Context-Aware Debugging Loops
Core Principle
Errors are symptoms. Context reveals causes. Feed the agent what it needs to reason, not just react.
The Diagnostic Triad
Every debug prompt needs three elements:
- What happened (error/symptom)
- What should have happened (intended behavior)
- What changed (recent modifications, if known)
Context Surfacing Patterns
Minimal Context (Quick Fixes)
"Error: Cannot read property 'map' of undefined in UserList.jsx line 42.
Expected: render array of users from API.
The API returns { data: { users: [...] } }"
Full Context (Complex Bugs)
"Authentication flow breaks after login redirect.
- Error: 401 on /api/profile after successful /api/login
- Expected: JWT in cookie should authorize subsequent requests
- Recent change: Added sameSite cookie attribute
- Relevant files: auth.middleware.js, login.controller.js
- Browser: Cookie IS being set (verified in DevTools)"
Diagnostic Questions to Answer First
Before prompting the agent:
- Can I reproduce it consistently?
- What's the smallest input that triggers it?
- What was the last working state?
- What assumptions am I making about the data?
Anti-Patterns
| Weak | Strong |
|---|---|
| "It doesn't work" | "Returns null instead of user object when..." |
| "Fix this error" + stack trace | Stack trace + expected behavior + recent changes |
| "Why is this broken?" | "This breaks when X but works when Y" |
Power Move
When stuck in a loop: "Explain three different hypotheses for why this could fail, then tell me what evidence would confirm or eliminate each."
The agent becomes a reasoning partner, not just a fix-generator.