@maintainabilityai/research-runner 0.1.12 → 0.1.14

package/dist/llm/llm-router.js

@@ -3,17 +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 = {
-    plan: { anthropic: 'claude-haiku-4-5', githubModels: 'openai/gpt-4o-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' },
 };
 /**
@@ -51,28 +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-4o-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/arxiv-search.js

@@ -35,7 +35,7 @@ async function runArxivSearch(opts) {
                     fromQuery: query,
                     title: r.title,
                     url: r.abstractUrl,
-                    content: r.summary.slice(0, 500),
+                    content: r.summary.slice(0, 2000),
                     // Position-derived score: arXiv returns by relevance, decay 0.9 → 0.5.
                     score: Math.max(0.5, 0.9 - j * 0.1),
                     publishedDate: r.published || undefined,

package/dist/runner/nodes/dedupe-and-rank.js

@@ -61,7 +61,7 @@ function dedupeAndRank(opts) {
             if (r.score > existing.scoreSum / existing.occurrences) {
                 existing.title = r.title || existing.title;
                 if (r.content) {
-                    existing.excerpt = r.content.slice(0, 500);
+                    existing.excerpt = r.content.slice(0, 2000);
                 }
             }
             if (!existing.publishedAt && r.publishedDate) {
@@ -76,7 +76,7 @@ function dedupeAndRank(opts) {
                 canonicalUrl: canonical,
                 provider: r.provider,
                 title: r.title || canonical,
-                excerpt: (r.content || '').slice(0, 500),
+                excerpt: (r.content || '').slice(0, 2000),
                 publishedAt: r.publishedDate,
                 authors: r.authors,
                 scoreSum: r.score,
@@ -134,7 +134,7 @@ function dedupeAndRank(opts) {
         url: entry.aggregated.canonicalUrl,
         retrieved_at: retrievedAt,
         salience_score: roundTo(entry.composite, 4),
-        excerpt: entry.aggregated.excerpt.slice(0, 500),
+        excerpt: entry.aggregated.excerpt.slice(0, 2000),
         ...(entry.aggregated.publishedAt ? { published_at: entry.aggregated.publishedAt } : {}),
         ...(entry.aggregated.authors && entry.aggregated.authors.length > 0 ? { authors: entry.aggregated.authors } : {}),
     }));

package/dist/runner/nodes/format-for-human.js

@@ -2,12 +2,14 @@
 Object.defineProperty(exports, "__esModule", { value: true });
 exports.formatForHuman = formatForHuman;
 /**
- * Truncates an excerpt for display in the issue comment. Sources can
- * carry up to 500 chars; we shorten to ~280 so the issue stays readable.
+ * Normalise the excerpt for blockquote display — collapse whitespace
+ * runs so newlines in arXiv abstracts don't break the markdown quote.
+ * No length cap here: the schema's 2000-char limit is what the agent
+ * downstream sees, and a downstream synthesis or PRD agent needs the
+ * full excerpt to write faithful citations.
  */
-function shortExcerpt(s, n = 280) {
-    const cleaned = s.replace(/\s+/g, ' ').trim();
-    return cleaned.length <= n ? cleaned : cleaned.slice(0, n - 1) + '…';
+function normaliseExcerpt(s) {
+    return s.replace(/\s+/g, ' ').trim();
 }
 function meshSummary(meshContext) {
     if (meshContext.bar) {
@@ -51,7 +53,7 @@ function providerSection(label, emoji, provider, sources, totalCount) {
         // The earlier form **[`S1`] [Title](url)** broke GitHub's markdown
         // parser (it tried to interpret the brackets as a reference link).
         lines.push(`- \`${s.id}\` **[${s.title}](${s.url})** — score ${s.salience_score.toFixed(2)}${date}${authors}`);
-        lines.push(`  > ${shortExcerpt(s.excerpt)}`);
+        lines.push(`  > ${normaliseExcerpt(s.excerpt)}`);
     }
     lines.push('');
     return lines;

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

@@ -46,7 +46,7 @@ async function runUsptoSearch(opts) {
                     fromQuery: query,
                     title: r.title,
                     url: r.url,
-                    content: r.abstract.slice(0, 500),
+                    content: r.abstract.slice(0, 2000),
                     score: Math.max(0.4, 0.85 - j * 0.1),
                     publishedDate: r.grantedAt || undefined,
                     authors: r.inventors,

package/dist/schemas/ranked-source.d.ts

@@ -15,7 +15,10 @@ export declare const RankedSource: z.ZodObject<{
     retrieved_at: z.ZodEffects<z.ZodString, string, string>;
     /** 0.0 - 1.0, higher = more relevant. Computed by dedupe_and_rank. */
     salience_score: z.ZodNumber;
-    /** ≤500-char excerpt the synthesis node may quote directly. */
+    /** ≤2000-char excerpt the synthesis agent quotes from. Sized for the
+     *  current flow where the agent (Copilot Coding Agent / @claude) reads
+     *  the issue comment — they have plenty of context budget, and 500
+     *  routinely truncated arXiv abstracts and patent summaries mid-thought. */
     excerpt: z.ZodString;
     /** Optional: pub date if the source has one (papers, news, patents). */
     published_at: z.ZodOptional<z.ZodEffects<z.ZodString, string, string>>;
@@ -52,7 +55,10 @@ export declare const RankedSourceList: z.ZodArray<z.ZodObject<{
     retrieved_at: z.ZodEffects<z.ZodString, string, string>;
     /** 0.0 - 1.0, higher = more relevant. Computed by dedupe_and_rank. */
     salience_score: z.ZodNumber;
-    /** ≤500-char excerpt the synthesis node may quote directly. */
+    /** ≤2000-char excerpt the synthesis agent quotes from. Sized for the
+     *  current flow where the agent (Copilot Coding Agent / @claude) reads
+     *  the issue comment — they have plenty of context budget, and 500
+     *  routinely truncated arXiv abstracts and patent summaries mid-thought. */
     excerpt: z.ZodString;
     /** Optional: pub date if the source has one (papers, news, patents). */
     published_at: z.ZodOptional<z.ZodEffects<z.ZodString, string, string>>;

package/dist/schemas/ranked-source.js

@@ -19,8 +19,11 @@ exports.RankedSource = zod_1.z.object({
     retrieved_at: primitives_1.IsoTimestamp,
     /** 0.0 - 1.0, higher = more relevant. Computed by dedupe_and_rank. */
     salience_score: zod_1.z.number().min(0).max(1),
-    /** ≤500-char excerpt the synthesis node may quote directly. */
-    excerpt: zod_1.z.string().max(500),
+    /** ≤2000-char excerpt the synthesis agent quotes from. Sized for the
+     *  current flow where the agent (Copilot Coding Agent / @claude) reads
+     *  the issue comment — they have plenty of context budget, and 500
+     *  routinely truncated arXiv abstracts and patent summaries mid-thought. */
+    excerpt: zod_1.z.string().max(2000),
     /** Optional: pub date if the source has one (papers, news, patents). */
     published_at: primitives_1.IsoTimestamp.optional(),
     /** Optional: authors (arxiv / news). */

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@maintainabilityai/research-runner",
-  "version": "0.1.12",
+  "version": "0.1.14",
   "description": "Research + PRD agent runner — orchestrates the Archeologist and PRD pipelines for the MaintainabilityAI governance mesh",
   "license": "MIT",
   "author": "MaintainabilityAI",