SuperAPI AI Directory

debug-tools

Coding & development

View repository

Updated 10-15 08:23

About this skill

Debug Tools

Iterative debugging workflow with flexible technique selection and escalation.

Triggers

  • Debug a bug ("debug this", "investigate", "trace issue", "fix bug", "why is X broken") → run the workflow below
  • Add debug logs ("add debug logs", "inject logs", "trace with logs") → enter at step 3
  • Cleanup logs ("remove debug logs", "cleanup logs") → enter at step 5
  • Pattern lookup ("debug patterns", "common bugs", "used to work") → enter at step 2

Workflow

investigate → fix → verify → done
  ^_______________________|  (max 3 attempts, then escalate)
  1. Load investigation.md and work its steps: understand the bug, analyze the code, enumerate hypotheses with confidence scores, rank them, propose a fix, verify, report. Enter at the step the current state calls for — a session already carrying evidence does not restart at Step 1.
  2. Load debugging-patterns.md when a symptom needs matching against a known bug shape, when analysis stalls and the broken code has to be diffed against a working example, or when the user reports that something used to work.
  3. Load log-injection.md when reading the code cannot show the mechanism and only observing the running system can. Not every session needs it.
  4. Fix and verify. Propose a fix only when the evidence names the mechanism; never as exploration. Run the reproduction after the fix is applied, and repeat it 3-5 times for a race condition or an intermittent bug.
  5. Load log-cleanup.md once the fix is verified, or on explicit request. Run it before changes go to version control.

A sensitive value never reaches an injected log — passwords, tokens, API keys, PII, session identifiers. This binds anywhere a log is added, including mid-investigation without step 3 loaded.

Anti-Pattern: Symptom Whack-a-Mole

Fixing the same symptom in multiple places signals an architectural issue, not a localized bug. When fix N introduces bug N+1, stop. The 4th attempt must escalate to architectural review: re-examine the abstraction, the missing layer, or the flawed assumption — not retry a deeper version of the same approach.