[Feature]: Optional Explicit Sub-Skill Registration to Mitigate Discovery Overhead

[taqtiqa-mark]

Add config flag for explicit sub-skill registration, preserving controlled linking, as fallback:

• if upstream changes course.
• for users who prefer auto-linking.

Description:
Given the open upstream issue #10238 on nested skill structures, and the viable workaround of explicit Markdown links in skills.md for auto-registration (per comment), I propose an opt-in mechanism to avoid recursive scanning of all skills/ subfolders, which risks fatal performance overhead in large repos.

Proposed Change:

• Introduce a config variable in opencode.json (or equivalent): "skill_discovery": "explicit" (default: "explicit" for backward compatability).
  • "auto": scan all skill subfolders recursively (current opencode-skills approach, still debated upstream (CC)).
  • "explicit": Consistent with current model driven behavior of upstream (CC) and a change from current opencode-skills nested skill registration —Only register sub-skills explicitly linked via yaml front matter in Markdown in parent skills.md files (e.g., currently done with instruction: "Read Async Event subskill for event handling"), in skills file with this there would be a tool skills_core_events_async_promise.

    # skills/core/SKILL.md
    skills:
      - events/
      - events/async
      - events/async/callback
      - events/async/promise

• Update docs: Add section in README.md under "Skill Organization" with example of the current model driven approach (i.e. currently no skills_core_events_async_promise):

  In skills/core/skills.md:
  ## Sub-Skills
  - [Events](events/skills.md): For event system implementation.
    - [Async Events](../skills/core/events/async/skills.md) subskill for async-event handling
    - [Async Events (Callbacks)](../skills/core/events/async/callback/skills.md) subskill for async-event handling using callbacks
  - [UI](ui/skills.md): For rendering logic.
  

  Note: Model followed links on parent load; no scanning needed.

Rationale:

• Supports controlled, low-overhead organization without depending on upstream changes.
• Consistent with current nested mimicry (e.g., skills/core/events/) via explicit refs, avoiding full-tree reads.
• Headroom: future fallback to auto-discovery still an option, ensuring minimal disruption.

Issues:

• What to do about reserved words, resources, scripts, etc.