paper-review

name	paper-review
description	Review an academic paper PDF — ingest, summarise, fan out multi-angle 锐评, polish user draft into publishable English reviewer comments, archive.
argument-hint	<path-to.pdf-or-slug>
allowed-tools	Read,Write,Bash,Glob,Grep,Agent,Skill,WebFetch,WebSearch,Workflow

/paper-review:new — Review a paper

You are the orchestrator. Given a PDF path (or a slug already in ongoing/), walk the pipeline below in order.

Pipeline

Step	File	What happens
0	00-language.md	⭐ Pick intermediate-file language (default English, switchable to Chinese) → review-config.md
1	01-ingest.md	marker-pdf → 1-paper-text/paper.md + md/ + img/sec* + INDEX.md
2	02-literature.md	Invoke workflow paper-review-literature — parallel extraction + IEEE search + author profiling
2b	02b-consensus.md	⭐ Write 2-review/summary.md — shared truth (folds in literature, + venue type + Obvious gaps), user confirms
3	03-angle-gate.md	⭐ Propose angles, user confirms
4	04-fanout.md	Invoke workflow paper-review-fanout — parallel reviewers (Sonnet agents + MCP-takeover for Codex/DeepSeek)
5	05-aggregate.md	Merge into 2-review/critiques.md
6	06-user-draft.md	⭐ Orchestrator generates complete draft, user edits (hard gate, session boundary)
7	07-polish.md	polisher-english → 3-response/final.md (plain text)
8	08-archive.md	Move to archived/YYMMDD/<slug>/, update global style + angles, optional postmortem

How to execute

At the start of each step, Read the corresponding sub-file and follow it. This file is the map — sub-files are the playbook.

Resume table (which step to enter on re-invocation)

Re-invoking /paper-review:new <slug> resumes from the latest non-empty artifact:

Existing artifact	Enter step
archived/*/<slug>/3-response/final.md (glob)	08 already done — confirm with user, use /paper-review:rerun instead
ongoing/<slug>/3-response/final.md	08
ongoing/<slug>/3-response/draft.md (non-empty)	07
ongoing/<slug>/2-review/critiques.md	06
ongoing/<slug>/2-review/critiques/*.md (all angles present)	05
ongoing/<slug>/2-review/critiques/*.md (partial — fewer files than angles in angles.md)	04 (incomplete fanout — offer retry/skip/continue)
ongoing/<slug>/2-review/angles.md	04
ongoing/<slug>/2-review/summary.md (non-empty)	03
ongoing/<slug>/2-review/literature.md (non-empty)	02b
ongoing/<slug>/1-paper-text/paper.md	02
ongoing/<slug>/review-config.md (no paper.md yet)	01
nothing	00

--restart-from <N> jumps to any step. --reingest forces step 01 re-run. Partial fanout (fewer critiques than angles) → offer retry/skip/continue; never silently enter step 05 incomplete.

Hard rules

• Core (see AGENT.md): no opinion without reading summary.md; polisher never invents.
• Literature before consensus: step 02 → 02b; summary folds in landscape + venue type.
• Venue type gates Obvious gaps: EE conference → no hardware/code/data gaps; journal → fair game.
• Intermediate-file language: steps writing prose read lang: from review-config.md (default en). final.md is always plain-text English.
• User gates: steps 03 (angles) and 06 (draft) — iterate until approved; never skip.
• Session boundary at 06: pipeline yields control; resume by re-invoking with the slug.
• Fanout is parallel via workflow: paper-review-fanout handles parallel() + schema validation; routing: angles.md override > agent frontmatter default. Sonnet reviewers run as direct agents, Codex/DeepSeek via MCP-takeover relay.
• Archive always updates style/profile.md and critiques-library/angles.md, with dedup + rolling caps.

Slug

Derive <slug> from the PDF filename (lowercase, non-alphanum → -, strip leading/trailing -). If the user passes an existing slug, use the resume table above.