@maintainabilityai/research-runner 0.1.2 → 0.1.4

package/dist/llm/anthropic-client.js

@@ -23,8 +23,10 @@ async function callAnthropic(opts) {
         throw new Error('ANTHROPIC_API_KEY missing — set the env var or pass apiKey directly');
     }
     const fetchImpl = opts.fetchImpl ?? globalThis.fetch;
+    // Match github-models default — 8K-token synth responses can run 60–90s.
+    const timeoutMs = opts.timeoutMs ?? 120_000;
     const controller = new AbortController();
-    const timer = setTimeout(() => controller.abort(), opts.timeoutMs ?? 60_000);
+    const timer = setTimeout(() => controller.abort(), timeoutMs);
     let response;
     try {
         response = await fetchImpl('https://api.anthropic.com/v1/messages', {
@@ -44,6 +46,12 @@ async function callAnthropic(opts) {
             signal: controller.signal,
         });
     }
+    catch (err) {
+        if (err instanceof Error && err.name === 'AbortError') {
+            throw new Error(`Anthropic request timed out after ${timeoutMs}ms (model=${opts.model}, max_tokens=${opts.maxTokens})`);
+        }
+        throw new Error(`Anthropic fetch failed (model=${opts.model}): ${err instanceof Error ? err.message : String(err)}`);
+    }
     finally {
         clearTimeout(timer);
     }

package/dist/llm/github-models-client.js

@@ -25,8 +25,11 @@ async function callGitHubModels(opts) {
     }
     const fetchImpl = opts.fetchImpl ?? globalThis.fetch;
     const endpoint = opts.endpoint ?? DEFAULT_ENDPOINT;
+    // Synthesis prompts can produce 8K-token responses on gpt-4.1, which
+    // routinely take 60–90s. Default to 120s so we don't abort mid-stream.
+    const timeoutMs = opts.timeoutMs ?? 120_000;
     const controller = new AbortController();
-    const timer = setTimeout(() => controller.abort(), opts.timeoutMs ?? 60_000);
+    const timer = setTimeout(() => controller.abort(), timeoutMs);
     const messages = [];
     if (opts.system) {
         messages.push({ role: 'system', content: opts.system });
@@ -50,6 +53,12 @@ async function callGitHubModels(opts) {
             signal: controller.signal,
         });
     }
+    catch (err) {
+        if (err instanceof Error && err.name === 'AbortError') {
+            throw new Error(`GitHub Models request timed out after ${timeoutMs}ms (model=${opts.model}, max_tokens=${opts.maxTokens})`);
+        }
+        throw new Error(`GitHub Models fetch failed (model=${opts.model}): ${err instanceof Error ? err.message : String(err)}`);
+    }
     finally {
         clearTimeout(timer);
     }

package/dist/llm/llm-router.js

@@ -10,7 +10,16 @@ const MODEL_BY_TIER = {
 };
 async function callLlm(opts) {
     const tierModels = MODEL_BY_TIER[opts.tier];
-    if (opts.provider === 'anthropic') {
+    // Hybrid routing: GitHub Models free tier caps requests at ~8K input
+    // tokens — too small for the synthesis step (full brief + every search
+    // result + mesh context routinely exceeds that). When the brief asks
+    // for github-models AND an Anthropic key is available, route synth →
+    // Anthropic and keep plan (small prompt) on github-models. Caller can
+    // force pure github-models by not setting anthropicApiKey.
+    const effectiveProvider = opts.provider === 'github-models' && opts.tier === 'synth' && opts.anthropicApiKey
+        ? 'anthropic'
+        : opts.provider;
+    if (effectiveProvider === 'anthropic') {
         if (!opts.anthropicApiKey) {
             throw new Error(`callLlm: provider=anthropic requires anthropicApiKey (set ANTHROPIC_API_KEY).`);
         }
@@ -33,7 +42,7 @@ async function callLlm(opts) {
             httpStatus: r.httpStatus,
         };
     }
-    if (opts.provider === 'github-models') {
+    if (effectiveProvider === 'github-models') {
         if (!opts.githubToken) {
             throw new Error(`callLlm: provider=github-models requires githubToken (set GITHUB_TOKEN; workflow needs \`permissions: models: read\`).`);
         }
@@ -56,5 +65,5 @@ async function callLlm(opts) {
             httpStatus: r.httpStatus,
         };
     }
-    throw new Error(`callLlm: provider "${opts.provider}" not yet implemented (phase 2c.1 ships anthropic + github-models; openai + azure-openai land later).`);
+    throw new Error(`callLlm: provider "${effectiveProvider}" not yet implemented (phase 2c.1 ships anthropic + github-models; openai + azure-openai land later).`);
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@maintainabilityai/research-runner",
-  "version": "0.1.2",
+  "version": "0.1.4",
   "description": "Research + PRD agent runner — orchestrates the Archeologist and PRD pipelines for the MaintainabilityAI governance mesh",
   "license": "MIT",
   "author": "MaintainabilityAI",