Implement just-in-time skill delivery (deprecate get_skills bulk loading)

[m2ux]

Summary

The current get_skills tool returns all skills for a scope in a single response (measured at 555 KB for the work-package workflow). Agents work around this by saving the response to disk and parsing it with python3 to extract specific skills — a three-step workaround for what should be a single targeted fetch. This work package replaces the bulk-loading model with just-in-time (JIT) skill delivery.

Requirements

1. Step-level skill scoping

Skills may be scoped at workflow and/or step level. Activity-level skill associations should be migrated to step-level associations, so each step declares exactly which skill it needs.

2. Deprecate get_skills

The get_skills tool causes context overload and should be deprecated. The get_skill (singular) tool should be maintained and enhanced as the primary skill retrieval mechanism.

3. Context-dependent get_skill via step_id

The get_skill tool should accept a new step_id parameter. The server resolves which skill to return based on the step's declared skill association, removing the need for the agent to know skill IDs upfront.

4. One skill per step

Each call to get_skill with a step_id returns exactly one skill. If a step currently requires multiple skills, the protocol should be further decomposed to ensure 1:1 step-to-skill mapping.

5. Unified workflow management skill

Workflow management (loading workflows, navigating activities, managing state) should be consolidated into a single skill. Loading a workflow or activity should require ingesting only one skill for the agent to be fully capable of workflow execution.

6. Just-in-time delivery

The primary aim is delivering skill content just-in-time to avoid agents performing extra operations on voluminous content to extract specific skills or resource data. Skills are fetched precisely when needed, not preloaded in bulk.

Motivation

• Context efficiency: 555 KB bulk response wastes agent context window
• Operational overhead: Agents currently save to disk + shell out to python3 to parse
• Simplicity: 1:1 step-to-skill mapping eliminates skill selection logic in agents
• Scalability: As workflows grow, bulk loading becomes increasingly wasteful

Scope

• Server-side: Modify get_skill tool, deprecate get_skills, update skill loader
• Workflow data: Migrate activity-level skill associations to step-level
• Skills: Consolidate workflow management skills, decompose multi-skill steps
• Schema: Update activity and workflow schemas for step-level skill declarations
• Bootstrap: Update discovery/bootstrap procedure to reflect new skill delivery model