@hiper2d/ai-agents 0.20.0 → 0.21.0

package/dist/index.d.mts

@@ -375,7 +375,6 @@ declare const LLM_CONSTANTS: {
     GEMINI_PRO: string;
     GEMINI_FLASH: string;
     GEMINI_LITE: string;
-    MISTRAL_MEDIUM: string;
     MISTRAL_SMALL: string;
     GROK: string;
     KIMI: string;

package/dist/index.d.ts

@@ -375,7 +375,6 @@ declare const LLM_CONSTANTS: {
     GEMINI_PRO: string;
     GEMINI_FLASH: string;
     GEMINI_LITE: string;
-    MISTRAL_MEDIUM: string;
     MISTRAL_SMALL: string;
     GROK: string;
     KIMI: string;

package/dist/index.js

@@ -984,7 +984,6 @@ var LLM_CONSTANTS = {
   GEMINI_PRO: "gemini-pro",
   GEMINI_FLASH: "gemini-flash",
   GEMINI_LITE: "gemini-lite",
-  MISTRAL_MEDIUM: "mistral-medium",
   MISTRAL_SMALL: "mistral-small",
   GROK: "grok",
   KIMI: "kimi",
@@ -1190,14 +1189,18 @@ var SupportedAiModels = {
     reasoningEffort: "medium",
     temperature: 0.7
   },
-  // Mistral models. Two hybrid entries since 2026-09-18: 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 `thinking` content chunk and works with json_schema structured output
-  // (verified live 2026-09-18 on both). Pinned to 'high' — the only level Mistral's docs
-  // describe ("full thinking chunk before the final answer"); Small is cheap enough that the
-  // extra tokens don't matter, Medium's are priced into its free-tier band via the hybrid
-  // multiplier. Live 2026-09-18 (one short schema ask each): Small 1.7s with a ~600-char
-  // trace, Medium 3.8s with ~2000 chars — Medium's trace ran ~15x its answer length.
+  // 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
+  // `thinking` content chunk and works with json_schema structured output (verified live
+  // 2026-09-18). Pinned to 'high' — the only level Mistral's docs describe ("full thinking
+  // chunk before the final answer"); Small is cheap enough that the extra tokens don't
+  // matter. Live 2026-09-18 (one short schema ask): 1.7s with a ~600-char trace.
+  //
+  // Mistral Medium 3.5 (`mistral-medium-3-5`, $1.5/$7.5, $0.15 cached) was removed in 0.21.0
+  // (2026-10-10): too slow and too expensive at the only reasoning level it takes (high|none).
+  // Production 2026-10-03..10: p50 15.9s, p90 35.1s, $0.036 a request (2x Gemini 3.1 Pro,
+  // 19x Small; reasoning spiked to 15-24k tokens), ~6% of calls failed (60s timeouts, empty
+  // replies, empty JSON), and a 15-player story pipeline took 277s. Wait for Medium 4.
   //
   // Mistral Large 3 and Magistral Medium 1.2 were dropped 2026-09-18. Large 3 was retired by
   // Mistral on 2026-08-31 (the `-latest` alias still answered, but its capabilities say
@@ -1205,7 +1208,7 @@ var SupportedAiModels = {
   // and `magistral-medium-latest` is now literally an alias of Medium 3.5 — the model list
   // returns `mistral-medium-3-5`, `mistral-medium-3`, `mistral-medium-latest` and
   // `magistral-medium-latest` as one entry — so Magistral seats had silently become Medium
-  // seats already. Consumers map both retired ids onto the surviving pair.
+  // seats already. Consumers map both retired ids onto Small (Medium itself is gone since 0.21.0).
   //
   // 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
@@ -1213,16 +1216,6 @@ var SupportedAiModels = {
   // 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
-    // now serves as an alias of the same model (Medium 3 itself retired 2026-08-31).
-    modelApiName: "mistral-medium-3-5",
-    apiKeyName: API_KEY_CONSTANTS.MISTRAL,
-    hasThinking: true,
-    reasoningEffort: "high",
-    tags: ["fast", "expensive"]
-  },
   [LLM_CONSTANTS.MISTRAL_SMALL]: {
     displayName: "Mistral 4 Small",
     // Resolves to mistral-small-2603 (Small 4), which also carries the magistral-small alias.
@@ -1610,11 +1603,6 @@ var MODEL_PRICING = {
   // 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.
-  [SupportedAiModels[LLM_CONSTANTS.MISTRAL_MEDIUM].modelApiName]: {
-    inputPrice: 1.5,
-    outputPrice: 7.5,
-    cacheHitPrice: 0.15
-  },
   [SupportedAiModels[LLM_CONSTANTS.MISTRAL_SMALL].modelApiName]: {
     inputPrice: 0.15,
     outputPrice: 0.6,
@@ -1695,8 +1683,7 @@ var HYBRID_THINKING_API_NAMES = /* @__PURE__ */ new Set([
   SupportedAiModels[LLM_CONSTANTS.QWEN_FLASH].modelApiName,
   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_MEDIUM].modelApiName,
+  // same definition as Qwen. Added 2026-09-18 when the Mistral entries went thinking-on.
   SupportedAiModels[LLM_CONSTANTS.MISTRAL_SMALL].modelApiName
 ]);
 function isHybridThinkingModel(modelApiName) {
@@ -5528,7 +5515,6 @@ 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_MEDIUM:
       case LLM_CONSTANTS.MISTRAL_SMALL:
         return new MistralAgent(name, instruction, model.modelApiName, key, shouldEnableThinking);
       case LLM_CONSTANTS.KIMI: