Feature Request: Per-agent isolation (skills, mode, LoRA, memory)

[OctoClaws]

Feature Request

Problem

MetaClaw currently operates as a single global instance with one shared skill library, one operating mode, one RL training pipeline, and one memory scope. This does not work well for multi-agent setups where different agents have fundamentally different roles and tasks.

For example, in an OpenClaw deployment with multiple agents:

• Agent A (research assistant) should accumulate academic skills and train a LoRA tuned for research tasks
• Agent B (daily assistant) should accumulate productivity skills and may only need skills_only mode
• Agent C (code reviewer) should have code-review-specific skills and its own RL policy

Currently, all agents share the same ~/.metaclaw/skills/ directory, the same mode, the same RL buffer, and the same LoRA checkpoint. Skills learned from one agent's conversations pollute another's prompt injections, and RL training mixes trajectories from incompatible task distributions.

Proposed Solution

Add per-agent isolation across four dimensions:

1. Agent Identity Propagation

The OpenClaw plugin (index.ts) should pass an X-Agent-Id header alongside the existing X-Session-Id and X-Turn-Type. The plugin's before_prompt_build callback has access to the session context, which should contain the agent identifier.

2. Per-Agent Skill Directories

~/.metaclaw/skills/
├── _shared/              # Skills shared across all agents (manually curated)
├── agent-a/              # Auto-evolved skills for Agent A
├── agent-b/              # Auto-evolved skills for Agent B
└── agent-c/              # Auto-evolved skills for Agent C

• SkillManager loads _shared/ + {agent_id}/ for each request
• SkillEvolver writes new skills to {agent_id}/ subdirectory
• Fallback: if no X-Agent-Id is provided, use a _default/ directory (backward compatible)

3. Per-Agent Mode Configuration

# ~/.metaclaw/config.yaml
agents:
  agent-a:
    mode: rl
  agent-b:
    mode: skills_only
  agent-c:
    mode: rl
  _default:
    mode: skills_only    # fallback for unknown agents

The API server routes each request based on the agent's configured mode:

• skills_only agents: skill injection only, no RL data collection
• rl agents: skill injection + conversation samples added to RL buffer

4. Per-Agent LoRA Training

This is critical for RL mode. If multiple agents use RL, they must train separate LoRA checkpoints, because:

• Different agents have different task distributions (research vs. chat vs. code review)
• Mixing trajectories from incompatible tasks degrades policy quality for all agents
• Each agent's PRM scoring context is different

Proposed structure:

agents:
  agent-a:
    mode: rl
    rl:
      model: "base-model-id"           # base model for this agent
      lora_output: "~/.metaclaw/lora/agent-a/"
      resume_from_ckpt: ""
  agent-c:
    mode: rl
    rl:
      model: "base-model-id"
      lora_output: "~/.metaclaw/lora/agent-c/"

Each agent gets:

• Its own RL sample buffer (filtered by agent_id)
• Its own LoRA checkpoint directory
• Its own training schedule (batch_size threshold per agent)

5. Per-Agent Memory Scope (already partially supported)

MetaClaw's memory layer already has user_id and workspace_id support via headers. The X-Agent-Id can be used as an additional scope dimension to ensure memory isolation, using the existing derive_memory_scope() infrastructure.

Backward Compatibility

• If no X-Agent-Id header is present, all behavior falls back to _default — identical to current single-agent behavior
• Existing ~/.metaclaw/skills/ flat structure continues to work as the _default agent's skill directory
• No config migration needed; the agents config section is optional

Estimated Changes

File	Change
index.ts (OpenClaw plugin)	Add X-Agent-Id header from session context (~5 lines)
config.py	Add AgentConfig dataclass and agents config section (~30 lines)
api_server.py	Read X-Agent-Id, route to per-agent skill dir and mode (~40 lines)
skill_manager.py	Support agent_id parameter for directory routing (~30 lines)
skill_evolver.py	Write skills to per-agent subdirectory (~10 lines)
data_formatter.py	Tag samples with agent_id for RL buffer filtering (~10 lines)
Total	~125 lines of production code

Environment

• MetaClaw version: 0.4.1
• OpenClaw with 6 configured agents
• Use case: multi-agent deployment with heterogeneous task profiles