phase-gated-debugging – skill Lập trình cho agent AI

Phase-Gated Debugging

Overview

AI coding agents see an error and immediately edit code. They guess at fixes, get it wrong, and spiral. This skill enforces a strict 5-phase protocol where you CANNOT edit source code until the root cause is identified and confirmed.

Based on [claude-debug](https://github.com/krabat-l/claude-debug) (full plugin with PreToolUse hook enforcement).

When to Use

Use this skill when:

• a bug keeps getting "fixed" without resolving the underlying issue
• you need to slow an agent down and force disciplined debugging before code edits
• the failure is intermittent, a regression, performance-related, or otherwise hard to isolate
• you want an explicit user confirmation checkpoint before any fix is applied

The Protocol

Phase 1: REPRODUCE

Run the failing command/test. Capture the exact error. Run 2-3 times for consistency.

• Inspect the command and its side effects before reproduction; do not execute unknown code blindly
• Do NOT hypothesize
• Do NOT edit any files

Phase 2: ISOLATE

Read code. Add diagnostic logging marked // DEBUG. Re-run with diagnostics. Binary search to narrow down.

• Only // DEBUG marked logging is allowed
• Do NOT fix the bug even if you see it

Phase 3: ROOT CAUSE

Analyze WHY at the isolated location. Use "5 Whys" technique. Remove debug logging.

State: "This is my root cause analysis: [explanation]. Do you agree, or should I investigate further?"

WAIT for user confirmation. Do NOT proceed without it.

Phase 4: FIX

Remove only the diagnostic lines added for this investigation; preserve pre-existing logging and user edits. Apply minimal change addressing confirmed root cause.

• Only edit files related to root cause
• Do NOT refactor unrelated code

Phase 5: VERIFY

Run original failing test — must pass. Run related tests. For intermittent bugs, run 5+ times.

If verification fails: root cause was wrong, go back to Phase 2.

Bug-Type Strategies

| Type | Technique |

|------|-----------|

| Crash/Panic | Stack trace backward — trace the bad value to its source |

| Wrong Output | Binary search — log midpoint, halve search space each iteration |

| Intermittent | Compare passing vs failing run logs — find ordering divergence |

| Regression | git bisect — find the offending commit |

| Performance | Timing at stage boundaries — find the bottleneck |

Key Rules

1. NEVER edit source code in phases 1-3 (except // DEBUG in phase 2)

2. NEVER proceed past phase 3 without user confirmation

3. ALWAYS reproduce before investigating

4. ALWAYS verify after fixing

Example

User request:

> a bug keeps getting "fixed" without resolving the underlying issue; diagnose it and return the concrete next action.

Limitations

• Use this skill only when the task clearly matches the scope described above.
• Do not treat the output as a substitute for environment-specific validation, testing, or expert review.
• Stop and ask for clarification if required inputs, permissions, safety boundaries, or success criteria are missing.

Execution boundaries

This Markdown guide does not bundle the upstream enforcement hook. Reproduction must use authorized fixtures and bounded runs; never repeat destructive or externally visible actions solely to meet a repetition count. User instructions about autonomy or stopping take precedence over the interview cadence. Failed verification calls for reassessing evidence, not assuming every failure proves the same root cause false.