bug-hunt（寻找虫子）

Bug Hunt Skill

Quick Ref: 4-phase investigation (Root Cause → Pattern → Hypothesis → Fix). Output: .agents/research/YYYY-MM-DD-bug-*.md

YOU MUST EXECUTE THIS WORKFLOW. Do not just describe it.

Systematic investigation to find root cause and design a complete fix — or proactive audit to find hidden bugs before they bite.

Requires:

• session-start.sh has executed (creates .agents/ directories for output)
• bd CLI (beads) for issue tracking if creating follow-up issues

Modes

Investigation mode uses the 4-phase structure below. Audit mode uses systematic read-and-classify — see Audit Mode.

The 4-Phase Structure (Investigation Mode)

For failure category taxonomy and the 3-failure rule, read skills/bug-hunt/references/failure-categories.md.

Execution Steps

Given /bug-hunt <symptom>:

Step 0: Load Prior Bug Context

Before investigating, check for prior learnings about this area of the codebase:

if command -v ao &>/dev/null; then
    ao lookup --query "<symptom-keywords> bug patterns" --limit 3 2>/dev/null || true
fi

Apply retrieved knowledge: If learnings are returned, check each for applicability to the current bug. For applicable learnings (e.g., prior bugs in same area, known fragile patterns), include as investigation leads and cite by filename. Record: ao metrics cite "<path>" --type applied 2>/dev/null || true

Prior bug reports, fix patterns, and known fragile areas reduce investigation time.

Phase 1: Root Cause Investigation

Step 1.1: Confirm the Bug

First, reproduce the issue:

• What's the expected behavior?
• What's the actual behavior?
• Can you reproduce it consistently?

Read error messages carefully. Do not skip or skim them.

If the bug can't be reproduced, gather more information before proceeding.

Step 1.2: Locate the Symptom

Find where the bug manifests:

# Search for error messages
grep -r "<error-text>" . --include="*.py" --include="*.ts" --include="*.go" 2>/dev/null | head -10

# Search for function/variable names
grep -r "<relevant-name>" . --include="*.py" --include="*.ts" --include="*.go" 2>/dev/null | head -10

Step 1.3: Git Archaeology

Find when/how the bug was introduced:

# When was the file last changed?
git log --oneline -10 -- <file>

# What changed recently?
git diff HEAD~10 -- <file>

# Who changed it and why?
git blame <file> | grep -A2 -B2 "<suspicious-line>"

# Search for related commits
git log --oneline --grep="<keyword>" | head -10

Step 1.4: Trace the Execution Path

USE THE TASK TOOL (subagent_type: "Explore") to trace the execution path:

• Find the entry point where the bug manifests
• Trace backward to find where bad data/state originates
• Identify all functions in the path and recent changes to them
• Return: execution path, likely root cause location, responsible changes

Step 1.5: Identify Root Cause

Based on tracing, identify:

• What is wrong (the actual bug)
• Where it is (file:line)
• When it was introduced (commit)
• Why it happens (the logic error)

Phase 2: Pattern Analysis

Step 2.1: Find Working Examples

Search the codebase for similar functionality that WORKS:

# Find similar patterns
grep -r "<working-pattern>" . --include="*.py" --include="*.ts" --include="*.go" 2>/dev/null | head -10

Step 2.2: Compare Against Reference

Identify ALL differences between:

• The broken code
• The working reference

Document each difference.

Phase 3: Hypothesis and Testing

Step 3.1: Form Single Hypothesis

State your hypothesis clearly:

"I think X is wrong because Y"

One hypothesis at a time. Do not combine multiple guesses.

Step 3.2: Test with Smallest Change

Make the SMALLEST possible change to test the hypothesis:

• If it works → proceed to Phase 4
• If it fails → record failure, form NEW hypothesis

Step 3.3: Check Failure Counter

Check failure count per skills/bug-hunt/references/failure-categories.md. After 3 countable failures, escalate to architecture review.

Phase 4: Implementation

Step 4.1: Design the Fix

Before writing code, design the fix:

• What needs to change?
• What are the edge cases?
• Will this fix break anything else?
• Are there tests to update?

Step 4.2: Create Failing Test (if possible)

Write a test that demonstrates the bug BEFORE fixing it.

Step 4.3: Implement Single Fix

Fix at the ROOT CAUSE, not at symptoms.

Step 4.4: Verify Fix

Run the failing test - it should now pass.

If the bug is in a high-complexity function, consider /refactor after fix to prevent recurrence.

Audit Mode

When invoked with --audit, bug-hunt switches to a proactive sweep. No symptom needed — you're hunting for bugs that haven't been reported yet.

/bug-hunt --audit cli/internal/goals/     # audit a package
/bug-hunt --audit src/auth/               # audit a directory
/bug-hunt --audit .                        # audit recent changes in repo

Audit Step 1: Scope

Identify target files from the scope argument:

# Find source files in scope
find <scope> -name "*.go" -o -name "*.py" -o -name "*.ts" -o -name "*.rs" | head -50

If scope is . or broad (>50 files), narrow to recently changed files:

git log --since="2 weeks ago" --name-only --pretty=format: -- <scope> | sort -u | head -30

Audit Step 2: Systematic Read

Read every file in scope line by line. For each file, check:

Key discipline: Read line by line. Do not skim. The proven methodology (5 bugs found, 0 hypothesis failures) came from careful reading, not heuristic scanning.

USE THE TASK TOOL (subagent_type: "Explore") for large scopes — split files across parallel agents.

Audit Step 3: Classify Findings

For each finding, assign severity:

Performance bugs (slow queries, memory leaks, N+1) → escalate to /perf for deeper analysis.

Audit Step 4: Write Audit Report

For audit report format, read skills/bug-hunt/references/audit-report-template.md.

Write to .agents/research/YYYY-MM-DD-bug-<scope-slug>.md.

Report to user with a summary table:

| # | Bug | Severity | File | Fix |
|---|-----|----------|------|-----|
| 1 | <description> | HIGH | <file:line> | <proposed fix> |

Include failure count (hypothesis tests that didn't confirm). Zero failures = clean audit.

Bug-Finding Pyramid Modes (BF1–BF5)

When running --audit, check for missing bug-finding test coverage:

BF4 — Chaos/Negative Testing (highest bug-finding power): For every file that makes external calls (APIs, databases, filesystems), verify:

• Timeout injection test exists
• Connection failure test exists
• Permission denied test exists
• Corrupt input test exists

If any boundary lacks failure injection → flag as finding (severity: significant).

BF5 — Script Functional Testing: For every.sh script that calls external tools (oc, kubectl, helm):

• Stub-based functional test exists
• JSON output schema validated
• Both healthy and unhealthy stub paths tested

If scripts lack functional tests → flag as finding (severity: moderate).

BF1 — Property-Based Testing: For every data transformation (parse/render/serialize):

• Property test with randomized inputs exists

Reference: the test pyramid standard in /standards for full BF level definitions and per-language tooling.

Step 5: Write Bug Report

For bug report template, read skills/bug-hunt/references/bug-report-template.md.

Step 6: Report to User

Tell the user:

1. Root cause identified (or not yet)
2. Location of the bug (file:line)
3. Proposed fix
4. Location of bug report
5. Failure count and types encountered
6. Next step: implement fix or gather more info

Key Rules

• Reproduce first - confirm the bug exists
• Use git archaeology - understand history
• Trace systematically - follow the execution path
• Identify root cause - not just symptoms
• Design before fixing - think through the solution
• Document findings - write the bug report

Quick Checks

Common bug patterns to check:

• Off-by-one errors
• Null/undefined handling
• Race conditions
• Type mismatches
• Missing error handling
• State not reset
• Cache issues

Examples

Investigating a Test Failure

User says: /bug-hunt "tests failing on CI but pass locally"

What happens:

1. Agent confirms bug by checking CI logs vs local test output
2. Agent uses git archaeology to find recent changes to test files
3. Agent traces execution path to identify environment-specific differences
4. Agent forms hypothesis about missing environment variable
5. Agent creates failing test locally by unsetting the variable
6. Agent implements fix by adding default value
7. Bug report written to .agents/research/2026-02-13-bug-test-failure.md

Result: Root cause identified as missing ENV variable in CI configuration. Fix applied and verified.

Tracking Down a Regression

User says: /bug-hunt "feature X broke after yesterday's deployment"

What happens:

1. Agent reproduces issue in current state
2. Agent uses git log --since="2 days ago" to find recent commits
3. Agent uses git bisect to identify exact breaking commit
4. Agent compares broken code against working examples in codebase
5. Agent forms hypothesis about introduced type mismatch
6. Agent implements minimal fix and verifies with existing tests
7. Bug report documents commit sha, root cause, and fix

Result: Regression traced to commit abc1234, type conversion error fixed at root cause in validation logic.

Proactive Code Audit

User says: /bug-hunt --audit cli/internal/goals/

What happens:

1. Agent scopes to all .go files in the goals package
2. Agent reads each file line by line, checking for resource leaks, string safety, dead code, etc.
3. Agent finds 5 bugs: zombie process groups (HIGH), UTF-8 truncation (MEDIUM), hardcoded paths (MEDIUM), lost paragraph breaks (LOW), dead branch (LOW)
4. All findings confirmed on first pass — 0 hypothesis failures
5. Audit report written to .agents/research/2026-02-24-bug-goals-go.md

Result: 5 concrete bugs with severity, file:line, and proposed fix — ready for implementation without debugging.

Troubleshooting

Reference Documents

• references/audit-report-template.md
• references/bug-report-template.md
• references/failure-categories.md

See Also

• refactor — Safe refactoring for complex bug-prone code
• perf — Performance profiling for performance-related bugs