@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.d.mts CHANGED
@@ -390,6 +390,7 @@ declare const LLM_CONSTANTS: {
390
390
  MISTRAL_SMALL: string;
391
391
  GROK: string;
392
392
  KIMI: string;
393
+ KIMI_HIGHSPEED: string;
393
394
  GLM: string;
394
395
  GLM_FLASH: string;
395
396
  FUGU_MAX: string;
@@ -1479,7 +1480,7 @@ declare class Gpt5Agent extends AbstractAgent {
1479
1480
  * nowhere to put reasoning items. The system prompt still travels ONLY as `instructions`.
1480
1481
  *
1481
1482
  * Reasoning is readable only within one model family and the API silently omits items it
1482
- * cannot read, so a bot moved between GPT-6 and GPT-5.6 Terra loses its reasoning and keeps
1483
+ * cannot read, so a bot moved between GPT-6 and a GPT-5.6 model loses its reasoning and keeps
1483
1484
  * working. A bot moved to another provider never reaches this code at all: every agent
1484
1485
  * reads only its own field, and a missing one falls back to a plain message.
1485
1486
  */
package/dist/index.d.ts CHANGED
@@ -390,6 +390,7 @@ declare const LLM_CONSTANTS: {
390
390
  MISTRAL_SMALL: string;
391
391
  GROK: string;
392
392
  KIMI: string;
393
+ KIMI_HIGHSPEED: string;
393
394
  GLM: string;
394
395
  GLM_FLASH: string;
395
396
  FUGU_MAX: string;
@@ -1479,7 +1480,7 @@ declare class Gpt5Agent extends AbstractAgent {
1479
1480
  * nowhere to put reasoning items. The system prompt still travels ONLY as `instructions`.
1480
1481
  *
1481
1482
  * Reasoning is readable only within one model family and the API silently omits items it
1482
- * cannot read, so a bot moved between GPT-6 and GPT-5.6 Terra loses its reasoning and keeps
1483
+ * cannot read, so a bot moved between GPT-6 and a GPT-5.6 model loses its reasoning and keeps
1483
1484
  * working. A bot moved to another provider never reaches this code at all: every agent
1484
1485
  * reads only its own field, and a missing one falls back to a plain message.
1485
1486
  */
package/dist/index.js CHANGED
@@ -1024,6 +1024,9 @@ var LLM_CONSTANTS = {
1024
1024
  MISTRAL_SMALL: "mistral-small",
1025
1025
  GROK: "grok",
1026
1026
  KIMI: "kimi",
1027
+ // Kimi K3.1 Pro Highspeed: the same model served 2-3x faster at 2x the price (added
1028
+ // 2026-10-23). Version-free like `kimi`, so a K3.2 repoint keeps both ids.
1029
+ KIMI_HIGHSPEED: "kimi-highspeed",
1027
1030
  GLM: "glm",
1028
1031
  GLM_FLASH: "glm-flash",
1029
1032
  FUGU_MAX: "fugu-max",
@@ -1145,6 +1148,9 @@ var SupportedAiModels = {
1145
1148
  // slot largely redundant — retiring it is a product call, not a catalog one.
1146
1149
  // Sol → GPT-6.1 Sol on 2026-09-29 (id verified against GET /v1/models; gpt-6-sol is still
1147
1150
  // served). Same price as GPT-6 Sol; its no-param default effort still measures `medium`.
1151
+ // Terra → GPT-6.1 Terra on 2026-10-26 (id verified against GET /v1/models; gpt-5.6-terra is
1152
+ // still served). Half the price of 5.6 Terra ($2/$12 → $1/$5) and back in its place below
1153
+ // Sol. Its no-param default effort measures `medium`, like its siblings.
1148
1154
  [LLM_CONSTANTS.GPT_SOL]: {
1149
1155
  displayName: "GPT-6.1 Sol",
1150
1156
  modelApiName: "gpt-6.1-sol",
@@ -1155,8 +1161,8 @@ var SupportedAiModels = {
1155
1161
  tags: ["expensive"]
1156
1162
  },
1157
1163
  [LLM_CONSTANTS.GPT]: {
1158
- displayName: "GPT-5.6 Terra",
1159
- modelApiName: "gpt-5.6-terra",
1164
+ displayName: "GPT-6.1 Terra",
1165
+ modelApiName: "gpt-6.1-terra",
1160
1166
  apiKeyName: API_KEY_CONSTANTS.OPENAI,
1161
1167
  reasoningEffort: "medium",
1162
1168
  hasThinking: true,
@@ -1291,21 +1297,40 @@ var SupportedAiModels = {
1291
1297
  reasoningEffort: "high",
1292
1298
  tags: ["very-fast", "cheap"]
1293
1299
  },
1294
- // Kimi models. Single always-reasoning entry: K3 reasons by default and the only way to stop
1295
- // it is the undocumented K2-era `thinking: disabled` toggle, which we no longer rely on.
1296
- // K3 always reasons at max effort; ~85-90% of its output tokens are reasoning tokens billed
1297
- // at the output rate, so real per-request cost runs well above the sticker output price.
1300
+ // Kimi models. Always-reasoning entries: K3-series models reason by default and the only way
1301
+ // to stop it is the undocumented K2-era `thinking: disabled` toggle, which we no longer rely on.
1302
+ // ~85-90% of output tokens are reasoning tokens billed at the output rate, so real
1303
+ // per-request cost runs well above the sticker output price.
1304
+ //
1305
+ // `kimi` repointed 2026-10-23 from kimi-k3 ($3/$15, 17-29s per turn, 146s for the
1306
+ // 15-character story pipeline) to Kimi K3.1 Pro ($2/$10, 1M context, same low|high|max
1307
+ // effort levels, default max). Live at high effort 2026-10-23: 3-sentence schema ask
1308
+ // 4.2-4.5s, 8-character scene (476-845 output tokens) 9.7-14.9s = 49-57 tok/s — the same
1309
+ // output speed as K3, so a full bot turn still lands in the 15-25s `slow` band.
1310
+ // Consumer policy (werewolf): 1 bot per free game (opt-in) until 100 turns are measured.
1298
1311
  [LLM_CONSTANTS.KIMI]: {
1299
- displayName: "Kimi K3",
1300
- modelApiName: "kimi-k3",
1312
+ displayName: "Kimi K3.1 Pro",
1313
+ modelApiName: "kimi-k3.1-pro",
1301
1314
  apiKeyName: API_KEY_CONSTANTS.MOONSHOT,
1302
1315
  hasThinking: true,
1303
- // Temperature is omitted from the request: kimi-k3 rejects any value other than 1.
1304
- // Speed samples: 17s (2026-08-04) and 28.9s (2026-08-05) — graded into the >25s tier.
1316
+ // Temperature is omitted from the request: K3 rejected any value other than 1.
1305
1317
  // Pinned 'high' 2026-09-22, down from Moonshot's 'max' default: K3 had no levels at
1306
1318
  // launch and the agent hardcoded max, so this is the first time the knob exists.
1307
1319
  reasoningEffort: "high",
1308
- tags: ["very-slow", "expensive"]
1320
+ tags: ["slow", "expensive"]
1321
+ },
1322
+ // Same model, Moonshot's high-speed serving tier ("roughly two to three times faster",
1323
+ // platform.kimi.ai K3.1 Pro quickstart, read 2026-10-23) at 2x the price ($4/$20 — Opus 5.5
1324
+ // money). Live the same day: schema ask 1.9-2.4s, 8-character scene (789-982 output tokens)
1325
+ // 7.9-8.4s = 100-117 tok/s, twice Pro. No speed tag: that is the unlabeled 6-13s middle.
1326
+ // Consumer policy (werewolf): paid tier only, deliberately — the estimate is 5.2¢ a turn.
1327
+ [LLM_CONSTANTS.KIMI_HIGHSPEED]: {
1328
+ displayName: "Kimi K3.1 Pro Highspeed",
1329
+ modelApiName: "kimi-k3.1-pro-highspeed",
1330
+ apiKeyName: API_KEY_CONSTANTS.MOONSHOT,
1331
+ hasThinking: true,
1332
+ reasoningEffort: "high",
1333
+ tags: ["expensive"]
1309
1334
  },
1310
1335
  // Z.AI models — thinking-only entry (non-thinking variant retired 2026-08-05)
1311
1336
  // reasoningEffort MUST be set: GLM-5.3 forces reasoning on and defaults the effort to 'max',
@@ -1542,30 +1567,37 @@ var MODEL_PRICING = {
1542
1567
  extendedContextCacheHitPrice: 2,
1543
1568
  extendedContextThresholdTokens: 272e3
1544
1569
  },
1545
- // OpenAI GPT-6 Sol and GPT-6 Luna (developers.openai.com/api/docs/pricing, read 2026-09-22),
1546
- // plus GPT-5.6 Terra, which has no GPT-6 successor and keeps its old rates.
1570
+ // OpenAI GPT-6 Sol and GPT-6 Luna (developers.openai.com/api/docs/pricing, read 2026-09-22).
1547
1571
  // Sol: $2/$10 short, $4/$15 long, cache hits $0.20/$0.40 — half what GPT-5.6 Sol cost.
1548
1572
  // Luna: $0.10/$0.50 short, $0.20/$0.75 long, cache hits $0.01/$0.02 — also halved.
1549
1573
  // As with Astra, OpenAI's table doesn't restate the short/long boundary, so we keep the
1550
1574
  // 272k threshold the 5.6 siblings use. Cache WRITES (Sol $2.50/$5, Luna $0.125/$0.25) are
1551
1575
  // not modelled: OpenAI caching is automatic and the API only reports hits.
1552
1576
  // GPT-6.1 Sol (2026-09-29) kept GPT-6 Sol's rates exactly, cache writes included.
1577
+ // Re-read 2026-10-26 with the Terra launch: the page now lists Sol's cache hit at $0.10
1578
+ // (5% of input, the same share as Terra), not the $0.20 recorded on 2026-09-22; Astra and
1579
+ // Luna still show 10%. Followed the page.
1553
1580
  [SupportedAiModels[LLM_CONSTANTS.GPT_SOL].modelApiName]: {
1554
1581
  inputPrice: 2,
1555
1582
  outputPrice: 10,
1556
- cacheHitPrice: 0.2,
1583
+ cacheHitPrice: 0.1,
1557
1584
  extendedContextInputPrice: 4,
1558
1585
  extendedContextOutputPrice: 15,
1559
- extendedContextCacheHitPrice: 0.4,
1586
+ extendedContextCacheHitPrice: 0.2,
1560
1587
  extendedContextThresholdTokens: 272e3
1561
1588
  },
1589
+ // GPT-6.1 Terra (developers.openai.com/api/docs/pricing, read 2026-10-26): $1/$5 short,
1590
+ // cache hits $0.05 (5% of input, like 6.1 Sol), 30m/12h cache writes $1.25/$2.50 — not
1591
+ // modelled, see above. The page's "Long" tab is rendered client-side and embeds no 6.x
1592
+ // rows, so the long-context rates follow the family rule every other entry here uses
1593
+ // (2x input, 1.5x output, 2x cache hit) over the same 272k threshold.
1562
1594
  [SupportedAiModels[LLM_CONSTANTS.GPT].modelApiName]: {
1563
- inputPrice: 2,
1564
- outputPrice: 12,
1565
- cacheHitPrice: 0.2,
1566
- extendedContextInputPrice: 4,
1567
- extendedContextOutputPrice: 18,
1568
- extendedContextCacheHitPrice: 0.4,
1595
+ inputPrice: 1,
1596
+ outputPrice: 5,
1597
+ cacheHitPrice: 0.05,
1598
+ extendedContextInputPrice: 2,
1599
+ extendedContextOutputPrice: 7.5,
1600
+ extendedContextCacheHitPrice: 0.1,
1569
1601
  extendedContextThresholdTokens: 272e3
1570
1602
  },
1571
1603
  [SupportedAiModels[LLM_CONSTANTS.GPT_MINI].modelApiName]: {
@@ -1597,11 +1629,18 @@ var MODEL_PRICING = {
1597
1629
  cacheHitPrice: 0.022,
1598
1630
  peakPricing: DEEPSEEK_PEAK_SCHEDULE
1599
1631
  },
1600
- // Kimi/Moonshot models
1632
+ // Kimi/Moonshot models (platform.kimi.ai/docs/pricing/chat, 2026-10-23). Cache writes
1633
+ // ($2/$4 per M for 5m/1h on Pro, $4/$8 on Highspeed) are not modelled: Moonshot caches
1634
+ // automatically and only reports hits.
1601
1635
  [SupportedAiModels[LLM_CONSTANTS.KIMI].modelApiName]: {
1602
- inputPrice: 3,
1603
- outputPrice: 15,
1604
- cacheHitPrice: 0.3
1636
+ inputPrice: 2,
1637
+ outputPrice: 10,
1638
+ cacheHitPrice: 0.2
1639
+ },
1640
+ [SupportedAiModels[LLM_CONSTANTS.KIMI_HIGHSPEED].modelApiName]: {
1641
+ inputPrice: 4,
1642
+ outputPrice: 20,
1643
+ cacheHitPrice: 0.2
1605
1644
  },
1606
1645
  // Z.AI models
1607
1646
  [SupportedAiModels[LLM_CONSTANTS.GLM].modelApiName]: {
@@ -2717,7 +2756,7 @@ ${instruction}`);
2717
2756
  * nowhere to put reasoning items. The system prompt still travels ONLY as `instructions`.
2718
2757
  *
2719
2758
  * Reasoning is readable only within one model family and the API silently omits items it
2720
- * cannot read, so a bot moved between GPT-6 and GPT-5.6 Terra loses its reasoning and keeps
2759
+ * cannot read, so a bot moved between GPT-6 and a GPT-5.6 model loses its reasoning and keeps
2721
2760
  * working. A bot moved to another provider never reaches this code at all: every agent
2722
2761
  * reads only its own field, and a missing one falls back to a plain message.
2723
2762
  */
@@ -5615,6 +5654,7 @@ var AgentFactory = class {
5615
5654
  case LLM_CONSTANTS.MISTRAL_SMALL:
5616
5655
  return new MistralAgent(name, instruction, model.modelApiName, key, shouldEnableThinking);
5617
5656
  case LLM_CONSTANTS.KIMI:
5657
+ case LLM_CONSTANTS.KIMI_HIGHSPEED:
5618
5658
  return new KimiAgent(name, instruction, model.modelApiName, key, 0, shouldEnableThinking);
5619
5659
  // Z.AI models — thinking-only since 2026-08-05
5620
5660
  case LLM_CONSTANTS.GLM: