sn-deep-research

What it does

The skill acts as a central controller for deep, multi-agent research tasks. It does not perform the research itself but dispatches specialized roles to carry out evidence collection, formatting, and reporting. It handles mode selection (quick, normal, heavy), coordinates between scouts, planners, researchers, and writers, and ensures language, file paths, and structured artifacts are passed through using absolute paths. It creates a centralized report directory structure and starts progress UI wired to a research workbench. It enforces a strict payload contract: each role must read its own role.md, receive the user query, language tag, and report directory, and produce or validate specific artifacts like briefing.json, format_proposal.json, plan.json, evidence.json, and final reports, with carefully defined failure/retry routes per stage.

How it works

• It maintains a hierarchical stage pipeline using §4 (档位选择器) and §5 (阶段库). It first selects a档位 (quick, normal, heavy) based on the user query and required validation, then dispatches roles accordingly.
• For quick mode, it creates a single-dimension research unit and follows a lean pipeline: evidence.json, quick_synthesis by report-writer, and render to sections/s_full.md without outline generation.
• For normal mode, it involves scout to produce a briefing and format_proposal, then plan.json, multi-dimension research, plan validation, outline planning, and parallel per-dimension work with potential supplements, followed by stitching and final render.
• For heavy mode, it extends normal with parallelization across more dimensions, including perspective and supplements, and enforces dependency-driven wave execution.
• It uses absolute file paths, communicates via structured JSON artifacts, and relies on a validator to enforce schemas. Language is fixed per request language and passed to all agents.
• It creates and updates a progress WebUI with a request.md and progress events, and maintains a single report_dir with a fixed, date-stamped structure for all outputs.

When to use it

• Triggered when the user request contains deep/systematic research keywords (e.g., “深度研究”, “竞品分析”, “趋势分析”“事实核查” or “research”, “deep research”).
• When the user asks for cross-source corroboration, multi-dimensional comparisons, or formal reports/whitepapers/due-diligence docs.
• When topics require latest policy/market/product/regulation vetting or source verification; quick single-point facts aren’t used.

What it can touch

• Externally, it uses the following tools indirectly through agents: Claude Code via the skill’s agent calls; all file paths are absolute; it uses a progress UI script and a workbench for web UI interactions.
• It expects role-specific artifacts under the {report_dir} such as {report_dir}/request.md, {report_dir}/briefing.json, {report_dir}/format_proposal.json, {report_dir}/plan.json, and sub_reports/{dimension}.evidence.json, among others, produced by the agents defined in §5.

Caveats

• It strictly enforces that only small fields are read by the controller; large documents are passed to roles via absolute file paths. The controller does not read big files directly.
• All language handling is centralized at the request level; subsequent messages to roles must carry language:{language}.
• If any required artifact is missing or invalid (e.g., format_proposal.json or plan.json), the workflow halts with a defined retry path.
• It uses a single report_dir per task; multiple runs create separate dated subdirectories to avoid overwriting.