@warmdrift/kgauto-compiler 2.0.0-alpha.87 → 2.0.0-alpha.88

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.
Files changed (14) hide show
  1. package/dist/{chunk-G3MKT7IR.mjs → chunk-DSHWJFM2.mjs} +1 -1
  2. package/dist/{chunk-GMZT4ZXD.mjs → chunk-GCOVYPJA.mjs} +1 -1
  3. package/dist/{chunk-YVKTTTTB.mjs → chunk-UNTUOZAG.mjs} +23 -21
  4. package/dist/glassbox-routes/index.js +23 -21
  5. package/dist/glassbox-routes/index.mjs +2 -2
  6. package/dist/index.d.mts +1 -1
  7. package/dist/index.d.ts +1 -1
  8. package/dist/index.js +24 -22
  9. package/dist/index.mjs +3 -3
  10. package/dist/key-health.js +1 -1
  11. package/dist/key-health.mjs +1 -1
  12. package/dist/profiles.js +23 -21
  13. package/dist/profiles.mjs +1 -1
  14. package/package.json +1 -1
package/dist/{chunk-G3MKT7IR.mjs → chunk-DSHWJFM2.mjs} RENAMED
@@ -2,7 +2,7 @@ import {
2
2
  allProfiles,
3
3
  resolveModelAlias,
4
4
  tryGetProfile
5
- } from "./chunk-YVKTTTTB.mjs";
5
+ } from "./chunk-UNTUOZAG.mjs";
6
6
 
7
7
  // src/brain-query.ts
8
8
  var FRESH_SNAPSHOT = {
package/dist/{chunk-GMZT4ZXD.mjs → chunk-GCOVYPJA.mjs} RENAMED
@@ -1,5 +1,5 @@
1
1
  // src/version.ts
2
- var LIBRARY_VERSION = "2.0.0-alpha.87";
2
+ var LIBRARY_VERSION = "2.0.0-alpha.88";
3
3
 
4
4
  // src/key-health.ts
5
5
  var JSON_HEADERS = { "Content-Type": "application/json" };
package/dist/{chunk-YVKTTTTB.mjs → chunk-UNTUOZAG.mjs} RENAMED
@@ -211,7 +211,7 @@ var PROFILES_RAW = [
211
211
  },
212
212
  {
213
213
  id: "claude-opus-4-7",
214
- verifiedAgainstDocs: "2026-08-14",
214
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
264
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
309
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
350
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
413
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
519
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
676
+ verifiedAgainstDocs: "2026-08-24",
677
677
  provider: "google",
678
678
  status: "preview",
679
679
  maxContextTokens: 1048576,
@@ -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: "2026-08-14",
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: 0.14,
755
- costOutputPer1m: 0.28,
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. PRICING MODEL CHANGE effective 2026-08-16: DeepSeek moves to peak/off-peak billing, off-peak at HALF the peak rate (docs, fetched 2026-08-14). Encoded prices are peak; cost math for off-peak traffic will overstate by up to 2x until time-of-day pricing is modeled.",
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 shipped the DEAD pre-transition prices ($0.14/$0.28 flash, $0.435/$0.87 pro) with a fresh verification stamp \u2014 corrected here (alpha.88). 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: "2026-08-14",
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: 0.435,
834
- costOutputPer1m: 0.87,
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. PRICING MODEL CHANGE effective 2026-08-16: DeepSeek moves to peak/off-peak billing, off-peak at HALF the peak rate (docs, fetched 2026-08-14). Encoded prices are peak; cost math for off-peak traffic will overstate by up to 2x until time-of-day pricing is modeled.",
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 shipped the DEAD pre-transition prices ($0.14/$0.28 flash, $0.435/$0.87 pro) with a fresh verification stamp \u2014 corrected here (alpha.88). 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: "2026-08-14",
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: 1.5,
1510
- costOutputPer1m: 7.5,
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 per 1M (catalog-sync two-agreeing-extraction same day \u2014 CHEAPER output than 3.5-flash at $9.00). 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; sat in the onboarding queue 18 days \u2014 drained by the s86 catalog-sync run.",
1541
+ notes: "Hand-onboarded s86 (2026-08-24; the session ran under a 10-day-slow machine clock, hence 08-14 stamps in kin artifacts), step-zero verified: 1,048,576 ctx / 65,536 out (Google model page). PRICING (page 2026-08-24): $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); ~31 days in the onboarding queue.",
1540
1542
  archetypePerf: {
1541
1543
  hunt: 9,
1542
1544
  // Flash-family parallel-tool tier (L-040); inherited from 3.5-flash
@@ -1570,7 +1572,7 @@ var PROFILES_RAW = [
1570
1572
  // 'preview' per the Fable precedent — no brain evidence yet; promotion to
1571
1573
  // 'current' is an explicit call. (L-049/L-081.)
1572
1574
  id: "claude-sonnet-5",
1573
- verifiedAgainstDocs: "2026-08-14",
1575
+ verifiedAgainstDocs: "2026-08-24",
1574
1576
  provider: "anthropic",
1575
1577
  status: "current",
1576
1578
  // s76 2026-07-25: GA per docs; was preview.
@@ -1583,7 +1585,7 @@ var PROFILES_RAW = [
1583
1585
  streaming: true,
1584
1586
  cliffs: [],
1585
1587
  costInputPer1m: 2,
1586
- // 2026-08-14: intro price made PERMANENT (docs: "the previously scheduled increase to $3/$15 on September 1, 2026 will not occur")
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: [
package/dist/glassbox-routes/index.js CHANGED
@@ -265,7 +265,7 @@ var PROFILES_RAW = [
265
265
  },
266
266
  {
267
267
  id: "claude-opus-4-7",
268
- verifiedAgainstDocs: "2026-08-14",
268
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
318
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
363
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
404
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
467
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
573
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
730
+ verifiedAgainstDocs: "2026-08-24",
731
731
  provider: "google",
732
732
  status: "preview",
733
733
  maxContextTokens: 1048576,
@@ -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: "2026-08-14",
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: 0.14,
809
- costOutputPer1m: 0.28,
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. PRICING MODEL CHANGE effective 2026-08-16: DeepSeek moves to peak/off-peak billing, off-peak at HALF the peak rate (docs, fetched 2026-08-14). Encoded prices are peak; cost math for off-peak traffic will overstate by up to 2x until time-of-day pricing is modeled.",
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 shipped the DEAD pre-transition prices ($0.14/$0.28 flash, $0.435/$0.87 pro) with a fresh verification stamp \u2014 corrected here (alpha.88). 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: "2026-08-14",
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: 0.435,
888
- costOutputPer1m: 0.87,
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. PRICING MODEL CHANGE effective 2026-08-16: DeepSeek moves to peak/off-peak billing, off-peak at HALF the peak rate (docs, fetched 2026-08-14). Encoded prices are peak; cost math for off-peak traffic will overstate by up to 2x until time-of-day pricing is modeled.",
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 shipped the DEAD pre-transition prices ($0.14/$0.28 flash, $0.435/$0.87 pro) with a fresh verification stamp \u2014 corrected here (alpha.88). 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: "2026-08-14",
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: 1.5,
1564
- costOutputPer1m: 7.5,
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 per 1M (catalog-sync two-agreeing-extraction same day \u2014 CHEAPER output than 3.5-flash at $9.00). 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; sat in the onboarding queue 18 days \u2014 drained by the s86 catalog-sync run.",
1595
+ notes: "Hand-onboarded s86 (2026-08-24; the session ran under a 10-day-slow machine clock, hence 08-14 stamps in kin artifacts), step-zero verified: 1,048,576 ctx / 65,536 out (Google model page). PRICING (page 2026-08-24): $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); ~31 days in the onboarding queue.",
1594
1596
  archetypePerf: {
1595
1597
  hunt: 9,
1596
1598
  // Flash-family parallel-tool tier (L-040); inherited from 3.5-flash
@@ -1624,7 +1626,7 @@ var PROFILES_RAW = [
1624
1626
  // 'preview' per the Fable precedent — no brain evidence yet; promotion to
1625
1627
  // 'current' is an explicit call. (L-049/L-081.)
1626
1628
  id: "claude-sonnet-5",
1627
- verifiedAgainstDocs: "2026-08-14",
1629
+ verifiedAgainstDocs: "2026-08-24",
1628
1630
  provider: "anthropic",
1629
1631
  status: "current",
1630
1632
  // s76 2026-07-25: GA per docs; was preview.
@@ -1637,7 +1639,7 @@ var PROFILES_RAW = [
1637
1639
  streaming: true,
1638
1640
  cliffs: [],
1639
1641
  costInputPer1m: 2,
1640
- // 2026-08-14: intro price made PERMANENT (docs: "the previously scheduled increase to $3/$15 on September 1, 2026 will not occur")
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: [
package/dist/glassbox-routes/index.mjs CHANGED
@@ -1,10 +1,10 @@
1
1
  import {
2
2
  ARCHETYPE_FLOOR_DEFAULT,
3
3
  getDefaultFallbackChain
4
- } from "../chunk-G3MKT7IR.mjs";
4
+ } from "../chunk-DSHWJFM2.mjs";
5
5
  import {
6
6
  tryGetProfile
7
- } from "../chunk-YVKTTTTB.mjs";
7
+ } from "../chunk-UNTUOZAG.mjs";
8
8
  import {
9
9
  subscribe,
10
10
  subscribeApp
package/dist/index.d.mts CHANGED
@@ -1250,7 +1250,7 @@ declare function runStrategyEvalWithAttribution(opts: Omit<GoldenEvalOptions, 'a
1250
1250
  * guard in `tests/version.test.ts` fails the suite (and therefore
1251
1251
  * `prepublishOnly`) when they diverge — a stale constant cannot reach npm.
1252
1252
  */
1253
- declare const LIBRARY_VERSION = "2.0.0-alpha.87";
1253
+ declare const LIBRARY_VERSION = "2.0.0-alpha.88";
1254
1254
 
1255
1255
  /**
1256
1256
  * Oracle contract — how an app tells the brain whether a response was good.
package/dist/index.d.ts CHANGED
@@ -1250,7 +1250,7 @@ declare function runStrategyEvalWithAttribution(opts: Omit<GoldenEvalOptions, 'a
1250
1250
  * guard in `tests/version.test.ts` fails the suite (and therefore
1251
1251
  * `prepublishOnly`) when they diverge — a stale constant cannot reach npm.
1252
1252
  */
1253
- declare const LIBRARY_VERSION = "2.0.0-alpha.87";
1253
+ declare const LIBRARY_VERSION = "2.0.0-alpha.88";
1254
1254
 
1255
1255
  /**
1256
1256
  * Oracle contract — how an app tells the brain whether a response was good.
package/dist/index.js CHANGED
@@ -583,7 +583,7 @@ var PROFILES_RAW = [
583
583
  },
584
584
  {
585
585
  id: "claude-opus-4-7",
586
- verifiedAgainstDocs: "2026-08-14",
586
+ verifiedAgainstDocs: "2026-08-24",
587
587
  provider: "anthropic",
588
588
  status: "legacy",
589
589
  // 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 +633,7 @@ var PROFILES_RAW = [
633
633
  },
634
634
  {
635
635
  id: "claude-opus-4-6",
636
- verifiedAgainstDocs: "2026-08-14",
636
+ verifiedAgainstDocs: "2026-08-24",
637
637
  provider: "anthropic",
638
638
  status: "legacy",
639
639
  maxContextTokens: 1e6,
@@ -678,7 +678,7 @@ var PROFILES_RAW = [
678
678
  },
679
679
  {
680
680
  id: "claude-sonnet-4-6",
681
- verifiedAgainstDocs: "2026-08-14",
681
+ verifiedAgainstDocs: "2026-08-24",
682
682
  provider: "anthropic",
683
683
  status: "legacy",
684
684
  // s76 2026-07-25: docs list as LEGACY (Sonnet 5 is current).
@@ -719,7 +719,7 @@ var PROFILES_RAW = [
719
719
  },
720
720
  {
721
721
  id: "claude-haiku-4-5",
722
- verifiedAgainstDocs: "2026-08-14",
722
+ verifiedAgainstDocs: "2026-08-24",
723
723
  provider: "anthropic",
724
724
  status: "current",
725
725
  maxContextTokens: 2e5,
@@ -782,7 +782,7 @@ var PROFILES_RAW = [
782
782
  // ── Google ──
783
783
  {
784
784
  id: "gemini-2.5-flash",
785
- verifiedAgainstDocs: "2026-08-14",
785
+ verifiedAgainstDocs: "2026-08-24",
786
786
  provider: "google",
787
787
  status: "current",
788
788
  maxContextTokens: 1048576,
@@ -888,7 +888,7 @@ var PROFILES_RAW = [
888
888
  // thresholds. The brain will validate/relax these as evidence accumulates
889
889
  // per (archetype, model) tuple. Currently ZERO brain rows for this model.
890
890
  id: "gemini-2.5-flash-lite",
891
- verifiedAgainstDocs: "2026-08-14",
891
+ verifiedAgainstDocs: "2026-08-24",
892
892
  provider: "google",
893
893
  status: "current",
894
894
  maxContextTokens: 1048576,
@@ -1045,7 +1045,7 @@ var PROFILES_RAW = [
1045
1045
  },
1046
1046
  {
1047
1047
  id: "gemini-3.1-pro-preview",
1048
- verifiedAgainstDocs: "2026-08-14",
1048
+ verifiedAgainstDocs: "2026-08-24",
1049
1049
  provider: "google",
1050
1050
  status: "preview",
1051
1051
  maxContextTokens: 1048576,
@@ -1105,7 +1105,7 @@ var PROFILES_RAW = [
1105
1105
  // deepseek-reasoner registered as aliases (see ALIASES below).
1106
1106
  {
1107
1107
  id: "deepseek-v4-flash",
1108
- verifiedAgainstDocs: "2026-08-14",
1108
+ verifiedAgainstDocs: "2026-08-24",
1109
1109
  provider: "deepseek",
1110
1110
  status: "current",
1111
1111
  maxContextTokens: 1e6,
@@ -1123,8 +1123,8 @@ var PROFILES_RAW = [
1123
1123
  reason: "Sequential tool calls only \u2014 L-040"
1124
1124
  }
1125
1125
  ],
1126
- costInputPer1m: 0.14,
1127
- costOutputPer1m: 0.28,
1126
+ costInputPer1m: 0.44,
1127
+ costOutputPer1m: 1.32,
1128
1128
  lowering: {
1129
1129
  system: { mode: "inline" },
1130
1130
  cache: { strategy: "unsupported" },
@@ -1143,7 +1143,7 @@ var PROFILES_RAW = [
1143
1143
  // baseQuality (it carries 'reasoning') with no latency counterweight,
1144
1144
  // leapfrogged gemini-2.5-flash as PB's summarize leader once reachable.
1145
1145
  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. PRICING MODEL CHANGE effective 2026-08-16: DeepSeek moves to peak/off-peak billing, off-peak at HALF the peak rate (docs, fetched 2026-08-14). Encoded prices are peak; cost math for off-peak traffic will overstate by up to 2x until time-of-day pricing is modeled.",
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 shipped the DEAD pre-transition prices ($0.14/$0.28 flash, $0.435/$0.87 pro) with a fresh verification stamp \u2014 corrected here (alpha.88). 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
1147
  // Master plan §6.2 anchor. Brain-validated tier 1 cross-provider for
1148
1148
  // classify (169 rows, 0% empty). Tier 0 for summarize-with-no-tools.
1149
1149
  // Falls off on hunt (sequential tools — L-040) and reasoning depth.
@@ -1178,7 +1178,7 @@ var PROFILES_RAW = [
1178
1178
  },
1179
1179
  {
1180
1180
  id: "deepseek-v4-pro",
1181
- verifiedAgainstDocs: "2026-08-14",
1181
+ verifiedAgainstDocs: "2026-08-24",
1182
1182
  provider: "deepseek",
1183
1183
  status: "current",
1184
1184
  maxContextTokens: 1e6,
@@ -1202,8 +1202,8 @@ var PROFILES_RAW = [
1202
1202
  // cache-hit in; no promo caveat). The old $1.74/$3.48 "regular post-promo"
1203
1203
  // was a projected revert that never happened (L-073/L-081) — it 4×-overstated
1204
1204
  // cost and suppressed V4-Pro in cost-aware ranking.
1205
- costInputPer1m: 0.435,
1206
- costOutputPer1m: 0.87,
1205
+ costInputPer1m: 1.32,
1206
+ costOutputPer1m: 3.96,
1207
1207
  lowering: {
1208
1208
  system: { mode: "inline" },
1209
1209
  cache: { strategy: "unsupported" },
@@ -1219,7 +1219,7 @@ var PROFILES_RAW = [
1219
1219
  // of the served set. Tag derivation would say 'medium'; the measurement says
1220
1220
  // otherwise.
1221
1221
  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. PRICING MODEL CHANGE effective 2026-08-16: DeepSeek moves to peak/off-peak billing, off-peak at HALF the peak rate (docs, fetched 2026-08-14). Encoded prices are peak; cost math for off-peak traffic will overstate by up to 2x until time-of-day pricing is modeled.",
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 shipped the DEAD pre-transition prices ($0.14/$0.28 flash, $0.435/$0.87 pro) with a fresh verification stamp \u2014 corrected here (alpha.88). 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
1223
  // Master plan §3.3: tier 3 cross-provider for plan chain. Reasoning
1224
1224
  // bumped one notch over V4-Flash; same parallel-tool ceiling.
1225
1225
  archetypePerf: {
@@ -1840,7 +1840,7 @@ var PROFILES_RAW = [
1840
1840
  // ─────────────────────────────────────────────────────────────────────────
1841
1841
  {
1842
1842
  id: "gemini-3.6-flash",
1843
- verifiedAgainstDocs: "2026-08-14",
1843
+ verifiedAgainstDocs: "2026-08-24",
1844
1844
  versionAdded: "2.0.0-alpha.87",
1845
1845
  provider: "google",
1846
1846
  status: "preview",
@@ -1878,8 +1878,10 @@ var PROFILES_RAW = [
1878
1878
  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
1879
  }
1880
1880
  ],
1881
- costInputPer1m: 1.5,
1882
- costOutputPer1m: 7.5,
1881
+ costInputPer1m: 0.75,
1882
+ // promo through 2026-12-31; $1.50 from 2027-01-01 (TIME_BOUNDED_PRICING_ROWS carries the transition)
1883
+ costOutputPer1m: 3.75,
1884
+ // promo through 2026-12-31; $7.50 from 2027-01-01
1883
1885
  lowering: {
1884
1886
  ...GOOGLE_LOWERING_BASE,
1885
1887
  cache: { ...GOOGLE_LOWERING_BASE.cache, discount: 0.1 },
@@ -1908,7 +1910,7 @@ var PROFILES_RAW = [
1908
1910
  ],
1909
1911
  strengths: ["agentic_loops", "coding", "1m_context", "parallel_tools", "thinking_mode"],
1910
1912
  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 per 1M (catalog-sync two-agreeing-extraction same day \u2014 CHEAPER output than 3.5-flash at $9.00). 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; sat in the onboarding queue 18 days \u2014 drained by the s86 catalog-sync run.",
1913
+ notes: "Hand-onboarded s86 (2026-08-24; the session ran under a 10-day-slow machine clock, hence 08-14 stamps in kin artifacts), step-zero verified: 1,048,576 ctx / 65,536 out (Google model page). PRICING (page 2026-08-24): $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); ~31 days in the onboarding queue.",
1912
1914
  archetypePerf: {
1913
1915
  hunt: 9,
1914
1916
  // Flash-family parallel-tool tier (L-040); inherited from 3.5-flash
@@ -1942,7 +1944,7 @@ var PROFILES_RAW = [
1942
1944
  // 'preview' per the Fable precedent — no brain evidence yet; promotion to
1943
1945
  // 'current' is an explicit call. (L-049/L-081.)
1944
1946
  id: "claude-sonnet-5",
1945
- verifiedAgainstDocs: "2026-08-14",
1947
+ verifiedAgainstDocs: "2026-08-24",
1946
1948
  provider: "anthropic",
1947
1949
  status: "current",
1948
1950
  // s76 2026-07-25: GA per docs; was preview.
@@ -1955,7 +1957,7 @@ var PROFILES_RAW = [
1955
1957
  streaming: true,
1956
1958
  cliffs: [],
1957
1959
  costInputPer1m: 2,
1958
- // 2026-08-14: intro price made PERMANENT (docs: "the previously scheduled increase to $3/$15 on September 1, 2026 will not occur")
1960
+ // 2026-08-24: intro price made PERMANENT (docs: "the previously scheduled increase to $3/$15 on September 1, 2026 will not occur")
1959
1961
  costOutputPer1m: 10,
1960
1962
  lowering: ANTHROPIC_LOWERING_BASE,
1961
1963
  recovery: [
@@ -6287,7 +6289,7 @@ function validateFinalFit(ir, profile, tokens) {
6287
6289
  }
6288
6290
 
6289
6291
  // src/version.ts
6290
- var LIBRARY_VERSION = "2.0.0-alpha.87";
6292
+ var LIBRARY_VERSION = "2.0.0-alpha.88";
6291
6293
 
6292
6294
  // src/pricing-brain.ts
6293
6295
  function isPricingRow(x) {
package/dist/index.mjs CHANGED
@@ -19,7 +19,7 @@ import {
19
19
  LIBRARY_VERSION,
20
20
  createKeyHealthRoute,
21
21
  keyFingerprint
22
- } from "./chunk-GMZT4ZXD.mjs";
22
+ } from "./chunk-GCOVYPJA.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-G3MKT7IR.mjs";
60
+ } from "./chunk-DSHWJFM2.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-YVKTTTTB.mjs";
76
+ } from "./chunk-UNTUOZAG.mjs";
77
77
  import {
78
78
  emitAdvisoryFired,
79
79
  emitCompileDone,
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.87";
31
+ var LIBRARY_VERSION = "2.0.0-alpha.88";
32
32
 
33
33
  // src/key-health.ts
34
34
  var JSON_HEADERS = { "Content-Type": "application/json" };
package/dist/key-health.mjs CHANGED
@@ -3,7 +3,7 @@ import {
3
3
  KEY_FINGERPRINT_LENGTH,
4
4
  createKeyHealthRoute,
5
5
  keyFingerprint
6
- } from "./chunk-GMZT4ZXD.mjs";
6
+ } from "./chunk-GCOVYPJA.mjs";
7
7
  export {
8
8
  KEY_FINGERPRINT_DOMAIN,
9
9
  KEY_FINGERPRINT_LENGTH,
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: "2026-08-14",
251
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
301
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
346
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
387
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
450
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
556
+ verifiedAgainstDocs: "2026-08-24",
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: "2026-08-14",
713
+ verifiedAgainstDocs: "2026-08-24",
714
714
  provider: "google",
715
715
  status: "preview",
716
716
  maxContextTokens: 1048576,
@@ -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: "2026-08-14",
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: 0.14,
792
- costOutputPer1m: 0.28,
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. PRICING MODEL CHANGE effective 2026-08-16: DeepSeek moves to peak/off-peak billing, off-peak at HALF the peak rate (docs, fetched 2026-08-14). Encoded prices are peak; cost math for off-peak traffic will overstate by up to 2x until time-of-day pricing is modeled.",
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 shipped the DEAD pre-transition prices ($0.14/$0.28 flash, $0.435/$0.87 pro) with a fresh verification stamp \u2014 corrected here (alpha.88). 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: "2026-08-14",
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: 0.435,
871
- costOutputPer1m: 0.87,
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. PRICING MODEL CHANGE effective 2026-08-16: DeepSeek moves to peak/off-peak billing, off-peak at HALF the peak rate (docs, fetched 2026-08-14). Encoded prices are peak; cost math for off-peak traffic will overstate by up to 2x until time-of-day pricing is modeled.",
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 shipped the DEAD pre-transition prices ($0.14/$0.28 flash, $0.435/$0.87 pro) with a fresh verification stamp \u2014 corrected here (alpha.88). 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: "2026-08-14",
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: 1.5,
1547
- costOutputPer1m: 7.5,
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 per 1M (catalog-sync two-agreeing-extraction same day \u2014 CHEAPER output than 3.5-flash at $9.00). 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; sat in the onboarding queue 18 days \u2014 drained by the s86 catalog-sync run.",
1578
+ notes: "Hand-onboarded s86 (2026-08-24; the session ran under a 10-day-slow machine clock, hence 08-14 stamps in kin artifacts), step-zero verified: 1,048,576 ctx / 65,536 out (Google model page). PRICING (page 2026-08-24): $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); ~31 days in the onboarding queue.",
1577
1579
  archetypePerf: {
1578
1580
  hunt: 9,
1579
1581
  // Flash-family parallel-tool tier (L-040); inherited from 3.5-flash
@@ -1607,7 +1609,7 @@ var PROFILES_RAW = [
1607
1609
  // 'preview' per the Fable precedent — no brain evidence yet; promotion to
1608
1610
  // 'current' is an explicit call. (L-049/L-081.)
1609
1611
  id: "claude-sonnet-5",
1610
- verifiedAgainstDocs: "2026-08-14",
1612
+ verifiedAgainstDocs: "2026-08-24",
1611
1613
  provider: "anthropic",
1612
1614
  status: "current",
1613
1615
  // s76 2026-07-25: GA per docs; was preview.
@@ -1620,7 +1622,7 @@ var PROFILES_RAW = [
1620
1622
  streaming: true,
1621
1623
  cliffs: [],
1622
1624
  costInputPer1m: 2,
1623
- // 2026-08-14: intro price made PERMANENT (docs: "the previously scheduled increase to $3/$15 on September 1, 2026 will not occur")
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
@@ -13,7 +13,7 @@ import {
13
13
  registerProfiles,
14
14
  resolveModelAlias,
15
15
  tryGetProfile
16
- } from "./chunk-YVKTTTTB.mjs";
16
+ } from "./chunk-UNTUOZAG.mjs";
17
17
  export {
18
18
  ALIASES,
19
19
  LATENCY_TIER_MS,
package/package.json CHANGED
@@ -1,6 +1,6 @@
1
1
  {
2
2
  "name": "@warmdrift/kgauto-compiler",
3
- "version": "2.0.0-alpha.87",
3
+ "version": "2.0.0-alpha.88",
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",