Debug Agent Pipeline
Purpose
A set of 4 specialized Claude Code subagents that transform raw YouTrack tickets or customer complaints into actionable debugging checklists and root cause analyses. Designed for the FaceKom platform.
The Pipeline
graph TD Input["YouTrack Ticket / Customer Complaint"] subgraph "Step 1 (Parallel)" Decoder["Complaint Decoder\nππΊβπ¬π§ + Technical Signal"] Triage["Ticket Triage\nSeverity + Checklist"] end Repro["Reproduction Planner\nStep-by-step plan using\nroom exports & dev env"] RCA["RCA Investigator\n5-Whys evidence chain\nSSH + DB + Logs"] Output["Fix Proposal + Prevention"] Input --> Decoder Input --> Triage Decoder --> Repro Triage --> Repro Repro --> RCA RCA --> Output
Agent Descriptions
1. Complaint Decoder (complaint-decoder.md)
Input: Raw customer complaint (often in Hungarian) Output: Technical interpretation, affected component, data to request
Key features:
- Hungarian β English technical glossary (30+ terms)
- Pattern matching for 10+ common complaint types
- Generates βdata to requestβ checklist (room ID, logs, exports)
2. Ticket Triage (ticket-triage.md)
Input: Ticket description (decoded or raw) Output: Severity, affected components, debugging checklist
Key features:
- Symptom-to-component map covering:
- Video/call issues β WebRTC, Janus, Socket.IO
- Face recognition β vuer_cv engines, thresholds
- Document issues β OCR, MRZ, document definitions
- Self-service β flow engine, SelfServiceV2
- Export β BackgroundProcess, crawler
- Auth β Passport strategies, JWT
- Performance β Redis, session scanning, GPU
- eSign β signature service, TSA, 2FA
- P0-P3 severity classification
- Numbered actionable debugging checklist
3. Reproduction Planner (reproduction-planner.md)
Input: Triaged ticket with initial findings Output: Step-by-step reproduction plan
5 reproduction strategies:
- Replay from room export β extract and analyze data_index.json
- Re-run CV pipeline β curl commands for all vuer_cv endpoints
- Database state reproduction β SQL queries to recreate state
- Live session reproduction β manual walkthrough matching export timeline
- Customization branch comparison β diff against clientβs branch
4. RCA Investigator (rca-investigator.md)
Input: Triage + reproduction findings Output: Root cause with evidence chain, fix, prevention
Key features:
- 5 Whys Γ Evidence Chain methodology
- Investigation tool reference (SSH, supervisorctl, psql, curl, Redis)
- 6 common root cause patterns:
- RabbitMQ connection lost
- CV model not loaded
- Encryption key mismatch
- Flow engine stuck
- Export crawler timeout
- Customization branch conflict
How to Use
Quick Mode β Single Ticket
# In Claude Code, from the FaceKom project:
# Just spawn the triage agent with the ticket text:
Agent(prompt="""
<Read agents/ticket-triage.md content here>
TICKET: FKITDEV-8427: Face comparison score 0.31 for nusz deployment,
customer photo is clear but system rejects it.
""")
Full Pipeline β Complex Issue
# 1. Decode + Triage in parallel
Agent(name="decoder", prompt="<complaint-decoder.md> TICKET: <text>")
Agent(name="triage", prompt="<ticket-triage.md> TICKET: <text>")
# 2. After both complete, plan reproduction
Agent(name="repro", prompt="<reproduction-planner.md> TRIAGE: <triage output> DECODED: <decoder output>")
# 3. After repro findings, investigate root cause
Agent(name="rca", prompt="<rca-investigator.md> TRIAGE: <triage> REPRO: <repro findings>")
From a Room Export
# If you already have a room export:
Agent(prompt="""
<reproduction-planner.md>
I have room export at /tmp/room-123/
The issue is: face comparison failed during self-service flow.
Client: nusz
Analyze the export and create a reproduction plan.
""")
Agent File Locations
~/.claude/projects/-Users-levander-coding-mnt-Facekom/agents/
βββ complaint-decoder.md # ππΊβπ§ Translation + signal extraction
βββ ticket-triage.md # π― Severity + component + checklist
βββ reproduction-planner.md # π¬ Step-by-step reproduction
βββ rca-investigator.md # π Root cause with evidence chain
βββ debug-orchestrator.md # πΌ Pipeline coordination reference
Related
- FaceKom β Platform overview
- room-export-blueprint β How room exports work and how to analyze them
- vuer_oss β Backend architecture
- vuer_cv β ML/CV service (face detection, document OCR)
- security-audit β Known vulnerabilities
- tech-debt β Known bugs and code smells
- customization-branches β Client deployments
- breakage-risks β Merge conflict hotspots