The Learning Code Doctor

What Problem It Addresses

Debugging often mixes two failure modes: low-level patterns such as if x = y, mutable default arguments, and bare except blocks, and context-dependent bugs that may break callers after a fix. This skill separates deterministic pattern fixes from a small Bash-driven repair loop, and records reflections to reduce repeated mistakes.

How It Works

• L1 deterministic fixes: matches known error patterns and applies direct repairs, such as changing append([x]) to extend([x]).
• L2 AI reasoning: uses commands like grep, sed, and python -c to locate, modify, and validate code without relying on complex framework APIs.
• Reflexion loop: when a command fails, tests do not pass, or multiple attempts fail, it generates a reflection, adds it to context, and tries a new strategy.
• Document memory: stores failures, lessons, and version changes in ERRORS.md, LEARNINGS.md, and VERSIONS.md.
• Verification guardrails: after a fix, it requires the original failure to disappear, runs a minimal test when possible, and checks call-site impact with grep -rn.

Boundaries And Notes

It defaults to review mode and does not perform irreversible production operations on behalf of the user. High-risk changes are marked for confirmation with rollback paths. If the input is ambiguous, external APIs are unavailable, or a suggestion conflicts with reality, it clarifies first, degrades gracefully, or marks the advice as not applicable. If no usable tests exist or too many call sites are affected, it flags the impact scope for confirmation instead of claiming safety.