fix(agent): flush messages before compression session rotation (#46567)

[Tranquil-Flow]

Summary

Fixes #46567 by flushing pending messages to the SQLite session DB before compression closes and rotates away from the current session.

The initial fix added the missing flush before end_session(...). During Path B review, I tightened the patch so the compression-time flush receives conversation_history, which preserves the existing _flush_messages_to_session_db(...) contract: already-persisted history rows are skipped, and only the pending mid-turn tail is written to the old session before rotation.

What changed

• compress_context(...) accepts optional conversation_history and passes it through to _flush_messages_to_session_db(...) before end_session(...).
• The main conversation loop supplies the active turn's conversation_history when auto-compression fires.
• Added regression coverage for:
  • flush is called before end_session
  • flush runs against the old session id before rotation
  • flush is present during compression rotation
  • conversation history is passed through to avoid duplicate SQLite rows

Verification

• RED proof from build artifact: 2/3 original regression tests failed on upstream/main because compression never called _flush_messages_to_session_db(...) before rotation.
• Focused tests after reviewer revision:
  • /Users/evinova-self/.hermes/hermes-agent/venv/bin/python3 -m pytest tests/agent/test_compression_flush_before_rotate.py -v -o "addopts=" --tb=short
  • Result: 4 passed
• Nearby compression tests:
  • /Users/evinova-self/.hermes/hermes-agent/venv/bin/python3 -m pytest tests/agent/test_compression*.py -v -o "addopts=" --tb=short
  • Result: 9 passed, 1 warning (discord.player audioop deprecation)
• Branch check: git rev-list --left-right --count upstream/main...HEAD → 0 1

Duplicate / competitor check

• Same-author PR search for 46567 in title/body: none
• Same-author branch search for fix/46567*: none
• Open PR issue-number search: none
• Open PR keyword search for compression flush session rotation: none

Auto-published by Moonsong via Path B automated pipeline.

[liuhao1024]

✅ Code Review — Clean

Reviewed the full diff (4 files, +214/-2). This is a well-scoped regression fix for #46567.

What was checked:

1. Flush ordering — _flush_messages_to_session_db is correctly placed before end_session in compress_context, and before session_id rotation. The test test_flush_called_before_end_session verifies this via call_order tracking.

2. conversation_history threading — The new parameter flows correctly from conversation_loop.py → _compress_context → compress_context → _flush_messages_to_session_db. The test test_flush_receives_conversation_history_to_avoid_duplicate_rows confirms the value passes through.

3. Dead variable check — conversation_history is used exactly once (passed to _flush_messages_to_session_db). No dead code.

4. Exception ordering — No new exception handling; the existing commit_memory_session call precedes the new flush, which is the correct order (memory extraction on the old session, then flush, then close).

5. Test coverage — 4 tests covering: call order, session ID used for flush, regression on unpatched code, and conversation_history forwarding. All use call_order tracking for verification.

No issues found. The fix correctly addresses the data loss in intermediate sessions during chained mid-turn compressions.

[teknium1]

Closing — the flush-before-rotate fix for #46567 landed via #48584 (salvaged from #47215, authorship preserved). That PR also fixed the sibling #47202 with the same root cause, plus a follow-up commit guarding the /compress unconditional rewrite (#44794/#39704). Your larger variant here is superseded by the merged minimal fix. Thanks for the detailed diagnosis and repro.