Skip to content

program.md

Latest commit

 

History

History
95 lines (70 loc) · 3.59 KB

program.md

File metadata and controls

95 lines (70 loc) · 3.59 KB
Raw

labrat program

This is the repo-level entrypoint for Codex or Claude Code.

The repository root is not a runnable lab. Use the root AGENTS.md or CLAUDE.md for repo maintenance, and use a lab directory when you want to operate the runtime itself.

Path 1: Evaluate the repo quickly

Use the flagship example.

python -m venv .venv
source .venv/bin/activate
pip install -e '.[nlp-sentiment]'
labrat doctor --lab-dir examples/nlp-sentiment/research_lab
labrat bootstrap --lab-dir examples/nlp-sentiment/research_lab
python -m http.server 8787 --directory examples/nlp-sentiment/research_lab
labrat status --lab-dir examples/nlp-sentiment/research_lab
labrat next-prompt --lab-dir examples/nlp-sentiment/research_lab --runner claude --phase auto

Codex uses the same flow with --runner codex.

Path 2: Start from a profile

Use this when the user arrives with a concrete research workload that matches an existing profile.

labrat new my_search --profile=transformer-arch
cd my_search
python -m pip install -r requirements.txt
python scripts/operator_helper.py doctor
python scripts/operator_helper.py check-readiness
python scripts/bootstrap.py
python scripts/operator_helper.py next-prompt --runner claude --phase auto

Profiles ship with filled Phase 0 files, a working run_experiment.py, AGENTS.md, CLAUDE.md, .claude/commands/, and agent_prompts/. transformer-arch is the first profile. See docs/PROFILES.md for the full list and contract.

Path 3: Create a real lab from scratch

Use this when no profile fits and the user wants a new runtime-backed research program.

  1. scaffold a new lab with labrat new
  2. finish Phase 0:
    • branches.yaml
    • dead_ends.md
    • research_brief.md
    • research_sources.md
    • evaluation.yaml
    • runtime.yaml
  3. confirm readiness
  4. bootstrap the runtime
  5. supervise the worker pool through the helper prompts
labrat new my_lab
cd my_lab
python scripts/operator_helper.py next-prompt --runner claude --phase design
python scripts/operator_helper.py check-readiness
python scripts/bootstrap.py
python scripts/operator_helper.py next-prompt --runner claude --phase auto

Use --runner codex if you are operating from Codex instead of Claude Code.

Interface contract

Every generated lab carries the operator instructions in version control:

  • AGENTS.md for Codex
  • .agents/skills/labrat-operator/SKILL.md for Codex's optional repeatable workflow
  • CLAUDE.md and .claude/commands/ for Claude Code
  • agent_prompts/ for the shared phase prompts

There is no hidden skill file. The goal is that a new user can open the lab in either interface and start from the files already present in the lab root.

For current frontier-model operating guidance, see docs/MODEL_GUIDANCE.md.

Defaults

  • The runtime is async and steady-state.
  • evaluator.py is authoritative for scores.
  • search_eval and selection_eval are separate by default.
  • prediction_tests define decisive held-out challenges that are not the same as the local hill-climb metric.
  • Cheap probes and audits come before frame break.
  • The static dashboard is the canonical UI.

Lineage

If you want the shortest explanation of where this repo is coming from: