diagnose

# Diagnose ## Purpose Use this skill for product, code, test, and runtime failures where the cause is not obvious. The core discipline is to build a fast, trusted pass/fail loop before fixing anything. Do not use this for agent self-recovery, repeated tool loops, or context drift. Use `agent-introspection-debugging` for those. ## Workflow ### 1. Build The Feedback Loop Find the smallest repeatable signal that shows the bug. Prefer, in order: 1. A failing test at the seam that reaches the bug. 2. A CLI command, fixture, or HTTP request that reproduces the symptom. 3. A browser automation script for UI, console, or network failures. 4. A replayed artifact such as a log, event payload, HAR file, trace, or saved input. 5. A throwaway harness that exercises the failing code path in isolation. 6. A stress, property, fuzz, or looped reproduction for nondeterministic failures. 7. A bisection or differential loop when the bug appeared between known commits, versions, datasets, or configs. 8. A human-in-the-loop script only when manual steps cannot be automated; adapt `scripts/hitl-loop.template.ps1`. Before moving on, name one command that has already been run and can catch this exact bug. A manual-loop script is acceptable only when automation is impossible. The loop must be red-capable, deterministic or high-reproduction for flaky bugs, fast enough to iterate, and runnable without hidden human judgment. If no credible loop can be built, stop and report what was tried. Ask ...