Unattended runs

Up to 4.x, Karajan also had a headless mode: kj run and kj autorun spawned coder, reviewer and judge agents as subprocesses and drove the loop themselves. That pipeline was retired in 5.0 (see Migrating to 5.0): the work is done by your agent, and Karajan governs it.

What stays for unattended use:

• --non-interactive (or KJ_NON_INTERACTIVE=1): safe gates auto-answer (warn findings continue, the decision goes to stderr); FAIL findings stop with exit code 1 instead of waiting for an answer.
• The gates in CI: kj harden seeds the workflows that re-check every PR (review verdicts, policy, privacy, PR size), so a tampered local hook does not get a diff through.
• Parallel work: one lane per card with kj worktree start <slug> and one agent session inside each. session.worktree_setup still declares the per-lane environment (ports, DBs, docker project names) so lanes never collide.

When to choose which

Situation	Use
You are working WITH an agent (Claude Code, Codex, Gemini, OpenCode)	The environment: this is the default
Several independent cards at once	One lane per card (kj worktree start), one agent session per lane
CI with no agent at all	The gates: kj review --check, kj policy check, kj privacy scan

The old pipeline

The subprocess pipeline is documented in the v3 (historical archive) section of this site. It describes how Karajan worked before 5.0; to keep using it, stay on karajan-code@4.