Agents - snapshot retention deletes newer numbered snapshots before older ones

[mlguys]

Area

Agents & chat

What happened

Session snapshot retention deletes numerically newer snapshots while preserving
older ones because numbered snapshot filenames are sorted lexicographically.

At tick 158 with a 100-snapshot cap, the retained tick ranges were 2-9,
16-99, and 151-158. The expected retained range was 59-158.

The complete sanitized observation is attached as
condor-snapshot-retention-evidence.txt.

Expected

Retention should keep the newest 100 numeric ticks and delete the oldest numeric
ticks, regardless of how many digits appear in the filename.

Likely source path

• MAX_SNAPSHOTS = 100
• _cleanup_old_snapshots() calls sorted() directly on snapshot_*.md paths, producing lexicographic rather than numeric tick order.

A minimal fix is to sort valid snapshot paths by the parsed integer tick before
pruning. A regression test should cross the filename-width boundary: with ticks
1 through 102 and a cap of 100, cleanup should retain exactly ticks 3 through
102. Existing snapshot listing tests are in
tests/test_sessions_index_snapshots.py; retention cleanup currently lacks this
boundary case.

Steps to reproduce

1. Create an isolated temporary session snapshots directory.
2. Create snapshot_1.md through snapshot_102.md.
3. Invoke JournalManager._cleanup_old_snapshots().
4. Observe that lexicographic cleanup removes snapshot_1.md and
   snapshot_10.md.
5. The correct numeric cleanup would remove snapshot_1.md and
   snapshot_2.md.

Environment

Observed in a private downstream checkout; no private branch or source is
published with this report. The same cleanup implementation was independently
confirmed in public upstream Condor commit
a3997a2b.

Logs

Attach: condor-snapshot-retention-evidence.txt

[mlguys]

condor-snapshot-retention-evidence.txt

[mlguys]

Some snapshots are skipped: