@hiper2d/ai-agents 0.39.0 → 0.41.0

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
@@ -403,6 +403,7 @@ declare const LLM_CONSTANTS: {
403
403
  GPT_SOL: string;
404
404
  GPT: string;
405
405
  GPT_MINI: string;
406
+ GEMINI_NEON: string;
406
407
  GEMINI_ARGON: string;
407
408
  GEMINI_PRO: string;
408
409
  GEMINI_FLASH: string;
package/dist/index.d.ts CHANGED
@@ -403,6 +403,7 @@ declare const LLM_CONSTANTS: {
403
403
  GPT_SOL: string;
404
404
  GPT: string;
405
405
  GPT_MINI: string;
406
+ GEMINI_NEON: string;
406
407
  GEMINI_ARGON: string;
407
408
  GEMINI_PRO: string;
408
409
  GEMINI_FLASH: string;
package/dist/index.js CHANGED
@@ -1031,6 +1031,7 @@ var LLM_CONSTANTS = {
1031
1031
  GPT: "gpt",
1032
1032
  GPT_MINI: "gpt-mini",
1033
1033
  // Gemini 4 Argon sits above Pro; version-free id so a later Argon repoint keeps persisted ids.
1034
+ GEMINI_NEON: "gemini-neon",
1034
1035
  GEMINI_ARGON: "gemini-argon",
1035
1036
  GEMINI_PRO: "gemini-pro",
1036
1037
  GEMINI_FLASH: "gemini-flash",
@@ -1149,8 +1150,8 @@ var SupportedAiModels = {
1149
1150
  // OpenAI documents the default for Sol, Luna and the 5.6 family but never states Astra's.
1150
1151
  // Pinning it changes no behaviour today; it stops the default moving under us silently.
1151
1152
  [LLM_CONSTANTS.GPT_ASTRA]: {
1152
- displayName: "GPT-6 Astra",
1153
- modelApiName: "gpt-6-astra",
1153
+ displayName: "GPT-6.2 Astra",
1154
+ modelApiName: "gpt-6.2-astra",
1154
1155
  apiKeyName: API_KEY_CONSTANTS.OPENAI,
1155
1156
  reasoningEffort: "medium",
1156
1157
  hasThinking: true,
@@ -1168,9 +1169,12 @@ var SupportedAiModels = {
1168
1169
  // Terra → GPT-6.1 Terra on 2026-10-26 (id verified against GET /v1/models; gpt-5.6-terra is
1169
1170
  // still served). Half the price of 5.6 Terra ($2/$12 → $1/$5) and back in its place below
1170
1171
  // Sol. Its no-param default effort measures `medium`, like its siblings.
1172
+ // Astra → GPT-6.2 Astra and Sol → GPT-6.2 Sol on 2026-11-17 (both ids verified against
1173
+ // GET /v1/models; gpt-6-astra and gpt-6.1-sol are still served). Both no-param defaults
1174
+ // measure `medium`. Astra got 20% cheaper; Sol kept its price.
1171
1175
  [LLM_CONSTANTS.GPT_SOL]: {
1172
- displayName: "GPT-6.1 Sol",
1173
- modelApiName: "gpt-6.1-sol",
1176
+ displayName: "GPT-6.2 Sol",
1177
+ modelApiName: "gpt-6.2-sol",
1174
1178
  apiKeyName: API_KEY_CONSTANTS.OPENAI,
1175
1179
  reasoningEffort: "medium",
1176
1180
  hasThinking: true,
@@ -1214,6 +1218,21 @@ var SupportedAiModels = {
1214
1218
  // werewolf day-2 vote fixture took 16-34s (380-415 output tokens). At medium the same
1215
1219
  // prompt took 4-11s (75-110 output tokens) with the same vote target and a reason just
1216
1220
  // as specific — the same trade Grok 4.7 made on 2026-10-06.
1221
+ // Gemini 4 Neon (added 2026-11-18, id verified against GET /v1beta/models): 1M in / 64K
1222
+ // out. Same thinkingLevel dialect — accepts low|medium|high, rejects 'minimal' and 'none'
1223
+ // (probed). Pinned LOW: it thinks even there (200-1,500 tokens a turn where Argon at medium
1224
+ // reports none). Werewolf day-3 prompt (10.5K tokens, cold cache, 2 runs, medians):
1225
+ // low = discussion 8.9s, vote 12.9s, night 10.7s, end-of-day notes 36s, 3.8-5.9¢ a turn;
1226
+ // medium = 15-16s on the short turns for no visible gain; high = 21-23s, notes 51s.
1227
+ // No `temperature` (see the Pro note). Speed tag owed until it has 100 production turns.
1228
+ [LLM_CONSTANTS.GEMINI_NEON]: {
1229
+ displayName: "Gemini 4 Neon",
1230
+ modelApiName: "gemini-4-neon",
1231
+ apiKeyName: API_KEY_CONSTANTS.GOOGLE,
1232
+ hasThinking: true,
1233
+ reasoningEffort: "low",
1234
+ tags: ["expensive"]
1235
+ },
1217
1236
  [LLM_CONSTANTS.GEMINI_ARGON]: {
1218
1237
  displayName: "Gemini 4 Argon",
1219
1238
  modelApiName: "gemini-4-argon",
@@ -1619,13 +1638,16 @@ var MODEL_PRICING = {
1619
1638
  // short context, $20/$75 cache-hit $2 long context. OpenAI's pricing table doesn't restate
1620
1639
  // the boundary; we assume the same 272k threshold as the GPT-5.6 siblings. Cache writes
1621
1640
  // ($12.50/$25) are not modelled — caching is automatic and we only see hits.
1641
+ // GPT-6.2 Astra (pricing page, read 2026-11-17, both Short and Long tabs): $8/$40 cache-hit
1642
+ // $0.40 short, $16/$60 cache-hit $0.80 long; cache writes $10/$20 (30m/12h), not modelled.
1643
+ // The cache hit is now 5% of input, like Sol and Terra (6 Astra was 10%).
1622
1644
  [SupportedAiModels[LLM_CONSTANTS.GPT_ASTRA].modelApiName]: {
1623
- inputPrice: 10,
1624
- outputPrice: 50,
1625
- cacheHitPrice: 1,
1626
- extendedContextInputPrice: 20,
1627
- extendedContextOutputPrice: 75,
1628
- extendedContextCacheHitPrice: 2,
1645
+ inputPrice: 8,
1646
+ outputPrice: 40,
1647
+ cacheHitPrice: 0.4,
1648
+ extendedContextInputPrice: 16,
1649
+ extendedContextOutputPrice: 60,
1650
+ extendedContextCacheHitPrice: 0.8,
1629
1651
  extendedContextThresholdTokens: 272e3
1630
1652
  },
1631
1653
  // OpenAI GPT-6 Sol and GPT-6 Luna (developers.openai.com/api/docs/pricing, read 2026-09-22).
@@ -1638,6 +1660,7 @@ var MODEL_PRICING = {
1638
1660
  // Re-read 2026-10-26 with the Terra launch: the page now lists Sol's cache hit at $0.10
1639
1661
  // (5% of input, the same share as Terra), not the $0.20 recorded on 2026-09-22; Astra and
1640
1662
  // Luna still show 10%. Followed the page.
1663
+ // GPT-6.2 Sol (read 2026-11-17, Short and Long tabs) is priced exactly like 6.1 Sol.
1641
1664
  [SupportedAiModels[LLM_CONSTANTS.GPT_SOL].modelApiName]: {
1642
1665
  inputPrice: 2,
1643
1666
  outputPrice: 10,
@@ -1755,6 +1778,14 @@ var MODEL_PRICING = {
1755
1778
  extendedContextThresholdTokens: 1e5
1756
1779
  },
1757
1780
  // Google models
1781
+ [SupportedAiModels[LLM_CONSTANTS.GEMINI_NEON].modelApiName]: {
1782
+ // ai.google.dev pricing page, read 2026-11-18: one flat tier over the whole 1M window
1783
+ // (no long-context rows). Implicit caching only, cached input at 5% of input. Not on
1784
+ // Google's free tier.
1785
+ inputPrice: 3,
1786
+ outputPrice: 15,
1787
+ cacheHitPrice: 0.15
1788
+ },
1758
1789
  [SupportedAiModels[LLM_CONSTANTS.GEMINI_ARGON].modelApiName]: {
1759
1790
  // Launch pricing (ai.google.dev pricing page, fetched 2026-10-13), one flat tier — no
1760
1791
  // >200K long-context rows, unlike 3.1 Pro. DOUBLES on 2026-12-01 to $4.00 / $20.00 /
@@ -5843,6 +5874,7 @@ var AgentFactory = class {
5843
5874
  case LLM_CONSTANTS.GPT:
5844
5875
  case LLM_CONSTANTS.GPT_MINI:
5845
5876
  return new Gpt5Agent(name, instruction, model.modelApiName, key, model.temperature, shouldEnableThinking);
5877
+ case LLM_CONSTANTS.GEMINI_NEON:
5846
5878
  case LLM_CONSTANTS.GEMINI_ARGON:
5847
5879
  case LLM_CONSTANTS.GEMINI_PRO:
5848
5880
  case LLM_CONSTANTS.GEMINI_FLASH: