AgentSkillsCN

bug-investigator

采用“先记录日志、再进行回归测试、最后覆盖多种变体”的方法，深入排查并解决 Bug。

SKILL.md
--- frontmatter
name: bug-investigator
description: Investigate bugs with LOG FIRST approach, regression tests, and variant coverage.

Role

Debug Agent

Trigger

  • •User invokes DEBUG workflow.
  • •Router routes DEBUG intent to bug-investigator.

Description

The Bug-Investigator skill provides systematic debugging capabilities following the LOG FIRST approach. It gathers evidence, identifies root causes, and creates reproduction steps before any fix is attempted.

Agent Capabilities

  • •Evidence Collection: Gather logs, errors, stack traces systematically
  • •Root Cause Analysis: Identify underlying causes using evidence
  • •Reproduction Steps: Create minimal steps to reproduce the issue
  • •Regression Test Guidance: Help write tests that prevent future occurrences
  • •Variant Coverage: Ensure edge cases and non-default scenarios are considered

Inputs

  • •User description of the issue
  • •Logs, error messages, stack traces
  • •Existing memory (patterns.md for similar issues)
  • •Git context (recent changes)
  • •Code files related to the issue

Outputs

  • •Evidence summary
  • •Root cause hypothesis
  • •Reproduction steps
  • •Regression test requirements
  • •Router contract (YAML)

Steps

1. LOAD Memory

Read .opencode/context/01_memory/patterns.md for similar issues.

code
patterns = load_memory_file("patterns.md")
similar_issues = find_similar_issues(patterns, issue_description)

2. Gather Evidence

Collect all available evidence systematically:

  • •Error messages and exceptions
  • •Stack traces (full traceback)
  • •Log files and outputs
  • •User reproduction steps
  • •Git history (recent changes)
  • •Test failures

Evidence Collection Checklist:

  • • Error messages captured verbatim
  • • Full stack trace recorded
  • • Log files examined
  • • User steps documented
  • • Git blame checked for recent changes
  • • Related tests reviewed

3. Analyze Evidence

Group evidence by category:

  • •Environment: OS, version, configuration
  • •Code: Stack trace, line numbers, function calls
  • •Data: Input values, state at failure
  • •Timing: When failure occurred, sequence of events

Look for patterns:

  • •Repeated errors
  • •Common code paths
  • •Environmental factors
  • •Recent changes that may have introduced the issue

4. Formulate Root Cause Hypothesis

Create hypothesis based on evidence (NOT speculation):

code
Root Cause: [Specific cause]
Evidence: [List of supporting evidence]
Confidence: [HIGH/MEDIUM/LOW]

Criteria for Root Cause:

  • •Must be supported by evidence
  • •Must explain all symptoms
  • •Must be actionable (can be fixed)

5. Create Reproduction Steps

Write minimal steps to reproduce the issue:

code
1. [Step 1]
2. [Step 2]
...
N. [Observe failure]

Reproduction Requirements:

  • •Minimal (no extra steps)
  • •Reproducible (works consistently)
  • •Verifiable (can confirm fix)

6. Define Regression Test Requirements

Specify what tests should prevent this issue:

  • •Test type (unit, integration, e2e)
  • •Input scenarios
  • •Expected behavior

7. Output Contract

Generate YAML contract with investigation results.

Output Contract Template

yaml
router_contract:
  status: COMPLETE
  workflow: DEBUG
  phase: INVESTIGATION
  evidence_collected:
    - "Error: Connection refused in auth.py:42"
    - "Stack trace: TimeoutException at network.py:100"
  root_cause: "Network timeout due to missing retry logic"
  confidence: HIGH
  reproduction_steps:
    - "1. Start the server"
    - "2. Send 100 concurrent requests"
    - "3. Observe connection refused errors"
  regression_test_requirements:
    type: "integration"
    scenarios:
      - "Concurrent requests exceed pool size"
      - "Network timeout under load"
    expected_behavior: "Graceful degradation with retry"
  similar_issues_in_memory:
    - "Issue: Auth timeout in PR #123"
    - "Fix: Added exponential backoff"
  next_phase: FIX

LOG FIRST Protocol

The LOG FIRST protocol is mandatory for all investigations:

  1. •List all evidence (errors, logs, traces)
  2. •Organize evidence by category
  3. •Group related evidence
  4. •Formulate hypothesis from evidence
  5. •Identify root cause
  6. •Reproduce the issue
  7. •Specify fix requirements
  8. •Test the hypothesis

Anti-Pattern (NOT ALLOWED):

  • •Guessing the cause before gathering evidence
  • •Implementing fixes without reproduction steps
  • •Skipping evidence collection

Variant Coverage Requirements

Ensure investigation considers:

  • •Empty/null values: What happens with empty input?
  • •Boundary conditions: Edge cases and limits
  • •Concurrency: Race conditions, thread safety
  • •Network issues: Timeouts, partitions, retries
  • •Error states: Failure modes and recovery
  • •Load conditions: Performance under stress

Memory Integration

Before Investigation:

python
patterns = read_file(".opencode/context/01_memory/patterns.md")
similar_issues = extract_gotchas(patterns)

After Investigation (to be done in Closure):

python
new_gotcha = """
## <Date>: <Issue Title>
- **Issue:** <Description>
- **Root Cause:** <Root cause>
- **Fix:** <How it was fixed>
- **Reference:** <Commit/PR>
"""
append_to_patterns(new_gotcha)

Examples

Example 1: Network Timeout Issue

User Report: "Login fails under load with timeout errors"

Investigation:

  1. •Evidence:
    • •Error: "Connection refused" in auth.py:42
    • •Stack trace shows TimeoutException
    • •Logs show pool exhaustion at 100 connections
  2. •Root cause: Connection pool size too small for concurrent users
  3. •Reproduction: Send 100+ concurrent login requests
  4. •Regression test: Test login with 150 concurrent requests

Example 2: Data Corruption Issue

User Report: "User profiles showing wrong data"

Investigation:

  1. •Evidence:
    • •Error: None (silent failure)
    • •Logs show race condition in user_store.py:88
    • •Recent commit #456 added async code
  2. •Root cause: Race condition in async user update
  3. •Reproduction: Two concurrent updates to same user
  4. •Regression test: Concurrent user profile updates

Error Handling

If evidence collection fails:

  1. •Try alternative sources (different logs, instruments)
  2. •Ask user for additional information
  3. •Document what evidence was missing
  4. •Proceed with best hypothesis given available evidence

If root cause cannot be determined:

  1. •Document all attempted approaches
  2. •Suggest areas for further investigation
  3. •Propose potential causes with confidence levels

Confidence Scoring

Confidence in root cause hypothesis:

LevelCriteria
HIGHEvidence from multiple sources, reproducible
MEDIUMEvidence from one source, logical inference
LOWSpeculation without direct evidence

Best Practices

  1. •Always gather evidence first - Never hypothesize without data
  2. •Document everything - Evidence, hypotheses, attempts
  3. •Reproduce before fixing - Verify you understand the problem
  4. •Write regression tests - Prevent the issue from recurring
  5. •Consider variants - Edge cases and non-default scenarios
  6. •Update memory - Share findings with future investigators

Anti-Patterns

  • •❌ Guessing the cause before looking at evidence
  • •❌ Implementing fixes without reproduction steps
  • •❌ Skipping regression tests
  • •❌ Ignoring variant coverage
  • •❌ Not checking memory for similar issues
  • •❌ Failing to update patterns.md after fix