fix: reuse pooled MCP client connections in pi extension

[EdwardIrby]

Summary

• reuse pooled You.com MCP client connections across Pi tool executions
• reconnect and retry once when a pooled connection drops
• close pooled clients on session_shutdown
• add tests for connection reuse, reconnect retry, and shutdown cleanup

Validation

• bun run --cwd packages/pi test

[factory-droid]

Droid finished @EdwardIrby's task —— View job

---

Droid review complete.

The pooled-connection design is correct in the serial case and the new lifecycle tests pass, but the reset/retry and session_shutdown teardown paths close shared connections without coordinating with concurrent in-flight callers, a real concurrency hazard. No security issues were found.

[factory-droid]

[P1] resetMcpClient closes a shared pooled connection out from under concurrent in-flight callers

mcpClients is module-level shared state, and getMcpClient returns the same Client object to every concurrent caller keyed by server. When two callTool invocations to the same server run concurrently and one throws, withPooledMcpClient's catch calls resetMcpClient, which calls closeMcpConnection (transport.terminateSession() + client.close()) on that shared connection while the other caller's callTool is still in flight on it. That in-flight call then fails or hangs on a torn-down transport; that caller's own catch then calls resetMcpClient again and can close the fresh connection the first retrier just cached, cascading failures. The same class of hazard exists in closeMcpClients on session_shutdown. There is no in-flight reference counting to prevent closing a connection that outstanding callers are using. Trigger: two concurrent tool calls to the same MCP URL where one fails. Fix: track in-flight callers per connection and defer close until the last caller drains, or have each caller snapshot its own connection and only close the old one once all outstanding callers have finished.

[factory-droid]

Droid encountered an error —— View job

---

Droid is reviewing code and running a security check…

[factory-droid]

Droid review complete.

LGTM — no new issues found. The previously reported concurrency hazard (resetMcpClient closing shared connections under concurrent in-flight callers) has been properly resolved by commit 01c5a9b, which adds activeCalls reference counting and closeWhenIdle deferral so connections are only torn down after the last caller drains. All six lifecycle tests pass and type checks are clean. The toToolResult details field change from full MCP result to structuredContent is a deliberate API change per the PR scope, with tests updated to match.

View job run

[factory-droid]

Droid finished @EdwardIrby's task —— View job

---

---

Pass 2 validation complete. No candidate comments were produced by Pass 1, so there are no inline review comments to post.

Summary: The pooled MCP client connection design is correct: activeCalls reference counting and closeWhenIdle deferral properly prevent closing shared connections while concurrent callers are in flight, resolving the concurrency hazard reported against an earlier commit on this branch. The retry-once logic, session_shutdown cleanup, toToolResult error/structuredContent handling, and auth-vs-public cache key isolation are all sound. No high-confidence bugs or security vulnerabilities were found.