Source Research

Source Research

Use this as the default source-gathering research router.

Supporting files

Open these only when relevant:

• references/specialist-subagents.md for focused single-specialist delegation briefs imported from the local Claude setup

Legacy tool mapping

• legacy web_search -> Codex web.search_query
• legacy web_fetch -> Codex web.open
• legacy task/session spawning -> Codex spawn_agent
• legacy exec -> Codex functions.exec_command

Live workspace routing

• Generic paper, article, preprint, DOI, ISBN, and book retrieval or sharing requests route to zotero first.
• Explicit Calibre-library ebook operations route to calibre.
• External paper or book retrieval routes to getscipapers-requester only after the relevant local-library workflow does not satisfy the request, or after the user explicitly says not to check/use the library. The word "download" alone is not an outside-library opt-out.
• Multi-agent discussion/review/research routes to agent-group-discuss by default and prose for more structured workflows.
• Annotated paper review routes to annotated-review only when the user explicitly asks for both annotation and review.
• Review-only paper tasks route to paper-review.
• Topic/news digests route to research-digest-wrapper or rss-news-digest.
• Digest-to-paper extraction routes to digest-bridge.
• Explicit phased deep research with structured source handoff routes to deep-research-workflow.
• Load-bearing lemmas that may need formalization: mention/tag as formal candidates and hand off to lean-formalization-intake / informal-to-lean runbook; do not run OpenGauss from discovery. Use opengauss only after intake proceed and (for live work) a manual Gauss install.
• Writing, report, review, digest, or final-answer workflows must load writing-style-settings.md; math or LaTeX writing must also load math-manuscript-style.md.
• Explicit TikZ drawing, refactoring, extraction, compile, or diagram-review requests route to tikz-draw.
• Small graph-theoretic verification routes to graph-verifier.
• Mathematical research tasks that need heavy graph-theoretic, combinatorial, algebraic, or spectral computation route to sagemath. Formal candidates that pass intake also run lean-research-library search before any new formalization is attempted.

Default workflow

1. If the request is current, says "latest", "today", "verify", or otherwise depends on recent facts, browse first.
2. Start with focused search queries, not long prompts.
3. Open the best primary sources, not just search snippets.
4. Extract only the parts needed for the answer.
5. Synthesize after evidence collection, not before.
6. Cite concrete sources and dates whenever the answer depends on live or specific facts.
7. For writing-producing outputs, record or surface style_profile_ref, policy_hash, active_overlays, active_requirement_ids, and style_applied; do not treat a bare style_applied: true assertion as sufficient evidence.
8. For report-only research requests, stop after the requested answer or findings. Do not create durable artifacts, patch reports, retrieve nonessential follow-up files, or begin remediation unless the user explicitly asks for those actions.

Research heuristics

• Prefer primary sources: official docs, vendor pages, papers, standards, government sites.
• Use domain filters when the source class is obvious.
• For PDFs, use web.open and web.screenshot when needed.
• For local files, use functions.exec_command instead of browsing.
• Separate sourced facts from your own inference.
• Prefer deep-research-workflow over this skill when the user explicitly wants phased search -> analysis -> writing with source preservation across stages.
• Prefer tikz-draw for explicit TikZ or structural-diagram requests, including research-derived figures that already have a narrowed brief.
• If a research task needs a diagram, keep the research in deep-research-workflow first and hand off to tikz-draw only after analysis identifies a concrete figure worth generating.
• For literature retrieval requests, check the relevant local library workflow before browsing or external download.
• For review tasks that need a paper/book, use the lookup order zotero -> calibre -> online fallback.
• For digest follow-up, use digest-bridge to extract identifiers before bulk paper retrieval.
• For explicit Calibre ebook-library workflows, prefer calibre over improvised filesystem search.
• For mathematical verification, prefer local Python for small checks and route to sagemath when the computation needs SageMath-native capabilities.
• Never use curl, wget, ad hoc browser fetches, or direct publisher-site HTTP requests as a substitute for the library tools on paper/book retrieval tasks.
• When a lookup returns multiple plausible paper/book matches, show a numbered list with title, authors, and year when available, then wait for the user's chosen index before --best, --index, add, attach, send, review, or external retrieval.

Parallelism

If the task has distinct independent threads, spawn sub-agents in parallel.

Good candidates:

• source discovery vs synthesis plan
• product A vs product B
• multiple independent subtopics

Do not spawn agents for the immediate blocking step if local work is faster. Ask for explicit confirmation before spawning subagents unless the user already requested multi-agent work, approved a workflow that includes subagents, or gave an explicit budget/depth instruction that makes parallel delegation expected.

Focused specialist delegation

When the task needs one narrow specialist rather than a full panel, reuse the briefs in references/specialist-subagents.md.

Good matches:

• literature survey or citation hunting -> literature-scout
• small-case exploration or counterexample search -> math-explorer
• single-reviewer manuscript critique -> paper-reviewer
• adversarial checking of one proof/lemma -> proof-checker

Route to narrower skills

• Use zotero for generic paper, DOI, ISBN, book, and paper-library retrieval or sharing requests, including "find/get/send/share this paper or book".
• Use docling for local parsing, conversion, chunking, and structure-aware analysis of PDFs, office documents, HTML, and image-backed documents.
• Use get-available-resources before heavy local parsing or compute when resource constraints may materially affect the approach.
• Use database-lookup for structured public scientific, regulatory, and economic database queries.
• Use paper-lookup for external literature metadata/discovery after the library-first workflow and before external retrieval.
• Use deep-research-workflow for single-agent phased deep research that needs explicit source handoff across search, analysis, and final writing.
• Use tikz-draw for explicit TikZ figure generation, refactoring, extraction, compile, or review work, especially when the output should follow a structure-first brief -> spec -> render flow.
• Use calibre for explicit Calibre library operations: ebook search/get by ID, sending Calibre-managed books, add/update, tags, shelves, sync, remove, convert, export, doctor, and clean.
• Use getscipapers-requester only for external retrieval or DOI/ISBN resolution after Zotero does not satisfy the request, or when the user explicitly says not to check/use the library and wants outside retrieval. For review tasks that need a document, check calibre after zotero and before this fallback.
• If an externally retrieved book file should be added to the ebook library, hand off to calibre for the final add/update step.
• Use annotated-review only for annotate+review paper flows.
• Use paper-review for single-agent review-only paper flows such as review, critique, hard review, or issue-finding requests.
• Use agent-group-discuss for conversational multi-agent discussion, review, or research panels.
• Use prose when the user wants explicit multi-agent research-and-synthesis orchestration.
• Use research-digest-wrapper and rss-news-digest for tracked-topic and RSS-based research updates.
• Use graph-verifier for lightweight graph sanity checks before escalating to SageMath.
• Use digest-bridge to turn digest results into identifiers and manifests for retrieval.
• Use sagemath for graph invariants, algebraic checks, spectral computations, exhaustive counterexample search, and other math tasks beyond lightweight local tooling.