openclaw-workspace
Agent workspace configuration files for a multi-agent OpenClaw deployment running 4 specialized agents.
Agents
| Agent | ID | Workspace | Role |
|---|---|---|---|
| Relay | relay | workspace-relay | Human-facing. All user communication. |
| Captain | main | workspace | Router/dispatcher. No direct execution. |
| Repo-Man | spec-github | workspace-spec-github | Infra, keys, backups, GitHub, model health |
| Quartermaster | spec-projects | workspace-spec-projects | Projects, decisions, auditing |
Key Concepts
Scoped Context
Every token in an agent's context must serve that agent's specific job. Shared documents use RACP (Role-Aware Context Projection) markers to generate per-agent targeted versions, reducing token waste by ~76% compared to blanket context.
Filesystem Communication Bus
Agents communicate with external tools (like Claude Code on the host VPS) via:
- Shared registry (
registry.json) — single source of truth for all IDs and constants - Task inbox/outbox — structured file-based handoff protocol
- SQLite ops database (
ops.db) — shared operational history
Workspace Files
Each agent workspace contains some combination of:
| File | Purpose |
|---|---|
AGENTS.md | Agent-specific instructions, tools, skills, rules |
SOUL.md | Personality, routing table, behavioral guidelines |
TOOLS.md | Available tools and how to use them |
IDENTITY.md | Agent identity and voice |
USER.md | User preferences and communication style |
HEARTBEAT.md | Periodic check instructions |
BOOTSTRAP.md | Session startup protocol |
DISCORD-REFERENCE.md | Targeted Discord capabilities (per-agent) |
AI Attribution
All code in this repository was written by Claude Code (Anthropic's CLI agent), powered by Claude Opus 4.6.
- Human: Robert Supernor — product direction, priorities, review
- AI: Claude Code — implementation, architecture, agent design
Every commit includes:
Co-Authored-By: Claude Opus 4.6 <[email protected]>
还没有评论。