Ruya-AI/cozempic v1.8.20

v1.8.20 — prune-counter integrity + prune-safety defense-in-depth

on GitHub

Highlights

🔢 Prune-counter integrity (stats fix + loop containment)

The guard recorded prune savings on projected numbers before the deferred write was confirmed — so a deferred, read-only, or looping cycle that never actually wrote still incremented the global prune/tokens counters every interval. This was the in-the-wild prune-counter spike (timed to the 1.8.18 guard rollout). Savings are now recorded exactly once per persisted prune, across both the terminate-first reload path and the overflow path; futile/looping cycles record nothing.

🛡️ Prune-safety defense-in-depth (PR #114, @ynaamane + C8)

Post-prune structural validation (validate_post_prune, checks C1–C8), floor preservation, and last-of-type metadata-singleton protection (incl. permission-mode, the original unresumable-session root cause):

• C1 baseline-relative parent resolution, C2 orphan-shell-aware root preservation, C4–C7 singleton protection.
• C8 (new): a surviving tool_use whose paired tool_result was pruned is rejected (a dangling tool_use is structurally valid but unresumable).

🪟 Windows guard fixes (@GravyaDev)

• #112: auto-prune os.replace onto a host-held transcript now defers gracefully instead of crashing (WinError 5).
• #113: pidfile publish uses os.replace so the guard starts cleanly over a stale .pid (WinError 183).

🧹 Review nits

• background-poll-collapse: per-message protection re-check (defense-in-depth).
• floor: preserve_last_k_turns counts real conversational turns, not tool-result envelopes.

Thanks to @ynaamane, @GravyaDev, and @stephenfeather. Full QA: multi-agent adversarial review + 1100+ tests green.