@hiper2d/ai-agents 0.26.3 → 0.28.0

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
@@ -846,6 +846,9 @@ var LLM_CONSTANTS = {
846
846
  MISTRAL_SMALL: "mistral-small",
847
847
  GROK: "grok",
848
848
  KIMI: "kimi",
849
+ // Kimi K3.1 Pro Highspeed: the same model served 2-3x faster at 2x the price (added
850
+ // 2026-10-23). Version-free like `kimi`, so a K3.2 repoint keeps both ids.
851
+ KIMI_HIGHSPEED: "kimi-highspeed",
849
852
  GLM: "glm",
850
853
  GLM_FLASH: "glm-flash",
851
854
  FUGU_MAX: "fugu-max",
@@ -967,6 +970,9 @@ var SupportedAiModels = {
967
970
  // slot largely redundant — retiring it is a product call, not a catalog one.
968
971
  // Sol → GPT-6.1 Sol on 2026-09-29 (id verified against GET /v1/models; gpt-6-sol is still
969
972
  // served). Same price as GPT-6 Sol; its no-param default effort still measures `medium`.
973
+ // Terra → GPT-6.1 Terra on 2026-10-26 (id verified against GET /v1/models; gpt-5.6-terra is
974
+ // still served). Half the price of 5.6 Terra ($2/$12 → $1/$5) and back in its place below
975
+ // Sol. Its no-param default effort measures `medium`, like its siblings.
970
976
  [LLM_CONSTANTS.GPT_SOL]: {
971
977
  displayName: "GPT-6.1 Sol",
972
978
  modelApiName: "gpt-6.1-sol",
@@ -977,8 +983,8 @@ var SupportedAiModels = {
977
983
  tags: ["expensive"]
978
984
  },
979
985
  [LLM_CONSTANTS.GPT]: {
980
- displayName: "GPT-5.6 Terra",
981
- modelApiName: "gpt-5.6-terra",
986
+ displayName: "GPT-6.1 Terra",
987
+ modelApiName: "gpt-6.1-terra",
982
988
  apiKeyName: API_KEY_CONSTANTS.OPENAI,
983
989
  reasoningEffort: "medium",
984
990
  hasThinking: true,
@@ -1113,21 +1119,40 @@ var SupportedAiModels = {
1113
1119
  reasoningEffort: "high",
1114
1120
  tags: ["very-fast", "cheap"]
1115
1121
  },
1116
- // Kimi models. Single always-reasoning entry: K3 reasons by default and the only way to stop
1117
- // it is the undocumented K2-era `thinking: disabled` toggle, which we no longer rely on.
1118
- // K3 always reasons at max effort; ~85-90% of its output tokens are reasoning tokens billed
1119
- // at the output rate, so real per-request cost runs well above the sticker output price.
1122
+ // Kimi models. Always-reasoning entries: K3-series models reason by default and the only way
1123
+ // to stop it is the undocumented K2-era `thinking: disabled` toggle, which we no longer rely on.
1124
+ // ~85-90% of output tokens are reasoning tokens billed at the output rate, so real
1125
+ // per-request cost runs well above the sticker output price.
1126
+ //
1127
+ // `kimi` repointed 2026-10-23 from kimi-k3 ($3/$15, 17-29s per turn, 146s for the
1128
+ // 15-character story pipeline) to Kimi K3.1 Pro ($2/$10, 1M context, same low|high|max
1129
+ // effort levels, default max). Live at high effort 2026-10-23: 3-sentence schema ask
1130
+ // 4.2-4.5s, 8-character scene (476-845 output tokens) 9.7-14.9s = 49-57 tok/s — the same
1131
+ // output speed as K3, so a full bot turn still lands in the 15-25s `slow` band.
1132
+ // Consumer policy (werewolf): 1 bot per free game (opt-in) until 100 turns are measured.
1120
1133
  [LLM_CONSTANTS.KIMI]: {
1121
- displayName: "Kimi K3",
1122
- modelApiName: "kimi-k3",
1134
+ displayName: "Kimi K3.1 Pro",
1135
+ modelApiName: "kimi-k3.1-pro",
1123
1136
  apiKeyName: API_KEY_CONSTANTS.MOONSHOT,
1124
1137
  hasThinking: true,
1125
- // Temperature is omitted from the request: kimi-k3 rejects any value other than 1.
1126
- // Speed samples: 17s (2026-08-04) and 28.9s (2026-08-05) — graded into the >25s tier.
1138
+ // Temperature is omitted from the request: K3 rejected any value other than 1.
1127
1139
  // Pinned 'high' 2026-09-22, down from Moonshot's 'max' default: K3 had no levels at
1128
1140
  // launch and the agent hardcoded max, so this is the first time the knob exists.
1129
1141
  reasoningEffort: "high",
1130
- tags: ["very-slow", "expensive"]
1142
+ tags: ["slow", "expensive"]
1143
+ },
1144
+ // Same model, Moonshot's high-speed serving tier ("roughly two to three times faster",
1145
+ // platform.kimi.ai K3.1 Pro quickstart, read 2026-10-23) at 2x the price ($4/$20 — Opus 5.5
1146
+ // money). Live the same day: schema ask 1.9-2.4s, 8-character scene (789-982 output tokens)
1147
+ // 7.9-8.4s = 100-117 tok/s, twice Pro. No speed tag: that is the unlabeled 6-13s middle.
1148
+ // Consumer policy (werewolf): paid tier only, deliberately — the estimate is 5.2¢ a turn.
1149
+ [LLM_CONSTANTS.KIMI_HIGHSPEED]: {
1150
+ displayName: "Kimi K3.1 Pro Highspeed",
1151
+ modelApiName: "kimi-k3.1-pro-highspeed",
1152
+ apiKeyName: API_KEY_CONSTANTS.MOONSHOT,
1153
+ hasThinking: true,
1154
+ reasoningEffort: "high",
1155
+ tags: ["expensive"]
1131
1156
  },
1132
1157
  // Z.AI models — thinking-only entry (non-thinking variant retired 2026-08-05)
1133
1158
  // reasoningEffort MUST be set: GLM-5.3 forces reasoning on and defaults the effort to 'max',
@@ -1364,30 +1389,37 @@ var MODEL_PRICING = {
1364
1389
  extendedContextCacheHitPrice: 2,
1365
1390
  extendedContextThresholdTokens: 272e3
1366
1391
  },
1367
- // OpenAI GPT-6 Sol and GPT-6 Luna (developers.openai.com/api/docs/pricing, read 2026-09-22),
1368
- // plus GPT-5.6 Terra, which has no GPT-6 successor and keeps its old rates.
1392
+ // OpenAI GPT-6 Sol and GPT-6 Luna (developers.openai.com/api/docs/pricing, read 2026-09-22).
1369
1393
  // Sol: $2/$10 short, $4/$15 long, cache hits $0.20/$0.40 — half what GPT-5.6 Sol cost.
1370
1394
  // Luna: $0.10/$0.50 short, $0.20/$0.75 long, cache hits $0.01/$0.02 — also halved.
1371
1395
  // As with Astra, OpenAI's table doesn't restate the short/long boundary, so we keep the
1372
1396
  // 272k threshold the 5.6 siblings use. Cache WRITES (Sol $2.50/$5, Luna $0.125/$0.25) are
1373
1397
  // not modelled: OpenAI caching is automatic and the API only reports hits.
1374
1398
  // GPT-6.1 Sol (2026-09-29) kept GPT-6 Sol's rates exactly, cache writes included.
1399
+ // Re-read 2026-10-26 with the Terra launch: the page now lists Sol's cache hit at $0.10
1400
+ // (5% of input, the same share as Terra), not the $0.20 recorded on 2026-09-22; Astra and
1401
+ // Luna still show 10%. Followed the page.
1375
1402
  [SupportedAiModels[LLM_CONSTANTS.GPT_SOL].modelApiName]: {
1376
1403
  inputPrice: 2,
1377
1404
  outputPrice: 10,
1378
- cacheHitPrice: 0.2,
1405
+ cacheHitPrice: 0.1,
1379
1406
  extendedContextInputPrice: 4,
1380
1407
  extendedContextOutputPrice: 15,
1381
- extendedContextCacheHitPrice: 0.4,
1408
+ extendedContextCacheHitPrice: 0.2,
1382
1409
  extendedContextThresholdTokens: 272e3
1383
1410
  },
1411
+ // GPT-6.1 Terra (developers.openai.com/api/docs/pricing, read 2026-10-26): $1/$5 short,
1412
+ // cache hits $0.05 (5% of input, like 6.1 Sol), 30m/12h cache writes $1.25/$2.50 — not
1413
+ // modelled, see above. The page's "Long" tab is rendered client-side and embeds no 6.x
1414
+ // rows, so the long-context rates follow the family rule every other entry here uses
1415
+ // (2x input, 1.5x output, 2x cache hit) over the same 272k threshold.
1384
1416
  [SupportedAiModels[LLM_CONSTANTS.GPT].modelApiName]: {
1385
- inputPrice: 2,
1386
- outputPrice: 12,
1387
- cacheHitPrice: 0.2,
1388
- extendedContextInputPrice: 4,
1389
- extendedContextOutputPrice: 18,
1390
- extendedContextCacheHitPrice: 0.4,
1417
+ inputPrice: 1,
1418
+ outputPrice: 5,
1419
+ cacheHitPrice: 0.05,
1420
+ extendedContextInputPrice: 2,
1421
+ extendedContextOutputPrice: 7.5,
1422
+ extendedContextCacheHitPrice: 0.1,
1391
1423
  extendedContextThresholdTokens: 272e3
1392
1424
  },
1393
1425
  [SupportedAiModels[LLM_CONSTANTS.GPT_MINI].modelApiName]: {
@@ -1419,11 +1451,18 @@ var MODEL_PRICING = {
1419
1451
  cacheHitPrice: 0.022,
1420
1452
  peakPricing: DEEPSEEK_PEAK_SCHEDULE
1421
1453
  },
1422
- // Kimi/Moonshot models
1454
+ // Kimi/Moonshot models (platform.kimi.ai/docs/pricing/chat, 2026-10-23). Cache writes
1455
+ // ($2/$4 per M for 5m/1h on Pro, $4/$8 on Highspeed) are not modelled: Moonshot caches
1456
+ // automatically and only reports hits.
1423
1457
  [SupportedAiModels[LLM_CONSTANTS.KIMI].modelApiName]: {
1424
- inputPrice: 3,
1425
- outputPrice: 15,
1426
- cacheHitPrice: 0.3
1458
+ inputPrice: 2,
1459
+ outputPrice: 10,
1460
+ cacheHitPrice: 0.2
1461
+ },
1462
+ [SupportedAiModels[LLM_CONSTANTS.KIMI_HIGHSPEED].modelApiName]: {
1463
+ inputPrice: 4,
1464
+ outputPrice: 20,
1465
+ cacheHitPrice: 0.2
1427
1466
  },
1428
1467
  // Z.AI models
1429
1468
  [SupportedAiModels[LLM_CONSTANTS.GLM].modelApiName]: {
@@ -2539,7 +2578,7 @@ ${instruction}`);
2539
2578
  * nowhere to put reasoning items. The system prompt still travels ONLY as `instructions`.
2540
2579
  *
2541
2580
  * Reasoning is readable only within one model family and the API silently omits items it
2542
- * cannot read, so a bot moved between GPT-6 and GPT-5.6 Terra loses its reasoning and keeps
2581
+ * cannot read, so a bot moved between GPT-6 and a GPT-5.6 model loses its reasoning and keeps
2543
2582
  * working. A bot moved to another provider never reaches this code at all: every agent
2544
2583
  * reads only its own field, and a missing one falls back to a plain message.
2545
2584
  */
@@ -5437,6 +5476,7 @@ var AgentFactory = class {
5437
5476
  case LLM_CONSTANTS.MISTRAL_SMALL:
5438
5477
  return new MistralAgent(name, instruction, model.modelApiName, key, shouldEnableThinking);
5439
5478
  case LLM_CONSTANTS.KIMI:
5479
+ case LLM_CONSTANTS.KIMI_HIGHSPEED:
5440
5480
  return new KimiAgent(name, instruction, model.modelApiName, key, 0, shouldEnableThinking);
5441
5481
  // Z.AI models — thinking-only since 2026-08-05
5442
5482
  case LLM_CONSTANTS.GLM: