"What is MetaClaw's defining architectural contribution?"	A continual meta-learning framework that sits as a proxy between ANY Claw-family agent and the upstream LLM. Two mechanisms: skill injection at inference time + gradient-based policy optimization (GRPO/LoRA) during idle hours. Makes any agent self-improving without agent-side changes.	harness-engineering::dd18::recall
"Is MetaClaw a governance proxy? Explain."	NO. MetaClaw is a LEARNING proxy, not a governance proxy. No content filters, rate limits, scope policies, or approval workflows. The actual governance proxies are CrabTrap (Brex) and IronCurtain (Niels Provos). MetaClaw occupies the same position but for augmentation, not control.	harness-engineering::dd18::analysis
"What is the dual-use proxy insight?"	Same HTTP interception architecture, opposite purposes. Learning proxy (MetaClaw): augments prompt + captures trajectory. Governance proxy (CrabTrap/IronCurtain): evaluates policy + blocks. The proxy position is neutral — what you do with intercepted traffic determines the function.	harness-engineering::dd18::analysis
"What does MetaClaw intercept and what does it NOT?"	Intercepts: chat-completion requests (both OpenAI and Anthropic wire formats) between agent and LLM. Mutates: prepends skills + memory. Captures: trajectories as ConversationSamples. Does NOT intercept: tool calls, filesystem access, network egress. Agent's outbound actions never touch MetaClaw.	harness-engineering::dd18::analysis
"How does the idle-time RL training work?"	OMLS (Opportunistic Meta-Learning Scheduler) fires during sleep/idle/calendar events. Collects ConversationSamples. LLM evolver summarizes failures into skills. GRPO RL with Process Reward Model scores quality. Cloud LoRA via Tinker/MinT/Weaver. Serving path never blocked. Reported: Kimi-K2.5 21.4% → 40.6%.	harness-engineering::dd18::recall
"How does MetaClaw connect to Claw-family agents?"	metaclaw setup patches the target agent's config file (~/.picoclaw/config.json, ~/.zeroclaw/config.toml, etc.) to point at the proxy port (30000), then restarts the agent's gateway. Supports OpenClaw, PicoClaw, ZeroClaw, Hermes, IronClaw, NanoClaw, CoPaw. Dual wire-format: OpenAI + Anthropic.	harness-engineering::dd18::recall
"What is MetaClaw's critical security limitation?"	The proxy routes user prompts through external endpoints by default — Issue #52 flags this as a risk surface. All agent-model traffic flows through MetaClaw, making it a surveillance/enrichment point. Cannot stop tool execution (doesn't intercept tool calls). For production: combine with a governance proxy (IronCurtain).	harness-engineering::dd18::analysis
"How does MetaClaw compare to Hermes (DD-08)?"	Both use self-evolving skills. Hermes: skills inside the agent, model-initiated writes. MetaClaw: skills at the proxy layer, injected without agent awareness. MetaClaw additionally does gradient-based RL training (Hermes does not). Hermes compounds skills; MetaClaw compounds skills AND model weights.	harness-engineering::dd18::analysis