[edgexyz]

Coordinators currently need to inspect individual worker CLIs to find model IDs before dispatching. This adds an on-demand MCP tool, list_models(agent, cwd, refresh=false), that returns a worker's advertised model catalog. list_agents also reports discovery support.

Changes

• Share ACP discovery between Kimi and DSH using a separate initialize / session/new handshake and the SDK's typed configuration options.
• Query Antigravity through agy models and Codex through paginated app-server model/list.
• Return model IDs, display names, advertised defaults, model-specific native effort levels, and cache metadata. Preserve null for unknown effort support and [] for explicitly empty levels.
• Cache catalogs in memory per worker/project for one hour by default. Support explicit refresh and [models] cache_sec / timeout_sec configuration.
• Add English/Chinese orchestration guidance, skill and MCP tool descriptions, and an isolated MCP smoke-test script.

Discovery is available to coordinator instances. ACP discovery denies permission requests and may leave an empty session in worker history. Lookup deadlines, output/page limits, diagnostic redaction, and subprocess cleanup cover worker failures and cancellation.

Validation

Tested on Windows with Python 3.14:

• Full regression suite: 737 passed.
• ruff check . and mypy: passed.
• Live queries through an isolated MCP server: Kimi 4, DSH 4, Antigravity 13, Codex 7 catalog entries; cache reuse passed for each worker.
• Tests cover SDK-backed flat/grouped ACP metadata, native effort semantics, pagination, malformed responses, timeout/cancellation cleanup, environment filtering, and cache concurrency.

The full suite reports one previously observed asyncio pipe-finalizer warning in test_quota.py::test_describe_quotas_covers_every_agent; all tests pass.