fix(agent-server): bypass lifecycle lock for cached event services

[jstar0]

HUMAN:

---

AGENT:

Why

ConversationService._get_or_load_event_service waits on the global lifecycle lock even when the requested conversation already has an open in-memory EventService. During a slow persistence/runtime preparation, those waits can consume the server's worker threads and block unrelated event endpoints.

Summary

• Return an open cached EventService before entering lifecycle serialization or disk hydration.
• Preserve the existing activity touch used to defer idle eviction.
• Add a regression test that holds the lifecycle lock and verifies cached access completes immediately.

Issue Number

Fixes #4514

How to Test

• uv run pytest -q tests/agent_server/test_conversation_service.py -k 'cached_event_service_bypasses_lifecycle_lock or get_event_service or conversation_lifecycle' --maxfail=1
• uv run pre-commit run --files openhands-agent-server/openhands/agent_server/conversation_service.py tests/agent_server/test_conversation_service.py
• make build

The focused regression/lifecycle tests pass, and the relevant pre-commit hooks pass: Ruff format/lint, pycodestyle, pyright, import rules, and tool registration. The full conversation/event service files were also run; unrelated workspace/git initialization tests fail in this local outer worktree because pytest temporary paths are detected as Git-backed and therefore use conversation worktrees. The changed code is not involved in those failures.

Video/Screenshots

Not applicable for this backend concurrency fix; the regression is covered by an in-process async test.

Type

• Bug fix
• Feature
• Refactor
• Breaking change
• Docs / chore

Notes

This is an internal agent-server scheduling change with no REST/WebSocket schema changes. The cached fast path is limited to open in-memory runtimes; closed or uncached conversations retain the existing lifecycle and hydration path.

[neubig]

This PR is superseded by merged PR #4570, which replaced the global lifecycle lock with per-conversation locking. That broader fix removes the cross-conversation blocking scenario this cached-event-service fast path targeted.

This comment was created by an AI agent (OpenHands) on behalf of the user.

[neubig]

Thank you for the focused investigation and implementation. Closing this PR as superseded by merged PR #4570, which addresses the broader lifecycle-lock problem with per-conversation locking.

This action was performed by an AI agent (OpenHands) on behalf of the user.