Epic: Make skills discoverable and directly invocable

[Aaronontheweb]

Outcome

Netclaw has one discoverable and invocable skill system.

Local skills, skill-server skills, and MCP prompts use one catalog.

Agents use the session skill index and skill_load.

First-party clients use an audience-filtered daemon API.

Scope

• Expose MCP prompts as transient skills.
• Expose MCP resources through bounded list and read operations.
• Define one invocable skill descriptor.
• Expose a skill catalog and completion HTTP API.
• Add slash-command completion to the TUI.
• Subscribe to MCP catalog and resource updates.
• Keep the catalog poll as a repair path.

Fixed decisions

• An MCP prompt uses mcp__<server>__<prompt>.
• Netclaw uses prompts/list for discovery.
• skill_load uses prompts/get.
• MCP prompt arguments stay in the MCP adapter.
• MCP resources remain remote reference content.
• The MCP server grant remains the permission boundary.
• A prompt does not grant a tool.
• Netclaw does not add a second prompt catalog.

Delivery order

1. Add MCP prompts to the skill system.
2. Define the common invocable descriptor.
3. Add the daemon catalog API.
4. Add TUI completion.
5. Add resource discovery and read operations.
6. Add proactive subscriptions and keep poll reconciliation.

Items can move when a contract dependency requires it.

Existing work

• Make system skills invokable via slash command and support skill dependencies #1631 covers system skill invocation and dependencies.
• MCP client never detects live catalog changes (no list_changed handling; tools only enumerated at connect/reconnect) #1769 and PR fix(mcp): detect live tool catalog changes on connected servers #1771 added live catalog polling.
• Current slash-command dispatch already exists.
• Current file skill metadata already includes invocable and argument-hint.

Sources

• MCP prompts, revision 2026-07-28
• MCP transport and subscriptions, revision 2026-07-28
• GitHub sub-issue guidance
• Approved Memorizer proposal
• Memorizer epic