grith run

grith run <TASK>

Run a single task through grith's built-in REPL non-interactively. The configured provider (Ollama / OpenAI / Anthropic / OpenRouter) is invoked with the task as a prompt; tool calls the model makes go through the filter pipeline before executing.

Synopsis

grith run [GLOBAL OPTIONS] <TASK>

Arguments

Behaviour

• Uses [llm] default_provider unless overridden via --provider (global flag).
• Runs the model to completion or until tool-call budget is exhausted.
• Every file/shell/network call the model issues is supervised exactly like grith exec.
• Exits when the task is complete. Exit code 0 on success.

Examples

A trivial completion:

$ grith run "what's the capital of New Zealand"
Wellington.

A task that requires a tool call:

$ grith run "summarise package.json"
[grith] file_read package.json → score 0.4 → allow
This project is "my-app", version 1.2.0, with dependencies on next, react...

A task that quarantines:

$ grith run "what's in my .ssh/config"
[grith] file_read ~/.ssh/config → score 4.2 → quarantine

[grith] Pending review. Approve in another terminal with:
      grith digest review

The CLI hangs until the digest item resolves, then continues with the model's response (or with the failure if denied).

When to use this vs grith exec

• grith run — single one-shot prompt, you want grith's built-in REPL plumbing.
• grith exec — wrapping an external agent (Claude Code, Codex, etc.) with grith's supervision. Almost always what you want for real work.

See also

• grith exec — supervise an external agent
• Connect a model — BYOK setup
• [llm] config