Headless automation

Use --print for a single non-interactive prompt:

mj --print "summarize the current diff"
git diff | mj --print -

If the prompt value is omitted or -, Mjolnir reads standard input.

Permission behavior

Headless mode defaults to --permission-mode manual, which rejects prompts instead of hanging. auto can approve supported file changes but rejects shell execution; yolo approves everything and belongs only in disposable scope.

mj --cwd /tmp/eval \
  --print \
  --permission-mode manual \
  "inspect the project without changing it"

Output formats

• text prints Thor’s final result.
• json emits one object with session_id, resumed, result, stop_reason, usage, council_usage, and error fields.
• stream-json emits newline-delimited typed records followed by a final result record.

Stream records can include connected, session_started, agent messages and thoughts, tool calls and updates, permissions, reviews, warnings, errors, and the result. Council activity carries actor labels so Thor, Eitri, and Loki remain attributable.

mj --print --output-format stream-json "summarize this repository" \
  | jq -c 'select(.type == "result" or .actor != null)'

Treat the machine-readable record shape as an integration contract for the current release, not an unversioned promise that fields will never grow.

One-shot role selection

--thor MODEL, --loki MODEL|disabled, and --eitri MODEL|disabled override the saved Council for one invocation. They require explicit IDs and are never written back to the config file.

For a controlled first run, use the 10-minute evaluation. For networked access to an interactive Council, continue with Remote control.