Headless mode

Before v4, Karajan WAS the orchestrator: kj run spawned coder, reviewer, tester, security and judge agents as subprocesses and drove the loop itself. That pipeline did not go away — it became the headless mode, for the cases where no host agent is at the wheel:

• Unattended delivery — kj autorun <spec> chains spec → plan → user stories → run to completion, with an arbiter resolving conflicts.
• Parallel lanes — kj run --parallel <n> executes a plan’s stories concurrently in per-story git worktrees. Each lane receives KJ_LANE_SLOT and KJ_PORT_OFFSET, and session.worktree_setup declares the per-lane environment (ports, DBs, docker project names) so lanes never collide.
• Unattended runs — --non-interactive (or KJ_NON_INTERACTIVE=1) auto-answers safe gates (warn findings continue, decision on stderr); FAIL findings stop the run with exit code 1 instead of blocking on a board modal.
• CI / scripted runs — anywhere without an interactive agent session.

Same gates, same verdicts

Headless runs are governed by the same contract as your agent: when the repo has the review gate, the pipeline records its internal reviewer’s approval as a verdict for the exact staged diff before committing. One gate, two ways of working — you can mix them freely in the same repo.

When to choose which

Situation	Use
You are working WITH an agent (Claude Code, Codex, Gemini, Cursor)	The environment — this is the v4 default
Batch of independent stories overnight	kj autorun / kj run --parallel
No agent CLI available (pure CI)	kj run

Full pipeline documentation

The subprocess pipeline (24 roles, triage policies, Solomon escalation, budget ceilings, HU Board integration, SonarQube stages) is documented in depth in the v3 (legacy) section of this site. “Legacy” means the docs describe the pre-v4 framing — the code is current and maintained; only the center of gravity moved to the environment.