fix(agent-server): lazy-load persisted conversations on access

[nehaaprasaad]

Why

Agent server startup eagerly hydrated every persisted conversation, creating full EventService and LocalConversation objects for all conversations on disk.

Summary

• Lazy-load persisted conversations only when runtime access is needed.
• Keep list/get/count paths lightweight by reading persisted metadata/state.
• Add regression coverage for lazy startup and single hydration.

Issue Number

fix : #3140

How to Test

• make build
• uv run pytest tests/agent_server/test_conversation_service.py -k "startup_does_not_eagerly or hydrates_persisted or stale_owner_cannot_append or restart_resumes"
• uv run pre-commit run --files openhands-agent-server/openhands/agent_server/conversation_service.py tests/agent_server/test_conversation_service.py

[all-hands-bot]

🔴 Needs improvement - Critical runtime error (missing import) and concurrency issues that need addressing before merge.

[CRITICAL ISSUES]

• Missing Import: importlib is used but not imported - will cause NameError at runtime
• Race Condition: Hydration lock creation in get_event_service() is not thread-safe
• Suboptimal Delete Path: delete_conversation() now hydrates before deleting, adding unnecessary overhead

[IMPROVEMENT OPPORTUNITIES]

• Memory Leak: _hydration_locks dict grows unbounded and is never cleaned up
• Cache Invalidation: _stored_conversations cache is never invalidated if metadata changes externally
• Test Naming: test_restart_resumes_conversations_after_non_graceful_shutdown name doesn't match its new assertion

[RISK ASSESSMENT]

• [Overall PR] ⚠️ Risk Assessment: 🟡 MEDIUM

Performance-critical path (startup, conversation access) with good test coverage, but the missing import is a production blocker. The race condition in lock management could cause subtle bugs under concurrent load. The lazy-loading approach is sound and valuable - just needs these issues fixed before merge.

VERDICT:
❌ Needs rework: Fix the missing import and race condition before merging

KEY INSIGHT:
The dual-state model (hydrated EventService vs metadata-only StoredConversation) is elegant, but the synchronization primitives need attention to prevent race conditions under concurrent access.

[csmith49]

Hey Neha, this is a great start! Do you mind addressing (or responding to) the review comments and checking on the failing CI?

[nehaaprasaad]

@csmith49 updated and addressed the comments please check thankyou.

[all-hands-bot]

🔴 Needs rework - Critical implementation bugs that will cause runtime failures.

Multiple methods are missing critical implementation code (return statements, loops, helper methods). These issues will cause immediate failures at runtime and must be fixed before merging.

[csmith49]

Looks like we had a bad merge or something? Suddenly lots of weird formatting issues and missing chunks.

[nehaaprasaad]

@csmith49 please check the updates!

[csmith49]

Hey Neha, you might want to look at the CI results. There's still some malformed code that is stopping our test suite from running properly.