Files
hermes-agent/hermes_cli
teknium1 c9f14820b6 fix(update): a git dir that cannot lock still repairs; wait out a live git before scanning
Review of the restore claim found three gaps:
- _lock_fd counted every OSError as "held", so on a filesystem without locks
  (ENOLCK on NFS without lockd, EOPNOTSUPP on some SMB shares) every launch
  waited 10 s, printed "another launch is repairing" and never restored, which
  worked before claims existed. Only the contention errnos mean "held" now;
  anything else proceeds unguarded, like the read-only git dir path.
- After a claim holder was kill -9'd, its surviving `git restore` child could
  still be writing when the next launch scanned the tree ("Could not check...
  could not open d/mNNN.py", 5/40 trials). The index.lock wait now runs
  before the scan instead of only before the restore.
- _held_open's docstring claimed /proc is exact and the claim guards an outside
  git; neither holds. It now states the real limits (and the Limits list names
  agent/__init__.py and agent/jiter_preload.py for the hermes-agent script).
2026-09-23 18:26:45 -07:00
..
…
…
…
…
…
…
…
…
…
…