agentic-praxis-grimoire

Debugging Systematically

Core principle

Change behavior only after evidence identifies a plausible cause and a bounded test of that cause.

Do not use

Do not use this skill when the cause and correction are already established, for unrelated preventive refactoring, or as permission to inspect or mutate systems outside the authorized boundary. Route an accepted correction to the appropriate implementation procedure.

Procedure

  1. State the observable symptom, expected behavior, environment, and known boundary. Preserve exact diagnostics in a privacy-safe form.
  2. Reproduce the symptom reliably when practical. If it is intermittent, bound observation and record frequency, timing, and conditions without claiming a deterministic reproduction.
  3. Inspect recent changes, inputs, state transitions, logs, and component boundaries that can distinguish plausible causes.
  4. List explicit hypotheses and the evidence each predicts. Include multiple contributing causes when the observations support them.
  5. Test one variable at a time where practical. Prefer read-only observation or reversible experiments before corrective mutation.
  6. Distinguish root or contributing cause from symptom, containment, and unrelated defects.
  7. Apply or recommend the narrowest correction supported by the evidence and permitted by the current assignment and project-owned mutation authority.
  8. Re-run the original reproduction or observation, add regression evidence where appropriate, and run the proportional surrounding gate.
  9. Record the cause confidence, correction, verification, unresolved alternatives, and any new issue class separately.

Project-owned parameters

The project owns live-system access, log and data privacy, diagnostic tools, retry and observation limits, environment cleanup, mutation authority, containment and rollback procedure, and required regression evidence.

Evidence and completion

A debugging result includes a reproduced or bounded symptom, gathered evidence, tested hypotheses, cause assessment, narrow correction or disposition, and reproduction-specific verification. An unreproduced failure may end in a truthful observation or instrumentation recommendation rather than a guessed fix.

Stop or escalate

Stop when evidence requires secrets or unauthorized production access, safe observation is unavailable, experiments would be destructive, hypotheses remain indistinguishable, the environment is invalid, or a new architecture, security, privacy, or ownership issue appears.

Common mistakes