deep-research

Deep Research

You are the researcher, not a wrapper around a search box. Scout the question first, let the report's structure follow from what retrieval actually returned, keep every source in a ledger, and write a report where each claim points back to something you retrieved.

Language Routing

• Use SKILL.zh.md when the user mainly writes in Chinese or explicitly requests Chinese output.
• Use this SKILL.md for all other requests.
• Keep code, commands, ledger field names, and API names in English in either workflow; only the prose switches language.

Deep Research produces two artifacts (v6 §1.2/§7.6)

Deep Research is not just an end feature (a report for humans). One DR run produces two artifacts off the same flow — a cited report, and an Evidence Ledger: the structured JSON the engine itself used to gate every claim (sources, claims, figures, datums, probes, outline, spend). The ledger is self-describing and reusable; what any downstream system does with it is that system's business, not the skill's. The skill does not know or name a consumer:

Deep Research Engine (scripts/evidence.py + scripts/aminer_open.py)
   ↓ one run
   ├── Evidence Ledger      (scripts/evidence.py state JSON — self-describing, reusable)
   └── Report + appendices  (evidence.py render --material → --renumber → --appendix)

self-check, no external consumer: evidence.py check · evaluation/evaluate.py

The ledger is the engine's own state — it is not hand-written, it falls out of the research loop (§7.6). It is a self-describing, versioned JSON: anyone downstream (a context store, a RAG index, a review pipeline, or nothing at all) may read it as-is. The skill produces it and stops — it does not export, convert, or adapt it to any other system's schema; that specialization is the external system's job, not the skill's.

Submodules (v6 §7.13)

• scripts/evidence.py — engine + evidence-ledger + report-renderer: the research loop, the machine-consumable ledger state ({version, topic, probes, outline, sources, claims, figures, datums, spend}), analyze() self-check, and render.
• scripts/aminer_open.py — AMiner Open Platform retrieval (stdlib urllib, 28 endpoints, price catalog, cost document). DR's own spend-tracking path (§7.14).
• scripts/chartrender.py — renders one registered figure to a PNG. Host-called (a sibling to aminer_open.py, never spawned by evidence.py, which stays a pure offline ledger): deterministic matplotlib templates (bar / hbar / line / pie / heatmap) or a host-written B script run in a best-effort sandbox (no network, locked cwd, 30 s timeout, forbidden-token scan, data on stdin); a B failure falls back to the matching template. The figure's numbers still come from the ledger, so check's data↔source gate holds either way.
• evaluation/evaluate.py — quality report from analyze() (§7.13 5th submodule, §7.15 internal validation).
• samples/patchtst_v3_ledger.json — v3-schema sample ledger (PatchTST).
• references/research-loop.md — the actual procedure (read it at task start).