Systematic Debugging

Systematic Debugging

A structured 4-phase approach to debugging that prevents wild goose chases and thrashing. Adapted from proven debugging methodologies for Forge workflows.

Why Systematic Debugging Matters

Anti-pattern (debugging thrash):

• See error
• Try random fix
• Different error appears
• Try another random fix
• Repeat for hours
• Eventually fix by accident

Systematic approach:

• Understand root cause before fixing
• Form hypothesis before acting
• Validate hypothesis with evidence
• Fix with confidence

The Four Phases

Phase 1: Root Cause Analysis

Goal: Understand WHY the error is happening, not just WHAT the error is.

Actions:

1. Read the full error message (don't skim)
2. Identify the exact line/function where failure occurs
3. Trace backward from failure point to find originating cause
4. Ask: "What assumption is being violated?"

For Python errors:

• Check stack trace for first application code (not library)
• Look for AttributeError, KeyError, TypeError patterns
• Verify models match migration state

For TypeScript/JavaScript errors:

• Read the type error explanation carefully
• Check for null/undefined access patterns
• Look for type mismatches in function calls

For build errors:

• Read the first error, not just the last one
• Check for missing imports or dependencies
• Look for version conflicts

Output:

Root Cause: {Specific reason the error occurs}
Evidence: {Error message, stack trace, or log showing this}

Phase 2: Pattern Recognition

Goal: Identify if this is a known pattern or new failure mode.

Actions:

1. Search codebase for similar error patterns
2. Check if this error has appeared in git history (might have been fixed before)
3. Look for existing error handling in related code
4. Ask: "Have we seen this category of error before?"

Common patterns:

• Path confusion (relative vs absolute paths)
• Context not loaded (accessing data before it exists)
• State mismatch (expecting something that isn't there)
• Git state confusion (expecting branch exists, but doesn't)

Output:

Pattern: {Known pattern name or "New failure mode"}
Similar cases: {Related errors or fixes in codebase}

Phase 3: Hypothesis Formation

Goal: Form testable hypothesis BEFORE making changes.

Actions:

1. Based on root cause and pattern, predict what change will fix it
2. Identify how to test the hypothesis
3. Consider alternative hypotheses
4. Choose most likely hypothesis first

Good hypothesis characteristics:

• Specific (not "something's wrong with X")
• Testable (can verify if correct)
• Explains all symptoms (not just one error)

Example hypotheses:

❌ Bad: "The file path is wrong"

✅ Good: "The cd command failed silently because directory
doesn't exist, so subsequent commands run in wrong location"

Output:

Hypothesis: {Specific prediction of what will fix it}
Test: {How to verify if hypothesis is correct}
Expected outcome: {What should happen if hypothesis is right}

Phase 4: Fix with Validation

Goal: Apply fix and verify it solves the problem without creating new issues.

Actions:

1. Make minimal change to test hypothesis
2. Run test or reproduce error scenario
3. Verify fix solves original problem
4. Check for side effects or new errors
5. If hypothesis wrong, return to Phase 3 with new information

Validation methods:

• Unit tests: Run specific test that failed
• Integration tests: Run full test suite for affected module
• Manual verification: Reproduce original error scenario
• Regression check: Ensure no new errors introduced

Anti-patterns:

• Changing multiple things at once (can't tell what fixed it)
• Not verifying the fix (claiming success without evidence)
• Fixing symptom instead of root cause (band-aid fix)

Output:

Fix applied: {What changed}
Validation: {Test/verification performed}
Result: {Pass/Fail with evidence}
Side effects: {Any new issues observed}

Debugging Workflow Example

Scenario: Agent fails to navigate to worktree

Phase 1: Root Cause

Error: bash: cd: _worktrees/my-feature: No such file or directory

Root Cause: Relative path check succeeded (from repo root),
but cd failed (running from inside worktree).

Evidence:
- test -d "_worktrees/my-feature" returned success
- cd _worktrees/my-feature failed
- pwd showed /home/user/project/_worktrees/other-feature

The agent is already INSIDE a different worktree, so relative
path doesn't exist from there.

Phase 2: Pattern

Pattern: Path confusion — relative vs absolute paths
Similar cases: Common issue when shell CWD isn't repo root
Related: Always use absolute paths for worktree navigation

Phase 3: Hypothesis

Hypothesis: Using absolute paths from git worktree list will
work from any location, unlike relative paths.

Test: Replace relative check with:
  path=$(git worktree list | grep "my-feature" | awk '{print $1}')
  cd "$path"

Expected outcome: cd succeeds regardless of current location

Phase 4: Fix

Fix applied: Updated to use absolute paths from git worktree list

Validation: Ran from three locations:
  - Repo root: ✅ navigated successfully
  - Different worktree: ✅ navigated successfully
  - Target worktree: ✅ detected "already there"

Result: PASS — all scenarios work
Side effects: None observed

When to Use This Pattern

Always use for:

• Test failures (don't guess why test failed)
• Runtime errors (understand before fixing)
• Build failures (trace compilation errors)
• Integration issues (API/contract mismatches)

Optional for:

• Typos or obvious mistakes (fix directly)
• Linter warnings (follow linter suggestion)
• Documentation updates (no debugging needed)

Integration with Forge Workflows

During implementation (implementation-worker agent):

• When tests fail, document debugging phases in commit message or PR
• Don't claim "fixed" without validation phase evidence
• Reference which phase revealed the solution

During code review (code-reviewer agent):

• Check if PRs include validation evidence for fixes
• Verify root cause is addressed, not just symptoms

In retrospectives:

• Track patterns that appear repeatedly
• Update context/guidelines when new patterns emerge
• Improve error messages based on root cause insights

Exceptions and Pragmatic Use

When to skip systematic debugging:

• Error message is self-explanatory ("file not found")
• Fix is one-line trivial change
• Debugging would take longer than trying obvious fix

When to use lightweight version:

• For simple errors: Root Cause + Fix (skip Pattern and Hypothesis)
• For known patterns: Pattern + Fix (leverage knowledge)

Full 4-phase process recommended for:

• Mysterious errors without obvious cause
• Errors that reappear after attempted fixes
• System integration failures (multi-component)

Documenting Debugging Sessions

When debugging reveals important insights, document them:

In commit messages:

fix(agent): Prevent worktree navigation failures

Root cause: Agents used relative paths which only work from
repo root.
Pattern: Path confusion (relative vs absolute).
Fix: Use absolute paths from git worktree list.
Validated: Tested from repo root, other worktree, and target.

In PR descriptions:

## Debugging Summary
Root cause: {brief}
Fix: {what changed}
Validation: {tests run, output observed}

Summary

Key principle: Understand before acting. Evidence before claiming success.