AutoProjectClaw

A 6-stage pipeline for grant proposal topic selection, from vague research direction to evaluable project framework with literature-backed evidence.

How It Works

AutoProjectClaw guides you through 6 stages, each producing artifacts consumed by the next:

Stage	Name	Purpose
0	Context Pack	Collect grant context, investigator profile, data inventory, and source materials
1	Project Init	Build a project framework with candidate directions and assumptions
2	Aim Decompose	Decompose the project into 3-4 coherent research aims
3	Multi-Dim Search	Create grant-oriented literature and evidence search strategies
4	Evidence Collect	Collect evidence cards tagged by aim, data source, method, and novelty
5	Candidate Evaluate	Evaluate candidate titles with FINER scoring, avoidance check, and HITL gate

Stage 5 blocks for human approval (approve / reject / resume). Rejection rolls back to Stage 3 by default.

Installation

pip install -e .

Requires Python >=3.11. Only runtime dependency: pyyaml.

Quick Start

Generate a config template:

autoprojectclaw setup --output my_config.yaml

Edit my_config.yaml with your project details, then run:

# Dry run (no LLM calls, deterministic output)
autoprojectclaw run --config my_config.yaml --to-stage 5 --llm-provider dry_run --run-id my-first-run

# Real LLM via local Claude CLI
autoprojectclaw run --config my_config.yaml --to-stage 5 --llm-provider claude_cli --run-id my-real-run

Review the recommendation and approve or reject:

autoprojectclaw status --config my_config.yaml --run-id my-real-run
autoprojectclaw approve --config my_config.yaml --run-id my-real-run --note "方向可继续"

CLI Commands

Command	Description
run	Execute pipeline stages
approve	Approve a blocked HITL gate
reject	Reject and optionally rerun from a rollback stage
resume	Resume from checkpoint
status	Show current run status
history	List all historical runs
report	Generate weekly summary report
setup	Generate a config template
llm	Test the LLM provider
cache clear	Clear literature search cache

All commands accept --verbose and --quiet flags.

Key Output Files

Path	Description
runs/<run_id>/pipeline_summary.json	Overall run status and stage counts
runs/<run_id>/checkpoint.json	Resume point (next/rollback stage)
runs/<run_id>/stage-01-project-init/project_framework.md	Candidate directions and framework
runs/<run_id>/stage-02-aim-decompose/research_aims.md	Research aim decomposition
runs/<run_id>/stage-05-candidate-evaluate/final_recommendation.md	Final recommended topic
runs/<run_id>/stage-05-candidate-evaluate/scoring_matrix.json	FINER + avoidance + evidence scores
runs/reports/weekly_<date>.md	Weekly summary report

LLM Providers

Provider	How it works	API key needed?
dry_run	Deterministic, no network	No
claude_cli	Shells out to local claude CLI	No (uses local auth)
codex_cli	Shells out to local codex exec	No (uses local auth)
openai	HTTP POST to OpenAI Responses API	Yes (OPENAI_API_KEY)
anthropic	HTTP POST to Anthropic Messages API	Yes (ANTHROPIC_API_KEY)

Literature Search

Disabled by default. Enable in config:

literature_search:
  enabled: true
  core_sources: [pubmed, openalex, semantic_scholar]
  preprint_sources: [arxiv, biorxiv, medrxiv]
  max_results_per_query: 5
  year_min: 2020

Six academic APIs supported: PubMed, OpenAlex, Semantic Scholar, arXiv, bioRxiv, medRxiv.

Testing

python3 -m unittest discover -s tests -v
# or
uv run pytest tests/ -q --tb=short

152 tests, 14 subtests. No network calls in tests (dry-run provider).

Documentation

• Architecture — module relationships and data flow
• Configuration Reference — complete YAML config options
• Contributing — development setup and conventions

Acknowledgements

Adapted from AutoResearchClaw. Original Stage 1-5 reference files preserved under vendor/autoresearchclaw_stage1_5/.

License

MIT