Research

Research Router

Route each request to the right source class; return a source-backed answer with explicit provenance, confidence, and gaps.

Routing policy

• Prefer active tools for common search, URL retrieval, repository inspection, and browsing.
• Use a specialized skill when its source or behavior better matches the request.
• Open-web questions: prefer live data.
• Choose the smallest route set sufficient to answer.
• Thin evidence: add a second source class for corroboration.
• Do not run multiple routes duplicating a provider and its evidence.

Route map

Code/docs

• Repository inspection/code search: available GitHub or repository tools, or gh.
• Repository docs/Q&A over owner/repo: deepwiki.
• Current library/API docs and examples: context7.
• Public OSS usage patterns: grep-app.

Web/live

• Ordinary live discovery: built-in web search.
• Use omp-search when its OMP CLI automatic provider chain, explicit provider selection, or structured headless output is useful.
• Known static pages: URL retrieval tool; browser tool only for interaction, authentication, or JavaScript-rendered pages.
• Direct Brave-backed scoping and recency checks: brave-search.

Sentiment/discussion

• Community sentiment, discussion trends, and user/topic signals: reddit.
• When feasible, corroborate high-impact claims with one non-Reddit source class.

X/Twitter

• Explicit X/Twitter post URLs or IDs, bounded timelines, post search, and conversations: x-research.
• X-plus-web news: use x-research for bounded X evidence, then an independent web route.
• X results are bounded, unofficial evidence; never infer public-opinion truth, deletion, or suspension from a missing result.

Model benchmark comparisons

• Use artificial-analysis-live, deepswe-live, livebench-live, and vals-live for their respective published results. Load only the sources needed for the question; use all four when the user requests that coverage.
• Follow each owning skill for commands, credentials, extraction, and cache recovery. Never substitute a remembered workaround or silently fall back to stale data.
• Preserve model, reasoning effort, provider, harness, configuration, benchmark version, task population, metric units, and freshness wherever published. Mark unknown fields and mismatches before comparing rows.
• Separate published values from derived deltas, ratios, or rankings. Compare costs only when their units and workload assumptions align. Missing observations remain missing, not zero or proof of non-evaluation.
• Do not deduplicate provider measurements solely by model or endpoint name. Do not average unrelated benchmark scales or declare a winner by source-count majority.
• Report confidence intervals and sample limits. Interval overlap alone is not a significance test; agreement across boards does not establish independent evidence.
• Explain disagreements using task relevance, configuration, methodology, and source quality. Preserve unresolved differences instead of forcing one ranking.

Knowledge base

• User-owned notebook sources and internal knowledge corpora: notebooklm.

Exclusion

• Do not substitute generic summarization for evidence retrieval.

Workflow

1. Classify the request: code/docs, web/live, sentiment, X/Twitter, or notebook KB.
2. Start with the active route for common operations.
3. Select a specialized skill when the route map gives it a better source or behavior match.
4. Execute the route and collect source evidence.
5. Add a different source class when confidence is low, sources conflict, or claim impact is high.
6. Return the answer with provenance, confidence, and explicit gaps.

Output contract

• Answer: concise result.
• Sources used: route and why selected.
• Confidence: high, medium, or low, with reason.
• Gaps: unresolved uncertainty or missing evidence.
• Next query: best follow-up when blocked.

Escalation rules

• Unclear library/API details: use context7; fall back to official docs.
• Unclear repository behavior: use repository inspection plus deepwiki when useful.
• Stale or shallow web evidence: try another search provider; use omp-search when its provider chain is available.
• Decision-driving sentiment claim: include reddit plus one non-Reddit source.