42 lines (34 loc) · 1.93 KB

KJ Default Workflow (Claude Code)

Objective

Use Karajan Code (KJ) as the default orchestrator for implementing tasks and fixing bugs in this project.

Default rule

When asked to implement, fix, or refactor code, use kj_run via MCP instead of editing manually:

If a Planning Game MCP is available and a task ID is provided, fetch the task context first.
Run kj_run with the task description and defaults below.
If neither KJ MCP nor PG MCP are available, implement directly.

Default execution parameters

For kj_run, use:

mode: "standard"
methodology: "tdd"
coder: "claude"
reviewer: "codex"
reviewerFallback: "claude"
maxIterations: 5
maxIterationMinutes: 5

When to change behavior

Maximum rigor requested: use mode: "paranoid".
User explicitly requests no TDD: use methodology: "standard".
If kj_run fails, diagnose with kj_doctor / kj_config and retry.
Edit manually only if the user asks or KJ cannot complete the task.

Troubleshooting and subprocess architecture

See docs/troubleshooting.md for common issues. Key points:

Claude as subprocess: Claude Code 2.x requires 3 workarounds when launching claude -p from Node.js: strip CLAUDECODE env var, stdin: "ignore", read from stderr (not stdout). See src/agents/claude-agent.js -> cleanExecaOpts() / pickOutput().
Interactive wizards: The coder runs without stdin. Tasks requiring pnpm create astro, npm init, etc. must use --yes/--no-input flags or report that they cannot complete.
Checkpoint: If elicitInput returns null, the session continues (does not stop). Only explicit "stop" or "4" stops it.
Resume: kj_resume accepts stopped, failed, and paused sessions.

Example

User input: "implement the next priority task" Expected action:

If PG MCP is available, fetch the priority task.
Run kj_run with that task and defaults above.
If no PG MCP, ask the user what to implement.

⚡