debuglisted

# Debugging Workflow Always work based on data. Never guess. ALWAYS run as subagent (opus high effort). ### 1. Understand the Error - Identify the failing test or error scenario from trace or user report - Pin down expected vs actual behavior in the user's own terms - Read the error message and stack trace if available ### 2. Build a Feedback Loop A loop you can run, not a plan to read code. **Pick a mechanism** — options are ordered by preference, pick the first that fits: 1. Failing test at the right seam (real code path, minimal mocking) 2. curl/HTTP script against a dev server 3. CLI invocation diffed against known-good output 4. Headless browser script (Playwright) 5. Replay a captured trace/payload 6. Throwaway minimal harness **Then optionally wrap it in a strategy** — each needs a mechanism underneath to drive: - Property/fuzz loop — for "sometimes wrong" bugs - Bisection for regressions (`git bisect run` needs a script that exits non-zero on the bug) - Differential loop — old vs new, or working vs broken input ### 3. Gate the Loop Do not proceed until all four hold: - **Red-capable** — asserts the user's exact symptom, not "didn't crash" - **Deterministic** — pin time, seed RNG, freeze network - **Fast** — seconds, not minutes - **Agent-runnable** — you can name one command you have already run at least once **Stop rule:** if you're reading code to build a theory before that command exists, stop. **Exemption — non-deterministic bugs:** when the bug is inheren