Protocol-driven agent loops for shipping real code with less babysitting.
Autoloop is a stateful orchestration runtime for repository work. It runs plan, implement, and test producer/verifier pairs with durable state, resumable runs, phase scoping, provider-neutral sessions, and git-backed checkpoints.
- Python 3.10+
- One supported provider CLI:
codexforprovider.name: codex(default)claudeforprovider.name: claude, with localclaude auth statusalready working
gitis optional when running with--no-git
Install from PyPI:
python3 -m pip install autoloop
autoloop --helpInstall from this repository:
python3 -m pip install -e .
autoloop --help
python3 -m autoloop --helpOr use the bundled installer:
./install_autoloop.shThe installer sets up Autoloop itself and the packaged skill, but it does not install provider CLIs for you. Install codex or claude separately based on the provider you plan to run.
Reruns are intentionally conservative: use ./install_autoloop.sh --dry-run to preview changes, --overwrite to replace an existing launcher or skill file, --recreate-venv to replace an existing virtualenv, and --skill-target both|codex|agents|none to control where the packaged skill is installed.
Put config in either:
~/.config/autoloop/autoloop.yaml(or$XDG_CONFIG_HOME/autoloop/autoloop.yaml) for global defaults/path/to/workspace/autoloop.yamlfor workspace-local overrides
Minimal Codex config:
provider:
name: codexMinimal Claude config:
provider:
name: claudeFirst run:
autoloop --workspace /path/to/workspace --intent "Describe the task you want shipped"Expected success artifacts under the workspace:
.autoloop/tasks/<task-id>/request.md.autoloop/tasks/<task-id>/runs/<run-id>/raw_phase_log.md.autoloop/tasks/<task-id>/runs/<run-id>/events.jsonl
Useful follow-up commands:
autoloop --workspace /path/to/workspace --list-tasks
autoloop --workspace /path/to/workspace --task-id <task-id> --resumeTroubleshooting:
| Symptom | Fix |
|---|---|
codex is missing for the default setup |
Install Codex with npm i -g @openai/codex, or switch to provider.name: claude and verify claude auth status. |
autoloop command is not found after running the installer |
Add the installer bin directory to PATH, or run python3 -m autoloop --help from the repo checkout. |
| Autoloop is not picking up your config | Put exactly one autoloop.yaml or autoloop.config file in the global config directory or workspace root. |
git is unavailable |
Install git, or run Autoloop with --no-git if you do not need checkpoints. |
Autoloop reads configuration from these directories, in this order:
$XDG_CONFIG_HOME/autoloop/or~/.config/autoloop/- the selected workspace root passed to
--workspace
Within each directory it looks for:
autoloop.yamlautoloop.config- legacy read-only fallback:
superloop.yaml - legacy read-only fallback:
superloop.config
If more than one config file exists in the same directory, Autoloop fails fast instead of guessing.
Use the nested provider shape for new configs:
provider:
name: codex
codex:
model: gpt-5.4
model_effort: null
claude:
model: null
effort: null
permission_strategy: inherit
runtime:
pairs: plan,implement,test
max_iterations: 15
phase_mode: single
intent_mode: preserve
full_auto_answers: false
no_git: falseNotes:
- Legacy flat Codex keys
provider.modelandprovider.model_effortare still read for backward compatibility. --modeland--model-effortare Codex-only CLI overrides.- Claude defaults intentionally inherit the operator's existing Claude Code environment. Default Claude runs do not add strict mode,
--bare, tool restriction flags, orstream-json. runtime.full_auto_answersnow uses the active provider instead of being Codex-only.
Fresh runs write state under .autoloop/.
For the first Autoloop release, resume and task listing can still read legacy .superloop/ workspaces when no .autoloop/ state exists for the requested task or run.
Session files now store provider-neutral provider and session_id fields. Legacy Codex session files that only contain thread_id remain load-compatible.
Maintainer live-validation steps for inherited-environment Claude repos and resume config-drift checks are recorded in docs/claude_rollout_checklist.md.
The packaged Autoloop skill lives at src/autoloop/skill/SKILL.md.