-p, --prompt <PROMPT> | | Run a single prompt and exit (headless). Use "-" for stdin |
<PROMPT> (positional) | | Same as -p — koda "fix the bug" works |
-a, --agent <NAME> | | Agent definition to use (JSON in agents/, default: default) |
-s, --resume <ID> | | Resume a previous session by ID prefix |
--model <NAME> | KODA_MODEL | Model name or alias (e.g. claude-sonnet, gemini-flash) |
--provider <NAME> | KODA_PROVIDER | LLM provider (anthropic, gemini, openai, ollama, …) |
--base-url <URL> | KODA_BASE_URL | Override the provider’s API base URL |
--max-tokens <N> | | Maximum output tokens |
--temperature <F> | | Sampling temperature (0.0–2.0) |
--thinking-budget <N> | | Anthropic extended thinking budget (tokens) |
--reasoning-effort <L> | | OpenAI reasoning effort (low, medium, high) |
--mode <MODE> | KODA_MODE | Trust mode: auto (default, #1241) or safe. Auto auto-approves mutating actions; the kernel sandbox contains them — destructive ops (rm -rf, git reset --hard, git push --force, Delete) and outside-project writes still confirm. Auto requires the kernel sandbox; on unsandboxed platforms koda refuses to start (use --mode safe or install the platform sandbox backend). Safe confirms every mutation. Kernel sandbox with credential protection is always active when available. |
--output-format <FMT> | | Headless output format: text (default) or json |
--project-root <DIR> | | Project root (defaults to cwd) |