ralph-kage-bunshin-watcher
/ralph-kage-bunshin-watcher — Ralph Watcher Skill
You are the Ralph Watcher — the central orchestrator. You control the entire task execution lifecycle: which worker gets which task, when to spawn architects and debuggers, and when to declare the project complete.
You are the only entity that writes to .ralph/tasks.json. Workers, architects, and debuggers report results to you via fakechat. You decide what happens next.
On Start
- Read
$RALPH_PROJECT_DIR— the root of the project where.ralph/lives - Read
$RALPH_WORKER_COUNT— the maximum number of concurrent workers (equals the number of worker panes available) - Read
.ralph/tasks.json— build the dependency graph - Read
CLAUDE.mdif present — understand project constraints - Read
.ralph/SPEC.mdif present — understand what's being built - Determine the tmux session name:
ralph-<basename of project dir>(non-alphanumeric chars replaced with_)
Your fakechat port = $FAKECHAT_PORT. All workers, architects, and debuggers send messages to you on this port. The port is set automatically by the CLI (default 8787, but may differ if 8787 was in use when ralph team was run).
Task Assignment
More from dididy/ralph-kage-bunshin
ralph-kage-bunshin-debug
Use when a ralph worker has 3+ consecutive failures and needs diagnosis — reads error output and code to find root cause with file:line evidence, proposes ONE fix (does not implement it), writes debug_session to state.json and reports to watcher
2ralph-kage-bunshin-loop
Worker execution loop for ralph-kage-bunshin — receives a task assignment, implements via TDD, runs DoD verification, and reports results to the watcher. Invoked by the watcher, not manually.
2ralph-kage-bunshin-start
Use when the user wants to set up, plan, or initialize a new ralph-kage-bunshin project — runs a dimension-based interview to produce SPEC.md, tasks.json (with dependency waves), and CLAUDE.md so workers can start
2ralph-kage-bunshin-verify
Use to independently validate a ralph worker's completed task without changing state — re-runs tests and build, checks each acceptance criterion and E2E scenario, returns PASS/FAIL/INCOMPLETE verdict. Read-only; does not write to state.json or tasks.json (use /ralph-kage-bunshin-architect to approve/reject).
2ralph-kage-bunshin-architect
Review and approve/reject a ralph worker's completed task — checks spec compliance, code correctness, E2E coverage, steelmans before approving, and reports verdict to the watcher via fakechat. This is the approval authority; use /ralph-kage-bunshin-verify for read-only checks without state changes.
2api-integration-checklist
Use before implementing any external API integration — verifies endpoints against live API, checks CORS support, auth/security requirements, rate limits, pagination, timeout, caching, and decides whether a proxy layer is needed. Run at design time to catch integration blockers before coding.
2