Workspace Files & Prompts

In DOSClaw, each agent workspace operates under a strict Separation of Concerns across its Markdown prompt and configuration files. This ensures high predictability, eliminates contradictory instructions, and maintains reliable tool-calling behavior on large and small language models alike.

---

Workspace Files Overview

/workspace/
├── openclaw.json       # Runtime configuration (models, channels, tool permissions)
├── SOUL.md             # Persona, voice, tone, and empathy boundaries ONLY (< 1KB)
├── IDENTITY.md         # Entity identity (name, role, avatar, business affiliation)
└── AGENTS.md           # Single rulebook for operating instructions and ## Tools

File	Primary Responsibility	Critical Rule
SOUL.md	Persona & Tone ONLY Defines who the agent is, how it speaks, formality, empathy, and pronouns.	Keep under 1KB. Never include operational procedures, tool lists, or RAG guidance here. Bloated personas suppress tool-calling instincts on smaller or quantized LLMs (e.g. Qwen 3.6).
IDENTITY.md	Entity Identity Defines the agent's display name, role, creature type, and business branding.	Updated dynamically when an operator renames the bot or customizes identity in the dashboard.
AGENTS.md	Single Rulebook & Operating Instructions Operating guidelines, safety boundaries, and local tool conventions under ## Tools.	Injected into 100% of sessions. Even when a merchant configures a custom_soul, the platform's AGENTS.md remains active to enforce tool safety and RAG rules.
openclaw.json	Runtime Configuration Declares LLM providers, model mappings, channel bindings, and tool authorization policies.	Managed and updated atomically by the DOS.AI Gateway. Overwritten safely on container restarts.

---

The Retirement of TOOLS.md

In earlier versions of the OpenClaw ecosystem, environment notes and tool usage guidelines were placed in a separate /workspace/TOOLS.md file.

Upstream OpenClaw Evolution

• Consolidation into AGENTS.md: Upstream OpenClaw officially retired TOOLS.md. Separate prompt files were found to fragment model attention and cause conflicting behavior.

• Single Source of Truth: All local tool instructions, usage conventions, and environment variables are now consolidated into the ## Tools section of AGENTS.md.

• Automated Migration (openclaw doctor --fix): OpenClaw includes a repair utility that automatically archives legacy TOOLS.md files, merges custom tool notes into AGENTS.md, and removes the retired file:

  openclaw doctor --fix

• DOS.AI Provisioning Policy: The DOS.AI Go Gateway never provisions or generates TOOLS.md. All tool orchestration instructions are rendered directly into AGENTS.md.

Note on Existing Containers: Because an agent's /workspace is mounted from a persistent Docker volume (agent-{id}-data), containers provisioned under earlier versions of OpenClaw may still have a legacy TOOLS.md file on disk. The file is ignored by modern OpenClaw runtimes.

---

Customizing Agent Personas (custom_soul)

Merchants and developers can customize their agent's personality through the DOS.AI Dashboard:

1. Pure Voice Customization: When you provide a custom personality, it replaces only SOUL.md. Your custom prompt dictates the tone (e.g. "You are an energetic TikTok live seller" or "You are a courteous financial assistant").

2. Preserved Platform Guardrails: Your custom agent still receives the full AGENTS.md rulebook automatically. This guarantees that:

   • The agent retains the ability to call search_knowledge for verified FAQ answers.

   • The agent retains connected e-commerce or CRM tools.

   • The agent adheres to anti-echo directives and tool-first action invariants.

---

Dual-Profile Architecture (agents.list)

Modern OpenClaw agent containers provisioned by DOS.AI define two execution profiles in agents.list:

1. main Profile (default: true):

   • Uses /workspace/IDENTITY.md for identity and persona.

   • Serves direct customer direct messages (DMs) and authenticated business interactions.

   • Has full access to enabled business MCP tools and connectors.

2. collaboration Profile:

   • Renders with a distinctive identity: <AgentName> (Collaboration 🤝).

   • Restricts tool execution (tools: { deny: ["*"] }).

   • Serves untrusted group mentions (@DOS or bot tag in public groups) without exposing write tools or private customer data to channel members.