---

name: fix-bug description: Intelligently debug and fix bugs with a repro-first approach. Use when the user reports a bug, unexpected behavior, or asks to fix an issue. Establishes reproducible test cases before fixing.

Bugfix - Intelligent Bug Debugging and Fixing

When to Use This Skill

Use this skill when:

• The user reports a bug or unexpected behavior
• The user asks to fix an issue or error
• The user says something isn't working correctly
• The user invokes /bugfix [description]

Core Philosophy

Understand, Reproduce, Fix: First consult documentation to understand expected behavior, then establish a reproducible test case, and finally fix the bug. If you can't reproduce it, you can't verify the fix.

Workflow

Step 1: Read Relevant Documentation

Use the read-docs skill to search for and read documentation in ./docs/:

• Search for docs related to the bug component/system
• Read all relevant documentation files
• Follow read-docs skill procedures: log access, score helpfulness, request missing docs

Decision Point - Are docs sufficient?

→ Docs are SUFFICIENT (complete, accurate, helpful):

• You understand the expected behavior
• You know what the bug is deviating from
• Proceed to Step 2

→ Docs are INSUFFICIENT (missing, incomplete, or outdated):

• You logged access with -1 or +0 score
• You added documentation request(s) to ./docs/documentation-requests.md
• Now evaluate: Is the missing information CRITICAL to understanding the bug?

CRITICAL means:

• You cannot determine expected behavior without it
• You don't understand how the component should work
• The bug involves behavior that isn't documented
• You would be guessing at what "correct" means

If CRITICAL → STOP and invoke docs-maintainer:

• DO NOT proceed to Step 2
• DO NOT "examine the code directly"
• DO NOT "work around the gap"
• Instead: Use Task tool with subagent_type=docs-maintainer
• Provide context about what documentation is needed
• Wait for documentation to be created
• Return to Step 1 (read the newly created docs)

If NOT CRITICAL (only minor details missing):

• The overall expected behavior is clear from docs
• You just need minor implementation details from code
• Proceed to Step 2 (use code inspection to fill gaps)

What documentation tells you:

• What the correct behavior should be
• What the bug is deviating from
• Where to look for the root cause
• What test cases to write

See .claude/skills/read-docs/SKILL.md for complete documentation access procedures.

Step 2: Establish a Reproducible Test Case

1. Check for existing failing tests

   • Run the test suite: go test ./...
   • Look for tests related to the bug description
   • If a test is already failing for this issue, use it as the repro

2. If no failing test exists, create one

   • Identify the affected package/module
   • Write a minimal test that demonstrates the bug
   • The test should:
     • Be focused and minimal (test one thing)
     • Have a clear assertion that fails due to the bug
     • Include a descriptive name explaining what should happen
   • Run the new test to confirm it fails: go test -v ./pkg/[package]/ -run TestName

3. If you cannot establish a repro

   • STOP immediately
   • Explain to the user what you've tried
   • Ask clarifying questions:
     • What exact steps trigger the bug?
     • What is the expected vs actual behavior?
     • Can they provide example inputs/outputs?
     • Are there specific scenes or parameters that trigger it?
   • Do NOT proceed to debugging without a repro

Step 3: Debug the Issue

1. Analyze the failing test

   • Examine the test output and error messages
   • Identify the root cause (not just symptoms)
   • Use the Read tool to examine relevant source files
   • Look for:
     • Logic errors
     • Edge cases not handled
     • Race conditions (if parallel code)
     • Numerical precision issues (common in raytracers)
     • Incorrect assumptions

2. Form a hypothesis

   • State clearly what you believe is causing the bug
   • Explain the reasoning
   • If uncertain, use debugging techniques:
     • Add temporary logging/prints
     • Check intermediate values
     • Verify assumptions with additional test assertions

Step 4: Implement the Fix

1. Make targeted changes

   • Fix the root cause, not symptoms
   • Keep changes minimal and focused
   • Follow the existing code style
   • Avoid over-engineering or adding unnecessary features

2. Consider edge cases

   • Does the fix handle boundary conditions?
   • Are there similar bugs elsewhere in the codebase?
   • Does the fix introduce new issues?

Step 5: Verify the Fix

1. Run the reproduction test

   • go test -v ./pkg/[package]/ -run TestName
   • The test that was failing must now pass
   • If it still fails, return to Step 3

2. Run the full test suite

   • go test ./...
   • Ensure no regressions were introduced
   • If any other tests fail, investigate and fix

3. Manual verification (if applicable)

   • If the bug affects visual output or interactive features:
     • Test with the CLI: ./raytracer --scene=... --max-samples=10
     • Or test with the web interface
   • Verify the expected behavior is now correct

Step 6: Clean Up

1. Remove debugging code

   • Remove any temporary logging, prints, or debug statements
   • Keep the test case you created

2. Document if needed

   • If the bug was subtle or non-obvious, add a comment explaining why the fix is necessary
   • Update any relevant documentation

Special Considerations for Raytracer Bugs

Rendering Issues:

• Always verify tests with visual output when fixing rendering bugs
• Be careful with floating-point comparisons (use epsilon for tolerance)
• Consider both path-tracing and BDPT integrators

Performance Issues:

• Use benchmarks, not just tests: go test -bench=. ./pkg/[package]/
• Profile if needed: --profile=cpu.prof then go tool pprof cpu.prof

Parallel/Concurrency Bugs:

• These are hard to reproduce - run tests multiple times
• Look for race conditions: go test -race ./...
• Check for proper synchronization in worker pools

BVH/Acceleration Structure Bugs:

• Test with scenes that stress the BVH: spheregrid, trianglemesh, dragon
• Verify both correctness and performance

Examples

Example 1: Visual Rendering Bug

CopyUser: "/bugfix The cornell box scene has black artifacts in the corners"

Response:
1. First, let me check documentation about the Cornell box scene and rendering...
   [Uses read-docs skill: searches, reads ./docs/scenes.md and ./docs/rendering.md]
   [Logs access per read-docs procedures]

2. Based on the docs, Cornell box should have uniform lighting. Let me check for
   existing tests...
   [Checks test files]

3. I don't see a failing test. Let me create a test that renders the cornell
   scene and checks for black pixels (RGB values below a threshold in expected
   areas):
   [Creates test in pkg/scene/cornell_test.go]

4. Running the test...
   [Test fails, confirms the bug]

5. Now analyzing the cornell scene setup and rendering code...
   [Investigates light sources, material properties, ray intersection]

6. Found the issue: [explains root cause]
   [Implements fix]

7. Re-running the test...
   [Test passes]

8. Running full test suite...
   [All tests pass]

The bug is fixed! The issue was [explanation].

Example 2: Cannot Reproduce

CopyUser: "/bugfix Random crashes when rendering"

Response:
1. Let me check documentation about the rendering system and known issues...
   [Uses read-docs skill: reads ./docs/renderer.md]
   [Logs access, scores +0 - found general overview only]

2. Documentation doesn't mention crashes. Let me check for existing tests and
   try to reproduce...
   [Attempts to create a test, runs various scenarios]

3. I cannot establish a reliable reproduction of this crash. I need more
   information:

   - Which scene(s) cause the crash?
   - What command line arguments are you using?
   - Does it happen after a certain number of passes/samples?
   - Are you using CLI or web interface?
   - What does the error message say (if any)?

   Once I can reproduce it, I'll be able to debug and fix it.

Example 3: Existing Failing Test

CopyUser: "/bugfix Tests are failing in pkg/geometry"

Response:
1. First, let me check documentation about the geometry package and BVH...
   [Uses read-docs skill: reads ./docs/bvh.md]
   [Logs access, scores +1 - found helpful info]

2. Documentation explains BVH should use median-split. Running the test suite
   to identify which test is failing...
   [go test ./pkg/geometry/]

3. Found failing test: TestBVHIntersection
   [Shows test output]

4. This test is our repro. Analyzing why it fails based on expected behavior
   from docs...
   [Investigates the code]

5. Root cause: [explanation]
   [Implements fix]

6. Verifying fix...
   [Test now passes, full suite passes]

Fixed! The issue was [explanation].

Example 4: Missing Documentation (Critical)

CopyUser: "/bugfix BDPT integrator produces incorrect brightness"

Response:
1. Let me check documentation about BDPT implementation...
   [Uses read-docs skill: reads ./docs/integrators.md]
   [Logs access, scores -1 - missing MIS and brightness details]
   [Adds documentation request per read-docs procedures]

2. The missing documentation on MIS weighting is critical to understanding the
   expected behavior. Let me invoke the docs-maintainer agent to create this
   documentation before proceeding...
   [Uses Task tool with subagent_type=docs-maintainer]
   [Agent creates documentation on BDPT MIS weighting]

3. Now reading the newly created documentation...
   [Reads updated ./docs/integrators.md]
   [Logs access, scores +1 - found needed MIS details]

4. Based on the documentation, I now understand the expected MIS calculation.
   Creating a test that verifies brightness values...
   [Creates test that checks brightness values]
   [Test fails, confirms bug]

5. Analyzing code against documented behavior...
   [Investigates BDPT implementation]

6. Found the issue: MIS weights not properly normalized per the documented formula
   [Implements fix]

7. Verifying fix...
   [Test passes, full suite passes]

Fixed! The bug was in the MIS weight normalization.

Common Pitfalls to Avoid

• Don't skip documentation - Always check docs first (follow read-docs skill procedures)
• DON'T PROCEED TO STEP 2 WHEN DOCS ARE CRITICAL - This is the most common mistake:
  • If you can't determine expected behavior from docs, STOP
  • If you say "let me examine the code directly" when docs are missing, you're doing it wrong
  • Invoke docs-maintainer first, wait for docs, then proceed
  • Code inspection is NOT a substitute for understanding expected behavior
• Don't guess at fixes - Always understand the root cause first
• Don't skip the repro - "It seems to work now" is not sufficient
• Don't introduce regressions - Always run the full test suite
• Don't over-complicate - Simple, targeted fixes are better
• Don't leave debug code - Clean up before finishing

Integration with Project Testing

This raytracer project has:

• Unit tests in each package: pkg/*/[package]_test.go
• Test helper utilities
• Benchmark support
• Race detection available

Use these existing tools and patterns when creating reproduction tests.