGoldenIrRow, buildLLMJudge, buildPairwiseJudgePrompt, buildShadowProbeRow, burstCaveat, call, chainProviderSpread, classifyEvidenceWindow, classifyStrategyOutcome, clearBrain, combineOrderSwappedVerdicts, compile, compileForAISDKv6, configureBrain, configureMeasuredFailureBrain, configurePromotionsBrain, countTokens, createDelegate, deriveFamilyFromModelId, deriveOwnership, estimateChainCostUsd, estimateModelCostUsd, execute, findBetterFit, flushBrainDeadLetter, formatEvidenceSpan, getActionableAdvisories, getAllStarterChains, getAllStarterChainsWithGrounding, getApplicablePromotion, getArchetypePerfScore, getDefaultFallbackChain, getDefaultFallbackChainWithGrounding, getMeasuredFailureVerdict, getModelCompatibility, getPerAxisMetrics, getReachabilityDiagnostic, getRecentRollback, getRecommendedPrimary, getSequentialStarterChain, getSequentialStarterChainWithGrounding, getStaleExclusionFindings, getStarterChain, getStarterChainWithGrounding, isAltStrategy, isAutoPromoteEnabledFromEnv, isBrainQueryActiveFor, isBrainSync, isDelegateEnabledFromEnv, isExclusionFindingsBrainActive, isMeasuredFailureBrainActive, isMeasuredFailureGateEnabledFromEnv, isModelReachable, isPromotionsBrainActive, isProviderReachable, judgeMeasuredFailure, loadAliasesFromBrain, loadArchetypePerfFromBrain, loadArchetypePerfNFromBrain, loadChainsFromBrain, loadModelsFromBrain, loadPricingFromBrain, mapMeasuredFailureRows, markAdvisoryResolved, markExclusionFindingHandled, markPromoteReadyHandled, parseJudgeVerdict, peekBrainDeadLetter, planDecomposition, prefetchMeasuredFailure, prefetchPromotions, probeShadow, profileToRow, readBrainReadEnv, record, recordGoldenIr, recordOutcome, recordShadowProbe, renderIrForJudge, resetTokenizer, resolveConventionsForProfile, resolveOutcomeSource, resolvePricingAt, resolveProviderKey, rowToAdvisory, rubricFor, runAdvisor, runGoldenEval, runStrategyEvalWithAttribution, setTokenizer, wilsonLowerBound, withAltBlindDisciplineContract, withAltDisciplineContract, withDisciplineContract };
|
package/dist/index.js
CHANGED
|
@@ -37,6 +37,7 @@ __export(index_exports, {
|
|
|
37
37
|
DECOMPOSITION_TEMPLATES: () => DECOMPOSITION_TEMPLATES,
|
|
38
38
|
DECOMPOSITION_TEMPLATES_VERSION: () => DECOMPOSITION_TEMPLATES_VERSION,
|
|
39
39
|
DEFAULT_FINDINGS_ENDPOINT: () => DEFAULT_FINDINGS_ENDPOINT,
|
|
40
|
+
DEFAULT_JUDGE_MODEL: () => DEFAULT_JUDGE_MODEL,
|
|
40
41
|
DEFAULT_MEASURED_FAILURE_ENDPOINT: () => DEFAULT_MEASURED_FAILURE_ENDPOINT,
|
|
41
42
|
DEFAULT_PROMOTIONS_ENDPOINT: () => DEFAULT_PROMOTIONS_ENDPOINT,
|
|
42
43
|
DELEGATE_TOOL_DEFINITION: () => DELEGATE_TOOL_DEFINITION,
|
|
@@ -583,7 +584,7 @@ var PROFILES_RAW = [
|
|
|
583
584
|
},
|
|
584
585
|
{
|
|
585
586
|
id: "claude-opus-4-7",
|
|
586
|
-
verifiedAgainstDocs: "2026-08-
|
|
587
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
587
588
|
provider: "anthropic",
|
|
588
589
|
status: "legacy",
|
|
589
590
|
// s76 2026-07-25: docs list as LEGACY (Opus 5 is current). Bundle said current while the brain already said legacy — the two disagreed.
|
|
@@ -633,7 +634,7 @@ var PROFILES_RAW = [
|
|
|
633
634
|
},
|
|
634
635
|
{
|
|
635
636
|
id: "claude-opus-4-6",
|
|
636
|
-
verifiedAgainstDocs: "2026-08-
|
|
637
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
637
638
|
provider: "anthropic",
|
|
638
639
|
status: "legacy",
|
|
639
640
|
maxContextTokens: 1e6,
|
|
@@ -678,7 +679,7 @@ var PROFILES_RAW = [
|
|
|
678
679
|
},
|
|
679
680
|
{
|
|
680
681
|
id: "claude-sonnet-4-6",
|
|
681
|
-
verifiedAgainstDocs: "2026-08-
|
|
682
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
682
683
|
provider: "anthropic",
|
|
683
684
|
status: "legacy",
|
|
684
685
|
// s76 2026-07-25: docs list as LEGACY (Sonnet 5 is current).
|
|
@@ -719,7 +720,7 @@ var PROFILES_RAW = [
|
|
|
719
720
|
},
|
|
720
721
|
{
|
|
721
722
|
id: "claude-haiku-4-5",
|
|
722
|
-
verifiedAgainstDocs: "2026-08-
|
|
723
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
723
724
|
provider: "anthropic",
|
|
724
725
|
status: "current",
|
|
725
726
|
maxContextTokens: 2e5,
|
|
@@ -782,7 +783,7 @@ var PROFILES_RAW = [
|
|
|
782
783
|
// ── Google ──
|
|
783
784
|
{
|
|
784
785
|
id: "gemini-2.5-flash",
|
|
785
|
-
verifiedAgainstDocs: "2026-08-
|
|
786
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
786
787
|
provider: "google",
|
|
787
788
|
status: "current",
|
|
788
789
|
maxContextTokens: 1048576,
|
|
@@ -888,7 +889,7 @@ var PROFILES_RAW = [
|
|
|
888
889
|
// thresholds. The brain will validate/relax these as evidence accumulates
|
|
889
890
|
// per (archetype, model) tuple. Currently ZERO brain rows for this model.
|
|
890
891
|
id: "gemini-2.5-flash-lite",
|
|
891
|
-
verifiedAgainstDocs: "2026-08-
|
|
892
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
892
893
|
provider: "google",
|
|
893
894
|
status: "current",
|
|
894
895
|
maxContextTokens: 1048576,
|
|
@@ -1045,7 +1046,7 @@ var PROFILES_RAW = [
|
|
|
1045
1046
|
},
|
|
1046
1047
|
{
|
|
1047
1048
|
id: "gemini-3.1-pro-preview",
|
|
1048
|
-
verifiedAgainstDocs: "2026-08-
|
|
1049
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
1049
1050
|
provider: "google",
|
|
1050
1051
|
status: "preview",
|
|
1051
1052
|
maxContextTokens: 1048576,
|
|
@@ -1143,7 +1144,7 @@ var PROFILES_RAW = [
|
|
|
1143
1144
|
// baseQuality (it carries 'reasoning') with no latency counterweight,
|
|
1144
1145
|
// leapfrogged gemini-2.5-flash as PB's summarize leader once reachable.
|
|
1145
1146
|
latencyTier: "slow",
|
|
1146
|
-
notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87
|
|
1147
|
+
notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87 was built + verified 2026-08-14 (prices then live), held at the publish gate 10 days, and published 2026-08-24 \u2014 by which time the 08-16 transition had killed them ($0.14/$0.28 flash, $0.435/$0.87 pro); corrected here (alpha.88). Build-then-hold is the mechanism: re-verify prices at publish when publish is not same-session as build. A single price field cannot represent a two-rate schedule; catalog-sync extractions oscillate between the rates day-to-day (peak read 08-21, off-peak read 08-22) \u2014 time-of-day pricing remains unmodeled, roadmap.",
|
|
1147
1148
|
// Master plan §6.2 anchor. Brain-validated tier 1 cross-provider for
|
|
1148
1149
|
// classify (169 rows, 0% empty). Tier 0 for summarize-with-no-tools.
|
|
1149
1150
|
// Falls off on hunt (sequential tools — L-040) and reasoning depth.
|
|
@@ -1219,7 +1220,7 @@ var PROFILES_RAW = [
|
|
|
1219
1220
|
// of the served set. Tag derivation would say 'medium'; the measurement says
|
|
1220
1221
|
// otherwise.
|
|
1221
1222
|
latencyTier: "slow",
|
|
1222
|
-
notes: "Pro tier. 1M context, 384k max output. $0.435/$0.87 per 1M (cache-hit in $0.003625) \u2014 verified live 2026-06-22; the launch promo became the standing price (no revert). Default mode = thinking. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87
|
|
1223
|
+
notes: "Pro tier. 1M context, 384k max output. $0.435/$0.87 per 1M (cache-hit in $0.003625) \u2014 verified live 2026-06-22; the launch promo became the standing price (no revert). Default mode = thinking. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87 was built + verified 2026-08-14 (prices then live), held at the publish gate 10 days, and published 2026-08-24 \u2014 by which time the 08-16 transition had killed them ($0.14/$0.28 flash, $0.435/$0.87 pro); corrected here (alpha.88). Build-then-hold is the mechanism: re-verify prices at publish when publish is not same-session as build. A single price field cannot represent a two-rate schedule; catalog-sync extractions oscillate between the rates day-to-day (peak read 08-21, off-peak read 08-22) \u2014 time-of-day pricing remains unmodeled, roadmap.",
|
|
1223
1224
|
// Master plan §3.3: tier 3 cross-provider for plan chain. Reasoning
|
|
1224
1225
|
// bumped one notch over V4-Flash; same parallel-tool ceiling.
|
|
1225
1226
|
archetypePerf: {
|
|
@@ -1910,7 +1911,7 @@ var PROFILES_RAW = [
|
|
|
1910
1911
|
],
|
|
1911
1912
|
strengths: ["agentic_loops", "coding", "1m_context", "parallel_tools", "thinking_mode"],
|
|
1912
1913
|
weaknesses: ["no_brain_evidence_yet"],
|
|
1913
|
-
notes: "Hand-onboarded s86 (2026-08-
|
|
1914
|
+
notes: "Hand-onboarded s86 (2026-08-14), step-zero verified: 1,048,576 ctx / 65,536 out (Google model page). PRICING re-verified 2026-08-24 at publish: $0.75/$3.75 promo through 2026-12-31, $1.50/$7.50 from 2027-01-01 \u2014 the alpha.87 profile carried the post-promo rate as if current; corrected alpha.88. Docs status Stable; enters kgauto as preview per standing rule (availability != lifecycle, promotion via brain evidence). Function calling + structured outputs + thinking per docs; parallel tools assumed per Flash family (docs silent). First flagged by the release watcher s73 (2026-07-24); ~21 days in the onboarding queue (onboarded 2026-08-14).",
|
|
1914
1915
|
archetypePerf: {
|
|
1915
1916
|
hunt: 9,
|
|
1916
1917
|
// Flash-family parallel-tool tier (L-040); inherited from 3.5-flash
|
|
@@ -1944,7 +1945,7 @@ var PROFILES_RAW = [
|
|
|
1944
1945
|
// 'preview' per the Fable precedent — no brain evidence yet; promotion to
|
|
1945
1946
|
// 'current' is an explicit call. (L-049/L-081.)
|
|
1946
1947
|
id: "claude-sonnet-5",
|
|
1947
|
-
verifiedAgainstDocs: "2026-08-
|
|
1948
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
1948
1949
|
provider: "anthropic",
|
|
1949
1950
|
status: "current",
|
|
1950
1951
|
// s76 2026-07-25: GA per docs; was preview.
|
|
@@ -6289,7 +6290,7 @@ function validateFinalFit(ir, profile, tokens) {
|
|
|
6289
6290
|
}
|
|
6290
6291
|
|
|
6291
6292
|
// src/version.ts
|
|
6292
|
-
var LIBRARY_VERSION = "2.0.0-alpha.
|
|
6293
|
+
var LIBRARY_VERSION = "2.0.0-alpha.89";
|
|
6293
6294
|
|
|
6294
6295
|
// src/pricing-brain.ts
|
|
6295
6296
|
function isPricingRow(x) {
|
|
@@ -6800,6 +6801,9 @@ function buildPayload(input, reg) {
|
|
|
6800
6801
|
// alpha.87 — explicit source wins; KGAUTO_SOURCE env fills the gap (the
|
|
6801
6802
|
// zero-code synthetic self-mark for harness/smoke processes).
|
|
6802
6803
|
source: resolveOutcomeSource(input.source),
|
|
6804
|
+
// 128-char truncation is the write-side contract (documented on the
|
|
6805
|
+
// type); an id that long is almost certainly a mistake, not identity.
|
|
6806
|
+
end_user_id: input.endUserId ? String(input.endUserId).slice(0, 128) : void 0,
|
|
6803
6807
|
error_type: input.errorType,
|
|
6804
6808
|
tools_called: input.toolsCalled,
|
|
6805
6809
|
oracle_score: input.oracleScore?.score,
|
|
@@ -8244,6 +8248,7 @@ async function call(ir, opts = {}) {
|
|
|
8244
8248
|
await record({
|
|
8245
8249
|
handle: initial.handle,
|
|
8246
8250
|
source: opts.source,
|
|
8251
|
+
endUserId: opts.endUserId,
|
|
8247
8252
|
tokensIn: 0,
|
|
8248
8253
|
tokensOut: 0,
|
|
8249
8254
|
latencyMs: latencyMs2,
|
|
@@ -8297,6 +8302,7 @@ async function call(ir, opts = {}) {
|
|
|
8297
8302
|
await record({
|
|
8298
8303
|
handle: initial.handle,
|
|
8299
8304
|
source: opts.source,
|
|
8305
|
+
endUserId: opts.endUserId,
|
|
8300
8306
|
tokensIn: 0,
|
|
8301
8307
|
tokensOut: 0,
|
|
8302
8308
|
latencyMs: latencyMs2,
|
|
@@ -8439,6 +8445,7 @@ async function call(ir, opts = {}) {
|
|
|
8439
8445
|
await record({
|
|
8440
8446
|
handle: initial.handle,
|
|
8441
8447
|
source: opts.source,
|
|
8448
|
+
endUserId: opts.endUserId,
|
|
8442
8449
|
tokensIn: validated.response.tokens.input,
|
|
8443
8450
|
tokensOut: validated.response.tokens.output,
|
|
8444
8451
|
latencyMs: latencyMs2,
|
|
@@ -8562,6 +8569,7 @@ async function call(ir, opts = {}) {
|
|
|
8562
8569
|
await record({
|
|
8563
8570
|
handle: initial.handle,
|
|
8564
8571
|
source: opts.source,
|
|
8572
|
+
endUserId: opts.endUserId,
|
|
8565
8573
|
tokensIn: lastErr?.tokens?.input ?? 0,
|
|
8566
8574
|
tokensOut: lastErr?.tokens?.output ?? 0,
|
|
8567
8575
|
latencyMs,
|
|
@@ -9035,6 +9043,7 @@ function extractKeptToolNames(result) {
|
|
|
9035
9043
|
|
|
9036
9044
|
// src/golden-eval.ts
|
|
9037
9045
|
var GENERIC_RUBRIC = "overall correctness, completeness against the request, clarity, and adherence to any requested format";
|
|
9046
|
+
var DEFAULT_JUDGE_MODEL = "claude-opus-5";
|
|
9038
9047
|
var JUDGE_RUBRICS = {
|
|
9039
9048
|
summarize: "faithfulness to the source (no fabricated facts), coverage of the key points, concision, and adherence to the requested output format",
|
|
9040
9049
|
classify: "assignment of the correct category from the allowed set, and nothing outside the allowed set",
|
|
@@ -9253,7 +9262,7 @@ async function runGoldenEval(opts) {
|
|
|
9253
9262
|
const threshold = opts.winOrTieThreshold ?? 0.8;
|
|
9254
9263
|
const latencyFloorRatio = opts.latencyFloorRatio ?? 3;
|
|
9255
9264
|
const minJudgeable = opts.minJudgeableCases ?? 5;
|
|
9256
|
-
const judgeModel = opts.judgeModel ??
|
|
9265
|
+
const judgeModel = opts.judgeModel ?? DEFAULT_JUDGE_MODEL;
|
|
9257
9266
|
const notes = [];
|
|
9258
9267
|
const axis = opts.axis ?? "model";
|
|
9259
9268
|
const strategyId = opts.strategy;
|
|
@@ -11067,6 +11076,7 @@ function compile2(ir, opts) {
|
|
|
11067
11076
|
DECOMPOSITION_TEMPLATES,
|
|
11068
11077
|
DECOMPOSITION_TEMPLATES_VERSION,
|
|
11069
11078
|
DEFAULT_FINDINGS_ENDPOINT,
|
|
11079
|
+
DEFAULT_JUDGE_MODEL,
|
|
11070
11080
|
DEFAULT_MEASURED_FAILURE_ENDPOINT,
|
|
11071
11081
|
DEFAULT_PROMOTIONS_ENDPOINT,
|
|
11072
11082
|
DELEGATE_TOOL_DEFINITION,
|
package/dist/index.mjs
CHANGED
|
@@ -19,7 +19,7 @@ import {
|
|
|
19
19
|
LIBRARY_VERSION,
|
|
20
20
|
createKeyHealthRoute,
|
|
21
21
|
keyFingerprint
|
|
22
|
-
} from "./chunk-
|
|
22
|
+
} from "./chunk-MKOOPD6S.mjs";
|
|
23
23
|
import {
|
|
24
24
|
ABSOLUTE_FLOOR,
|
|
25
25
|
ARCHETYPE_FLOOR_DEFAULT,
|
|
@@ -57,7 +57,7 @@ import {
|
|
|
57
57
|
policySetHas,
|
|
58
58
|
readBrainReadEnv,
|
|
59
59
|
resolveProviderKey
|
|
60
|
-
} from "./chunk-
|
|
60
|
+
} from "./chunk-D7RX3UD3.mjs";
|
|
61
61
|
import {
|
|
62
62
|
ALIASES,
|
|
63
63
|
LATENCY_TIER_MS,
|
|
@@ -73,7 +73,7 @@ import {
|
|
|
73
73
|
registerProfiles,
|
|
74
74
|
resolveModelAlias,
|
|
75
75
|
tryGetProfile
|
|
76
|
-
} from "./chunk-
|
|
76
|
+
} from "./chunk-OK2TMFRR.mjs";
|
|
77
77
|
import {
|
|
78
78
|
emitAdvisoryFired,
|
|
79
79
|
emitCompileDone,
|
|
@@ -3913,6 +3913,9 @@ function buildPayload(input, reg) {
|
|
|
3913
3913
|
// alpha.87 — explicit source wins; KGAUTO_SOURCE env fills the gap (the
|
|
3914
3914
|
// zero-code synthetic self-mark for harness/smoke processes).
|
|
3915
3915
|
source: resolveOutcomeSource(input.source),
|
|
3916
|
+
// 128-char truncation is the write-side contract (documented on the
|
|
3917
|
+
// type); an id that long is almost certainly a mistake, not identity.
|
|
3918
|
+
end_user_id: input.endUserId ? String(input.endUserId).slice(0, 128) : void 0,
|
|
3916
3919
|
error_type: input.errorType,
|
|
3917
3920
|
tools_called: input.toolsCalled,
|
|
3918
3921
|
oracle_score: input.oracleScore?.score,
|
|
@@ -5056,6 +5059,7 @@ async function call(ir, opts = {}) {
|
|
|
5056
5059
|
await record({
|
|
5057
5060
|
handle: initial.handle,
|
|
5058
5061
|
source: opts.source,
|
|
5062
|
+
endUserId: opts.endUserId,
|
|
5059
5063
|
tokensIn: 0,
|
|
5060
5064
|
tokensOut: 0,
|
|
5061
5065
|
latencyMs: latencyMs2,
|
|
@@ -5109,6 +5113,7 @@ async function call(ir, opts = {}) {
|
|
|
5109
5113
|
await record({
|
|
5110
5114
|
handle: initial.handle,
|
|
5111
5115
|
source: opts.source,
|
|
5116
|
+
endUserId: opts.endUserId,
|
|
5112
5117
|
tokensIn: 0,
|
|
5113
5118
|
tokensOut: 0,
|
|
5114
5119
|
latencyMs: latencyMs2,
|
|
@@ -5251,6 +5256,7 @@ async function call(ir, opts = {}) {
|
|
|
5251
5256
|
await record({
|
|
5252
5257
|
handle: initial.handle,
|
|
5253
5258
|
source: opts.source,
|
|
5259
|
+
endUserId: opts.endUserId,
|
|
5254
5260
|
tokensIn: validated.response.tokens.input,
|
|
5255
5261
|
tokensOut: validated.response.tokens.output,
|
|
5256
5262
|
latencyMs: latencyMs2,
|
|
@@ -5374,6 +5380,7 @@ async function call(ir, opts = {}) {
|
|
|
5374
5380
|
await record({
|
|
5375
5381
|
handle: initial.handle,
|
|
5376
5382
|
source: opts.source,
|
|
5383
|
+
endUserId: opts.endUserId,
|
|
5377
5384
|
tokensIn: lastErr?.tokens?.input ?? 0,
|
|
5378
5385
|
tokensOut: lastErr?.tokens?.output ?? 0,
|
|
5379
5386
|
latencyMs,
|
|
@@ -5847,6 +5854,7 @@ function extractKeptToolNames(result) {
|
|
|
5847
5854
|
|
|
5848
5855
|
// src/golden-eval.ts
|
|
5849
5856
|
var GENERIC_RUBRIC = "overall correctness, completeness against the request, clarity, and adherence to any requested format";
|
|
5857
|
+
var DEFAULT_JUDGE_MODEL = "claude-opus-5";
|
|
5850
5858
|
var JUDGE_RUBRICS = {
|
|
5851
5859
|
summarize: "faithfulness to the source (no fabricated facts), coverage of the key points, concision, and adherence to the requested output format",
|
|
5852
5860
|
classify: "assignment of the correct category from the allowed set, and nothing outside the allowed set",
|
|
@@ -6065,7 +6073,7 @@ async function runGoldenEval(opts) {
|
|
|
6065
6073
|
const threshold = opts.winOrTieThreshold ?? 0.8;
|
|
6066
6074
|
const latencyFloorRatio = opts.latencyFloorRatio ?? 3;
|
|
6067
6075
|
const minJudgeable = opts.minJudgeableCases ?? 5;
|
|
6068
|
-
const judgeModel = opts.judgeModel ??
|
|
6076
|
+
const judgeModel = opts.judgeModel ?? DEFAULT_JUDGE_MODEL;
|
|
6069
6077
|
const notes = [];
|
|
6070
6078
|
const axis = opts.axis ?? "model";
|
|
6071
6079
|
const strategyId = opts.strategy;
|
|
@@ -7503,6 +7511,7 @@ export {
|
|
|
7503
7511
|
DECOMPOSITION_TEMPLATES,
|
|
7504
7512
|
DECOMPOSITION_TEMPLATES_VERSION,
|
|
7505
7513
|
DEFAULT_FINDINGS_ENDPOINT,
|
|
7514
|
+
DEFAULT_JUDGE_MODEL,
|
|
7506
7515
|
DEFAULT_MEASURED_FAILURE_ENDPOINT,
|
|
7507
7516
|
DEFAULT_PROMOTIONS_ENDPOINT,
|
|
7508
7517
|
DELEGATE_TOOL_DEFINITION,
|
|
@@ -1174,6 +1174,12 @@ interface CallOptions {
|
|
|
1174
1174
|
* {@link OutcomeSource}.
|
|
1175
1175
|
*/
|
|
1176
1176
|
source?: OutcomeSource;
|
|
1177
|
+
/**
|
|
1178
|
+
* alpha.89 — per-end-user metering: passed through to every outcome row
|
|
1179
|
+
* this call records, including failure rows (a failed call still spent
|
|
1180
|
+
* money attributable to the customer). See {@link RecordInput.endUserId}.
|
|
1181
|
+
*/
|
|
1182
|
+
endUserId?: string;
|
|
1177
1183
|
/**
|
|
1178
1184
|
* alpha.84 — fires once per FAILED attempt during the fallback walk, before
|
|
1179
1185
|
* the walk continues. IC-Cairn's filing, 2026-07-29.
|
|
@@ -1576,6 +1582,16 @@ interface RecordInput {
|
|
|
1576
1582
|
* {@link OutcomeSource}.
|
|
1577
1583
|
*/
|
|
1578
1584
|
source?: OutcomeSource;
|
|
1585
|
+
/**
|
|
1586
|
+
* alpha.89 (migration 065) — per-end-user metering. Your own OPAQUE
|
|
1587
|
+
* customer id (hash it upstream if your ids are sensitive; kgauto stores
|
|
1588
|
+
* it verbatim, truncated to 128 chars). With it, the brain becomes a
|
|
1589
|
+
* per-customer COGS ledger: `kgauto_end_user_ledger_v` aggregates calls /
|
|
1590
|
+
* tokens / cost_usd per (app, end user, month) over organic traffic only —
|
|
1591
|
+
* the substrate for usage-based billing and per-customer quota checks.
|
|
1592
|
+
* Omit for traffic that has no end user.
|
|
1593
|
+
*/
|
|
1594
|
+
endUserId?: string;
|
|
1579
1595
|
/** Actual tokens consumed (post-call). */
|
|
1580
1596
|
tokensIn: number;
|
|
1581
1597
|
tokensOut: number;
|
|
@@ -1174,6 +1174,12 @@ interface CallOptions {
|
|
|
1174
1174
|
* {@link OutcomeSource}.
|
|
1175
1175
|
*/
|
|
1176
1176
|
source?: OutcomeSource;
|
|
1177
|
+
/**
|
|
1178
|
+
* alpha.89 — per-end-user metering: passed through to every outcome row
|
|
1179
|
+
* this call records, including failure rows (a failed call still spent
|
|
1180
|
+
* money attributable to the customer). See {@link RecordInput.endUserId}.
|
|
1181
|
+
*/
|
|
1182
|
+
endUserId?: string;
|
|
1177
1183
|
/**
|
|
1178
1184
|
* alpha.84 — fires once per FAILED attempt during the fallback walk, before
|
|
1179
1185
|
* the walk continues. IC-Cairn's filing, 2026-07-29.
|
|
@@ -1576,6 +1582,16 @@ interface RecordInput {
|
|
|
1576
1582
|
* {@link OutcomeSource}.
|
|
1577
1583
|
*/
|
|
1578
1584
|
source?: OutcomeSource;
|
|
1585
|
+
/**
|
|
1586
|
+
* alpha.89 (migration 065) — per-end-user metering. Your own OPAQUE
|
|
1587
|
+
* customer id (hash it upstream if your ids are sensitive; kgauto stores
|
|
1588
|
+
* it verbatim, truncated to 128 chars). With it, the brain becomes a
|
|
1589
|
+
* per-customer COGS ledger: `kgauto_end_user_ledger_v` aggregates calls /
|
|
1590
|
+
* tokens / cost_usd per (app, end user, month) over organic traffic only —
|
|
1591
|
+
* the substrate for usage-based billing and per-customer quota checks.
|
|
1592
|
+
* Omit for traffic that has no end user.
|
|
1593
|
+
*/
|
|
1594
|
+
endUserId?: string;
|
|
1579
1595
|
/** Actual tokens consumed (post-call). */
|
|
1580
1596
|
tokensIn: number;
|
|
1581
1597
|
tokensOut: number;
|
package/dist/key-health.js
CHANGED
|
@@ -28,7 +28,7 @@ __export(key_health_exports, {
|
|
|
28
28
|
module.exports = __toCommonJS(key_health_exports);
|
|
29
29
|
|
|
30
30
|
// src/version.ts
|
|
31
|
-
var LIBRARY_VERSION = "2.0.0-alpha.
|
|
31
|
+
var LIBRARY_VERSION = "2.0.0-alpha.89";
|
|
32
32
|
|
|
33
33
|
// src/key-health.ts
|
|
34
34
|
var JSON_HEADERS = { "Content-Type": "application/json" };
|
package/dist/key-health.mjs
CHANGED
package/dist/profiles.d.mts
CHANGED
package/dist/profiles.d.ts
CHANGED
package/dist/profiles.js
CHANGED
|
@@ -248,7 +248,7 @@ var PROFILES_RAW = [
|
|
|
248
248
|
},
|
|
249
249
|
{
|
|
250
250
|
id: "claude-opus-4-7",
|
|
251
|
-
verifiedAgainstDocs: "2026-08-
|
|
251
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
252
252
|
provider: "anthropic",
|
|
253
253
|
status: "legacy",
|
|
254
254
|
// s76 2026-07-25: docs list as LEGACY (Opus 5 is current). Bundle said current while the brain already said legacy — the two disagreed.
|
|
@@ -298,7 +298,7 @@ var PROFILES_RAW = [
|
|
|
298
298
|
},
|
|
299
299
|
{
|
|
300
300
|
id: "claude-opus-4-6",
|
|
301
|
-
verifiedAgainstDocs: "2026-08-
|
|
301
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
302
302
|
provider: "anthropic",
|
|
303
303
|
status: "legacy",
|
|
304
304
|
maxContextTokens: 1e6,
|
|
@@ -343,7 +343,7 @@ var PROFILES_RAW = [
|
|
|
343
343
|
},
|
|
344
344
|
{
|
|
345
345
|
id: "claude-sonnet-4-6",
|
|
346
|
-
verifiedAgainstDocs: "2026-08-
|
|
346
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
347
347
|
provider: "anthropic",
|
|
348
348
|
status: "legacy",
|
|
349
349
|
// s76 2026-07-25: docs list as LEGACY (Sonnet 5 is current).
|
|
@@ -384,7 +384,7 @@ var PROFILES_RAW = [
|
|
|
384
384
|
},
|
|
385
385
|
{
|
|
386
386
|
id: "claude-haiku-4-5",
|
|
387
|
-
verifiedAgainstDocs: "2026-08-
|
|
387
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
388
388
|
provider: "anthropic",
|
|
389
389
|
status: "current",
|
|
390
390
|
maxContextTokens: 2e5,
|
|
@@ -447,7 +447,7 @@ var PROFILES_RAW = [
|
|
|
447
447
|
// ── Google ──
|
|
448
448
|
{
|
|
449
449
|
id: "gemini-2.5-flash",
|
|
450
|
-
verifiedAgainstDocs: "2026-08-
|
|
450
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
451
451
|
provider: "google",
|
|
452
452
|
status: "current",
|
|
453
453
|
maxContextTokens: 1048576,
|
|
@@ -553,7 +553,7 @@ var PROFILES_RAW = [
|
|
|
553
553
|
// thresholds. The brain will validate/relax these as evidence accumulates
|
|
554
554
|
// per (archetype, model) tuple. Currently ZERO brain rows for this model.
|
|
555
555
|
id: "gemini-2.5-flash-lite",
|
|
556
|
-
verifiedAgainstDocs: "2026-08-
|
|
556
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
557
557
|
provider: "google",
|
|
558
558
|
status: "current",
|
|
559
559
|
maxContextTokens: 1048576,
|
|
@@ -710,7 +710,7 @@ var PROFILES_RAW = [
|
|
|
710
710
|
},
|
|
711
711
|
{
|
|
712
712
|
id: "gemini-3.1-pro-preview",
|
|
713
|
-
verifiedAgainstDocs: "2026-08-
|
|
713
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
714
714
|
provider: "google",
|
|
715
715
|
status: "preview",
|
|
716
716
|
maxContextTokens: 1048576,
|
|
@@ -808,7 +808,7 @@ var PROFILES_RAW = [
|
|
|
808
808
|
// baseQuality (it carries 'reasoning') with no latency counterweight,
|
|
809
809
|
// leapfrogged gemini-2.5-flash as PB's summarize leader once reachable.
|
|
810
810
|
latencyTier: "slow",
|
|
811
|
-
notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87
|
|
811
|
+
notes: "Cheap workhorse. 1M context, 384k max output. Cache-hit input $0.0028/M (1/50\xD7 of miss). Aliased as `deepseek-chat` (non-thinking) and `deepseek-reasoner` (thinking) \u2014 see ALIASES. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87 was built + verified 2026-08-14 (prices then live), held at the publish gate 10 days, and published 2026-08-24 \u2014 by which time the 08-16 transition had killed them ($0.14/$0.28 flash, $0.435/$0.87 pro); corrected here (alpha.88). Build-then-hold is the mechanism: re-verify prices at publish when publish is not same-session as build. A single price field cannot represent a two-rate schedule; catalog-sync extractions oscillate between the rates day-to-day (peak read 08-21, off-peak read 08-22) \u2014 time-of-day pricing remains unmodeled, roadmap.",
|
|
812
812
|
// Master plan §6.2 anchor. Brain-validated tier 1 cross-provider for
|
|
813
813
|
// classify (169 rows, 0% empty). Tier 0 for summarize-with-no-tools.
|
|
814
814
|
// Falls off on hunt (sequential tools — L-040) and reasoning depth.
|
|
@@ -884,7 +884,7 @@ var PROFILES_RAW = [
|
|
|
884
884
|
// of the served set. Tag derivation would say 'medium'; the measurement says
|
|
885
885
|
// otherwise.
|
|
886
886
|
latencyTier: "slow",
|
|
887
|
-
notes: "Pro tier. 1M context, 384k max output. $0.435/$0.87 per 1M (cache-hit in $0.003625) \u2014 verified live 2026-06-22; the launch promo became the standing price (no revert). Default mode = thinking. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87
|
|
887
|
+
notes: "Pro tier. 1M context, 384k max output. $0.435/$0.87 per 1M (cache-hit in $0.003625) \u2014 verified live 2026-06-22; the launch promo became the standing price (no revert). Default mode = thinking. PEAK/OFF-PEAK since 2026-08-16 (page fetched 2026-08-24): peak = weekdays 01:00-04:00 + 06:00-10:00 UTC, off-peak = all other hours at HALF peak. Encoded prices ARE the peak rate (conservative: never understate cost); off-peak traffic costs half what the estimate says. alpha.87 was built + verified 2026-08-14 (prices then live), held at the publish gate 10 days, and published 2026-08-24 \u2014 by which time the 08-16 transition had killed them ($0.14/$0.28 flash, $0.435/$0.87 pro); corrected here (alpha.88). Build-then-hold is the mechanism: re-verify prices at publish when publish is not same-session as build. A single price field cannot represent a two-rate schedule; catalog-sync extractions oscillate between the rates day-to-day (peak read 08-21, off-peak read 08-22) \u2014 time-of-day pricing remains unmodeled, roadmap.",
|
|
888
888
|
// Master plan §3.3: tier 3 cross-provider for plan chain. Reasoning
|
|
889
889
|
// bumped one notch over V4-Flash; same parallel-tool ceiling.
|
|
890
890
|
archetypePerf: {
|
|
@@ -1575,7 +1575,7 @@ var PROFILES_RAW = [
|
|
|
1575
1575
|
],
|
|
1576
1576
|
strengths: ["agentic_loops", "coding", "1m_context", "parallel_tools", "thinking_mode"],
|
|
1577
1577
|
weaknesses: ["no_brain_evidence_yet"],
|
|
1578
|
-
notes: "Hand-onboarded s86 (2026-08-
|
|
1578
|
+
notes: "Hand-onboarded s86 (2026-08-14), step-zero verified: 1,048,576 ctx / 65,536 out (Google model page). PRICING re-verified 2026-08-24 at publish: $0.75/$3.75 promo through 2026-12-31, $1.50/$7.50 from 2027-01-01 \u2014 the alpha.87 profile carried the post-promo rate as if current; corrected alpha.88. Docs status Stable; enters kgauto as preview per standing rule (availability != lifecycle, promotion via brain evidence). Function calling + structured outputs + thinking per docs; parallel tools assumed per Flash family (docs silent). First flagged by the release watcher s73 (2026-07-24); ~21 days in the onboarding queue (onboarded 2026-08-14).",
|
|
1579
1579
|
archetypePerf: {
|
|
1580
1580
|
hunt: 9,
|
|
1581
1581
|
// Flash-family parallel-tool tier (L-040); inherited from 3.5-flash
|
|
@@ -1609,7 +1609,7 @@ var PROFILES_RAW = [
|
|
|
1609
1609
|
// 'preview' per the Fable precedent — no brain evidence yet; promotion to
|
|
1610
1610
|
// 'current' is an explicit call. (L-049/L-081.)
|
|
1611
1611
|
id: "claude-sonnet-5",
|
|
1612
|
-
verifiedAgainstDocs: "2026-08-
|
|
1612
|
+
verifiedAgainstDocs: "2026-08-14",
|
|
1613
1613
|
provider: "anthropic",
|
|
1614
1614
|
status: "current",
|
|
1615
1615
|
// s76 2026-07-25: GA per docs; was preview.
|
package/dist/profiles.mjs
CHANGED
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { t as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, n as CallAttempt } from './ir-
|
|
1
|
+
import { t as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, n as CallAttempt } from './ir-CTx026t0.js';
|
|
2
2
|
|
|
3
3
|
/**
|
|
4
4
|
* Glass-Box observability types (alpha.17).
|
|
@@ -1,4 +1,4 @@
|
|
|
1
|
-
import { t as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, n as CallAttempt } from './ir-
|
|
1
|
+
import { t as MutationApplied, B as BestPracticeAdvisory, F as FallbackReason, n as CallAttempt } from './ir-DeYMLWge.mjs';
|
|
2
2
|
|
|
3
3
|
/**
|
|
4
4
|
* Glass-Box observability types (alpha.17).
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "@warmdrift/kgauto-compiler",
|
|
3
|
-
"version": "2.0.0-alpha.
|
|
3
|
+
"version": "2.0.0-alpha.89",
|
|
4
4
|
"description": "Prompt compiler with executable provider knowledge for multi-model AI apps: normalized multi-provider transport with fallback chains, compile-time cliff guards, a curated model registry, and a telemetry flight recorder. Swap models without rewriting prompts.",
|
|
5
5
|
"main": "./dist/index.js",
|
|
6
6
|
"module": "./dist/index.mjs",
|