OpenClaw Repo Autopilot

Deploy and run a production-style autopilot for any GitHub repository.

Use bundled scripts and templates instead of rebuilding orchestrators from scratch.

Core Workflow

1. Deploy templates.
2. Diagnose environment and CLI availability.
3. Patch config for the actual host.
4. Run one round (--once) and inspect logs.
5. Run unattended mode.

Deploy

Use this when the user asks to initialize a new workspace.

Fastest path (wrapper):

python3 scripts/setup_autopilot.py \
  --output-dir /path/to/workdir \
  --repo-url https://github.com/<owner>/<repo>.git \
  --config-profile production \
  --token "<github_token>" \
  --run-doctor \
  --run-once \
  --force

python3 scripts/deploy_autopilot.py \
  --output-dir /path/to/workdir \
  --config-profile production \
  --repo-url https://github.com/<owner>/<repo>.git \
  --project-name "<project-name>" \
  --caller-name "你的自动化系统名" \
  --task-requirement "按你的目标执行迭代需求（可以是性能、功能、重构或修复）" \
  --git-user-name ai \
  --git-user-email ai@local \
  --require-code-changes \
  --branch dev \
  --init-env \
  --auto-disable-missing-clis \
  --doctor-after-deploy \
  --print-diagnose-cmd

Notes:

• --config-profile production uses production baseline (includes require_code_changes=true) and then applies CLI args overrides.
• production profile also enables stronger substance gates by default: require_non_doc_code_changes=true, minimum_non_doc_files_changed=2, minimum_non_doc_lines_changed=30.
• --auto-disable-missing-clis prevents immediate failures on hosts without claude, open-code, or gemini.
• --init-env creates .env from .env.example when missing.
• --task-requirement is injected into runtime prompt and can be any concrete user requirement.
• --caller-name controls how the orchestrator is referenced inside CLI prompts; avoid hardcoding caller identity.
• --repo-url should always be set to a real repo URL. Placeholder values will fail doctor/runtime checks.
• --doctor-after-deploy runs doctor immediately and returns non-zero on health failure.
• --cli-order lets you customize priority order.
• --only-cli restricts execution to specific CLI(s).
• --interactive-cli enters selected CLI interactive session mode for one task.
• --interactive-max-turns controls max redo turns in interactive session mode.
• Codex interactive mode auto-loads the round prompt as initial context and uses --no-alt-screen for better terminal compatibility.
• --require-code-changes enables a hard gate: no_changes is treated as audit failure and will trigger redo/switch logic.
• Init phase is enabled by default: each CLI runs a one-time repository initialization before real task execution.
• Init behavior now explicitly asks CLI to run /init first (or perform equivalent initialization if slash command is unsupported).
• Codex defaults to init_required_paths=[".codex"]; if .codex is not created, init is treated as failed.
• Runtime keeps .codex/ (and other configured init directories) during git clean, and excludes them from commits.
• --disable-init-phase turns off init phase; --force-reinit forces init to run again.

Diagnose Before First Run

Use this before --once and after any host/toolchain change.

python3 scripts/doctor_autopilot.py --config openclaw_config.json

For token validation:

python3 scripts/doctor_autopilot.py --config openclaw_config.json --check-github-token

Doctor checks:

• required config keys
• CLI binary existence and version probing
• Codex/Gemini login status hints
• optional GitHub token validity

Run Modes

One-round validation:

source .env
python3 openclaw_autopilot.py --config openclaw_config.json --once --token-env GITHUB_TOKEN

Unattended loop:

./start_openclaw.sh

Runtime Contract

Keep these behaviors unless the user explicitly asks to change policy:

• Default CLI priority: Codex CLI -> Gemini CLI -> Open Code CLI -> Claude Code CLI.
• Before every CLI attempt, runtime fetches and resets to remote latest origin/dev to avoid drift with external commits.
• Timeout policy:
  • 30s no output: terminate and switch.
  • 15min runtime: send progress probe.
  • 5min no useful probe response: terminate and switch.
  • 30min hard cap: terminate and switch.
• Loop anomaly: repeated same operation output >3 times -> terminate.
• Audit policy: report-only (run_status, test_pass_rate, summary fields).
• Approval gate: run_status successful and test_pass_rate >= 90.
• Optional gate: when require_code_changes=true, rounds with no staged code changes are treated as failed and retried/switched.
• Optional gate: when require_non_doc_code_changes=true, rounds with docs-only changes are treated as failed and retried/switched.
• Substantive gate: when minimum_non_doc_files_changed / minimum_non_doc_lines_changed are set, rounds below threshold are treated as failed and retried/switched.
• Init gate: when enabled, CLI must pass one-time init phase (with timeout control) before entering task execution.
• Init artifact gate: required init directories (such as .codex) must exist after init or the tool is switched.
• Git policy: commit/push only to dev.
• Git identity default: ai <ai@local>.
• Commit message includes concrete change summary (A/M/D + file paths).
• On paused state, writes logs/PAUSED_REASON.txt for external monitors.

Runtime override examples:

python3 openclaw_autopilot.py --config openclaw_config.json --cli-order codex,gemini,open-code,claude --once
python3 openclaw_autopilot.py --config openclaw_config.json --only-cli gemini --once
python3 openclaw_autopilot.py --config openclaw_config.json --interactive-cli codex --interactive-max-turns 5

Fallback Report Policy

If CLI exits without writing report JSON, autopilot can generate fallback report.

Default fallback behavior:

• run validation command candidates (pytest first)
• infer pass rate from output
• write standard report JSON
• continue audit pipeline using fallback report

Config keys live in fallback_report inside openclaw_config.json.

Minimal-Change Tuning Order

When users ask to improve behavior, patch in this order:

1. openclaw_config.json thresholds and CLI commands.
2. fallback test command candidates.
3. CLI ordering/enabled flags.
4. only then edit orchestrator logic.

Bundled Files

• scripts/deploy_autopilot.py: deploy templates + apply practical overrides.
• scripts/doctor_autopilot.py: host and config diagnostics.
• scripts/setup_autopilot.py: one-command setup wrapper for deploy + optional doctor/once run.
• scripts/log_summary.py: summarize round and event logs quickly.
• scripts/install_skill.py: install this skill repo into local Codex skills directory.
• scripts/smoke_test_deploy.py: offline smoke test for deploy/runtime templates.
• assets/templates/openclaw_autopilot.py: orchestrator template.
• assets/templates/openclaw_config.json: default runtime config.
• assets/templates/openclaw_config.production.json: production baseline config.
• assets/templates/start_openclaw.sh: unattended launcher.
• assets/templates/.env.example: token env template.

References

Read these only when needed:

• Operational details: references/operations.md
• Failure signatures and fixes: references/troubleshooting.md