lsm / neokai / 24995864733

82%

Build Type

push

github

Committed by web-flow

Commit Message

fix(daemon): preserve in-process MCP servers across runtime mutations (Task #140) (#1664)

Implements P0+P1 fixes from docs/research/node-agent-mcp-loss-root-cause.md
so node agents stop losing their handoff tools mid-turn.

- Renames AgentSession.setRuntimeMcpServers to replaceAllRuntimeMcpServers
  and converts the remaining production callers (NeoAgentManager,
  RoomRuntimeService) to mergeRuntimeMcpServers, eliminating the
  replace-all races that silently dropped concurrent SDK servers.
- Adds a structured mcp.attach telemetry log emitted by every runtime
  merge/detach/replace, joinable on sessionId/spaceId/taskId so MCP
  loss regressions can be reconstructed from logs alone.
- Promotes the query-runner mcpServers debug line to a query.mcp.snapshot
  payload that includes sessionType + workflow sub-session metadata.
- Adds tests for config.mcp.* in-process preservation, reinject query
  restart, session-reuse node-agent rebuild, and the mcp.attach payload
  shape. Updates affected mock sessions to expose mergeRuntimeMcpServers.

Co-authored-by: Claude Opus 4.7 <noreply@anthropic.com>

Coverage Stats

9391 of 12586 branches covered (74.61%)

Branch coverage included in aggregate %.

39 of 46 new or added lines in 4 files covered. (84.78%)

1 existing line in 1 file now uncovered.

66818 of 79264 relevant lines covered (84.3%)

304.06 hits per line