@hiper2d/ai-agents 0.19.0 → 0.19.2

package/dist/index.mjs

@@ -806,16 +806,12 @@ var LLM_CONSTANTS = {
   GEMINI_PRO: "gemini-pro",
   GEMINI_FLASH: "gemini-flash",
   GEMINI_LITE: "gemini-lite",
-  // Reused 2026-10-06 for Mistral Large 4: the old id (Large 3) was retired 2026-09-18 and
-  // consumers had mapped it away, so the version-free id was free again.
-  MISTRAL_LARGE: "mistral-large",
   MISTRAL_MEDIUM: "mistral-medium",
   MISTRAL_SMALL: "mistral-small",
   GROK: "grok",
   KIMI: "kimi",
   GLM: "glm",
   GLM_FLASH: "glm-flash",
-  FUGU_ULTRA: "fugu-ultra",
   FUGU_MAX: "fugu-max",
   // Qwen (QwenCloud/DashScope). Stable picker ids without the version, matching the gpt/gemini
   // pattern, so future repoints don't orphan persisted ids.
@@ -965,6 +961,10 @@ var SupportedAiModels = {
     apiKeyName: API_KEY_CONSTANTS.GOOGLE,
     hasThinking: true,
     reasoningEffort: "high",
+    // Pre-3.6: still honours sampling params, so it keeps the 0.2 GoogleAgent always sent
+    // before 2026-10-07. 3.6+ ignore them and future Gemini models reject them — leave
+    // `temperature` unset on every newer Gemini entry (GoogleAgent then sends none).
+    temperature: 0.2,
     tags: ["expensive"]
   },
   [LLM_CONSTANTS.GEMINI_FLASH]: {
@@ -984,20 +984,26 @@ var SupportedAiModels = {
     apiKeyName: API_KEY_CONSTANTS.GOOGLE,
     hasThinking: true,
     reasoningEffort: "minimal",
+    // Pre-3.6, same as Pro: keeps the 0.2 it always had.
+    temperature: 0.2,
     tags: ["fast", "cheap"]
   },
   // Always-on reasoning — no non-thinking sibling. Bumped 4.6 → 4.7 on 2026-09-22 (id verified
   // against GET /v1/models). Same price, same 200K extended-context boundary; the context
-  // window is 500K. NOTE 4.7 newly accepts `reasoning_effort` (low | medium | high | xhigh,
-  // default high) where 4.6 had no knob at all. GrokAgent still sends none, so behaviour is
-  // unchanged — but Grok is the app's most expensive model per turn (94% of its output is
-  // reasoning), so a lower pin is the obvious lever to measure before anything else.
+  // window is 500K. 4.7 accepts `reasoning: {effort}` (low | medium | high | xhigh, default
+  // high) where 4.6 had no knob at all; GrokAgent sends the pin below.
+  //
+  // Pinned medium 2026-10-06. At high, production day summaries ran ~2 min and the werewolf
+  // day-2 fixture measured votes 27-55s and summaries 44-55s (1.5-3k output tokens) — too
+  // close to the 60s per-call timeout. At medium the same prompts took votes 5-8s and
+  // summaries 23-25s (3-5x fewer output tokens) with the same vote target and summaries just
+  // as specific.
   [LLM_CONSTANTS.GROK]: {
     displayName: "Grok 4.7",
     modelApiName: "grok-4.7",
     apiKeyName: API_KEY_CONSTANTS.GROK,
     hasThinking: true,
-    reasoningEffort: "high",
+    reasoningEffort: "medium",
     temperature: 0.7
   },
   // Mistral models. Two hybrid entries since 2026-09-18: reasoning is off by default on the
@@ -1017,20 +1023,12 @@ var SupportedAiModels = {
   // `magistral-medium-latest` as one entry — so Magistral seats had silently become Medium
   // seats already. Consumers map both retired ids onto the surviving pair.
   //
-  // Mistral Large 4 added 2026-10-06 under the freed `mistral-large` picker id. The model list
-  // reports reasoning: true (Large 3 said false), 512K context; same reasoning_effort switch,
-  // but only 'high' and 'none' are accepted (medium/low are a 400). Live 2026-10-06 at high:
-  // a chat-turn schema ask took 14-21s with 1.3-2k output tokens (~$0.003), one constrained
-  // pick took 49s — it thinks long, so it is cheap but slow.
-  [LLM_CONSTANTS.MISTRAL_LARGE]: {
-    displayName: "Mistral Large 4",
-    // `mistral-large-4-0` is the same entry; `mistral-large-latest` still points at Large 3.
-    modelApiName: "mistral-large-4",
-    apiKeyName: API_KEY_CONSTANTS.MISTRAL,
-    hasThinking: true,
-    reasoningEffort: "high",
-    tags: ["slow", "cheap"]
-  },
+  // Mistral Large 4 (`mistral-large-4`, $0.68/$2.09, $0.07 cached) was added in 0.19.0 and
+  // removed in 0.19.1 (2026-10-06/07): too slow to play. Like Small/Medium it accepts only
+  // reasoning_effort high|none, and unlike them it reasons when the field is omitted. At high
+  // a realistic day-2 vote took 25-154s (half past the 60s timeout) and its parallel story
+  // batches finished ~100s apart (Mistral served them one at a time: ~6 min per story). Off,
+  // it answers in ~1s with no trace. Re-test before re-adding.
   [LLM_CONSTANTS.MISTRAL_MEDIUM]: {
     displayName: "Mistral Medium 3.5",
     // The docs' explicit 3.5 id. It was `mistral-medium-3` until 2026-09-18, which Mistral
@@ -1115,9 +1113,9 @@ var SupportedAiModels = {
   // real against $0.85 tracked, a 5.7x undercharge. It is a router with no published price,
   // so the rate is not even guaranteed stable, and its cache hit rate was 9.3% — effectively
   // zero, since every hit came from a duplicate call seconds apart rather than turn-to-turn
-  // prefix reuse. Ultra costs the same and is predictable.
+  // prefix reuse.
   //
-  // reasoningEffort is PINNED on both entries because the server default differs per model:
+  // reasoningEffort was PINNED on both entries because the server default differs per model:
   // fugu-ultra defaults to `xhigh` (console.sakana.ai/models), fugu-max to `high`. Until
   // 2026-09-20 the agent sent no effort at all, so every Ultra turn ran at xhigh. Measured that
   // day on a 2k-token three-day game prompt (two samples each, chat completions, json mode):
@@ -1127,17 +1125,17 @@ var SupportedAiModels = {
   //   max   xhigh   41s /  47s   $0.02
   // `max_tokens` is not a latency lever for Ultra: Sakana applies it to the final response only,
   // the orchestrator "still uses maximum token limit". Effort is the only knob.
-  [LLM_CONSTANTS.FUGU_ULTRA]: {
-    displayName: "Sakana Fugu Ultra",
-    modelApiName: "fugu-ultra",
-    apiKeyName: API_KEY_CONSTANTS.FUGU,
-    hasThinking: false,
-    reasoningEffort: "high",
-    tags: ["extremely-slow", "expensive"]
-  },
+  //
+  // Fugu Ultra (`fugu-ultra`, $5/$30, $0.50 cached; ×~2 above 272K) was REMOVED in 0.19.1
+  // (2026-10-07): too slow and too expensive for what it played. At `high` a day-2 vote took
+  // 39s and a turn 65-78s (past the 60s request timeout), and as Game Master its four parallel
+  // story batches ran past 240s each. It also billed "orchestration" tokens (its internal expert
+  // calls) on top of the visible counts; FuguAgent still folds those in, in case Max ever
+  // reports them. Consumers map the id to Max.
   // Fugu Max (2026-09-20, fugu-max → fugu-max-v1.0): Sakana's "largest pool of models on the
-  // cost–performance Pareto frontier" — the everyday Fugu at 5x less than Ultra's output rate
-  // and a fraction of its latency. It reasons (reasoning_tokens 80-500 a turn) but, like
+  // cost–performance Pareto frontier" — 5x less than Ultra's output rate and a fraction of its
+  // latency; the only Fugu since 0.19.1. Live 2026-10-06: day-2 vote 11s, 15-character story
+  // pipeline 100s. It reasons (reasoning_tokens 80-500 a turn) but, like
   // Ultra, never returns reasoning_content, hence hasThinking: false. No orchestration
   // tokens observed on any call.
   [LLM_CONSTANTS.FUGU_MAX]: {
@@ -1420,13 +1418,7 @@ var MODEL_PRICING = {
   // Mistral models (mistral.ai/pricing/api, verified 2026-09-18). Cached tokens bill at 10% of
   // the input price (documented on the prompt_cache_key param; no per-model cached prices
   // published). Reasoning tokens are counted inside completion_tokens — no separate
-  // reasoning_tokens field — so the output rate already covers the trace. Large 4 added
-  // 2026-10-06 at its published rates ($0.07 cached, not an exact 10%).
-  [SupportedAiModels[LLM_CONSTANTS.MISTRAL_LARGE].modelApiName]: {
-    inputPrice: 0.68,
-    outputPrice: 2.09,
-    cacheHitPrice: 0.07
-  },
+  // reasoning_tokens field — so the output rate already covers the trace.
   [SupportedAiModels[LLM_CONSTANTS.MISTRAL_MEDIUM].modelApiName]: {
     inputPrice: 1.5,
     outputPrice: 7.5,
@@ -1450,21 +1442,10 @@ var MODEL_PRICING = {
     extendedContextCacheHitPrice: 1,
     extendedContextThresholdTokens: 2e5
   },
-  // Sakana Fugu models (console.sakana.ai/pricing, read 2026-09-20). Base `fugu` was retired
-  // 2026-08-04 — it had no published price and measured out at these same ultra rates, so it
-  // has no pricing entry. fugu-ultra: above 272K context the rates roughly double. Ultra also
-  // reports "orchestration" tokens (its internal expert calls) in prompt_tokens_details /
-  // completion_tokens_details, billed at these same input/output rates — FuguAgent folds them
-  // into the token counts before pricing.
-  [SupportedAiModels[LLM_CONSTANTS.FUGU_ULTRA].modelApiName]: {
-    inputPrice: 5,
-    outputPrice: 30,
-    cacheHitPrice: 0.5,
-    extendedContextInputPrice: 10,
-    extendedContextOutputPrice: 45,
-    extendedContextCacheHitPrice: 1,
-    extendedContextThresholdTokens: 272e3
-  },
+  // Sakana Fugu (console.sakana.ai/pricing, read 2026-09-20). Base `fugu` (retired 2026-08-04)
+  // and `fugu-ultra` (removed 0.19.1) have no entries. Any "orchestration" tokens reported in
+  // prompt_tokens_details / completion_tokens_details bill at these same rates — FuguAgent
+  // folds them into the token counts before pricing.
   // fugu-max: flat rates, no context tiers.
   [SupportedAiModels[LLM_CONSTANTS.FUGU_MAX].modelApiName]: {
     inputPrice: 2,
@@ -1524,7 +1505,6 @@ var HYBRID_THINKING_API_NAMES = /* @__PURE__ */ new Set([
   SupportedAiModels[LLM_CONSTANTS.MINIMAX].modelApiName,
   // Mistral reasons only when asked (`reasoning_effort`), and we always ask — hybrid by the
   // same definition as Qwen. Added 2026-09-18 when the two entries went thinking-on.
-  SupportedAiModels[LLM_CONSTANTS.MISTRAL_LARGE].modelApiName,
   SupportedAiModels[LLM_CONSTANTS.MISTRAL_MEDIUM].modelApiName,
   SupportedAiModels[LLM_CONSTANTS.MISTRAL_SMALL].modelApiName
 ]);
@@ -1593,7 +1573,7 @@ var GLM_REASONING_EFFORTS = ["low", "high", "max"];
 var DEEPSEEK_REASONING_EFFORTS = ["low", "high", "max"];
 var FUGU_REASONING_EFFORTS = ["high", "xhigh", "max"];
 var META_REASONING_EFFORTS = ["minimal", "low", "medium", "high", "xhigh", "max"];
-var MISTRAL_REASONING_EFFORTS = ["minimal", "low", "medium", "high", "xhigh"];
+var MISTRAL_REASONING_EFFORTS = ["high"];
 var XAI_REASONING_EFFORTS = ["low", "medium", "high", "xhigh"];
 var KIMI_REASONING_EFFORTS = ["low", "high", "max"];
 function clampReasoningEffort(effort, allowed) {
@@ -3083,6 +3063,13 @@ function throwForEmptyGeminiResponse(model, response, fallback) {
 }
 var GoogleAgent = class extends AbstractAgent {
   client;
+  /**
+   * Sampling temperature to send, or undefined to send none. Google deprecated custom sampling
+   * (notice of 2026-10): since Gemini 3.6 Flash temperature/top_p/top_k are ignored, and its
+   * upcoming models reject them with a 400. So it is sent only for a catalog entry that sets
+   * `temperature` (the pre-3.6 models, to keep their behaviour) — never by default.
+   */
+  samplingTemperature;
   defaultConfig = {
     responseMimeType: "application/json"
   };
@@ -3098,7 +3085,9 @@ var GoogleAgent = class extends AbstractAgent {
     unsupportedRole: (role) => `Unsupported role type: ${role}`
   };
   constructor(name, instruction, model, apiKey, enableThinking = false, agentLoggingConfig = DEFAULT_LOGGING_CONFIG.agents) {
-    super(name, instruction, model, 0.2, enableThinking, agentLoggingConfig);
+    const samplingTemperature = getModelConfigByApiName(model)?.temperature;
+    super(name, instruction, model, samplingTemperature ?? 0, enableThinking, agentLoggingConfig);
+    this.samplingTemperature = samplingTemperature;
     this.client = new GoogleGenAI3({
       apiKey
     });
@@ -3197,7 +3186,7 @@ var GoogleAgent = class extends AbstractAgent {
     try {
       const googleSchema = ZodSchemaConverter.toGoogleSchema(zodSchema);
       const config = {
-        temperature: this.temperature,
+        ...this.samplingTemperature !== void 0 ? { temperature: this.samplingTemperature } : {},
         responseMimeType: "application/json",
         responseSchema: googleSchema,
         maxOutputTokens: this.maxOutputTokens,
@@ -3287,7 +3276,7 @@ var GoogleAgent = class extends AbstractAgent {
     const contents = this.convertToContents(messages);
     try {
       const config = {
-        temperature: this.temperature,
+        ...this.samplingTemperature !== void 0 ? { temperature: this.samplingTemperature } : {},
         maxOutputTokens: this.maxOutputTokens,
         systemInstruction: this.instruction
       };
@@ -5346,7 +5335,6 @@ var AgentFactory = class {
       case LLM_CONSTANTS.DEEPSEEK_PRO:
         return new DeepSeekV2Agent(name, instruction, model.modelApiName, key, model.temperature ?? 0, shouldEnableThinking);
       // Mistral models
-      case LLM_CONSTANTS.MISTRAL_LARGE:
       case LLM_CONSTANTS.MISTRAL_MEDIUM:
       case LLM_CONSTANTS.MISTRAL_SMALL:
         return new MistralAgent(name, instruction, model.modelApiName, key, shouldEnableThinking);
@@ -5357,7 +5345,6 @@ var AgentFactory = class {
       case LLM_CONSTANTS.GLM_FLASH:
         return new GlmAgent(name, instruction, model.modelApiName, key, model.temperature, shouldEnableThinking);
       // Sakana Fugu models — always-on reasoning, no temperature (ignored by the model)
-      case LLM_CONSTANTS.FUGU_ULTRA:
       case LLM_CONSTANTS.FUGU_MAX:
         return new FuguAgent(name, instruction, model.modelApiName, key, shouldEnableThinking);
       // Qwen models — thinking-only (enable_thinking always sent)