lsm / neokai / 27457810412

82%

push

github

Proxy Space MCP tools for ACP (#2105)

* feat: proxy Space MCP tools for ACP

Bridge in-process Space MCP servers through a per-session Unix socket so ACP agents can receive stdio MCP configs without losing runtime tool handlers.

* fix: harden ACP MCP proxy bridge

Use private per-bridge sockets, authenticate proxy requests, always clean stale-query bridges, and preserve JSON schemas for raw SDK tool shapes.

* fix: complete ACP MCP proxy hardening

Use registered MCP callbacks, route compiled binaries through an embedded proxy entrypoint, add MCP ping support, and guard proxy calls/teardown against hangs.

* fix: polish ACP MCP proxy edge cases

Add Windows IPC paths, negotiate MCP protocol versions, and return tool handler failures as model-visible MCP tool results.

* fix: expose Space MCP tool definitions

Keep production Space MCP tool metadata on runtime servers so the ACP proxy can collect schemas and callbacks without depending on test-only internals.

* test: cover ACP MCP proxy edge cases

Exercise invalid proxy tokens and tools-array fallback, align mocks with the real SDK handler field, and move proxy tool catalogs out of argv.

* fix: scope ACP proxy tool catalogs

Write per-server tool catalog files so each stdio proxy advertises only the tools it can route.

* fix: expose agent memory proxy tools

Preserve agent-memory tool metadata for ACP proxy fallback and decode source proxy paths via fileURLToPath for cross-platform launches.

* fix: support registered MCP callbacks

Accept both registered callback and handler fields when collecting proxied SDK tools, with coverage for callback-backed registrations.

* fix: proxy db-query for ACP sessions

Include db-query in the ACP MCP proxy surface and avoid loading normal server modules before embedded proxy mode starts.

* fix: close ACP proxy bridge before retry

Tear down proxy sockets before retrying ACP startup and guard delayed bridge responses against closed peer sockets.

10449 of 15296 branches covered (68.31%)

Branch coverage included in aggregate %.

57 of 285 new or added lines in 6 files covered. (20.0%)

89649 of 107423 relevant lines covered (83.45%)

288.67 hits per line