AgentSkillsCN

omo-agent-metadata-and-model-picker

当配置UI需要更丰富的内置代理描述（目的、访问权限/只读、注意事项），以及一个可扩展的模型选择器（搜索+筛选+分组结果），而非庞大的“其他”列表时使用此功能。 适用于OmO-Agent-Config：将本地AGENT_PROFILES与上游oh-my-opencode代理元数据合并，随后在WebUI以及代理→模型选择模态框中呈现。

SKILL.md
--- frontmatter
name: omo-agent-metadata-and-model-picker
description: |
  Use when a config UI needs richer built-in agent descriptions (purpose, access/read-only, caveats)
  and a scalable model picker (search + filters + grouped results) instead of a giant “other” list.
  Applies to OmO-Agent-Config: merge local AGENT_PROFILES with upstream oh-my-opencode agent metadata,
  then surface it in WebUI and the agent→model selection modal.

OmO agent metadata + scalable model picker

Problem

  • •Agent lists are hard to understand with only a name and a minimal description.
  • •Model selection dialogs don’t scale when they dump a long-tail “other” bucket of 100–200+ models.

Context / Trigger Conditions

  • •You have an agent→model assignment UI.
  • •Users complain that the picker is a flat list / “other” bucket.
  • •You already have (or can fetch) upstream agent definitions (system prompt metadata) and want to enrich local UI.

Solution

1) Merge two sources of truth

  • •Local static profile (fast, curated UX fields): AGENT_PROFILES in lib/constants.js.
    • •Add fields like: access (read-only|limited|write), usage[], caveats[], plus existing preferred[], minContext.
  • •Upstream parsed agent metadata (authoritative intent): lib/core/agents.js.
    • •Fetches TypeScript from code-yeongyu/oh-my-opencode and parses description, role/behaviors/constraints.
    • •Infer an access mode from common constraint language (read-only/plan-only), but prefer the local curated AGENT_PROFILES.access when present.

Expose merged metadata via the API:

  • •In lib/server.js GET /api/agents and GET /api/agents/:name, attach:
    • •summary (curated description)
    • •access, usage[], caveats[], preferred[]
    • •recommendedModels[] (top N) plus recommendedModel (single best)

2) Replace “other models” with a picker that scales

In the WebUI model selector modal:

  • •Add an inline search box and lightweight filters:
    • •provider select
    • •min context select
    • •capability chips (reasoning/image/pdf/thinking/fast)

Render results in sections:

  1. •Recommended for this agent (top N IDs from API; map to model objects in the client)
  2. •In use in this config (models currently assigned to any agent)
  3. •All models (N) grouped by provider (optionally capped to avoid huge renders)

This removes the UX anti-pattern of a giant “other” bucket while still offering full access.

Verification

  • •GET /api/agents returns merged fields (summary, access, usage, caveats, preferred, recommendedModels).
  • •WebUI agent cards show richer summaries and caveats.
  • •WebUI “Change model” modal supports search + filters and groups results (no flat “other” dump).
  • •Basic runtime check: node -e "require('./lib/constants'); require('./lib/core/agents'); require('./lib/server');".

Example

  • •User clicks Change Model on an agent card.
  • •Modal shows agent summary + requirement hints.
  • •User types claude and selects provider anthropic.
  • •Results narrow instantly; recommended models are pinned at the top.

Notes

  • •Keep the UI dependency-free (plain DOM + event listeners + re-render) to match this repo’s zero-npm approach.
  • •Cap “All models” rendering (e.g., first 120) and prompt users to refine filters for large catalogs.