[bandzalkin]

Why

ACP rooms need workspace-specific external MCP servers, resolved on the connection that will use them. Shared setup must stay owned through caller cancellation and cleanup.

What

Add optional resolve_mcp_servers(workspace) to the ACP client adapter and its Copilot/Cursor constructors. Validate native descriptors against negotiated transports, preserve Band routing, and reuse one copied snapshot for restore/fresh-session fallback. Join shared setup through RetainedRoom.wait_owned, observe abandoned errors once without exposing exception contents, and retain process custody until child exit and stdout EOF are confirmed.

Scope

ACP resolver, session setup, process cleanup, documentation and regressions under INT-1889. Uses the shared Python retention contract merged in #911 (INT-1905), inherited from main; this PR contains no generic retention changes. TypeScript parity and planning guidance remain separate follow-ups INT-1906 and INT-1907. No idle-retention enablement or resolver rollout to other Python adapters.

Also repairs a pre-existing resync test's synchronization and async activity fixture exposed during CI validation; generic runtime behavior is unchanged.

[linear-code]

INT-1889

[AlexanderZ-Band]

Two lifecycle defects reproduced at the ACP setup boundary. The review also notes one documentation formatting nit.

[AlexanderZ-Band]

Proposed next steps: split the shared lifecycle change into a separate core PR and keep this PR open for ACP MCP resolution.

1. Core PR: agree on the smallest cleanup ownership contract, then implement and test it independently. Protect a caller because it is actively waiting for owned child work, rather than assuming task creation establishes that dependency. Cover awaited work, detached work, cancellation, and unrelated sibling tasks, and check the existing Codex and Claude SDK lifecycle flows.
2. Update this PR: move its src/band/core/retention/room.py changes and corresponding core regression coverage into the core PR, base this PR on that work, and use the agreed ownership mechanism for ACP session setup. Keep MCP resolution, descriptor validation, and shared initializer management in ACP. Also fix the open abandoned-setup exception finding so a late resolver failure is consumed or reported before the initializer is discarded.
3. Review and merge the core PR first, then complete this PR against the merged core behavior. Keep feat: resolve room-scoped MCP servers per connection #896 open to preserve its review history and keep its final scope focused on ACP.

The core design still needs to be settled before coding. This split does not add MCP resolver features to other adapters.

[AlexanderZ-Band]

One verified process-closure regression at the reviewed head.

[AlexanderZ-Band]

One nonblocking duplication nit. The buffered-stdout teardown regression is already recorded in its existing thread and is not duplicated here.

[AlexanderZ-Band]

The requested ownership split is fulfilled at a2af1c4a: shared active-wait cleanup ownership merged separately in #911 (INT-1905), and #896 inherits it from main at b671f899 without generic retention changes. Both ACP setup waits use wait_owned; the ACP-specific observer retrieves terminal errors and reports abandoned failures without exception contents.

The PR description matches the resulting ACP scope. INT-1906 and INT-1907 remain independent TypeScript follow-ups. Verification in this review: 25 focused ownership/setup/error tests passed, and the buffered-stdout reproduction was independently confirmed against the exact base. That inline finding remains open; this is not a readiness or merge approval.

[AlexanderZ-Band]

The earlier ownership-split request is complete: #911 merged into main, and #896 inherits its completion-only RetainedRoom.wait_owned contract without generic retention changes in its own diff. Both live and abandoned setup waits use that contract, with result consumption only on the live path. The remaining review findings are fixed and resolved in published head 193665fc42ab2002d4600a7cb5b9dfab22cde6d1; validation evidence is recorded in INT-1889.