wolverine-ai 4.0.3 → 4.0.5

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "wolverine-ai",
-  "version": "4.0.3",
+  "version": "4.0.5",
   "description": "Self-healing Node.js server framework powered by AI. Catches crashes, diagnoses errors, generates fixes, verifies, and restarts — automatically.",
   "main": "src/index.js",
   "bin": {

package/src/brain/brain.js

@@ -133,6 +133,10 @@ const SEED_DOCS = [
     text: "Configuration: hybrid-always architecture — no provider selection. Users pick the best model for each of 8 task roles directly in settings.json 'models' section. Mix and match: wolverine for audit, claude for reasoning, gpt for coding. Provider auto-detected from model name. Embedding is separate ('embedding' key) — always wolverine-embedding-1 billed through credits (proxies to text-embedding-3-small at 2x markup). Secrets in .env.local. Config priority: env vars > settings.json > defaults.",
     metadata: { topic: "configuration" },
   },
+  {
+    text: "AI client prompt caching: all 3 providers cache automatically. Anthropic: system prompt marked cache_control:ephemeral, 90% cheaper on repeat calls within 5 min TTL. OpenAI: automatic prefix caching for >=1024 token prefixes, 50% cheaper on cached input, tracked via usage.prompt_tokens_details.cached_tokens. Wolverine/llama.cpp: cache_prompt:true in request body reuses KV cache for identical prefixes between requests, near-zero TTFT on second+ call in a heal pipeline. Cache savings tracked in analytics: cacheCreation (tokens written to cache) and cacheRead (tokens served from cache).",
+    metadata: { topic: "prompt-caching" },
+  },
   {
     text: "Platform telemetry: lightweight background process, zero-config. Default platform: api.wolverinenode.xyz. Auto-registers on first run (retries every 60s until platform responds), saves key to .wolverine/platform-key. Heartbeat payload matches PLATFORM.md spec: instanceId, server (name/port/uptime/status/pid), process (memoryMB/cpuPercent), routes, repairs, usage (tokens/cost/calls/byCategory), brain, backups. Offline-resilient: queues up to 1440 heartbeats locally, drains on reconnect. No chalk dependency, cached version/key in memory, minimal IO. Opt out: WOLVERINE_TELEMETRY=false. Override URL: WOLVERINE_PLATFORM_URL.",
     metadata: { topic: "platform-telemetry" },

package/src/brain/embedder.js

@@ -51,7 +51,7 @@ async function embed(text) {
     response = await client.embeddings.create({ model, input: text });
   } catch (err) {
     // If wolverine proxy is down (startup, crash loop), fall back to OpenAI direct
-    if (provider === "wolverine" && /ECONNREFUSED|ECONNRESET|ETIMEDOUT|fetch failed/i.test(err.message || "")) {
+    if (provider === "wolverine" && /ECONNREFUSED|ECONNRESET|ETIMEDOUT|fetch failed|Connection error/i.test(err.message || "")) {
       const directClient = getClient("openai");
       response = await directClient.embeddings.create({ model: "text-embedding-3-small", input: text });
     } else {
@@ -98,7 +98,7 @@ async function embedBatch(texts) {
   try {
     response = await client.embeddings.create({ model, input: uncached });
   } catch (err) {
-    if (provider === "wolverine" && /ECONNREFUSED|ECONNRESET|ETIMEDOUT|fetch failed/i.test(err.message || "")) {
+    if (provider === "wolverine" && /ECONNREFUSED|ECONNRESET|ETIMEDOUT|fetch failed|Connection error/i.test(err.message || "")) {
       const directClient = getClient("openai");
       response = await directClient.embeddings.create({ model: "text-embedding-3-small", input: uncached });
     } else {

package/src/core/ai-client.js

@@ -21,8 +21,9 @@ function _extractTokens(usage) {
     output: usage.completion_tokens || usage.output_tokens || 0,
     // Anthropic cache fields
     cacheCreation: usage.cache_creation_input_tokens || usage.cache_write_tokens || 0,
-    // OpenAI uses cache_read_tokens, Anthropic uses cache_read_input_tokens
-    cacheRead: usage.cache_read_input_tokens || usage.cache_read_tokens || 0,
+    // OpenAI prompt_tokens_details.cached_tokens + Anthropic cache_read_input_tokens
+    cacheRead: usage.cache_read_input_tokens || usage.cache_read_tokens
+      || usage.prompt_tokens_details?.cached_tokens || 0,
   };
 }
 
@@ -232,7 +233,7 @@ async function aiCall({ model, systemPrompt, userPrompt, maxTokens = 2048, tools
         result = await _chatCall(_getWolverineClient(), { model, systemPrompt, userPrompt, maxTokens, tools, toolChoice });
       } catch (proxyErr) {
         // If billing proxy is down (server crashing), fall back to direct GPU
-        const isConnErr = /ECONNREFUSED|ECONNRESET|ETIMEDOUT|fetch failed/i.test(proxyErr.message || "");
+        const isConnErr = /ECONNREFUSED|ECONNRESET|ETIMEDOUT|fetch failed|Connection error/i.test(proxyErr.message || "");
         const directClient = _getWolverineDirectClient();
         if (isConnErr && directClient) {
           console.log(chalk.yellow("  ⚠️  Billing proxy down — using direct GPU (unbilled)"));
@@ -269,7 +270,7 @@ async function aiCallWithHistory({ model, messages, tools, maxTokens = 4096, cat
       try {
         result = await _chatCallWithHistory(_getWolverineClient(), { model, messages, tools, maxTokens });
       } catch (proxyErr) {
-        const isConnErr = /ECONNREFUSED|ECONNRESET|ETIMEDOUT|fetch failed/i.test(proxyErr.message || "");
+        const isConnErr = /ECONNREFUSED|ECONNRESET|ETIMEDOUT|fetch failed|Connection error/i.test(proxyErr.message || "");
         const directClient = _getWolverineDirectClient();
         if (isConnErr && directClient) {
           console.log(chalk.yellow("  ⚠️  Billing proxy down — using direct GPU (unbilled)"));
@@ -520,19 +521,20 @@ async function _chatCall(openai, { model, systemPrompt, userPrompt, maxTokens, t
   if (systemPrompt) messages.push({ role: "system", content: systemPrompt });
   messages.push({ role: "user", content: userPrompt });
 
-  // No temperature for o-series and gpt-5+ (forbidden, causes error)
   const noTemp = /^(o[1-9]|gpt-5)/.test(model);
+  const isWolverine = detectProvider(model) === "wolverine";
   const params = {
     model, messages,
     ...(!noTemp ? { temperature: 0 } : {}),
     ...tokenParam(model, maxTokens),
     ..._reasoningParams(model),
+    // Prompt caching: llama.cpp reuses KV cache for identical prefixes
+    ...(isWolverine ? { cache_prompt: true } : {}),
   };
 
   if (tools && tools.length > 0) {
     params.tools = tools;
     params.tool_choice = toolChoice || "auto";
-    // Disable parallel calls for reliability — sequential is more predictable for healing
     params.parallel_tool_calls = false;
   }
 
@@ -589,11 +591,14 @@ async function _responsesCallWithHistory(openai, { model, messages, tools, maxTo
 
 async function _chatCallWithHistory(openai, { model, messages, tools, maxTokens }) {
   const noTemp = /^(o[1-9]|gpt-5)/.test(model);
+  const isWolverine = detectProvider(model) === "wolverine";
   const params = {
     model, messages,
     ...(!noTemp ? { temperature: 0 } : {}),
     ...tokenParam(model, maxTokens),
     ..._reasoningParams(model),
+    // Prompt caching: llama.cpp KV cache reuse for multi-turn agent conversations
+    ...(isWolverine ? { cache_prompt: true } : {}),
   };
   if (tools && tools.length > 0) {
     params.tools = tools;