@hiper2d/ai-agents 0.38.0 → 0.40.0

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/index.mjs CHANGED
@@ -966,8 +966,8 @@ var SupportedAiModels = {
966
966
  // OpenAI documents the default for Sol, Luna and the 5.6 family but never states Astra's.
967
967
  // Pinning it changes no behaviour today; it stops the default moving under us silently.
968
968
  [LLM_CONSTANTS.GPT_ASTRA]: {
969
- displayName: "GPT-6 Astra",
970
- modelApiName: "gpt-6-astra",
969
+ displayName: "GPT-6.2 Astra",
970
+ modelApiName: "gpt-6.2-astra",
971
971
  apiKeyName: API_KEY_CONSTANTS.OPENAI,
972
972
  reasoningEffort: "medium",
973
973
  hasThinking: true,
@@ -985,9 +985,12 @@ var SupportedAiModels = {
985
985
  // Terra → GPT-6.1 Terra on 2026-10-26 (id verified against GET /v1/models; gpt-5.6-terra is
986
986
  // still served). Half the price of 5.6 Terra ($2/$12 → $1/$5) and back in its place below
987
987
  // Sol. Its no-param default effort measures `medium`, like its siblings.
988
+ // Astra → GPT-6.2 Astra and Sol → GPT-6.2 Sol on 2026-11-17 (both ids verified against
989
+ // GET /v1/models; gpt-6-astra and gpt-6.1-sol are still served). Both no-param defaults
990
+ // measure `medium`. Astra got 20% cheaper; Sol kept its price.
988
991
  [LLM_CONSTANTS.GPT_SOL]: {
989
- displayName: "GPT-6.1 Sol",
990
- modelApiName: "gpt-6.1-sol",
992
+ displayName: "GPT-6.2 Sol",
993
+ modelApiName: "gpt-6.2-sol",
991
994
  apiKeyName: API_KEY_CONSTANTS.OPENAI,
992
995
  reasoningEffort: "medium",
993
996
  hasThinking: true,
@@ -1279,16 +1282,25 @@ var SupportedAiModels = {
1279
1282
  // Speed tags from the 2026-08-05 live measurements (two samples each): plus 17.4s/14.5s,
1280
1283
  // flash 14.3s/16.4s (both slow); max 30.6s/100.5s — its latency tracks how long it decides
1281
1284
  // to think (4.2K reasoning tokens on the slow run), hence the budget cap below.
1285
+ //
1286
+ // `qwen-max` repointed 2026-11-16 from qwen3.8-max ($2/$6, hits $0.25) to qwen4-max
1287
+ // ($1.50/$4.50, hits $0.075; not yet in GET /models, but the chat endpoint serves it and
1288
+ // echoes the name). Probe on a real day-4 prompt (32K-char system prompt, ~10K tokens in,
1289
+ // cold cache, budget 1024, one call at a time): discussion 7-9s vs 5-7s, vote 9s vs 24s
1290
+ // (3.8 Max spent ~1,000 output tokens on a vote, 4 Max ~220), night action 12s vs 17s,
1291
+ // end-of-day notes 33s vs 34s; every answer valid through the shared schema. Three
1292
+ // identical calls sent AT ONCE timed out on both models (night action, notes) — the account
1293
+ // under a burst, not the model. The 'slow' tag is 3.8 Max's production grade (p90 26.2s);
1294
+ // re-grade from production turns once 4 Max has 100+.
1282
1295
  [LLM_CONSTANTS.QWEN_MAX]: {
1283
- displayName: "Qwen3.8 Max",
1284
- modelApiName: "qwen3.8-max",
1296
+ displayName: "Qwen4 Max",
1297
+ modelApiName: "qwen4-max",
1285
1298
  apiKeyName: API_KEY_CONSTANTS.QWEN,
1286
1299
  hasThinking: true,
1287
1300
  temperature: 0.7,
1288
- // Caps `thinking_budget` to bound the 30–100s latency variance. The same knob works on
1289
- // the 3.7 models (verified live) — add it to their entries if they ever need taming.
1301
+ // Caps `thinking_budget` to bound the latency variance (3.8 Max swung 30–100s without
1302
+ // it). Uncapped, 4 Max's vote went 9s → 14s.
1290
1303
  thinkingBudgetTokens: 1024,
1291
- // Capped it measures 25-26s → the >25s tier.
1292
1304
  tags: ["slow"]
1293
1305
  },
1294
1306
  // qwen3.8-flash replaced qwen3.7-flash on 2026-08-30 (same 1M context, 128k max output);
@@ -1427,13 +1439,16 @@ var MODEL_PRICING = {
1427
1439
  // short context, $20/$75 cache-hit $2 long context. OpenAI's pricing table doesn't restate
1428
1440
  // the boundary; we assume the same 272k threshold as the GPT-5.6 siblings. Cache writes
1429
1441
  // ($12.50/$25) are not modelled — caching is automatic and we only see hits.
1442
+ // GPT-6.2 Astra (pricing page, read 2026-11-17, both Short and Long tabs): $8/$40 cache-hit
1443
+ // $0.40 short, $16/$60 cache-hit $0.80 long; cache writes $10/$20 (30m/12h), not modelled.
1444
+ // The cache hit is now 5% of input, like Sol and Terra (6 Astra was 10%).
1430
1445
  [SupportedAiModels[LLM_CONSTANTS.GPT_ASTRA].modelApiName]: {
1431
- inputPrice: 10,
1432
- outputPrice: 50,
1433
- cacheHitPrice: 1,
1434
- extendedContextInputPrice: 20,
1435
- extendedContextOutputPrice: 75,
1436
- extendedContextCacheHitPrice: 2,
1446
+ inputPrice: 8,
1447
+ outputPrice: 40,
1448
+ cacheHitPrice: 0.4,
1449
+ extendedContextInputPrice: 16,
1450
+ extendedContextOutputPrice: 60,
1451
+ extendedContextCacheHitPrice: 0.8,
1437
1452
  extendedContextThresholdTokens: 272e3
1438
1453
  },
1439
1454
  // OpenAI GPT-6 Sol and GPT-6 Luna (developers.openai.com/api/docs/pricing, read 2026-09-22).
@@ -1446,6 +1461,7 @@ var MODEL_PRICING = {
1446
1461
  // Re-read 2026-10-26 with the Terra launch: the page now lists Sol's cache hit at $0.10
1447
1462
  // (5% of input, the same share as Terra), not the $0.20 recorded on 2026-09-22; Astra and
1448
1463
  // Luna still show 10%. Followed the page.
1464
+ // GPT-6.2 Sol (read 2026-11-17, Short and Long tabs) is priced exactly like 6.1 Sol.
1449
1465
  [SupportedAiModels[LLM_CONSTANTS.GPT_SOL].modelApiName]: {
1450
1466
  inputPrice: 2,
1451
1467
  outputPrice: 10,
@@ -1642,10 +1658,11 @@ var MODEL_PRICING = {
1642
1658
  // $0.016. Neither has input-length tiers (the tier column is "-" for both). These
1643
1659
  // published cached rates supersede the 20%-of-input rule charged before 2026-08-30; we
1644
1660
  // still don't send explicit cache_control.
1661
+ // qwen4-max (2026-11-16, same page): $1.50/$4.50, implicit-cache hits $0.075, no tiers.
1645
1662
  [SupportedAiModels[LLM_CONSTANTS.QWEN_MAX].modelApiName]: {
1646
- inputPrice: 2,
1647
- outputPrice: 6,
1648
- cacheHitPrice: 0.25
1663
+ inputPrice: 1.5,
1664
+ outputPrice: 4.5,
1665
+ cacheHitPrice: 0.075
1649
1666
  },
1650
1667
  [SupportedAiModels[LLM_CONSTANTS.QWEN_FLASH].modelApiName]: {
1651
1668
  inputPrice: 0.15,