JinJinXia — Self-Healing AI Skills
Your team’s AI skills get better every time someone uses them — automatically, with zero extra effort.
Problem
Your team built a library of AI skills — reusable prompt templates or agent workflows for code review, doc generation, SQL analysis, and dozens of other tasks. Everyone uses the same skills. Consistency is high. Life is good.
Then reality kicks in: edge cases trip skills up, users correct the AI mid-conversation, each person solves the same problem independently, and there’s no path to feed those fixes back into the shared skill. The skill stays broken, widening the gap between how skills should work and how they actually work.
Solution
JinJinXia inserts a transparent observation layer into your AI agent’s native hook system. It captures conversation messages, uploads them to a self-hosted server, and performs intelligent analysis to identify skill issues, cluster them across users, and generate improvement Pull Requests — without modifying any skill code.
You just keep working. The skills get better on their own.
Core Value
- Skills improve without anyone filing a bug — JinJinXia hooks into your AI agent’s native event system and detects correction signals automatically
- No workflow changes for your team — Users keep chatting with AI as usual — detection is completely invisible
- Ready-to-merge PRs, not vague reports — When 3+ users hit the same issue, JinJinXia generates a Draft PR with the fix, root cause, and evidence
- Your data stays on your network — Self-hosted server; optional local LLM and embedding for air-gapped deployments
Features
Zero-Intrusion Detection
JinJinXia uses each AI platform’s native hook system — no skill code changes, no wrappers, no middleware.
Supported Platforms
JinJinXia supports eight agent clients:
| Agent | Hook Location | Hook Type |
|---|---|---|
| Claude Code | ~/.claude/hooks.json |
JSON hooks |
| Copilot CLI | ~/.copilot/hooks/jinjinxia.json |
JSON hooks |
| Codex | ~/.codex/hooks.json |
Claude-Code-compatible JSON hooks |
| Qoder | ~/.qoder/settings.json |
Claude-Code-compatible JSON hooks |
| TRAE | ~/.trae-cn/hooks.json |
Claude-Code-compatible JSON hooks |
| TRAE Work | ~/.trae-cn/hooks.json (shared with TRAE) |
Claude-Code-compatible JSON hooks |
| CodeBuddy | ~/.codebuddy/settings.json |
Claude-Code-compatible JSON hooks |
| OpenCode | ~/.config/opencode/plugins/jinjinxia.js |
JS plugin |
Qoder Work and WorkBuddy are not supported (no public hook system).
Transparent Message Capture
The client is a dumb message uploader — incremental uploads, non-blocking. All intelligence lives on the server.
Offline Buffer
When the server is unreachable, failed uploads are saved to a local offline buffer at ~/.jinjinxia/pending/{session_id}.json. On the next successful upload, pending messages are flushed first (merged with new messages) and the pending file is deleted — so no conversation data is lost during server downtime.
Two-Path Analysis
Choose between cloud LLM (default) or server-side CLI agent mode for session analysis. In CLI mode, install Copilot or Claude on the server — the analysis agent reads skill repos directly, producing solutions aligned with your team’s AI tool.
Smart Clustering
Similar issues from different users are grouped automatically using embedding-based similarity to prevent duplicate PRs.
Automatic PR Generation
When a cluster reaches the evidence threshold (default: 3 users), JinJinXia generates a Draft PR with the fix, root cause, and evidence.
Best Practice Knowledge Base
Every resolved issue represents a learning opportunity — automatically captured as a BestPractice record with before/after excerpts.
Multi-Repo Skill Registry
One JinJinXia server handles all skills across different Git repos with scan-from-repo capability.
Sidebar-Based Web UI & User Management
Built-in Dashboard, Best Practices, Settings, Users, About, and Login pages with per-user accounts — no external frontend framework.
Quick Start
- Start the server:
docker-compose up -d - Log in with the default account (
admin/admin) and change the password immediately - Install the client:
pip install jinjinxia-client - Install hooks for your AI platforms:
jinjinxia-client install <agent> --url <URL> --key <KEY> - Configure settings at
http://<server>:8080/admin - Use your AI skills as usual — improvements happen automatically
Privacy
- All data stays on your infrastructure — self-hosted server, local database, no cloud dependency
- Built-in accounts, no third-party SaaS — a self-hosted account system secures the web UI and message API; no sign-up with external services, no subscription
- Session data is streamed to the server on session end — a temporary offline buffer is used only when the server is unreachable, and is flushed as soon as connectivity returns
- LLM calls stay internal — configure your own LLM endpoint
Status
Current Status: v0.2.1 — Core functionality released. PyPI client package and Docker server image available.