@maintainabilityai/research-runner 0.1.13 → 0.1.15

package/dist/llm/llm-router.js

@@ -3,23 +3,32 @@ Object.defineProperty(exports, "__esModule", { value: true });
 exports.callLlm = callLlm;
 const anthropic_client_1 = require("./anthropic-client");
 const github_models_client_1 = require("./github-models-client");
-/** Per-tier per-provider model id lookup. */
+/**
+ * Per-tier per-provider model id lookup. `githubModelsFallback` (when
+ * present) is tried automatically on 401/403/404 from the primary —
+ * how we let users with a GMT (Copilot-Pro user PAT) reach the
+ * "custom"-tier `gpt-5-chat` while users on a workflow bot token
+ * still work via the "low"-tier fallback.
+ */
 const MODEL_BY_TIER = {
-    // gpt-4.1-mini outperforms gpt-4o-mini on the per-provider query-plan
-    // task. Verified empirically against the celeb-api brief: 4.1-mini
-    // produces more on-topic arxiv phrases ("celebrity identity disambig-
-    // uation" vs 4o-mini's generic "API integration challenges"), tighter
-    // patent AND-clauses, and stays inside the spec's word counts more
-    // reliably. Same "low" rate-limit tier as 4o-mini, so no infra change.
-    plan: { anthropic: 'claude-haiku-4-5', githubModels: 'openai/gpt-4.1-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
+    // Plan tier — small structured-JSON output, one shot, fits any cap.
+    // Try gpt-5-chat first (custom tier, only reachable with a Copilot-
+    // enrolled PAT like GMT). Fall back to gpt-4.1-mini (low tier, works
+    // on every token including the Actions bot). Empirically equivalent
+    // output quality on the V3-anchor prompt — the upgrade is "if free,
+    // why not"; the fallback keeps everyone working.
+    plan: {
+        anthropic: 'claude-haiku-4-5',
+        githubModels: 'openai/gpt-5-chat',
+        githubModelsFallback: 'openai/gpt-4.1-mini',
+    },
+    // Synth tier — 200K context, non-reasoning (verified live with
+    // 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).
+    // visible markdown. Synth needs predictable structured output. No
+    // fallback model here — synth runs on the agent side now (Copilot
+    // Coding Agent / @claude), so the runner doesn't fire synth itself.
     synth: { anthropic: 'claude-sonnet-4-6', githubModels: 'openai/gpt-5-chat' },
 };
 /**
@@ -57,28 +66,54 @@ async function callAnthropicTier(opts, tierModels) {
         httpStatus: r.httpStatus,
     };
 }
+/**
+ * GitHub Models returns 401 / 403 / 404 when the token can't reach the
+ * requested model (typically the workflow bot token hitting a "custom"-
+ * tier model like gpt-5-chat). These are recoverable via fallback;
+ * everything else (timeouts, 5xx, 413 cap, parse errors) should
+ * propagate.
+ */
+function isModelAccessError(err) {
+    if (!(err instanceof Error)) {
+        return false;
+    }
+    return /GitHub Models returned 40[134]:/.test(err.message);
+}
 async function callGitHubModelsTier(opts, tierModels) {
     if (!opts.githubToken) {
         throw new Error(`callLlm: provider=github-models requires githubToken (set GITHUB_TOKEN; workflow needs \`permissions: models: read\`).`);
     }
-    const r = await (0, github_models_client_1.callGitHubModels)({
-        token: opts.githubToken,
-        model: tierModels.githubModels,
-        system: opts.system,
-        prompt: opts.prompt,
-        maxTokens: opts.maxTokens,
-        temperature: opts.temperature,
-        fetchImpl: opts.fetchImpl,
-    });
-    return {
-        provider: 'github-models',
-        model: tierModels.githubModels,
-        text: r.text,
-        inputTokens: r.inputTokens,
-        outputTokens: r.outputTokens,
-        costUsd: r.costUsd,
-        httpStatus: r.httpStatus,
+    const callOne = async (model) => {
+        const r = await (0, github_models_client_1.callGitHubModels)({
+            token: opts.githubToken,
+            model,
+            system: opts.system,
+            prompt: opts.prompt,
+            maxTokens: opts.maxTokens,
+            temperature: opts.temperature,
+            fetchImpl: opts.fetchImpl,
+        });
+        return {
+            provider: 'github-models',
+            model,
+            text: r.text,
+            inputTokens: r.inputTokens,
+            outputTokens: r.outputTokens,
+            costUsd: r.costUsd,
+            httpStatus: r.httpStatus,
+        };
     };
+    try {
+        return await callOne(tierModels.githubModels);
+    }
+    catch (err) {
+        if (tierModels.githubModelsFallback && isModelAccessError(err)) {
+            const cause = err instanceof Error ? err.message : String(err);
+            process.stderr.write(`[research-runner] ⚠ github-models ${tierModels.githubModels} access-denied, falling back to ${tierModels.githubModelsFallback}. Cause: ${cause.slice(0, 200)}\n`);
+            return await callOne(tierModels.githubModelsFallback);
+        }
+        throw err;
+    }
 }
 async function callLlm(opts) {
     const tierModels = MODEL_BY_TIER[opts.tier];

package/dist/runner/archeologist.js

@@ -515,7 +515,9 @@ async function runArcheologist(opts) {
                 provider: brief.llm_provider,
                 // plan_queries is the only LLM hop we run now (synth handed off
                 // to the assigned agent). Surface that model in the Hatter's Tag.
-                model: 'openai/gpt-4.1-mini',
+                // Plan-tier primary (router falls back to gpt-4.1-mini on access denial).
+                // Synth runs on the agent side, so this is the only LLM hop the runner makes.
+                model: 'openai/gpt-5-chat',
                 input_tokens: totalInputTokens,
                 output_tokens: totalOutputTokens,
                 cost_usd: roundUsd(totalCostUsd),

package/dist/runner/nodes/hackernews-search.js

@@ -40,7 +40,10 @@ async function runHackerNewsSearch(opts) {
                     fromQuery: query,
                     title: r.title,
                     url,
-                    content: '', // HN search returns no abstract
+                    // Show HN / Ask HN posts carry a self-post body. URL-only
+                    // submissions have no body — leave empty; the synth agent
+                    // can read the linked URL if it needs more context.
+                    content: r.storyText,
                     score: pointsToScore(r.points),
                     publishedDate: r.createdAt || undefined,
                     authors: r.author ? [r.author] : undefined,

package/dist/search/hackernews-client.d.ts

@@ -15,6 +15,8 @@ export interface HackerNewsResult {
     points: number;
     numComments: number;
     createdAt: string;
+    /** Self-post body for Show HN / Ask HN; empty for URL submissions. */
+    storyText: string;
 }
 export interface HackerNewsSearchOpts {
     query: string;

package/dist/search/hackernews-client.js

@@ -39,6 +39,29 @@ async function hackerNewsSearch(opts) {
         points: h.points ?? 0,
         numComments: h.num_comments ?? 0,
         createdAt: h.created_at ?? '',
+        // Algolia returns the self-post body for Show HN / Ask HN. URL-only
+        // submissions have this empty. Strip basic HTML tags so the excerpt
+        // is readable in the issue comment.
+        storyText: stripBasicHtml(h.story_text ?? '').slice(0, 2000),
     })).filter(r => r.objectId && r.title);
     return { query: opts.query, results, responseBytes: Buffer.byteLength(rawText, 'utf8'), httpStatus };
 }
+/**
+ * HN Algolia returns Show HN / Ask HN bodies with light HTML
+ * (`<p>`, `<i>`, etc.). Strip tags + decode the common entities so
+ * the excerpt blockquote in the issue body reads cleanly. No need for
+ * a full HTML parser — these posts are plain text with a sprinkle of
+ * inline tags.
+ */
+function stripBasicHtml(s) {
+    return s
+        .replace(/<\/?(?:p|br|i|b|em|strong|code|pre|a|ul|ol|li)[^>]*>/gi, ' ')
+        .replace(/&nbsp;/g, ' ')
+        .replace(/&amp;/g, '&')
+        .replace(/&lt;/g, '<')
+        .replace(/&gt;/g, '>')
+        .replace(/&quot;/g, '"')
+        .replace(/&#x27;/g, "'")
+        .replace(/\s+/g, ' ')
+        .trim();
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@maintainabilityai/research-runner",
-  "version": "0.1.13",
+  "version": "0.1.15",
   "description": "Research + PRD agent runner — orchestrates the Archeologist and PRD pipelines for the MaintainabilityAI governance mesh",
   "license": "MIT",
   "author": "MaintainabilityAI",