| name | debug |
| description | Find the root cause of a bug or failing test. Spawns 1 debugger subagent that reproduces, ladders hypotheses, isolates with evidence, and applies the minimal fix. Use when a test fails, /build reports blocked, or "X stopped working". Does NOT design features (that's /plan). |
Skill: /debug
ORCHESTRATOR. One debugger subagent. It diagnoses before it touches anything.
Invocation
/debug "<error text / failing test / what broke>" (optionally /debug <slug> to tie it
to a feature build)
Steps
- Gather the symptom: the error/stack/failing-test name from the user, plus the build
report
swarm-report/<slug>-build.md if a slug was given.
- Spawn 1 debugger (
Task, general-purpose). Prompt:
Answer TERSE. Read .claude/agents/debugger.md and follow it exactly.
Symptom: <verbatim error / failing test>. Feature slug (if any): .
- Report the debugger's return verbatim: root cause, evidence, fix, and the
re-run output.
- Route the outcome:
fixed → suggest /review <slug> to check the fix.root_cause_found (found but not fixed) or stuck → show the hypothesis ladder;
the user decides the fix, or re-run /build <slug> with the root cause as input.cannot_reproduce → ask the user for exact repro steps; do not guess-patch.
Where it sits in the loop
/plan → /build → /review. If a test fails or review says rework → /debug →
back to /build. The Stop test-gate hook blocks "done" until tests actually run.
