Systematic Debugging

Start a methodical debugging session using the four-phase investigation framework with iterative review.

Instructions

• Consider the recent conversation history, your context, and the problem to be debugged.
• Specific user instructions should be followed unless they are contradictory to the task at hand. $ARGUMENTS

Steps

1. Load — Use the Skill tool and load the baselayer:debugging-and-diagnosis skill
2. Consider — Ultrathink and analyze the problem, consider available evidence, potential causes, and investigation approach
3. Dispatch or Execute — Choose execution path based on available tools:
   • If Task tool available: Run the debug loop (see below)
   • If Task tool unavailable: Execute the debugging methodology directly using the loaded skill

Debug Loop (when Task tool available)

Run iterative cycles until the issue is resolved:

┌─────────────────────────────────────────────────────────┐
│  1. INVESTIGATE — Dispatch baselayer:debugger           │
│     └─ Collect evidence, form hypothesis, propose fix   │
│                          ↓                              │
│  2. REVIEW — Dispatch baselayer:ranger                  │
│     └─ Validate fix, check for regressions, verify      │
│        root cause addressed (not just symptoms)         │
│                          ↓                              │
│  3. EVALUATE — Check review outcome                     │
│     └─ If approved → Done                               │
│     └─ If issues found → Back to step 1 with feedback   │
└─────────────────────────────────────────────────────────┘

Loop Execution

1. Dispatch debugger (background) — Pass error context, evidence, any prior feedback
2. Retrieve results — Use TaskOutput to get proposed fix and rationale
3. Dispatch ranger (background) — Pass the proposed fix and debugger's findings for code review
4. Retrieve review — Use TaskOutput to get validation results
5. Evaluate:
   • Approved: Report success, document root cause and fix
   • Issues found: Loop back to debugger with reviewer feedback
   • Max iterations (3): Escalate to user with findings so far

Context Accumulation

Each loop iteration passes forward:

• Original error context
• Investigation findings from debugger
• Review feedback from reviewer
• Cumulative hypotheses tested

The Iron Law

NO FIXES WITHOUT ROOT CAUSE INVESTIGATION FIRST

If you catch yourself wanting to "just try something" — stop. Return to investigation.

Context Handoff (for initial dispatch)

When dispatching to the debugger subagent, include:

• Exact error messages or unexpected behavior
• Stack traces if available
• Recent changes (git diff context)
• Reproduction steps if known
• Any hypotheses already formed
• Prior loop feedback (if iterating)