Finding / omp

2026-10-13-omp-invalid-approval-flags-and-late-ast-guards-were-fixed-but-eval-still-sits-outside-those

Invalid approval flags and late AST guards were fixed, but eval still sits outside those barriers. Since v18.4.3 an invalid --approval-mode exits 2 before ordinary or ACP operation instead of quietly using configured policy, normally yolo. Since v18.4.1 directly dispatched AST-conditioned TTSR guards are awaited before execution. Previously the late check deterministically allowed writes on chat-completions end-of-stream paths; other transports were timing-dependent. Eval-bridged calls still receive only a reminder after execution. bash.patterns deny rules do not apply to commands spawned through eval, which keeps ambient filesystem, network and subprocess access. The eval computer API is disabled by default; its read_only guard is not a general JavaScript sandbox.

Channel: tagged-release. Half: defect. Date: 2026-09-28 (v18.4.1); 2026-09-29 (v18.4.3).

Operator consequence: Re-run launch scripts to find invalid approval flags. Gate eval with tools.approval.eval: prompt or deny when bash patterns or AST rules matter. Prompting removes unattended eval autonomy; a deny removes that execution surface entirely.

Receipt

Finding metadata

Run: 2026-10-13-profile-refresh-2026-09-21_2026-10-13-frontier-v0

Finding ID: 2026-10-13-omp-invalid-approval-flags-and-late-ast-guards-were-fixed-but-eval-still-sits-outside-those

Profile citations

  • omp / claim / eval-orchestration-ungated-by-patterns

Source links

Primary links, including exact changelog lines when available.