@maintainabilityai/research-runner 0.1.6 → 0.1.8

package/dist/llm/anthropic-client.js

@@ -23,8 +23,12 @@ 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;
+    // Sonnet at ~50–80 tok/s × 8K output = 100–160s, plus prompt-processing.
+    // 120s aborted real synth runs mid-stream; 240s gives headroom for
+    // slow days. (Real fix is streaming so the connection stays alive,
+    // but the cost/benefit isn't worth it yet — single-shot is fine
+    // until we hit 240s legitimately.)
+    const timeoutMs = opts.timeoutMs ?? 240_000;
     const controller = new AbortController();
     const timer = setTimeout(() => controller.abort(), timeoutMs);
     let response;

package/dist/llm/github-models-client.d.ts

@@ -30,7 +30,7 @@
  * remains the preferred synth target when an Anthropic key is set (see
  * llm-router.ts hybrid routing).
  */
-export type GitHubModelsModel = 'openai/gpt-4o' | 'openai/gpt-4o-mini' | 'openai/gpt-4.1' | 'openai/gpt-4.1-mini' | 'openai/gpt-5' | 'openai/gpt-5-mini';
+export type GitHubModelsModel = 'openai/gpt-4o' | 'openai/gpt-4o-mini' | 'openai/gpt-4.1' | 'openai/gpt-4.1-mini' | 'openai/gpt-5' | 'openai/gpt-5-mini' | 'openai/gpt-5-chat';
 export interface CallGitHubModelsOpts {
     /** Workflow GITHUB_TOKEN. The model server checks the `models:read` permission scope. */
     token: string;

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

@@ -25,10 +25,13 @@ async function callGitHubModels(opts) {
     }
     const fetchImpl = opts.fetchImpl ?? globalThis.fetch;
     const endpoint = opts.endpoint ?? DEFAULT_ENDPOINT;
-    // Synthesis prompts can produce 8K-token responses (and the "custom"
-    // tier models like gpt-5-mini can return much more), which routinely
-    // take 60–120s. Default to 120s so we don't abort mid-stream.
-    const timeoutMs = opts.timeoutMs ?? 120_000;
+    // Synthesis prompts can produce 8K+ output tokens (and "custom"-tier
+    // models like gpt-5-chat can return much more), which routinely take
+    // 100–180s. Match the Anthropic client at 240s for headroom on slow
+    // days. (Real fix is streaming so the connection stays alive, but
+    // the cost/benefit isn't worth it yet — single-shot is fine until
+    // we hit 240s legitimately.)
+    const timeoutMs = opts.timeoutMs ?? 240_000;
     const controller = new AbortController();
     const timer = setTimeout(() => controller.abort(), timeoutMs);
     const messages = [];
@@ -48,7 +51,12 @@ async function callGitHubModels(opts) {
             body: JSON.stringify({
                 model: opts.model,
                 messages,
-                max_tokens: opts.maxTokens,
+                // `max_completion_tokens` is the current Chat Completions param;
+                // `max_tokens` is rejected outright by gpt-5 family models with
+                // HTTP 400. Verified gpt-4o-mini accepts the new name too, so we
+                // use one code path. Temperature stays optional (gpt-5 reasoning
+                // models may ignore it but accept it).
+                max_completion_tokens: opts.maxTokens,
                 temperature: opts.temperature ?? 0,
             }),
             signal: controller.signal,

package/dist/llm/llm-router.js

@@ -6,11 +6,15 @@ const github_models_client_1 = require("./github-models-client");
 /** Per-tier per-provider model id lookup. */
 const MODEL_BY_TIER = {
     plan: { anthropic: 'claude-haiku-4-5', githubModels: 'openai/gpt-4o-mini' },
-    // gpt-5-mini is in the "custom" GH-Models tier — 200K input context,
-    // 100K output, reasoning + tool-calling. Bypasses the 8K cap that
-    // hits "high"-tier models like gpt-4.1. Requires the caller's token
-    // to have Models access through a Copilot subscription (GMT path).
-    synth: { anthropic: 'claude-sonnet-4-6', githubModels: 'openai/gpt-5-mini' },
+    // gpt-5-chat is in the "custom" GH-Models tier (200K input / 100K
+    // output) and is NON-reasoning — verified end-to-end with a live API
+    // call (reasoning_tokens=0, finish_reason=stop). Picked over gpt-5
+    // and gpt-5-mini because those are reasoning models that consume the
+    // completion budget on hidden chain-of-thought before producing any
+    // visible markdown — bad for the synthesis step where we need
+    // predictable structured output. Requires the caller's token to have
+    // Models access through a Copilot subscription (GMT path).
+    synth: { anthropic: 'claude-sonnet-4-6', githubModels: 'openai/gpt-5-chat' },
 };
 async function callLlm(opts) {
     const tierModels = MODEL_BY_TIER[opts.tier];

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@maintainabilityai/research-runner",
-  "version": "0.1.6",
+  "version": "0.1.8",
   "description": "Research + PRD agent runner — orchestrates the Archeologist and PRD pipelines for the MaintainabilityAI governance mesh",
   "license": "MIT",
   "author": "MaintainabilityAI",