agent-introspection

Agent Introspection — Systematic Self-Diagnosis

Iron Law

NO UNSUPPORTED SELF-HEALING CLAIMS. Only assert recovery actions you can actually perform with available tools.

When to Trigger

• 3+ consecutive failures on the same task
• Repeating the same approach with different parameters (loop detection)
• Context window exceeds 70% without resolution
• Error messages repeat with no progress
• User explicitly says "you're stuck" or "try a different approach"

Process

1. Failure Capture

Stop immediately and record:

• Error type: syntax_error / type_error / test_fail / lint_fail / build_fail / runtime_error / permission_denied / not_found
• Last 3 tool calls: what was attempted and what failed
• Context pressure: approximate context usage percentage
• What changed: what was the last successful state

2. Root Cause Diagnosis

Match against known patterns:

Pattern	Symptoms	Root Cause
Loop trap	Same error 3+ times	Wrong approach, not wrong parameters
Context overflow	Increasingly confused responses	Too much information, need compaction
Environment drift	"Works locally" failures	Missing env var, different tool version
Cascade failure	Fix A breaks B	Underlying assumption is wrong
Tool mismatch	Wrong tool for the job	Need a different approach entirely

3. Controlled Recovery

Execute ONLY the smallest safe action:

• Loop trap → Abandon current approach entirely. Try a fundamentally different strategy.
• Context overflow → Run /compact or summarize current state, then continue.
• Environment drift → Verify environment with explicit checks (which, --version).
• Cascade failure → Revert to last known good state. Re-analyze assumptions.
• Tool mismatch → Switch tools. If Edit fails 3 times, try Write. If Bash fails, try Read first.

4. Introspection Report

Generate a structured report:

## Introspection Report
- **Failure type**: [type]
- **Root cause**: [diagnosis]
- **Recovery action**: [what was done]
- **Confidence**: [high/medium/low]
- **Next step**: [what to do if this happens again]