@hiper2d/ai-agents 0.26.2 → 0.26.3

package/dist/index.js

@@ -1115,7 +1115,9 @@ 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;
@@ -1195,7 +1197,9 @@ 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",
@@ -1246,7 +1250,10 @@ 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