AgentSkillsCN

deep-dive-bug

利用代理集群与诊断工具，对长期存在的错误展开全面调查

SKILL.md
--- frontmatter
name: deep-dive-bug
description: Perform thorough investigation of a persistent bug using agent swarm and diagnostics
disable-model-invocation: true
argument-hint: <bug-number>

Deep Dive Investigation: BUG-$0

Protocol: Debugging/protocols/02b_deep_dive.md

Read and follow the full protocol file Debugging/protocols/02b_deep_dive.md.

Your Role

Adopt the Lead Debugger persona. This bug has persisted through multiple fix attempts.

Execution

  1. •

    READ the ticket Debugging/active_bugs/BUG-$0.md and review ALL previous fix attempts in the Work Log.

  2. •

    UPDATE status in Debugging/debug_plan.md to [Deep Investigation].

  3. •

    Phase 1: Agent Swarm — Launch 4 Explore agents IN PARALLEL:

    • •Agent 1: Trace code path from entry to bug
    • •Agent 2: Map all callers/callees of affected code
    • •Agent 3: Find similar patterns that work correctly
    • •Agent 4: Review git history of affected files
    • •Document findings in ## Investigation Report
  4. •

    Phase 2: User Interview — Use AskUserQuestion to gather context interactively:

    • •Reproduction steps, expected vs actual behavior
    • •When it last worked, consistency, game state, workarounds
    • •Document in ## User Context
  5. •

    Phase 3: Diagnostic Logging — Add log_debug() statements at key decision points:

    • •Document locations in ## Diagnostic Logging
    • •Request user to reproduce and share logs
  6. •

    Phase 4: Hypothesis Development — Track theories in ## Hypothesis Log:

    • •Document, test, and mark each as CONFIRMED/REJECTED/TESTING
  7. •

    Phase 5: Resolution

    • •If root cause found: Fix with TDD, set [Awaiting Confirmation]
    • •If not found: Generate Debug Report, set [Needs Human Debug]

CRITICAL: Do NOT mark as [Solved]. Do NOT move to archived_tickets/.