Reference

agent run

Drive a coding agent (Claude Code, OpenAI Codex, Gemini CLI, or OpenCode) to completion inside an ephemeral project sandbox (read-only /project; read-write /output, /code, or /ir, plus /error; persistent per-project agent memory); --allow-host runs it as a host process instead. The backend is --agent (a named agent config or backend id), else the project's first configured agent recipe, else $OVERPLANE_AGENT. The prompt is --prompt, else the trailing positional words, else piped stdin. Live assistant text streams to stdout and a one-line usage/cost summary goes to stderr; with --json, stdout carries one canonical result object (backend, model, text, session_id, usage, cost_usd, cost_source, wall_ms, exit_code). Cost is reported directly by the backend where available (claude-code, opencode) else computed from an advisory price table (codex, gemini-cli). The agent's exit code is propagated.

Optional

--agent

string

Named agent config or backend id: claude-code, codex, gemini-cli, or opencode.

--prompt

string

Task prompt (-p); else positional words, else piped stdin.

--model

string

Model id (-m); defaults to the backend's default model.

--output

string

Extract the /output fileset to this host directory on success (-o).

--system-prompt-file

string

Shared cache-anchor system-prefix file routed per backend.

--workdir

string

Container working directory (required: /output, /code, or /ir).

--json

boolean default false

Emit a canonical JSON result instead of live text.

--allow-host

boolean default false

Run on the host instead of inside a sandbox.