Expose MCP prompts as transient Netclaw skills

[Aaronontheweb]

Outcome

An agent discovers an MCP prompt in the normal skill index.

The agent loads that prompt through skill_load.

Requirements

• Call prompts/list for a server that declares prompt support.
• Publish prompt descriptors in the immutable server generation.
• Map each prompt to mcp__<server>__<prompt>.
• Keep local and remote skills in one atomic registry snapshot.
• Add an optional MCP argument map to skill_load.
• Validate required and unknown arguments before prompts/get.
• Preserve prompt roles and source attribution in the result.
• Apply the existing MCP server permission to discovery and use.
• Fail visibly for stale generations and server faults.
• Keep tool-only servers healthy.
• Do not add load_mcp_prompt.

Proof

• Add a parameterized prompt to the smoke MCP server.
• Add focused unit and integration tests.
• Add model evals for relevant and unrelated requests.
• Update the MCP PRD, specs, and netclaw-operations skill.
• Run the required repository quality gates.

Sources

• MCP prompts, revision 2026-07-28
• Approved design