[Feature Request] Skill extension system with semantic-diff overrides, dependency injection, skill-composition (composition over inheritance, modularity)

Please move skills and agents toward:

• Modularity & SRP
• composability micro-skills with composition over inheritance
• Overridden parts where modularity is not provided: (using diff/patch or semantic diff) - keep using community skills "except that one thing".
• Micro-skills - steering users toward micro-skills and composition.

Strong +1 on composition over inheritance for skills. I maintain a personal agent framework with 100+ skills in daily use, and I hit exactly the gap you describe: skills that wanted to extend another skill had no mechanism except copy-paste, which diverges silently.

I ran a dependency-graph scan over my skill library to classify hub skills (referenced by many others) and orphan skills (referenced by nothing), and the edge taxonomy that fell out of that exercise matches your proposal almost one-to-one: compose (aggregate), override (replace a section), inject (provide a dependency at load time).

One practical question from my experience: for semantic-diff overrides, how do you propose resolving conflicts when the base skill is updated upstream — rebase-and-flag, or re-derive from the intent anchor? Happy to share my graph-scan taxonomy if it's useful as a data point.

I expect that the skill override would be represented as a patch file, and behave like a merge operation, making both the base skill and derived skill available.

The derived skill would have a patch representation, which is auto applied to the original resulting in a generated file; OR the patch could be simply appended with a short cookie cutter instruction for an LLM.

This could be compared to any other updates. Ideally skills would have formatting normalization into bullet lists to facilitate merging. But regardless of whether they do, the semantic diff merge could be represented like a merge conflict for human or AI to resolve.

Even if the patch is appended to the skill, the AI consuming it could still potentially understand the intent.