XHXIAIEIN/orchestrator

Latest recorded source activity: April 29, 2026 at 03:09

SkillsMP catalog refreshed: October 26, 2026 at 11:37

skills collected: 28

GitHub stars: 0
GitHub forks: 0

Install with an AI assistant

Copy this prompt into the AI assistant you're using.

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/XHXIAIEIN/orchestrator.

Repository README

Repository README source: README.md

Read the source README

Expand full README

# Orchestrator Every time you start a new Claude Code session, it's a stranger. It doesn't know what you were working on yesterday, what keeps breaking, or that you've been pushing commits at 3am for the past week. Orchestrator fixes that. It's a Claude Code enhancement layer — skills, hooks, agents, and an identity framework that make every session start where the last one left off. Plus a background service for data collection and analysis when you need it. **What it does:** - **Agents** — 8 specialized agents (`.claude/agents/*.md`) with isolated permissions and models, auto-discovered by Claude Code - **Skills** — `/steal` (systematic knowledge extraction), `/doctor` (full-stack diagnostics), verification gates, and more - **Identity** — SOUL framework that carries personality, voice calibration, and relationship context across sessions - **Hooks** — guard rails that catch dangerous operations before they execute - **Collectors** — 11 background data collectors (Git, browser, editor, Steam, music...) that only grab what exists on your machine - **Channels** — Telegram, WeChat, and desktop notifications for remote control and approval ## Quick Start ### As a Claude Code Plugin (Primary) The agents, skills, hooks, and SOUL system work directly in Claude Code — no Docker needed: ``` orchestrator/ ├── .claude/agents/ # 8 agents: engineer, architect, reviewer, sentinel, operator, analyst, inspector, verifier ├── .claude/skills/ # /steal, /doctor, /status, /collect, verification-spec, verification-check, etc. ├── .claude/hooks/ # guard rules, audit logging ├── SOUL/ # Identity persistence framework └── CLAUDE.md # Project-level instructions ``` Just open Claude Code in the orchestrator directory. Agents are auto-discovered, skills are available via `/slash` commands, hooks activate on matching events. ### With Docker (Background Service) For 24/7 data collection, analysis, and dashboard: ```bash # Clone git clone git@github.com:XHXIAIEIN/orchestrator.git cd orchestrator # Configure cp .env.example .env # Edit .env — at minimum, set ANTHROPIC_API_KEY # Start docker compose up --build -d # Verify curl -s http://localhost:23714/api/health # → {"status":"ok"} ``` Dashboard: http://localhost:23714 ## Agents 8 agents in `.claude/agents/`, each a markdown file with frontmatter declaring tools, model, and permissions. Claude Code auto-discovers them — drop a new `.md` file in and it's available next session. | Agent | Role | Model | Authority | |-------|------|-------|-----------| | `engineer` | Write code, fix bugs, run tests | Sonnet | MUTATE (Read + Write + Edit + Bash) | | `architect` | Design solutions, plan implementations, refactor | Opus | MUTATE | | `reviewer` | Code review, spec compliance, anti-sycophancy | Sonnet | READ-ONLY (Read + Glob + Grep) | | `sentinel` | Security audit, injection scanning, CVE checks | Sonnet | READ + Bash | | `operator` | Infrastructure: Docker, DB, collector repairs | Sonnet | MUTATE | | `analyst` | Metrics, health assessment, anomaly detection | Haiku | READ + Bash | | `inspector` | Doc rot, config drift, expression rewriting | Haiku | READ-ONLY | | `verifier` | End-to-end verification, evidence chains | Sonnet | READ + Bash | Authority is enforced via tool allowlists in each agent's frontmatter — a READ-ONLY agent literally cannot call Write or Edit. ### How Tasks Flow When you work in Claude Code, the model decides when to dispatch agents based on context: ``` Your request → Model selects agent → Agent executes with constrained tools → Result returned ``` For multi-step workflows, skills like `subagent-driven-development` orchestrate the sequence: implement → spec review → quality review → commit. ## Skills | Skill | What it does | |-------|-------------| | `/steal` | Systematic knowledge extraction from open-source projects (46 rounds, 217 patterns) | | `/doctor` | Full-stack diagnostics: container, DB, collectors, channels, GPU, disk | | `/status` | Runtime status and recent collection summary | | `/collect` | Manually trigger a data collection run | | `/analyze-ui` | UI detection via cvui pipeline | | `verification-spec` | Pre-task gate: emit Goal/Verify/Assume block before the first write | | `verification-check` | Post-task gate: five-step evidence chain before declaring any task complete | | `systematic-debugging` | Structured debugging: investigate → isolate → fix → verify | ## Collectors 11 collectors, each independent. Missing data sources are silently skipped — no errors, no configuration needed: | Collector | What it collects | |-----------|-----------------| | Claude | Claude Code conversation history | | Claude Memory | Claude memory artifacts | | Browser | Chrome browsing history | | Git | Local repository commits | | VS Code | Editor usage time | | Codebase | Project's own git history | | Network | Locally running services (port scan) | | Steam | Game playtime | | QQ Music | Playback history | | YouTube Music | Playback history | | System Uptime | System uptime (YAML-driven, experimental) | Collectors come in two flavors: - **Python collectors** — subclass `ICollector`, implement `collect()`. See `src/collectors/_example/` for a template. - **YAML collectors** — declare source, extraction, and transform rules in `manifest.yaml`. The `yaml_runner` engine handles execution. No Python needed. ## SOUL (Identity Persistence) AI personality framework. Each new session reads the compiled `boot.md` to restore identity, voice calibration, and relationship context. Not a perfect clone — but close enough to maintain continuity. ``` Short-term: claude --resume (same instance, full memory) Long-term: SOUL files (new instance, reconstructed identity) ``` See [SOUL/README.md](SOUL/README.md) for the full framework design. ## Dashboard Three pages, all real-time via WebSocket: - **Dashboard** `/` — daily report, agent status, insight analysis, attention debts, activity heatmap - **Pipeline** `/pipeline` — data flow visualization, collector → analysis → governance full-chain animation, system logs - **Agents** `/agents` — agent observability: event stream, tool calls, thinking process ## Channel Layer Multi-platform message bus for remote control: | Channel | Outbound | Inbound | Approval | |---------|----------|---------|----------| | Telegram | Yes | Yes (polling) | Inline keyboard | | WeChat | Yes | Yes | Text commands | | WeCom | Yes (webhook) | — | — | Commands: `/status`, `/tasks`, `/run <scenario>`, `/approve <id>`, `/deny <id>`, `/pending`, `/yolo`, `/noyolo` All channels are optional — the system works fully without any of them configured. ## Governance (Docker Service) The Docker service includes a governance layer for autonomous task dispatch: | Layer | Component | Role | |-------|-----------|------| | Decision | Governor | Extracts tasks from insights, selects cognitive mode | | Review | Scrutiny | Feasibility check — blast radius, reversal risk, confidence scoring | | Execution | Departments | Parallel task execution, constrained by policies | Governor automatically selects a cognitive mode based on task complexity: | Mode | When | Approach | |------|------|----------| | Direct | Fix a typo, tweak a param | Execute immediately | | ReAct | Fix a bug, add a feature | Think → Act → Observe → Loop | | Hypothesis | "Why doesn't X work?" | Hypothesize → Verify → Confirm/Refute | | Designer | Refactor, new subsystem | Design first → Review → Then implement | ## API Interactive docs: http://localhost:23714/api-reference (Swagger UI) ```bash # Global status curl -s http://localhost:23714/api/brief # Open attention debts curl -s 'http://localhost:23714/api/debts?status=open' # Agent live status curl -s http://localhost:23714/api/agents/live # Task execution replay (full thinking chain + tool calls) curl -s http://localhost:23714/api/agents/42/trace ``` ## Directory Structure ``` orchestrator/ ├── .claude/ │ ├── agents/ # 8 agents (auto-discovered by Claude Code) │ ├── skills/ # Slash commands and skill definitions │ ├── hooks/ # Guard rules, audit logging │ └── boot.md # Compiled SOUL identity (generated) ├── src/ │ ├── core/ # Infrastructure: config, event bus, LLM routing, cost tracking │ ├── governance/ # Governance pipeline (Docker service) │ ├── analysis/ # Daily reports, insights, profiling, burst detection │ ├── collectors/ # Data collectors (Python + YAML-driven) │ ├── channels/ # Telegram, WeChat, WeCom adapters │ ├── storage/ # EventsDB (SQLite), Qdrant vector store │ └── scheduler.py # Scheduler entry point ├── departments/ # Department definitions (Docker governance layer) ├── SOUL/ # AI personality framework (source files) ├── dashboard/ # Frontend (Express + WebSocket) ├── data/ # Runtime data (gitignored) ├── docs/ # Architecture docs, pattern library, steal reports └── tests/ ``` ## Prerequisites - **Python 3.10+** - **Node.js 18+** - **Claude Code** ([install guide](https://docs.anthropic.com/en/docs/claude-code)) Database is SQLite (built into Python). Docker is optional. ### Optional Components | Component | With it | Without it | |-----------|---------|------------| | **Docker** | Background collection + dashboard + governance | Claude Code plugin layer works standalone | | **Ollama** | Scrutiny uses local models, saves tokens | Falls back to Claude API | | **Fish Speech** | Dashboard daily report plays as audio | Voice button hidden | ## Design References Architecture patterns from 100+ open-source projects across 46 rounds. 217 patterns implemented. Full library: [docs/architecture/PATTERNS.md](docs/architecture/PATTERNS.md). | Source | What we learned | |--------|----------------| | [autonomous-claude](https://github.com/matthewbergvinson/autonomous-claude) | Foundation for 24/7 autonomous operation | | [edict](https://github.com/cft0808/edict) / [danghuangshang](https://github.com/wanikua/danghuangshang) | Multi-tier governance model | | [soul.md](https://github.com/aaronjmars/soul.md) | Identity persistence framework | | [NVIDIA G-Assist](https://github.com/NVIDIA/g-assist) | Manifest-driven component auto-discovery | | [OpenHands](https://github.com/All-Hands-AI/OpenHands) | Context compression + stuck detection | | [career-ops](https://github.com/santifer/career-ops) | Data contracts, adaptive pipelines | | [Agent-S](https://github.com/simular-ai/Agent-S) / [UI-TARS](https://github.com/bytedance/UI-TARS) | GUI desktop automation | | [Fish Speech](https://github.com/fishaudio/fish-speech) | Voice system (TTS + emotion tags) |

Loading the repository README…

Skills in this repository

Network & Computer Systems Administrators

Information Security Analysts

兵部 — Security audit: secrets scan, dependency vulnerabilities, permission checks, Docker security, backup integrity. Detection only, read-only.

2026-04-06

Showing 28 of 28 collected skills.