On this page

Commands and permissions

How are commands, permissions, and model routes configured?

In one minute

  • Commands are argument arrays with a repository-relative working directory and a timeout; no shell is inserted.
  • Trusted mode uses provider permission-bypass flags. Restricted mode uses each provider's own controls.
  • Trusted mode is not operating-system containment.
  • Routes name a provider and model per stage, with an ordered list of fallbacks.

Commands use argument arrays, a repository-relative working directory, and a bounded timeout. No shell is inserted automatically.

{
  "argv": ["python3", "-m", "unittest", "discover", "-s", "tests", "-v"],
  "cwd": ".",
  "timeout_seconds": 300
}

If a command deliberately invokes a shell, set shell: true. Keep credentials in the host or named env_refs, never literal environment values in tracked configuration.

Permission modes

Provider Trusted Restricted
Codex --dangerously-bypass-approvals-and-sandbox workspace-write with unattended approval refusal
Claude --dangerously-skip-permissions dontAsk with an explicit allowed_tools list
Cursor --force, --sandbox disabled, --trust --sandbox enabled, --trust

The runner checks installed flags before execution. Unsupported controls stop preflight. Restricted Claude requires execution.provider_settings.claude.allowed_tools. Use only the tools the task actually needs.

These providers offer different controls; restricted mode is not an identical cross-provider sandbox. Trusted mode relies on configured access and post-stage outcome checks. It is not operating-system containment.

Routes

A route names the provider, model, and optional reasoning setting. A stage can have an ordered fallback list.

{
  "pipeline": {
    "routes": {
      "default": [
        {"provider": "codex", "model": "YOUR_CODEX_MODEL", "reasoning": "high"}
      ]
    }
  }
}

Setup records your chosen models. Preflight checks command compatibility; model/account availability is established when the provider receives the request. Cursor effort belongs in its supported model expression.

Model and cost choices · Evidence formats

Checkpoints EXE-1 EXE-4 ORG-3

Edit this page on GitHub