Environment: Linux web development app, Chromium preview, 1280x800, light theme.
Codex CLI 0.169.0; model gpt-6-luna; trusted SessionStart + Stop command hooks.
Fresh independent T3 databases; synthetic git project; same Codex home.
Control: plain codex exec completed with ok, SessionStart + Stop recorded.
Each server chat turn completed with ok, SessionStart + Stop recorded in app-server (expected).
Base 73cbc6b1ede1692c2642b01cf55d9ef75c37c5f4:
  Regenerate title, --enable hooks: SessionStart + Stop from codex exec.
  Repeated recorded regeneration, --enable hooks: SessionStart + Stop from codex exec.
  Regenerate title, --enable=hooks --enable codex_hooks: SessionStart + Stop from codex exec.
  Regenerate title, empty launch args: SessionStart + Stop from codex exec.
Head 69710bb30f2963d4ce170569401ad9e8d40a78d8:
  Regenerate title, --enable hooks: Title kept; no hook entries added.
  Regenerate title, --enable=hooks --enable codex_hooks: Title kept; no hook entries added.
  Regenerate title, empty launch args: Title kept; no hook entries added.
The terminal panel runs show-hooks.py against the actual hook log. It counts exec hooks after the case offset.
The chat count is cumulative across the two independent servers sharing the synthetic hook log (2 before, 4 after).
Title kept is the successful generation result: the model retained the already-fitting title.
The recordings cover the --enable hooks regeneration on each commit.
No tests or fake provider were used as evidence.