@maintainabilityai/research-runner 0.1.11 → 0.1.13

package/dist/llm/llm-router.js

@@ -5,7 +5,13 @@ const anthropic_client_1 = require("./anthropic-client");
 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-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

package/dist/runner/archeologist.js

@@ -515,7 +515,7 @@ 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',
+                model: 'openai/gpt-4.1-mini',
                 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

@@ -30,6 +30,19 @@ function canonicalizeUrl(rawUrl) {
         return rawUrl.trim().toLowerCase();
     }
 }
+/**
+ * Per-provider quota for the top-N output. Without these floors, Tavily
+ * (normalized scores 0.9–1.0) crushes every other provider in pure
+ * global ranking — synth would see zero HN signal and zero patent
+ * coverage. Quotas sum to topN's default (20). Any unused slack
+ * spills over to the highest-scoring non-quota entries across providers.
+ */
+const PROVIDER_QUOTA = {
+    tavily: 8,
+    arxiv: 5,
+    uspto: 4,
+    hackernews: 3,
+};
 function dedupeAndRank(opts) {
     const topN = opts.topN ?? 20;
     const retrievedAt = opts.retrievedAt ?? new Date().toISOString();
@@ -48,7 +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) {
@@ -63,7 +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,
@@ -72,12 +85,46 @@ function dedupeAndRank(opts) {
             });
         }
     }
-    const ranked = [...bucket.values()]
-        .map(a => {
+    const allEntries = [...bucket.values()].map(a => {
         const recall = 1 + 0.15 * (a.queries.size - 1);
         const composite = Math.min(1, a.scoreSum * recall / Math.max(1, a.occurrences));
         return { aggregated: a, composite };
-    })
+    });
+    // Phase 1 — per-provider quota: take each provider's top-K (K from PROVIDER_QUOTA).
+    // Phase 2 — spillover: fill the remaining budget with the next-highest entries
+    //          from anywhere, including providers that have already filled their quota.
+    // Phase 3 — re-sort the combined set by composite score for stable display order.
+    const used = new Set();
+    const picks = [];
+    for (const provider of Object.keys(PROVIDER_QUOTA)) {
+        const k = PROVIDER_QUOTA[provider];
+        if (k === 0) {
+            continue;
+        }
+        const fromProvider = allEntries
+            .filter(e => e.aggregated.provider === provider)
+            .sort((a, b) => b.composite - a.composite)
+            .slice(0, k);
+        for (const e of fromProvider) {
+            if (used.has(e.aggregated.canonicalUrl)) {
+                continue;
+            }
+            picks.push(e);
+            used.add(e.aggregated.canonicalUrl);
+        }
+    }
+    const remainingBudget = Math.max(0, topN - picks.length);
+    if (remainingBudget > 0) {
+        const spillover = allEntries
+            .filter(e => !used.has(e.aggregated.canonicalUrl))
+            .sort((a, b) => b.composite - a.composite)
+            .slice(0, remainingBudget);
+        for (const e of spillover) {
+            picks.push(e);
+            used.add(e.aggregated.canonicalUrl);
+        }
+    }
+    const ranked = picks
         .sort((a, b) => b.composite - a.composite)
         .slice(0, topN);
     return ranked.map((entry, i) => ({
@@ -87,7 +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) {
@@ -46,8 +48,12 @@ function providerSection(label, emoji, provider, sources, totalCount) {
     for (const s of sources) {
         const authors = s.authors && s.authors.length > 0 ? ` — _${s.authors.slice(0, 3).join(', ')}${s.authors.length > 3 ? ' et al.' : ''}_` : '';
         const date = s.published_at ? ` _(${s.published_at.slice(0, 10)})_` : '';
-        lines.push(`- **[\`${s.id}\`] [${s.title}](${s.url})** — score ${s.salience_score.toFixed(2)}${date}${authors}`);
-        lines.push(`  > ${shortExcerpt(s.excerpt)}`);
+        // Render the citation id as standalone inline code so the synth
+        // agent (and a downstream PRD agent) can grep `\bS\d+\b` cleanly.
+        // The earlier form **[`S1`] [Title](url)** broke GitHub's markdown
+        // parser (it tried to interpret the brackets as a reference link).
+        lines.push(`- \`${s.id}\` **[${s.title}](${s.url})** — score ${s.salience_score.toFixed(2)}${date}${authors}`);
+        lines.push(`  > ${normaliseExcerpt(s.excerpt)}`);
     }
     lines.push('');
     return lines;
@@ -159,6 +165,6 @@ function formatForHuman(opts) {
     lines.push('');
     lines.push('---');
     lines.push('');
-    lines.push(`🤖 Generated by \`research-runner archeologist\` v${'pkg'}. Run id: \`${runId}\`.`);
+    lines.push(`🤖 Generated by \`research-runner archeologist\`. Run id: \`${runId}\` (see the Hatter's Tag for agent version + audit chain).`);
     return { body: lines.join('\n') };
 }

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": "0.1.11",
+  "version": "0.1.13",
   "description": "Research + PRD agent runner — orchestrates the Archeologist and PRD pipelines for the MaintainabilityAI governance mesh",
   "license": "MIT",
   "author": "MaintainabilityAI",