---
name: cad-debug
description: "Systematic debugging - scientific method with hypothesis state persisted across /clear, and a user-gated second-model consult at dead-ends. Single pass, no session-manager layer"
argument-hint: "[list | status <slug> | continue <slug> | --diagnose] [symptom]"
allowed-tools:
  - Read
  - Write
  - Edit
  - Bash
  - Grep
  - Glob
  - AskUserQuestion
---

<objective>
Debug a symptom by the scientific method - form hypotheses, run the cheapest
discriminating test for each, confirm a root cause, fix it, verify the symptom
is gone. The investigation is persisted to `.planning/debug/<slug>.md` after
every step, so a `/clear` or a new session resumes it losslessly. At a genuine
dead-end (repeated failed fixes, exhausted hypotheses), OFFER a second-model
consult (references/consult.md) - user-gated, decision-support only.

Routing (parse `$ARGUMENTS` first):
- `list` - list open debug sessions (`.planning/debug/*.md` not resolved).
- `status <slug>` - print that session's current state.
- `continue <slug>` - resume that session from its state file.
- `--diagnose` - stop at the Root Cause Report; do not apply a fix.
- otherwise - the remainder is the symptom; start a new session.
</objective>

<execution_context>
@$HOME/.claude/cadence-core/workflows/debug.md
</execution_context>

<process>
Route on `$ARGUMENTS`, then run the debug workflow single-pass end-to-end.
Persist the state file after every hypothesis test and fix attempt. Never
auto-loop fixes and never auto-consult - both are deliberate, and the consult
is always user-gated.
</process>