@hiper2d/ai-agents 0.4.0 → 0.4.2

package/dist/index.js

@@ -1023,9 +1023,11 @@ var SupportedAiModels = {
   // knob exists) both models emitted ~8 reasoning tokens per answer token in prod
   // (requestStats 30d: flash p50 8.9s / p90 36s, pro p50 18.9s / p90 56s) and a 15-bot story
   // took 68-105s. Latency tracks reasoning length ~linearly, so effort is the only lever.
+  // Flash moved to V4.1 2026-09-12: the API id is now the version-free alias `deepseek-flash`
+  // (the retired `deepseek-v4-flash` still resolves to V4.1 server-side at the Flash price).
   [LLM_CONSTANTS.DEEPSEEK_FLASH]: {
-    displayName: "DeepSeek V4 Flash",
-    modelApiName: "deepseek-v4-flash",
+    displayName: "DeepSeek V4.1 Flash",
+    modelApiName: "deepseek-flash",
     apiKeyName: API_KEY_CONSTANTS.DEEPSEEK,
     hasThinking: true,
     reasoningEffort: "low",
@@ -1273,14 +1275,18 @@ var SupportedAiModels = {
   // encrypted reasoning items replayed across turns like Grok. 'medium' is pinned as the
   // game default: turns are short and every reasoning token bills as output.
   // Temperature: Meta documents the model as tuned to its 1.0 default.
-  // Speed/tags: unmeasured until the first live run — no tag rather than a guess.
+  // Speed: measured 2026-09-12 at medium effort — 9.6s on a full-context day-2 vote, 4-8s on
+  // short turns (reasoning ≈ 90% of output tokens). That is the untagged middle band by the
+  // grading above; tagged 'slow' anyway by decision so players expect a wait. Price-wise it
+  // is neither cheap nor expensive.
   [LLM_CONSTANTS.MUSE_SPARK]: {
     displayName: "Muse Spark 1.3",
     modelApiName: "muse-spark-1.3",
     apiKeyName: API_KEY_CONSTANTS.META,
     hasThinking: true,
     temperature: 1,
-    reasoningEffort: "medium"
+    reasoningEffort: "medium",
+    tags: ["slow"]
   }
 };
 function createCatalog(overrides = {}) {
@@ -1392,10 +1398,12 @@ var MODEL_PRICING = {
   // (api-docs.deepseek.com/quick_start/pricing, fetched 2026-08-13; rates re-confirmed
   // 2026-08-30). Since 2026-08-23 00:00 Beijing (UTC+8) the surcharge is weekdays-only:
   // Saturday and Sunday Beijing time bill at the off-peak rate all day (DeepSeek notice email).
+  // 2026-09-12: Flash is V4.1 at lower rates ($0.15 / $0.60, cache hit $0.003 off-peak);
+  // Pro rates unchanged and DeepSeek confirmed Pro stays available past 2026-09-14.
   [SupportedAiModels[LLM_CONSTANTS.DEEPSEEK_FLASH].modelApiName]: {
-    inputPrice: 0.22,
-    outputPrice: 0.66,
-    cacheHitPrice: 7e-3,
+    inputPrice: 0.15,
+    outputPrice: 0.6,
+    cacheHitPrice: 3e-3,
     peakPricing: DEEPSEEK_PEAK_SCHEDULE
   },
   [SupportedAiModels[LLM_CONSTANTS.DEEPSEEK_PRO].modelApiName]: {
@@ -3298,7 +3306,7 @@ var GoogleAgent = class extends AbstractAgent {
 
 // src/agents/mistral-agent.ts
 var import_mistralai = require("@mistralai/mistralai");
-var import_http = require("@mistralai/mistralai/lib/http");
+var import_http = require("@mistralai/mistralai/lib/http.js");
 var MistralAgent = class extends AbstractAgent {
   client;
   // A getter, not a field: `maxOutputTokens` can be raised after construction, and a field
@@ -5022,7 +5030,16 @@ ${schemaDescription}`;
         try {
           const reasoningItems = JSON.parse(msg.metaEncryptedReasoning);
           if (Array.isArray(reasoningItems)) {
-            input.push(...reasoningItems);
+            for (const item of reasoningItems) {
+              if (item?.encrypted_content) {
+                input.push({
+                  type: "reasoning",
+                  ...item.id ? { id: item.id } : {},
+                  summary: Array.isArray(item.summary) ? item.summary : [],
+                  encrypted_content: item.encrypted_content
+                });
+              }
+            }
           }
         } catch {
           this.logger(`Failed to parse stored encrypted reasoning, replaying message without it`);