fix(memory): keep max_msg_id valid when an overwrite rollback empties the cache

[Linxiushen]

Change Summary

DefaultMemory._delete_single() (ms_agent/memory/default_memory.py) recomputes max_msg_id with max(self.cache_messages.keys()) after popping the entry that held the maximum. With history_mode='overwrite', add() rolls back every cached block starting from the first one whose hash no longer matches the incoming history. When the very first block mismatches — a second conversation reusing the same memory directory, or an edited/cleared history — the last cached entry is deleted as well, the dict is empty, and max() raises:

File "ms_agent/memory/default_memory.py", line 449, in add
    self._delete_single(msg_id=msg_id)
File "ms_agent/memory/default_memory.py", line 181, in _delete_single
    self.max_msg_id = max(self.cache_messages.keys())
ValueError: max() iterable argument is empty

The exception leaves add() uncaught, reaches run_loop's handler as ErrorRaised(recoverable=False) and terminates the whole agent run. It is not hit with the default history_mode='add'; it needs history_mode: overwrite in the memory config (the mode has its own test, test_overwrite_with_tool, which only covers a partial rollback where block 0 still matches).

Fix: max(self.cache_messages.keys(), default=-1). -1 is the sentinel load_cache() already uses for an empty store, so add_single()'s self.max_msg_id += 1 numbers the next block from 0 exactly as on a fresh store; verified that "full rollback, then add a block" ends in the same state as "add the same block to a new store" (cache keys, max_msg_id, block hash, memory_snapshot).

Tests: tests/memory/test_default_memory_rollback.py — two cases (second session with a different first turn; rollback that empties the cache and re-adds), with mem0 stubbed since only the cache bookkeeping is under test. Both fail on main with the ValueError above and pass with the change. A differential over the well-formed overwrite/add scenarios (build 3 blocks, resubmit same history, partial rollback, tail rollback, reload from disk, delete a non-max id) produces identical state dumps before and after; pytest tests/memory goes from 176 to 178 passed with the same pre-existing environment failures.

flake8 / isort / yapf (repo setup.cfg) clean on the changed lines (E126/W604 reported elsewhere in the file pre-exist on main).

Related issue number

None found (searched max_msg_id, _delete_single, history_mode, overwrite, max() iterable).

Checklist

• The PR title is a changelog-ready summary of the change
• Unit tests added for the change
• pre-commit checks (flake8 / isort / yapf from setup.cfg) pass on the changed files
• Documentation updated — not needed, no user-facing behaviour change beyond not crashing

Written with Claude Code (AI-assisted) and submitted under the account owner's authorization.