Digest auto-detects research-style vs project-context (no new flag)

[ranjiao]

Summary

User feedback: '我时不时需要参考学术论文 / 行业技术报告。现在的 digest 适合处理这类信息吗？'

Analysis: current digest schema (TL;DR + Key facts + Open questions + What PMO must remember + Section map) was tuned for project-context docs — term sheets, constraints, screenshots, internal notes. It underserves academic / industry research content where the reuse value lives in Method / Limitations / Reproducibility / Applicability, not in 'key facts.'

User direction: NO new flag. Have the agent auto-detect the document type from signals in the source itself, fall back to one AskUserQuestion when ambiguous.

Auto-detection logic (new § in digests.md)

Two classes:

• project-context — existing default; term sheets, internal docs, constraints, screenshots
• research-style — papers, industry tech reports, white papers, formal technical notes

Six weighted signals computed at digest time:

Signal	Counts toward research-style if …
Length	≥ 5000 words / ≥ 8 pages
Structural markers	Abstract / References / numbered figure captions / equations / DOI / arXiv ID
Citations density	≥ 10 distinct external citations
File source / URL	arxiv / acm / ieee / nature / springer / ssrn etc.; or <author>-<year>-<topic>.pdf pattern
Writing style	First-person plural method narration, hypothesis testing
Section structure	≥ 3 of: Introduction / Background / Method / Experiments / Results / Discussion / Conclusion / Limitations

Threshold: ≥ 3 signals = high-confidence research-style (proceed silently). ≤ 1 = high-confidence project-context (proceed silently). = 2 = ambiguous → one AskUserQuestion with the user picking once.

Classification logged in digest's Doc type: front-matter for audit + manual override via --refresh.

Two schemas

Both share: front-matter convention, placement under knowledge/<topic>/, Status: active | archived | eternal | superseded lifecycle, archive-candidate detection, INDEX registration.

They differ in body sections only.

Project-context (unchanged)

state/digest_TEMPLATE.md — TL;DR + Key facts + Open questions + What PMO must remember + Section map.

Research-style (NEW)

state/digest_paper_TEMPLATE.md:

• TL;DR — plain user-language, not abstract verbatim
• Method — sample / setup / baseline / eval (judgeable + transferable)
• Key results — number + context (not raw claim)
• Limitations — author-acknowledged + reader-observed (critical reading)
• Reproducibility — data / code / clarity / barriers
• Applicable to this project? — verdict (use / partially / don't) + reason + landing point
• Followup citations — 2-3 max with one-line reason each
• Open questions + What PMO must remember (carried over)
• Extra front-matter: Authors, Published, Source venue

Each section has a stated rationale in a ### Why a different schema for research paragraph — so future edits don't drift the boundary in either direction.

Why this is better than --paper flag

• One mental model: drop file in inputs/, run /pmo digest <path>, agent does the right thing
• Document-type evidence lives IN the document, not in the user's memory of which flag to pass
• Borderline cases (long-form blog post that reads academic, or a short methodology memo) get explicit user confirmation rather than a wrong silent default

Files changed

• pmo/reference/digests.md: +144 lines (auto-detection signals, two-schema description, rationale)
• pmo/state/digest_paper_TEMPLATE.md: new file, 76 lines
• pmo/state/digest_TEMPLATE.md: unchanged (still the project-context template)
• No SKILL.md change (digest reference is loaded only on /pmo digest; type-detection is a reference-internal concern)
• No new flag, no new subcommand syntax

Test plan

• Drop an arXiv PDF (e.g., a paper with arxiv in URL, 12 pages, abstract+references) → confirm agent silently picks research-style and writes the 6-section schema.
• Drop a 2-page screenshot or term sheet → confirm agent silently picks project-context.
• Drop a borderline doc (e.g., a 6-page technical memo from a company blog) → confirm one AskUserQuestion fires and the user's pick is honored.
• Check the resulting digest's Doc type: front-matter field is set correctly.

🤖 Generated with Claude Code