fix(agent): persist repaired-turn responses

[teknium1]

Summary

SessionDB flushing now persists newly appended assistant responses even when message-sequence repair has shortened the live messages list below conversation_history length.

Changes

• run_agent.py: replace positional flush slicing with per-session message identity tracking, so history dicts are skipped by identity and newly appended dicts are written once.
• tests/run_agent/test_identity_flush.py: cover repair-shrunk history, stale cached-agent cursor, same-turn dedup, and cached-agent turn reset.
• tests/run_agent/test_compression_persistence.py: flip the stale-history regression from documenting the drop to asserting persistence.

Validation

Check	Result
python3 -m py_compile run_agent.py tests/run_agent/test_identity_flush.py tests/run_agent/test_compression_persistence.py	pass
scripts/run_tests.sh tests/run_agent/test_identity_flush.py tests/run_agent/test_compression_persistence.py tests/run_agent/test_860_dedup.py	18 passed

Fixes #46053.

Infographic

[github-actions]

🔎 Lint report: hermes/hermes-82ce03d4 vs origin/main

ruff

Total: 0 on HEAD, 0 on base (➖ 0)

🆕 New issues: none

✅ Fixed issues: none

Unchanged: 0 pre-existing issues carried over.

ty (type checker)

Total: 10895 on HEAD, 10893 on base (🆕 +2)

🆕 New issues (2):

Rule	Count
unresolved-attribute	2

First entries

tests/run_agent/test_credits_notices_toggle.py:76: [unresolved-attribute] unresolved-attribute: Unresolved attribute `_credits_session_start_micros` on type `AIAgent`
run_agent.py:2920: [unresolved-attribute] unresolved-attribute: Object of type `Self@get_credits_spent_micros` has no attribute `_credits_session_start_micros`

✅ Fixed issues (1):

Rule	Count
invalid-assignment	1

First entries

tests/run_agent/test_credits_notices_toggle.py:76: [invalid-assignment] invalid-assignment: Object of type `None` is not assignable to attribute `_credits_session_start_micros` of type `int`

Unchanged: 5724 pre-existing issues carried over.

Diagnostics are surfaced as warnings — this check never fails the build.

[liuhao1024]

Verification: LGTM — solid identity-based flush fix

Reviewed the full diff. The positional-slice approach (messages[flush_from:]) was fragile when repair_message_sequence merged/shrunk the history — len(conversation_history) > len(messages) made the slice empty, silently dropping assistant responses (#46053).

The replacement using id(msg) identity tracking is correct:

• History dicts are skipped by identity (same object in both messages and conversation_history)
• New dicts appended during the turn are written once, even if repair compacts the list around them
• Repeated flush calls preserve bug: SQLite session transcript accumulates duplicate messages (3-4x token inflation) #860 dedup behavior via the flushed_ids set
• Session-id scoping resets the identity window on session rotation

Test coverage is thorough — 4 new tests in test_identity_flush.py covering the exact regression scenario (stale cursor, repeated flush, cursor reset on new turn). The existing test_flush_with_stale_history_loses_messages is updated from "verifies the bug exists" to "verifies the fix works" — clean transition.

One note: _last_flushed_db_idx is still set to len(messages) at the end (line ~1633), but it's no longer used for slicing — only for the session-id reset gate. This is harmless but could be cleaned up in a follow-up if desired.

[teknium1]

Live testing added after CI because this touches the core persistence path:

Scenario	Result
Direct real SessionDB + real AIAgent._flush_messages_to_session_db with len(conversation_history)=8, len(messages)=3	Old positional flush_from=8 would be empty; PR persisted the new user + assistant rows and repeat flush kept row count stable at 11
Headless real Hermes CLI from this worktree with isolated HERMES_HOME	Provider returned LIVE_PERSISTENCE_OK; state.db contained both user and assistant rows, assistant finish_reason=stop
Real AIAgent.run_conversation() with poisoned DB-style history of 7 consecutive persisted user rows	repair_message_sequence merged the history path, model returned LIVE_REPAIR_FLOW_OK, state.db persisted the assistant row; repeat _persist_session kept row count stable at 9

The repair-flow test is the sensitive case from #46053: persisted conversation_history is longer than the repaired working messages list. The assistant row still landed in state.db.