Academic repo analyzer

Academic Repo Analyzer

Produce a 仓库快速理解文档 for downstream figure planning.

Keywords: → keywords.md
Missing info: → ../docs/missing-info-policy.md

Input Contract

• Prefer: repo path, README, deps, entry scripts, model files, configs
• Minimum: any one of README / entry script / model file
• Missing: partial analysis with 推断 / 待确认

Output Contract — Quick Understanding Doc

• overview (name, task, framework, architecture one-liner)
• completeness block
• stack details
• model / algorithm notes
• train / inference flow (or “evidence insufficient”)
• figure suggestions for paper-analyzer

Steps

Step 1: Scan structure

Locate README, dependency files, entry scripts (train|main|eval|inference), configs/, models|networks|src/, data loaders.

Done when: tree of key paths exists and each must-read class is read or marked missing.

Step 2: Task + stack

Use keywords.md. Classify task type and framework from imports, deps, and paths.

Done when: task type + primary framework are stated with file evidence.

Step 3: Architecture + algorithms

From model files: backbone family, key modules, losses, training tricks. Prefer evidence over naming guesses.

Done when: architecture summary cites concrete classes/files, or is marked 推断.

Step 4: Emit quick-understanding doc

# 仓库快速理解文档
## 仓库概览
| 项目 | 内容 |
| 仓库名称 / 任务类型 / 核心框架 / 主要架构 / 一句话描述 | ... |
## 信息完整度说明
## 技术栈详情
## 模型架构分析
## 工作流程
## 配图建议（→ paper-analyzer）

Done when: Output Contract fields are filled; figure suggestions list concrete types (framework / arch / module / …).

Sparse-input cases

gap	action
no README	infer from code; label as structure-inferred
no entry scripts	module-level understanding only
no model files	stack/task only; soft architecture language
huge repo	sample top-level + 3–5 core files; mark limited sample
almost nothing	pre-analysis + minimum materials list (README → deps → entry → model → config)

Tooling cues

Prefer the environment’s file/search tools. Typical digs: dependency files, class.*Model|Network|Transformer, loss|criterion, model package entrypoints.

Stop

Stop when the quick-understanding doc is delivered. Suggest paper-analyzer only if the user wants figure planning next.