AGENT SKILL

openmuse

Use when improving MDE's existing AI Concierge chat, OpenMuse-style composer, Stop/Retry, queued messages, genuine tool progress, context-aware answers, grounded cards/map pins, search race recovery, or responsive chat/mobile UX. Adapt CopilotKit v2 patterns, not the OpenMuse app or a second chatbot.

License 许可
MIT
所属仓库
amoai-tech/mdeai
最近更新
2026-11-13
Stars
★ 2
Forks
1

01 前言 Front Matter

name
openmuse
description
Use when improving MDE's existing AI Concierge chat, OpenMuse-style composer, Stop/Retry, queued messages, genuine tool progress, context-aware answers, grounded cards/map pins, search race recovery, or responsive chat/mobile UX. Adapt CopilotKit v2 patterns, not the OpenMuse app or a second chatbot.

02 资源文件 3

03 SKILL.md 正文

OpenMuse — reliable MDE Concierge interaction

Owner and boundaries

Own the user-visible conversation/search/map experience. The user sees one MDE AI Concierge, while authorized Mastra specialist agents/tools may operate behind it. CopilotKit runtime/transport belongs to copilotkit, agents/memory to mastra, rental data to real-estate, and DB/RLS to supabase. Do not duplicate the runtime, provider, thread store, search engine, maps, or domain workflows.

Phase A: never add listing eligibility, approval, ownership, photo-rights or publishing prerequisites. Keep source attribution, server-side authorization, private thread isolation and truthful coordinates, prices and availability.

Implementation procedure

  1. Verify the real path. Read the SAN task, active PR base, installed CopilotKit v2/AG-UI types and current MDE chat, fast-path hooks, tool renders and map. Compare only the relevant pinned upstream behavior using the source map.
  2. One UI, correctly scoped context. Reuse the existing chat-center-panel and canonical /api/copilotkit runtime. Keep user/thread identity server-owned; selected-page IDs/filters are untrusted hints. Selecting A, switching to B, closing the page or logging out must not leak old context.
  3. Actual progress, not a timer. Show running/completed/failed/aborted stages only from real run/tool lifecycle events or returned data. Correlate thread + run + tool call; elapsed time may animate or dismiss a label, never assert completed steps or counts. If only coarse states exist, show "Searching…" instead of invented substeps. Owner: SAN-609.
  4. Never lose a second question. When a fast-path search is busy, queue the next message, visibly keep it for retry, or return unhandled for Concierge fallback. Never claim "handled" while dropping it. Distinguish in-memory follow-up queues from durable server tasks; Stop preserves unsent drafts and does not imply external work is canceled. Owners: SAN-1559 / SAN-1407.
  5. Only current results own the map. Old runs, prior chat's tool calls or virtualized result re-renders must not override newer selected results. Render pins only for verified coordinates. Owner: SAN-1429 for tool render effects and SAN-1407 for ordering.
  6. Smallest change and proof. Reuse → configure → model upstream UX → minimal adaptation; never copy React Native. Test rapid overlapping searches (older finishes last), page A→B→clear, long chat virtualization, Stop/Retry, source/card/map agreement, disconnect, keyboard and 390px. Run focused tests, applicable lint/typecheck and exact-head floor. Report observed behavior, failed checks and exact SHA.

Example

"Find Laureles apartments" then "Under COP 4M" before results return: both requests remain accounted for. The current map cannot be replaced by a stale result; the progress strip cannot claim ranking finished without a real ranking-tool completion.

04 同仓库更多 Skills