Genor's Project Orchestration

Core Principles

1. Repowise First — Never touch code without codebase intelligence
2. Plan Before Act — update_plan for anything beyond one edit
3. Verify Before Claim — No completion claim without fresh evidence
4. Self-Review — Every output gets audited before delivery
5. Fail Gracefully — Every step has a fallback chain
6. Document Everything — Log sessions, decisions, and architecture

Data Directory

All private data (models, sessions, prices) lives outside the skill. Set ORCHESTRATOR_DATA_DIR to point to your data directory. Default: ../../orchestrator-data/ from the skill root (i.e. <workspace>/orchestrator-data/)

File Naming Convention

All generated files use consistent, sortable naming for cross-session compatibility:

Cross-Session Resume

Session state files in sessions/ persist across OpenClaw sessions. When a conversation is compacted or a new session starts, the LLM can read the last session state file to continue where it left off — no dependency on conversation history.

🆕 Fresh Installation — First-Run Onboarding

If this is a fresh skill install with no data yet, run onboarding:

bash {baseDir}/scripts/onboard.sh

or ask the LLM: "Set up orchestrator" or "/project-orchestration onboard"

This walks through:

1. Data directory setup — where to store config
2. Provider discovery — what services you have
3. Model cataloguing — what models you use
4. Project discovery — existing projects
5. Routing rules — initial routing table
6. Cron setup — nightly price checks
7. Logging — first session recorded

If you're in an LLM session, the LLM will guide you through each step.

🆕 Project Onboarding

When starting a new project or adding an existing one:

bash {baseDir}/scripts/init-project.sh <project-path> <Project Name> [stack]

or ask the LLM: "Start project "

This creates .planning/ with:

• CONFIG.md — project config, ports, tech stack
• STATE.md — current state tracker
• ROADMAP.md — planned work
• REQUIREMENTS.md — requirements
• ADRs/ — architecture decisions

For existing projects, the LLM will also run Repowise/scan, detect tech stack, and initialise the project in the orchestrator.

Slash Commands

When invoked via /project-orchestration, interpret the subcommand:

Quick Scripts

Active Tools (do real work)

Guidance Scripts (print instructions for OpenClaw session tools)

On-Demand Reading

When you need detailed instructions, read the relevant file:

• {baseDir}/README.md — Full documentation (all procedures)
• {baseDir}/ROUTING.md — Model routing table template
• $ORCHESTRATOR_DATA_DIR/models.json — Your model inventory
• $ORCHESTRATOR_DATA_DIR/session_log.md — Quick reference session history
• $ORCHESTRATOR_DATA_DIR/sessions/ — Detailed session state files (for resume)

Using External Skills

For specialised tasks, use whatever skills are available in your instance. Always fall back to built-in tools (edit, exec, read, web_search, web_fetch, browser) when specialised skills are unavailable.

Prompt Injection

Every spawned sub-agent or Cursor session MUST include:

IMPORTANT: Follow Genor's Project Orchestration conventions:
- Repowise first (or exec find as fallback)
- Plan before coding (update_plan or mental plan)
- Verify before claiming (build, test, screenshot, vision Q&A)
- Self-review your output before returning
- Fallback chains: if a tool is unavailable, use the next fallback
- Document decisions, log sessions