Jintech OMG Dev — MCP Server de Revisão de Código
ahsan0444/jintech-omg-dev
Pipeline SDLC completo com grafo semântico e enforcement automático de convenções
Ver no GitHubServidor MCP que integra um pipeline completo de SDLC com skills de investigação, implementação e auditoria de código. Inclui um grafo semântico de 9k+ nós do codebase OMG, enforcement hooks automáticos, e agentes especializados para roteamento inteligente de tickets e implementação com TDD.
Instalação
npx convez-ai ahsan0444/jintech-omg-dev1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Auditoria automática de PRs antes de merge com Perl::Critic e convenções de camadas
- Investigação e rastreamento de bugs com análise de hipóteses e roteamento inteligente
- Geração de planos de desenvolvimento validados contra convenções OMG e TDD
Veredito

Ferramenta robusta e bem-estruturada para times que usam OMG SDLC, com automação real de enforcement e grafo semântico do codebase. Exige configuração substantiva (Perl, Bitbucket, Jira) e é sensível ao contexto específico do OMG — não é um MCP genérico, mas sim especializado em seu ecossistema.
README
A tradução para o português deste README ainda está na fila.
jintech-omg-dev
Claude Code plugin for Jintech OMG development. Bundles the full SDLC skill pipeline, the code-review-graph MCP server, and enforcement hooks into a single installable unit.
Supports macOS, Linux, and Windows (CI-tested on ubuntu and windows).
What's included
| Component | What it does |
|---|---|
/ticket | Investigates a Jira ticket via subagents, produces a Plan Mode proposal |
/implement | Executes an approved plan via subagents with TDD and layer compliance |
/prepr | Pre-PR audit: Perl::Critic, OMG layer conventions, TT/JS/SQL/SCSS |
/pr | Creates a Bitbucket draft PR, posts PR link back to Jira |
/grill-me | Stress-tests a design/plan before /ticket runs |
/debug | Root cause analysis — traces, ranks hypotheses, routes to /implement or /ticket |
code-review-graph MCP | Semantic graph of 9k+ nodes, 78k+ edges across the OMG codebase — 14 tools wired into skills |
| Hooks | skill-router (UserPromptSubmit), enforce-mcp-search + enforce-skill-usage (PreToolUse), post-edit-update (PostToolUse), session-start-status (SessionStart) — all registered automatically via hooks/hooks.json |
| Commands | /learn, /save-session, /resume-session, /embed-graph, /router-stats |
| Routing manifest | skill-routing-manifest.json — regex intents → skills/inline procedures (see Hooks below) |
| Agents | omg-investigator (read-only locator — no Read/Edit/Write tools, MCP-graph-first) and omg-implementer (single-step executor — OMG layer rules + TDD baked into its system prompt). Skills spawn these instead of generic Explore/general-purpose, turning "no file reads in subagents" from a prompt rule into a tool permission. |
Prerequisites
| Requirement | macOS | Windows |
|---|---|---|
| Claude Code (latest stable) | ✅ | ✅ |
| Python 3.12+ (3.14 recommended) | ✅ | ✅ — install from python.org or Microsoft Store; ensure python3 is on PATH. code-review-graph requires Python 3.12 minimum — Python 3.9 will fail at install. |
| Git | ✅ | ✅ — Git for Windows (includes Git Bash) |
| Perl + Perl::Critic | ✅ | ✅ — Strawberry Perl recommended |
| Bitbucket credentials | ✅ | ✅ |
| Jira/Atlassian MCP | ✅ | ✅ |
Windows note: All hook scripts and the CRG server are written in Python — no bash required.
python3must be on your PATH (verify withpython3 --versionin a terminal).
Installation
1. Add the Jintech plugin marketplace
/plugin marketplace add https://github.com/ahsan0444/claude-marketplace
2. Install this plugin
/plugin install jintech-omg-dev
This will:
- Copy skills and commands into your Claude Code installation
- Register the
code-review-graphMCP server (auto-installs via pip on first use) - Wire up the lifecycle hooks
3. Set required environment variables
macOS / Linux — add to ~/.zshrc or ~/.bashrc
export BITBUCKET_USER="your-bitbucket-username"
export BITBUCKET_TOKEN="your-bitbucket-app-password"
Then run source ~/.zshrc.
Security: Use a Bitbucket app password scoped to Pull Requests read/write only — never your account password. The plugin's curl calls pass credentials via stdin (
curl -K -), so the token never appears in the process list. Rotate the token if a session transcript is ever shared.
Windows — add to your PowerShell profile ($PROFILE) or System Environment Variables
$env:BITBUCKET_USER = "your-bitbucket-username"
$env:BITBUCKET_TOKEN = "your-bitbucket-app-password"
To persist across sessions, set them as System Environment Variables:
Start → "Edit the system environment variables" → Environment Variables → New
Optional environment overrides (all platforms)
Defaults match the standard Jintech setup — set these only if your machine differs. This is the single place to relocate the workspace; no skill files need editing.
export OMG_WORKSPACE_ROOT="/Users/Shared/Code" # parent dir of omg/omg_db/omg_ice checkouts (Windows Git Bash: /c/Code)
export OMG_BITBUCKET_WORKSPACE="zlalani" # Bitbucket Cloud workspace slug
export CLAUDE_SKILL_ROUTER_DISABLED=1 # kill switch for prompt routing (unset = enabled)
4. Initialise the code-review-graph (first time per repo)
Navigate to your OMG checkout and run:
macOS
cd /Users/Shared/Code/omg
code-review-graph register . --alias omg # register the repo with CRG
code-review-graph build # parse all files (~5-10 min)
Windows
cd C:\Code\omg # adjust to your checkout path
code-review-graph register . --alias omg
code-review-graph build
Then enable semantic search (one-time, ~2-5 min):
After the build completes, restart Claude Code (so the MCP server picks up the new graph), then run:
/embed-graph
This generates vector embeddings locally using all-MiniLM-L6-v2 (no API key required). Without this step, MCP search falls back to keyword matching only.
Auto-embed: The plugin detects missing embeddings on startup and triggers embedding in the background automatically. If the
/embed-graphcommand reports "already embedded", no action needed.
Subsequent updates run automatically after each file edit via the PostToolUse hook.
Usage
Skills are invoked by short name within an active OMG project session:
/grill-me OMGXI-1234 # optional pre-investigation alignment
/ticket OMGXI-1234 # investigate + produce approved plan
/implement # execute the approved plan
/prepr # pre-PR audit
/prepr fix # audit + auto-fix warnings
/pr # create Bitbucket draft PR
/debug OMGXI-1234 # root cause analysis
Or with the fully-qualified plugin namespace from any directory:
/jintech-omg-dev:ticket OMGXI-1234
Pipeline
/grill-me (optional, Session 1)
↓ saves .planning/grill-TICKET.md
/ticket (Session 2)
↓ saves .planning/approved-plan-TICKET.md
/implement (Session 3)
↓ deletes approved-plan after completion
/prepr (Session 4)
↓ no blockers
/pr (Session 5)
Each skill is a fresh session. See WORKFLOW.md for full architecture diagrams.
Code-review-graph
The MCP server starts automatically when the plugin activates. It self-installs code-review-graph from PyPI into a plugin-local venv on first run — no manual pip install needed.
It requires a built graph at .code-review-graph/graph.db in your project root (Step 4 above).
Manual commands (run from inside your OMG checkout):
code-review-graph status # check graph health and node count
code-review-graph update # incremental update (changed files only)
code-review-graph build # full rebuild (~5–10 min)
Auto-rebuild schedule:
| Platform | Mechanism | Schedule |
|---|---|---|
| macOS | launchd (com.omg.graphify-rebuild.plist) | Every Monday 3am |
| Windows | Task Scheduler | Set up manually — see below |
Windows: set up weekly auto-rebuild
Run once in PowerShell (Admin):
$action = New-ScheduledTaskAction -Execute "code-review-graph" -Argument "build" -WorkingDirectory "C:\Code\omg"
$trigger = New-ScheduledTaskTrigger -Weekly -DaysOfWeek Monday -At 3am
Register-ScheduledTask -TaskName "OMG-GraphRebuild" -Action $action -Trigger $trigger -RunLevel Highest
Adjust -WorkingDirectory to match your OMG checkout path.
Hooks
All hooks register automatically from hooks/hooks.json when the plugin is enabled. Every hook fails open — a broken or missing script never blocks your tools.
| Hook | Event | What it does |
|---|---|---|
skill-router | UserPromptSubmit | Matches natural-language prompts against skill-routing-manifest.json and injects a routing instruction (e.g. "create a PR" → /pr). Kill switch: export CLAUDE_SKILL_ROUTER_DISABLED=1. Override the manifest by copying it to ~/.claude/skill-routing-manifest.json or setting SKILL_ROUTER_MANIFEST_OVERRIDE. Prefix a prompt with \ to bypass routing once. |
enforce-mcp-search | PreToolUse (Grep/Bash) | Denies grep inside lib/, public/javascripts/, t/ of a repo with a built code-review-graph — directs Claude to MCP tools instead. views/ is exempt. |
enforce-skill-usage | PreToolUse (Bash) | Denies gh pr create (OMG uses Bitbucket) and points to /pr. |
post-edit-update | PostToolUse (Edit/Write) | Incremental code-review-graph update after source-file edits. Skips docs/config edits. |
session-start-status | SessionStart | Prints graph health (node count, staleness) into context at session start. |
Permissions template:
settings.jsonat the plugin root is a template — Claude Code does not load it from a plugin. Copy itspermissions.allowblock into your project's.claude/settings.jsonto pre-approve the read-only Bash and MCP calls the skills make.
Design note — router limits: regex intent matching has a permanent false-positive tail (it matches mentions, not intent — e.g. a prompt discussing the prepr skill can route to it). Run
/router-statsperiodically and tighten noisy patterns. Longer term, prefer letting Claude pick skills from theirdescriptiontriggers and keep the router only for the inline procedures that have no skill equivalent; if a routing instruction clearly doesn't fit the request, Claude is instructed (CLAUDE.md) to say so and proceed correctly rather than force-run the skill.
Updating the plugin
/plugin update jintech-omg-dev
Troubleshooting
| Problem | Fix |
|---|---|
python3: command not found (Windows) | Hooks invoke python3 literally. python.org installs ship only python.exe — create a shim once: copy "%LocalAppData%\Programs\Python\Python314\python.exe" "%LocalAppData%\Programs\Python\Python314\python3.exe" (adjust path; the Microsoft Store build already provides python3). Verify with python3 --version. |
code-review-graph: command not found | The plugin venv may not have activated — restart Claude Code (the MCP server self-installs on first start), or reinstall via /plugin → jintech-omg-dev |
code-review-graph install fails on Python 3.9/3.10/3.11 | Requires Python 3.12+. On macOS: brew install python@3.14. On Windows: download from python.org and ensure it's on PATH |
| Hook not firing | Verify hooks are registered: check .claude/settings.json in your project |
| Graph DB missing | Run code-review-graph build from inside the OMG repo root |
| MCP tools return 0 results | Ensure graph is built and registered: code-review-graph register /path/to/omg --alias omg then verify with code-review-graph status |
Contributing
Changes to skills, hooks, or the CRG server go through the standard OMG development pipeline. Raise a PR against github.com/ahsan0444/jintech-omg-dev.
For questions: #dev-tooling on Slack or raise a Jira ticket in the TOOLS project.