ClawLab

ClawLab is a persistent Auto Research Loop system built on top of a Claude Code CLI source snapshot. It is designed for real repo work: topic-driven research, existing-project improvement, SSH/GPU experiments, and now a practical rebuttal workflow.

What works now

These are the parts that are actually implemented and runnable in this repo today:

• new_project and existing_project_improvement research missions via /research start
• explicit state-machine-driven research loop under src/research/**
• local + SSH executor support for experiments
• model routing for auto, anthropic_oauth, anthropic_api_key, and openai_compatible
• native integration inspection/scaffolding for Codex, Claude Code, and OpenClaw
• a local rebuttal pipeline that reads paper/review files, scans repo evidence, applies venue policies, and drafts rebuttal artifacts
• a small executable local skill catalog that can be listed, shown, and run

Important reality check

These are deliberately not overstated:

• /research team ... is currently team scaffolding and role guidance, not a full embedded OMX runtime
• the rebuttal pipeline is artifact-first and locally runnable, but model-assisted drafting still depends on whatever model auth/provider is actually configured on your machine
• the repo-wide TypeScript baseline is still noisy because this source snapshot contains many unrelated upstream issues; the focused ClawLab tests are the reliable verification path right now

Install

bun install
bun src/entrypoints/cli.tsx

Or use the binary alias:

clawlab

If you want Anthropic OAuth-backed model access inside the CLI:

/login

Setup

Initialize the local scaffold once:

/research setup

This creates:

• .clawlab/tasks/
• .clawlab/docs/
• .clawlab/memory/
• .clawlab/team/
• .clawlab/skills/
• .clawlab/rebuttal/
• .clawlab/integrations/

Research workflows

1. Start from a topic

/research start --mode new "test-time adaptation for multimodal agents"

2. Improve an existing project

/research start \
  --mode improve \
  --repo /path/to/project \
  --problem "validation F1 is stuck around 0.72 after epoch 3" \
  --target-metric f1 \
  --current-metric f1=0.72 \
  --goal "push F1 beyond 0.76 without a large inference-cost regression"

3. Summarize only when you explicitly want it

/research summarize report
/research summarize summary
/research summarize paper

Native integrations

ClawLab now has a real integration layer for three external ecosystems:

• codex
• claude-code
• openclaw

Commands:

/research integration status
/research integration doctor
/research integration doctor codex
/research integration init codex
/research integration init claude-code
/research integration init openclaw

What it does:

• detects CLI availability on PATH
• checks user-level config locations
• checks whether project-local adapter files exist
• performs conservative auth detection where that is statically safe
• writes project-local adapter templates under .codex/, .claude/, and .openclaw/

Current auth detection policy is intentionally conservative:

• Codex: detects auth.json or OPENAI_API_KEY
• Claude Code: can confirm env-backed auth, but does not claim an interactive Claude login is valid from static files alone
• OpenClaw: checks config/profile signals, not live gateway liveness

Rebuttal workflow

ClawLab now includes a runnable rebuttal path.

Initialize rebuttal workspace

/research rebuttal init

Build a rebuttal plan

/research rebuttal plan \
  --paper /path/to/paper.pdf \
  --review /path/to/review1.pdf \
  --review /path/to/review2.txt \
  --repo /path/to/repo \
  --venue neurips

Draft from an existing rebuttal run

/research rebuttal draft --run-dir /path/to/.clawlab/rebuttal/runs/run_...

Validate a draft against venue rules

/research rebuttal validate --draft /path/to/rebuttal_draft.md --venue neurips

Current built-in venue presets:

• cvpr
• neurips
• iclr
• acl_arr
• generic

Artifacts written per rebuttal run:

• inputs.json
• paper.txt
• reviews.txt
• venue_policy.json
• concerns.json
• repo_evidence.json
• rebuttal_plan.json
• rebuttal_plan.md
• rebuttal_draft.md
• rebuttal_validation.json

Executable local skills

ClawLab now exposes a small executable skill catalog instead of a giant fake list.

Commands:

/research skills list
/research skills show integration-doctor
/research skills run integration-doctor
/research skills run review-concern-extract --review /path/to/review.pdf
/research skills run venue-policy-check --draft /path/to/draft.md --venue neurips

Current executable built-ins:

• integration-doctor
• review-concern-extract
• venue-policy-check
• repo-evidence-scan
• rebuttal-plan

Curated external references are also listed, but they are clearly marked as references rather than pretending to be built-in local skills.

Team scaffolding

The current /research team ... surface is still useful, but be clear about what it is:

• role scaffolding
• team memory templates
• role switch/status commands
• role-oriented playbook recommendations

It is not the same thing as a fully embedded OMX $team runtime.

Available commands:

/research team init
/research team status
/research team roles
/research team switch reviewer
/research team skills --stage experiment

Validation

These are the verification commands that currently give reliable signal for ClawLab work in this repo:

bun run lint:clawlab
bun run test:clawlab
bun run check:clawlab

What I have actually verified in this environment:

• bun run test:clawlab passes
• bun run lint:clawlab passes with complexity warnings only

The focused typecheck:clawlab script is still limited by inherited upstream TypeScript graph issues from this repository snapshot, so I do not treat it as the main pass/fail gate yet.

External references used for this direction

• OpenAI Codex docs
• OpenAI Codex GitHub repo
• Claude Code docs
• Anthropic Claude Code docs
• OpenClaw docs
• OpenClaw GitHub repo
• oh-my-codex
• oh-my-claudecode
• Paper2Rebuttal

More documentation

• English guide: docs/auto-research-loop/README.en.md
• 中文指南: docs/auto-research-loop/README.zh-CN.md
• Docs landing page: docs/auto-research-loop/README.md