@maintainabilityai/research-runner 0.1.5 → 0.1.7

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

@@ -11,16 +11,26 @@
  * their result types.
  *
  * Model names use GitHub Models namespacing — e.g. `openai/gpt-4o`,
- * `openai/gpt-4o-mini`, `openai/gpt-4.1`. The router (in
+ * `openai/gpt-4o-mini`, `openai/gpt-5-mini`. The router (in
  * llm-router.ts) maps internal logical model tiers (`plan` / `synth`) to
  * the concrete provider-specific id.
  */
 /**
  * Subset of GitHub Models model ids we use. Extend as new tiers land.
- * GitHub Models does not currently host Anthropic Claude — synth tier
- * uses `openai/gpt-4.1` (the "outperforms gpt-4o across the board" tier).
+ *
+ * GitHub Models has two relevant rate-limit tiers:
+ *   - "high" — gpt-4o, gpt-4o-mini, gpt-4.1 etc. Per-request input is
+ *     capped at ~8K tokens regardless of subscription. Fine for our
+ *     plan-tier (small structured-JSON prompt).
+ *   - "custom" — gpt-5 family, o-series. Per-request input scales to
+ *     the model's advertised limit (200K for gpt-5-mini). Routed through
+ *     Copilot-billed access, so the token-owner needs Copilot.
+ *
+ * Synth tier uses gpt-5-mini for the larger context window. Anthropic
+ * 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';
+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

@@ -12,7 +12,7 @@
  * their result types.
  *
  * Model names use GitHub Models namespacing — e.g. `openai/gpt-4o`,
- * `openai/gpt-4o-mini`, `openai/gpt-4.1`. The router (in
+ * `openai/gpt-4o-mini`, `openai/gpt-5-mini`. The router (in
  * llm-router.ts) maps internal logical model tiers (`plan` / `synth`) to
  * the concrete provider-specific id.
  */
@@ -25,8 +25,9 @@ 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.
+    // 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;
     const controller = new AbortController();
     const timer = setTimeout(() => controller.abort(), timeoutMs);
@@ -47,7 +48,12 @@ 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

@@ -6,7 +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' },
-    synth: { anthropic: 'claude-sonnet-4-6', githubModels: 'openai/gpt-4.1' },
+    // 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": "0.1.5",
+  "version": "0.1.7",
   "description": "Research + PRD agent runner — orchestrates the Archeologist and PRD pipelines for the MaintainabilityAI governance mesh",
   "license": "MIT",
   "author": "MaintainabilityAI",