feat: per-agent isolation for skills, mode, and evolution

[OctoClaws]

Summary

Adds per-agent isolation to MetaClaw, enabling multi-agent deployments where each agent has independent skills, operating mode, and evolution pipeline.

Closes #49

Changes

1. extensions/metaclaw-openclaw/index.ts (+2 lines)

• Propagates X-Agent-Id header from OpenClaw's ctx.agentId alongside existing X-Session-Id and X-Turn-Type

2. metaclaw/config.py (+10 lines)

• Adds agents: dict field to MetaClawConfig for per-agent overrides
• Supports mode, skills_dir, lora_output per agent

3. metaclaw/api_server.py (+133 lines, -25 lines)

• Reads X-Agent-Id header in chat_completions endpoint
• New _get_agent_skill_manager(agent_id) method:
  • Creates/caches per-agent SkillManager instances
  • Loads skills from {skills_dir}/{agent_id}/ + {skills_dir}/_shared/
  • Falls back to global skill_manager when no agents config
• Per-agent mode routing: resolves skills_only vs rl per agent from config
• All 4 skill evolution trigger sites pass agent_id for correct directory targeting
• _inject_skills() and _inject_augmentation() accept optional skill_manager parameter

Config Example

agents:
  research-agent:
    mode: rl
  assistant-agent:
    mode: skills_only

Directory Structure

~/.metaclaw/skills/
├── _shared/              # Shared across all agents
├── research-agent/       # Auto-evolved skills for research-agent
└── assistant-agent/      # Auto-evolved skills for assistant-agent

Backward Compatibility

• No X-Agent-Id header → falls back to _default agent → existing behavior
• No agents config section → all agents share global config → existing behavior
• Existing flat skills/ directory works as-is for single-agent setups

[richard-peng-xia]

Hi @OctoClaws,

Thanks for the PR! The overall design is solid. Two issues before we can merge:

1. config_store.py not updated
   MetaClawConfig now has an agents field, but to_metaclaw_config() in config_store.py never reads the agents: block from config.yaml and passes it in. This means per-agent mode overrides in config will silently have no effect — MetaClawConfig.agents will always be an empty dict regardless of what the user configures. This needs to be fixed before merging.

2. ctx.agentId is not exposed by the OpenClaw plugin SDK
   The extension reads ctx.agentId in the before_prompt_build handler, but this field doesn't exist in the current OpenClaw context type. At runtime it will always be undefined, so X-Agent-Id will never be sent and all agents will fall back to _default — meaning the isolation feature won't actually work end-to-end. Could you confirm whether OpenClaw already exposes agentId on the event context, or if a companion PR to OpenClaw is needed to add it first?

Happy to help review a follow-up once these are addressed.