@warmdrift/kgauto-compiler 2.0.0-alpha.87 → 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-G3MKT7IR.mjs → chunk-D7RX3UD3.mjs} +1 -1
- package/dist/{chunk-GMZT4ZXD.mjs → chunk-MKOOPD6S.mjs} +1 -1
- package/dist/{chunk-YVKTTTTB.mjs → chunk-OK2TMFRR.mjs} +15 -13
- 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 +15 -13
- 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 +27 -15
- 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 +15 -13
- 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
|
@@ -733,7 +733,7 @@ var PROFILES_RAW = [
|
|
|
733
733
|
// deepseek-reasoner registered as aliases (see ALIASES below).
|
|
734
734
|
{
|
|
735
735
|
id: "deepseek-v4-flash",
|
|
736
|
-
verifiedAgainstDocs: "2026-08-
|
|
736
|
+
verifiedAgainstDocs: "2026-08-24",
|
|
737
737
|
provider: "deepseek",
|
|
738
738
|
status: "current",
|
|
739
739
|
maxContextTokens: 1e6,
|
|
@@ -751,8 +751,8 @@ var PROFILES_RAW = [
|
|
|
751
751
|
reason: "Sequential tool calls only \u2014 L-040"
|
|
752
752
|
}
|
|
753
753
|
],
|
|
754
|
-
costInputPer1m: 0.
|
|
755
|
-
costOutputPer1m:
|
|
754
|
+
costInputPer1m: 0.44,
|
|
755
|
+
costOutputPer1m: 1.32,
|
|
756
756
|
lowering: {
|
|
757
757
|
system: { mode: "inline" },
|
|
758
758
|
cache: { strategy: "unsupported" },
|
|
@@ -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.
|
|
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.
|
|
@@ -806,7 +806,7 @@ var PROFILES_RAW = [
|
|
|
806
806
|
},
|
|
807
807
|
{
|
|
808
808
|
id: "deepseek-v4-pro",
|
|
809
|
-
verifiedAgainstDocs: "2026-08-
|
|
809
|
+
verifiedAgainstDocs: "2026-08-24",
|
|
810
810
|
provider: "deepseek",
|
|
811
811
|
status: "current",
|
|
812
812
|
maxContextTokens: 1e6,
|
|
@@ -830,8 +830,8 @@ var PROFILES_RAW = [
|
|
|
830
830
|
// cache-hit in; no promo caveat). The old $1.74/$3.48 "regular post-promo"
|
|
831
831
|
// was a projected revert that never happened (L-073/L-081) — it 4×-overstated
|
|
832
832
|
// cost and suppressed V4-Pro in cost-aware ranking.
|
|
833
|
-
costInputPer1m:
|
|
834
|
-
costOutputPer1m:
|
|
833
|
+
costInputPer1m: 1.32,
|
|
834
|
+
costOutputPer1m: 3.96,
|
|
835
835
|
lowering: {
|
|
836
836
|
system: { mode: "inline" },
|
|
837
837
|
cache: { strategy: "unsupported" },
|
|
@@ -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.
|
|
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: {
|
|
@@ -1468,7 +1468,7 @@ var PROFILES_RAW = [
|
|
|
1468
1468
|
// ─────────────────────────────────────────────────────────────────────────
|
|
1469
1469
|
{
|
|
1470
1470
|
id: "gemini-3.6-flash",
|
|
1471
|
-
verifiedAgainstDocs: "2026-08-
|
|
1471
|
+
verifiedAgainstDocs: "2026-08-24",
|
|
1472
1472
|
versionAdded: "2.0.0-alpha.87",
|
|
1473
1473
|
provider: "google",
|
|
1474
1474
|
status: "preview",
|
|
@@ -1506,8 +1506,10 @@ var PROFILES_RAW = [
|
|
|
1506
1506
|
reason: "Inherited from 2.5-flash s11 cliff: summarize+tools \u2192 empty response. Preemptive family guard until brain evidence on 3.6-flash specifically."
|
|
1507
1507
|
}
|
|
1508
1508
|
],
|
|
1509
|
-
costInputPer1m:
|
|
1510
|
-
|
|
1509
|
+
costInputPer1m: 0.75,
|
|
1510
|
+
// promo through 2026-12-31; $1.50 from 2027-01-01 (TIME_BOUNDED_PRICING_ROWS carries the transition)
|
|
1511
|
+
costOutputPer1m: 3.75,
|
|
1512
|
+
// promo through 2026-12-31; $7.50 from 2027-01-01
|
|
1511
1513
|
lowering: {
|
|
1512
1514
|
...GOOGLE_LOWERING_BASE,
|
|
1513
1515
|
cache: { ...GOOGLE_LOWERING_BASE.cache, discount: 0.1 },
|
|
@@ -1536,7 +1538,7 @@ var PROFILES_RAW = [
|
|
|
1536
1538
|
],
|
|
1537
1539
|
strengths: ["agentic_loops", "coding", "1m_context", "parallel_tools", "thinking_mode"],
|
|
1538
1540
|
weaknesses: ["no_brain_evidence_yet"],
|
|
1539
|
-
notes: "Hand-onboarded s86 (2026-08-14), step-zero verified: 1,048,576 ctx / 65,536 out (Google model page), $1.50/$7.50
|
|
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).",
|
|
1540
1542
|
archetypePerf: {
|
|
1541
1543
|
hunt: 9,
|
|
1542
1544
|
// Flash-family parallel-tool tier (L-040); inherited from 3.5-flash
|
|
@@ -1583,7 +1585,7 @@ var PROFILES_RAW = [
|
|
|
1583
1585
|
streaming: true,
|
|
1584
1586
|
cliffs: [],
|
|
1585
1587
|
costInputPer1m: 2,
|
|
1586
|
-
// 2026-08-
|
|
1588
|
+
// 2026-08-24: intro price made PERMANENT (docs: "the previously scheduled increase to $3/$15 on September 1, 2026 will not occur")
|
|
1587
1589
|
costOutputPer1m: 10,
|
|
1588
1590
|
lowering: ANTHROPIC_LOWERING_BASE,
|
|
1589
1591
|
recovery: [
|
|
@@ -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
|
/**
|
|
@@ -787,7 +787,7 @@ var PROFILES_RAW = [
|
|
|
787
787
|
// deepseek-reasoner registered as aliases (see ALIASES below).
|
|
788
788
|
{
|
|
789
789
|
id: "deepseek-v4-flash",
|
|
790
|
-
verifiedAgainstDocs: "2026-08-
|
|
790
|
+
verifiedAgainstDocs: "2026-08-24",
|
|
791
791
|
provider: "deepseek",
|
|
792
792
|
status: "current",
|
|
793
793
|
maxContextTokens: 1e6,
|
|
@@ -805,8 +805,8 @@ var PROFILES_RAW = [
|
|
|
805
805
|
reason: "Sequential tool calls only \u2014 L-040"
|
|
806
806
|
}
|
|
807
807
|
],
|
|
808
|
-
costInputPer1m: 0.
|
|
809
|
-
costOutputPer1m:
|
|
808
|
+
costInputPer1m: 0.44,
|
|
809
|
+
costOutputPer1m: 1.32,
|
|
810
810
|
lowering: {
|
|
811
811
|
system: { mode: "inline" },
|
|
812
812
|
cache: { strategy: "unsupported" },
|
|
@@ -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.
|
|
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.
|
|
@@ -860,7 +860,7 @@ var PROFILES_RAW = [
|
|
|
860
860
|
},
|
|
861
861
|
{
|
|
862
862
|
id: "deepseek-v4-pro",
|
|
863
|
-
verifiedAgainstDocs: "2026-08-
|
|
863
|
+
verifiedAgainstDocs: "2026-08-24",
|
|
864
864
|
provider: "deepseek",
|
|
865
865
|
status: "current",
|
|
866
866
|
maxContextTokens: 1e6,
|
|
@@ -884,8 +884,8 @@ var PROFILES_RAW = [
|
|
|
884
884
|
// cache-hit in; no promo caveat). The old $1.74/$3.48 "regular post-promo"
|
|
885
885
|
// was a projected revert that never happened (L-073/L-081) — it 4×-overstated
|
|
886
886
|
// cost and suppressed V4-Pro in cost-aware ranking.
|
|
887
|
-
costInputPer1m:
|
|
888
|
-
costOutputPer1m:
|
|
887
|
+
costInputPer1m: 1.32,
|
|
888
|
+
costOutputPer1m: 3.96,
|
|
889
889
|
lowering: {
|
|
890
890
|
system: { mode: "inline" },
|
|
891
891
|
cache: { strategy: "unsupported" },
|
|
@@ -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.
|
|
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: {
|
|
@@ -1522,7 +1522,7 @@ var PROFILES_RAW = [
|
|
|
1522
1522
|
// ─────────────────────────────────────────────────────────────────────────
|
|
1523
1523
|
{
|
|
1524
1524
|
id: "gemini-3.6-flash",
|
|
1525
|
-
verifiedAgainstDocs: "2026-08-
|
|
1525
|
+
verifiedAgainstDocs: "2026-08-24",
|
|
1526
1526
|
versionAdded: "2.0.0-alpha.87",
|
|
1527
1527
|
provider: "google",
|
|
1528
1528
|
status: "preview",
|
|
@@ -1560,8 +1560,10 @@ var PROFILES_RAW = [
|
|
|
1560
1560
|
reason: "Inherited from 2.5-flash s11 cliff: summarize+tools \u2192 empty response. Preemptive family guard until brain evidence on 3.6-flash specifically."
|
|
1561
1561
|
}
|
|
1562
1562
|
],
|
|
1563
|
-
costInputPer1m:
|
|
1564
|
-
|
|
1563
|
+
costInputPer1m: 0.75,
|
|
1564
|
+
// promo through 2026-12-31; $1.50 from 2027-01-01 (TIME_BOUNDED_PRICING_ROWS carries the transition)
|
|
1565
|
+
costOutputPer1m: 3.75,
|
|
1566
|
+
// promo through 2026-12-31; $7.50 from 2027-01-01
|
|
1565
1567
|
lowering: {
|
|
1566
1568
|
...GOOGLE_LOWERING_BASE,
|
|
1567
1569
|
cache: { ...GOOGLE_LOWERING_BASE.cache, discount: 0.1 },
|
|
@@ -1590,7 +1592,7 @@ var PROFILES_RAW = [
|
|
|
1590
1592
|
],
|
|
1591
1593
|
strengths: ["agentic_loops", "coding", "1m_context", "parallel_tools", "thinking_mode"],
|
|
1592
1594
|
weaknesses: ["no_brain_evidence_yet"],
|
|
1593
|
-
notes: "Hand-onboarded s86 (2026-08-14), step-zero verified: 1,048,576 ctx / 65,536 out (Google model page), $1.50/$7.50
|
|
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).",
|
|
1594
1596
|
archetypePerf: {
|
|
1595
1597
|
hunt: 9,
|
|
1596
1598
|
// Flash-family parallel-tool tier (L-040); inherited from 3.5-flash
|
|
@@ -1637,7 +1639,7 @@ var PROFILES_RAW = [
|
|
|
1637
1639
|
streaming: true,
|
|
1638
1640
|
cliffs: [],
|
|
1639
1641
|
costInputPer1m: 2,
|
|
1640
|
-
// 2026-08-
|
|
1642
|
+
// 2026-08-24: intro price made PERMANENT (docs: "the previously scheduled increase to $3/$15 on September 1, 2026 will not occur")
|
|
1641
1643
|
costOutputPer1m: 10,
|
|
1642
1644
|
lowering: ANTHROPIC_LOWERING_BASE,
|
|
1643
1645
|
recovery: [
|
|
@@ -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,
|
|
@@ -1105,7 +1106,7 @@ var PROFILES_RAW = [
|
|
|
1105
1106
|
// deepseek-reasoner registered as aliases (see ALIASES below).
|
|
1106
1107
|
{
|
|
1107
1108
|
id: "deepseek-v4-flash",
|
|
1108
|
-
verifiedAgainstDocs: "2026-08-
|
|
1109
|
+
verifiedAgainstDocs: "2026-08-24",
|
|
1109
1110
|
provider: "deepseek",
|
|
1110
1111
|
status: "current",
|
|
1111
1112
|
maxContextTokens: 1e6,
|
|
@@ -1123,8 +1124,8 @@ var PROFILES_RAW = [
|
|
|
1123
1124
|
reason: "Sequential tool calls only \u2014 L-040"
|
|
1124
1125
|
}
|
|
1125
1126
|
],
|
|
1126
|
-
costInputPer1m: 0.
|
|
1127
|
-
costOutputPer1m:
|
|
1127
|
+
costInputPer1m: 0.44,
|
|
1128
|
+
costOutputPer1m: 1.32,
|
|
1128
1129
|
lowering: {
|
|
1129
1130
|
system: { mode: "inline" },
|
|
1130
1131
|
cache: { strategy: "unsupported" },
|
|
@@ -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.
|
|
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.
|
|
@@ -1178,7 +1179,7 @@ var PROFILES_RAW = [
|
|
|
1178
1179
|
},
|
|
1179
1180
|
{
|
|
1180
1181
|
id: "deepseek-v4-pro",
|
|
1181
|
-
verifiedAgainstDocs: "2026-08-
|
|
1182
|
+
verifiedAgainstDocs: "2026-08-24",
|
|
1182
1183
|
provider: "deepseek",
|
|
1183
1184
|
status: "current",
|
|
1184
1185
|
maxContextTokens: 1e6,
|
|
@@ -1202,8 +1203,8 @@ var PROFILES_RAW = [
|
|
|
1202
1203
|
// cache-hit in; no promo caveat). The old $1.74/$3.48 "regular post-promo"
|
|
1203
1204
|
// was a projected revert that never happened (L-073/L-081) — it 4×-overstated
|
|
1204
1205
|
// cost and suppressed V4-Pro in cost-aware ranking.
|
|
1205
|
-
costInputPer1m:
|
|
1206
|
-
costOutputPer1m:
|
|
1206
|
+
costInputPer1m: 1.32,
|
|
1207
|
+
costOutputPer1m: 3.96,
|
|
1207
1208
|
lowering: {
|
|
1208
1209
|
system: { mode: "inline" },
|
|
1209
1210
|
cache: { strategy: "unsupported" },
|
|
@@ -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.
|
|
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: {
|
|
@@ -1840,7 +1841,7 @@ var PROFILES_RAW = [
|
|
|
1840
1841
|
// ─────────────────────────────────────────────────────────────────────────
|
|
1841
1842
|
{
|
|
1842
1843
|
id: "gemini-3.6-flash",
|
|
1843
|
-
verifiedAgainstDocs: "2026-08-
|
|
1844
|
+
verifiedAgainstDocs: "2026-08-24",
|
|
1844
1845
|
versionAdded: "2.0.0-alpha.87",
|
|
1845
1846
|
provider: "google",
|
|
1846
1847
|
status: "preview",
|
|
@@ -1878,8 +1879,10 @@ var PROFILES_RAW = [
|
|
|
1878
1879
|
reason: "Inherited from 2.5-flash s11 cliff: summarize+tools \u2192 empty response. Preemptive family guard until brain evidence on 3.6-flash specifically."
|
|
1879
1880
|
}
|
|
1880
1881
|
],
|
|
1881
|
-
costInputPer1m:
|
|
1882
|
-
|
|
1882
|
+
costInputPer1m: 0.75,
|
|
1883
|
+
// promo through 2026-12-31; $1.50 from 2027-01-01 (TIME_BOUNDED_PRICING_ROWS carries the transition)
|
|
1884
|
+
costOutputPer1m: 3.75,
|
|
1885
|
+
// promo through 2026-12-31; $7.50 from 2027-01-01
|
|
1883
1886
|
lowering: {
|
|
1884
1887
|
...GOOGLE_LOWERING_BASE,
|
|
1885
1888
|
cache: { ...GOOGLE_LOWERING_BASE.cache, discount: 0.1 },
|
|
@@ -1908,7 +1911,7 @@ var PROFILES_RAW = [
|
|
|
1908
1911
|
],
|
|
1909
1912
|
strengths: ["agentic_loops", "coding", "1m_context", "parallel_tools", "thinking_mode"],
|
|
1910
1913
|
weaknesses: ["no_brain_evidence_yet"],
|
|
1911
|
-
notes: "Hand-onboarded s86 (2026-08-14), step-zero verified: 1,048,576 ctx / 65,536 out (Google model page), $1.50/$7.50
|
|
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).",
|
|
1912
1915
|
archetypePerf: {
|
|
1913
1916
|
hunt: 9,
|
|
1914
1917
|
// Flash-family parallel-tool tier (L-040); inherited from 3.5-flash
|
|
@@ -1955,7 +1958,7 @@ var PROFILES_RAW = [
|
|
|
1955
1958
|
streaming: true,
|
|
1956
1959
|
cliffs: [],
|
|
1957
1960
|
costInputPer1m: 2,
|
|
1958
|
-
// 2026-08-
|
|
1961
|
+
// 2026-08-24: intro price made PERMANENT (docs: "the previously scheduled increase to $3/$15 on September 1, 2026 will not occur")
|
|
1959
1962
|
costOutputPer1m: 10,
|
|
1960
1963
|
lowering: ANTHROPIC_LOWERING_BASE,
|
|
1961
1964
|
recovery: [
|
|
@@ -6287,7 +6290,7 @@ function validateFinalFit(ir, profile, tokens) {
|
|
|
6287
6290
|
}
|
|
6288
6291
|
|
|
6289
6292
|
// src/version.ts
|
|
6290
|
-
var LIBRARY_VERSION = "2.0.0-alpha.
|
|
6293
|
+
var LIBRARY_VERSION = "2.0.0-alpha.89";
|
|
6291
6294
|
|
|
6292
6295
|
// src/pricing-brain.ts
|
|
6293
6296
|
function isPricingRow(x) {
|
|
@@ -6798,6 +6801,9 @@ function buildPayload(input, reg) {
|
|
|
6798
6801
|
// alpha.87 — explicit source wins; KGAUTO_SOURCE env fills the gap (the
|
|
6799
6802
|
// zero-code synthetic self-mark for harness/smoke processes).
|
|
6800
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,
|
|
6801
6807
|
error_type: input.errorType,
|
|
6802
6808
|
tools_called: input.toolsCalled,
|
|
6803
6809
|
oracle_score: input.oracleScore?.score,
|
|
@@ -8242,6 +8248,7 @@ async function call(ir, opts = {}) {
|
|
|
8242
8248
|
await record({
|
|
8243
8249
|
handle: initial.handle,
|
|
8244
8250
|
source: opts.source,
|
|
8251
|
+
endUserId: opts.endUserId,
|
|
8245
8252
|
tokensIn: 0,
|
|
8246
8253
|
tokensOut: 0,
|
|
8247
8254
|
latencyMs: latencyMs2,
|
|
@@ -8295,6 +8302,7 @@ async function call(ir, opts = {}) {
|
|
|
8295
8302
|
await record({
|
|
8296
8303
|
handle: initial.handle,
|
|
8297
8304
|
source: opts.source,
|
|
8305
|
+
endUserId: opts.endUserId,
|
|
8298
8306
|
tokensIn: 0,
|
|
8299
8307
|
tokensOut: 0,
|
|
8300
8308
|
latencyMs: latencyMs2,
|
|
@@ -8437,6 +8445,7 @@ async function call(ir, opts = {}) {
|
|
|
8437
8445
|
await record({
|
|
8438
8446
|
handle: initial.handle,
|
|
8439
8447
|
source: opts.source,
|
|
8448
|
+
endUserId: opts.endUserId,
|
|
8440
8449
|
tokensIn: validated.response.tokens.input,
|
|
8441
8450
|
tokensOut: validated.response.tokens.output,
|
|
8442
8451
|
latencyMs: latencyMs2,
|
|
@@ -8560,6 +8569,7 @@ async function call(ir, opts = {}) {
|
|
|
8560
8569
|
await record({
|
|
8561
8570
|
handle: initial.handle,
|
|
8562
8571
|
source: opts.source,
|
|
8572
|
+
endUserId: opts.endUserId,
|
|
8563
8573
|
tokensIn: lastErr?.tokens?.input ?? 0,
|
|
8564
8574
|
tokensOut: lastErr?.tokens?.output ?? 0,
|
|
8565
8575
|
latencyMs,
|
|
@@ -9033,6 +9043,7 @@ function extractKeptToolNames(result) {
|
|
|
9033
9043
|
|
|
9034
9044
|
// src/golden-eval.ts
|
|
9035
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";
|
|
9036
9047
|
var JUDGE_RUBRICS = {
|
|
9037
9048
|
summarize: "faithfulness to the source (no fabricated facts), coverage of the key points, concision, and adherence to the requested output format",
|
|
9038
9049
|
classify: "assignment of the correct category from the allowed set, and nothing outside the allowed set",
|
|
@@ -9251,7 +9262,7 @@ async function runGoldenEval(opts) {
|
|
|
9251
9262
|
const threshold = opts.winOrTieThreshold ?? 0.8;
|
|
9252
9263
|
const latencyFloorRatio = opts.latencyFloorRatio ?? 3;
|
|
9253
9264
|
const minJudgeable = opts.minJudgeableCases ?? 5;
|
|
9254
|
-
const judgeModel = opts.judgeModel ??
|
|
9265
|
+
const judgeModel = opts.judgeModel ?? DEFAULT_JUDGE_MODEL;
|
|
9255
9266
|
const notes = [];
|
|
9256
9267
|
const axis = opts.axis ?? "model";
|
|
9257
9268
|
const strategyId = opts.strategy;
|
|
@@ -11065,6 +11076,7 @@ function compile2(ir, opts) {
|
|
|
11065
11076
|
DECOMPOSITION_TEMPLATES,
|
|
11066
11077
|
DECOMPOSITION_TEMPLATES_VERSION,
|
|
11067
11078
|
DEFAULT_FINDINGS_ENDPOINT,
|
|
11079
|
+
DEFAULT_JUDGE_MODEL,
|
|
11068
11080
|
DEFAULT_MEASURED_FAILURE_ENDPOINT,
|
|
11069
11081
|
DEFAULT_PROMOTIONS_ENDPOINT,
|
|
11070
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
|
@@ -770,7 +770,7 @@ var PROFILES_RAW = [
|
|
|
770
770
|
// deepseek-reasoner registered as aliases (see ALIASES below).
|
|
771
771
|
{
|
|
772
772
|
id: "deepseek-v4-flash",
|
|
773
|
-
verifiedAgainstDocs: "2026-08-
|
|
773
|
+
verifiedAgainstDocs: "2026-08-24",
|
|
774
774
|
provider: "deepseek",
|
|
775
775
|
status: "current",
|
|
776
776
|
maxContextTokens: 1e6,
|
|
@@ -788,8 +788,8 @@ var PROFILES_RAW = [
|
|
|
788
788
|
reason: "Sequential tool calls only \u2014 L-040"
|
|
789
789
|
}
|
|
790
790
|
],
|
|
791
|
-
costInputPer1m: 0.
|
|
792
|
-
costOutputPer1m:
|
|
791
|
+
costInputPer1m: 0.44,
|
|
792
|
+
costOutputPer1m: 1.32,
|
|
793
793
|
lowering: {
|
|
794
794
|
system: { mode: "inline" },
|
|
795
795
|
cache: { strategy: "unsupported" },
|
|
@@ -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.
|
|
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.
|
|
@@ -843,7 +843,7 @@ var PROFILES_RAW = [
|
|
|
843
843
|
},
|
|
844
844
|
{
|
|
845
845
|
id: "deepseek-v4-pro",
|
|
846
|
-
verifiedAgainstDocs: "2026-08-
|
|
846
|
+
verifiedAgainstDocs: "2026-08-24",
|
|
847
847
|
provider: "deepseek",
|
|
848
848
|
status: "current",
|
|
849
849
|
maxContextTokens: 1e6,
|
|
@@ -867,8 +867,8 @@ var PROFILES_RAW = [
|
|
|
867
867
|
// cache-hit in; no promo caveat). The old $1.74/$3.48 "regular post-promo"
|
|
868
868
|
// was a projected revert that never happened (L-073/L-081) — it 4×-overstated
|
|
869
869
|
// cost and suppressed V4-Pro in cost-aware ranking.
|
|
870
|
-
costInputPer1m:
|
|
871
|
-
costOutputPer1m:
|
|
870
|
+
costInputPer1m: 1.32,
|
|
871
|
+
costOutputPer1m: 3.96,
|
|
872
872
|
lowering: {
|
|
873
873
|
system: { mode: "inline" },
|
|
874
874
|
cache: { strategy: "unsupported" },
|
|
@@ -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.
|
|
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: {
|
|
@@ -1505,7 +1505,7 @@ var PROFILES_RAW = [
|
|
|
1505
1505
|
// ─────────────────────────────────────────────────────────────────────────
|
|
1506
1506
|
{
|
|
1507
1507
|
id: "gemini-3.6-flash",
|
|
1508
|
-
verifiedAgainstDocs: "2026-08-
|
|
1508
|
+
verifiedAgainstDocs: "2026-08-24",
|
|
1509
1509
|
versionAdded: "2.0.0-alpha.87",
|
|
1510
1510
|
provider: "google",
|
|
1511
1511
|
status: "preview",
|
|
@@ -1543,8 +1543,10 @@ var PROFILES_RAW = [
|
|
|
1543
1543
|
reason: "Inherited from 2.5-flash s11 cliff: summarize+tools \u2192 empty response. Preemptive family guard until brain evidence on 3.6-flash specifically."
|
|
1544
1544
|
}
|
|
1545
1545
|
],
|
|
1546
|
-
costInputPer1m:
|
|
1547
|
-
|
|
1546
|
+
costInputPer1m: 0.75,
|
|
1547
|
+
// promo through 2026-12-31; $1.50 from 2027-01-01 (TIME_BOUNDED_PRICING_ROWS carries the transition)
|
|
1548
|
+
costOutputPer1m: 3.75,
|
|
1549
|
+
// promo through 2026-12-31; $7.50 from 2027-01-01
|
|
1548
1550
|
lowering: {
|
|
1549
1551
|
...GOOGLE_LOWERING_BASE,
|
|
1550
1552
|
cache: { ...GOOGLE_LOWERING_BASE.cache, discount: 0.1 },
|
|
@@ -1573,7 +1575,7 @@ var PROFILES_RAW = [
|
|
|
1573
1575
|
],
|
|
1574
1576
|
strengths: ["agentic_loops", "coding", "1m_context", "parallel_tools", "thinking_mode"],
|
|
1575
1577
|
weaknesses: ["no_brain_evidence_yet"],
|
|
1576
|
-
notes: "Hand-onboarded s86 (2026-08-14), step-zero verified: 1,048,576 ctx / 65,536 out (Google model page), $1.50/$7.50
|
|
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).",
|
|
1577
1579
|
archetypePerf: {
|
|
1578
1580
|
hunt: 9,
|
|
1579
1581
|
// Flash-family parallel-tool tier (L-040); inherited from 3.5-flash
|
|
@@ -1620,7 +1622,7 @@ var PROFILES_RAW = [
|
|
|
1620
1622
|
streaming: true,
|
|
1621
1623
|
cliffs: [],
|
|
1622
1624
|
costInputPer1m: 2,
|
|
1623
|
-
// 2026-08-
|
|
1625
|
+
// 2026-08-24: intro price made PERMANENT (docs: "the previously scheduled increase to $3/$15 on September 1, 2026 will not occur")
|
|
1624
1626
|
costOutputPer1m: 10,
|
|
1625
1627
|
lowering: ANTHROPIC_LOWERING_BASE,
|
|
1626
1628
|
recovery: [
|
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",
|