Documentation

Harness

Use the Gajae Code harness to run coding-agent sessions with observable lifecycle, submit gates, and workspace-aware control.

Why a harness?

A coding model is useful only when it can be steered, observed, paused, and verified. The harness turns an agent run into a managed session with lifecycle state, owner liveness, branch context, and submission gates.

Design intent: use lifecycle state, owner liveness, and workspace/branch context to coordinate work. Record evidence through the supported Coordinator MCP or SDK interfaces rather than terminal output.

Typical loop

gjc
/skill:deep-interview
/skill:ralplan
gjc ultragoal create-goals --brief "implement, verify, and report evidence"
gjc ultragoal complete-goals

Default loop is interview → plan → prove. Run /skill:autoresearch only when evidence must ground that loop — before the interview or after the spec — never as a fourth sequential step.

Visible interactive sessions

For human-operated, visible-TUI work, scripts/gjc-session/create.sh <session-name> <worktree-path> can start or reuse a named worktree-backed session. Use this only to set up the interactive workspace; it is not an automation or evidence interface.

Automations must use Coordinator MCP or the SDK for lifecycle control, task submission, status, and evidence. Do not inject prompts, scrape panes, or treat terminal output or tmux ownership as authoritative.

Operational guardrails

Use dedicated worktrees for risky implementation and review work.
Keep branch metadata aligned with the actual checkout.
Use Coordinator MCP or the SDK for automated task routing and lifecycle checks.
Keep interactive sessions human-operated and visible when using the TUI.
Retire completed workspaces and sessions instead of leaving stale state behind.