Headless & output formats

--headless runs the same agent with no terminal UI — for scripts, CI, and automation. This page is the output contract; for end-to-end patterns see the CI recipe.

empryo --headless "explain the auth middleware"
echo "find unused exports" | empryo --headless        # prompt from stdin

Assistant text streams to stdout; progress (model, tools, footer) goes to stderr, so pipes stay clean. Output is one of three shapes.

Plain (default)

Assistant text streams straight to stdout. When stdout is a TTY it is Markdown-rendered; piped, it is raw text. Force either way with --render / --no-render. Add --quiet to drop the stderr progress lines, or --diff to print the changed files at the end.

JSON — --json

One pretty-printed object after the run completes. No streaming.

{
  "model": "anthropic/claude-opus-4-8",
  "mode": "default",
  "prompt": "list all TODOs",
  "output": "…assistant's final answer…",
  "steps": 7,
  "tokens": { "input": 18432, "output": 1204, "cacheRead": 16000 },
  "toolCalls": ["grep", "read", "find"],
  "filesEdited": ["src/auth/session.ts"],
  "duration": 38211
}

The error key is present only on failure.

Events — --events

A JSONL stream: one JSON object per line, emitted in real time. Each has a type:

empryo --headless --events "refactor the store" | while read -r line; do
  case "$(jq -r .type <<<"$line")" in
    tool-call) echo "→ $(jq -r .tool <<<"$line")" ;;
    done)      echo "done in $(jq -r .duration <<<"$line")ms" ;;
  esac
done

Chat — --chat

A multi-turn REPL over stdin. Each line is a turn; a trailing \ continues to the next line; a blank line is skipped. --json and --events apply per turn. On exit (EOF or Ctrl+C) the session is auto-saved and a resume hint is printed.

empryo --headless --chat <<'EOF'
what does the session store do?
now add a TTL field to it
EOF

Chat mode emits ready before each prompt, turn-done per turn, and chat-done at the end.

Exit codes

What headless disables

• The TUI, editor panel, and splash screen.
• Interactive approvals — destructive actions auto-allow.
• User steering — there is no live stdin during a run (except --chat).