@hiper2d/ai-agents 0.26.2 → 0.27.0

package/dist/index.mjs

@@ -846,6 +846,9 @@ var LLM_CONSTANTS = {
   MISTRAL_SMALL: "mistral-small",
   GROK: "grok",
   KIMI: "kimi",
+  // Kimi K3.1 Pro Highspeed: the same model served 2-3x faster at 2x the price (added
+  // 2026-10-23). Version-free like `kimi`, so a K3.2 repoint keeps both ids.
+  KIMI_HIGHSPEED: "kimi-highspeed",
   GLM: "glm",
   GLM_FLASH: "glm-flash",
   FUGU_MAX: "fugu-max",
@@ -937,7 +940,9 @@ var SupportedAiModels = {
     reasoningEffort: "low",
     // Reasoning tokens share the output budget, so leave room for both CoT and answer.
     maxOutputTokens: 65536,
-    tags: ["cheap"]
+    // slow: werewolf prod 2026-10-21, 785 calls over 30 days, p50 10s, p90 31s, max 156s;
+    // 7 of 79 bot turns past the 60s timeout in a 16-player game at days 8-9.
+    tags: ["slow", "cheap"]
   },
   // Models with always-on reasoning
   // GPT-6 Astra (2026-09-03): OpenAI's frontier tier above Sol. No `none` reasoning effort;
@@ -1017,7 +1022,9 @@ var SupportedAiModels = {
     apiKeyName: API_KEY_CONSTANTS.GOOGLE,
     hasThinking: true,
     reasoningEffort: "medium",
-    tags: ["expensive"]
+    // slow: werewolf prod 2026-10-21, 193 calls over 30 days, p50 12s, p90 34s, max 207s;
+    // 3 of 50 night actions past the 60s timeout even at medium effort.
+    tags: ["slow", "expensive"]
   },
   [LLM_CONSTANTS.GEMINI_PRO]: {
     displayName: "Gemini 3.1 Pro Preview",
@@ -1068,7 +1075,10 @@ var SupportedAiModels = {
     apiKeyName: API_KEY_CONSTANTS.GROK,
     hasThinking: true,
     reasoningEffort: "medium",
-    temperature: 0.7
+    temperature: 0.7,
+    // very-slow: werewolf prod 2026-10-21, 29 calls over 30 days, p50 48s, p90 192s, max 244s
+    // at medium effort; the 15-character story pipeline took 213s on 2026-10-07.
+    tags: ["very-slow"]
   },
   // Mistral models. One hybrid entry: reasoning is off by default on the API and switched on
   // per request with `reasoning_effort` (see mistral-agent.ts); the trace comes back as a
@@ -1106,21 +1116,40 @@ var SupportedAiModels = {
     reasoningEffort: "high",
     tags: ["very-fast", "cheap"]
   },
-  // Kimi models. Single always-reasoning entry: K3 reasons by default and the only way to stop
-  // it is the undocumented K2-era `thinking: disabled` toggle, which we no longer rely on.
-  // K3 always reasons at max effort; ~85-90% of its output tokens are reasoning tokens billed
-  // at the output rate, so real per-request cost runs well above the sticker output price.
+  // Kimi models. Always-reasoning entries: K3-series models reason by default and the only way
+  // to stop it is the undocumented K2-era `thinking: disabled` toggle, which we no longer rely on.
+  // ~85-90% of output tokens are reasoning tokens billed at the output rate, so real
+  // per-request cost runs well above the sticker output price.
+  //
+  // `kimi` repointed 2026-10-23 from kimi-k3 ($3/$15, 17-29s per turn, 146s for the
+  // 15-character story pipeline) to Kimi K3.1 Pro ($2/$10, 1M context, same low|high|max
+  // effort levels, default max). Live at high effort 2026-10-23: 3-sentence schema ask
+  // 4.2-4.5s, 8-character scene (476-845 output tokens) 9.7-14.9s = 49-57 tok/s — the same
+  // output speed as K3, so a full bot turn still lands in the 15-25s `slow` band.
+  // Consumer policy (werewolf): 1 bot per free game (opt-in) until 100 turns are measured.
   [LLM_CONSTANTS.KIMI]: {
-    displayName: "Kimi K3",
-    modelApiName: "kimi-k3",
+    displayName: "Kimi K3.1 Pro",
+    modelApiName: "kimi-k3.1-pro",
     apiKeyName: API_KEY_CONSTANTS.MOONSHOT,
     hasThinking: true,
-    // Temperature is omitted from the request: kimi-k3 rejects any value other than 1.
-    // Speed samples: 17s (2026-08-04) and 28.9s (2026-08-05) — graded into the >25s tier.
+    // Temperature is omitted from the request: K3 rejected any value other than 1.
     // Pinned 'high' 2026-09-22, down from Moonshot's 'max' default: K3 had no levels at
     // launch and the agent hardcoded max, so this is the first time the knob exists.
     reasoningEffort: "high",
-    tags: ["very-slow", "expensive"]
+    tags: ["slow", "expensive"]
+  },
+  // Same model, Moonshot's high-speed serving tier ("roughly two to three times faster",
+  // platform.kimi.ai K3.1 Pro quickstart, read 2026-10-23) at 2x the price ($4/$20 — Opus 5.5
+  // money). Live the same day: schema ask 1.9-2.4s, 8-character scene (789-982 output tokens)
+  // 7.9-8.4s = 100-117 tok/s, twice Pro. No speed tag: that is the unlabeled 6-13s middle.
+  // Consumer policy (werewolf): paid tier only, deliberately — the estimate is 5.2¢ a turn.
+  [LLM_CONSTANTS.KIMI_HIGHSPEED]: {
+    displayName: "Kimi K3.1 Pro Highspeed",
+    modelApiName: "kimi-k3.1-pro-highspeed",
+    apiKeyName: API_KEY_CONSTANTS.MOONSHOT,
+    hasThinking: true,
+    reasoningEffort: "high",
+    tags: ["expensive"]
   },
   // Z.AI models — thinking-only entry (non-thinking variant retired 2026-08-05)
   // reasoningEffort MUST be set: GLM-5.3 forces reasoning on and defaults the effort to 'max',
@@ -1412,11 +1441,18 @@ var MODEL_PRICING = {
     cacheHitPrice: 0.022,
     peakPricing: DEEPSEEK_PEAK_SCHEDULE
   },
-  // Kimi/Moonshot models
+  // Kimi/Moonshot models (platform.kimi.ai/docs/pricing/chat, 2026-10-23). Cache writes
+  // ($2/$4 per M for 5m/1h on Pro, $4/$8 on Highspeed) are not modelled: Moonshot caches
+  // automatically and only reports hits.
   [SupportedAiModels[LLM_CONSTANTS.KIMI].modelApiName]: {
-    inputPrice: 3,
-    outputPrice: 15,
-    cacheHitPrice: 0.3
+    inputPrice: 2,
+    outputPrice: 10,
+    cacheHitPrice: 0.2
+  },
+  [SupportedAiModels[LLM_CONSTANTS.KIMI_HIGHSPEED].modelApiName]: {
+    inputPrice: 4,
+    outputPrice: 20,
+    cacheHitPrice: 0.2
   },
   // Z.AI models
   [SupportedAiModels[LLM_CONSTANTS.GLM].modelApiName]: {
@@ -5430,6 +5466,7 @@ var AgentFactory = class {
       case LLM_CONSTANTS.MISTRAL_SMALL:
         return new MistralAgent(name, instruction, model.modelApiName, key, shouldEnableThinking);
       case LLM_CONSTANTS.KIMI:
+      case LLM_CONSTANTS.KIMI_HIGHSPEED:
         return new KimiAgent(name, instruction, model.modelApiName, key, 0, shouldEnableThinking);
       // Z.AI models — thinking-only since 2026-08-05
       case LLM_CONSTANTS.GLM: