v0.11.2
Fail-loud safety and documentation correction for the upstream Codex Hook-dispatch regression.
Important compatibility notice
Some current Codex Desktop and non-interactive CLI builds can discover a Hook and show it as Active/Trusted, yet never invoke its handler. This is tracked upstream at openai/codex#21639.
Registration, Trust, restart, and policy self-tests do not prove that the current session is dispatching Hooks. Before write-capable md-swarm / md-iterate work, both real actions printed by probe_live_hooks.py --prepare must be explicitly DENIED.
Changes
- Live-probe failure now stops write-capable work and points to the upstream regression.
- If Codex already shows Active/Trusted but either probe action lands, the guidance no longer loops reinstall / Trust / restart.
- The safe fallback is to switch to a host that passes both live DENY checks, currently Claude Code, and retest Codex only after its runtime or configuration actually changes.
- New self-test assertions keep the upstream link, no-loop instruction, and safe-host fallback from silently disappearing.
- README, INSTALL, CHANGELOG, and the user guide now clarify that v0.11.1 fixed payload-shape compatibility only when the host actually dispatches Hooks.
Upgrade
- Pull/sync v0.11.2.
- Re-run the hook installer only if the adapter is missing or untrusted.
- Run the current-session live probe.
- If both actions are not explicitly DENIED, stop write-capable md-paper work in that host and use a verified host.
All portable CI checks pass.
本版本针对 Codex 上游 Hook 调度回归做“响亮停机”和文档勘误:Active/Trusted 不等于处理器真的执行。写入型 md-swarm / md-iterate 之前,只认 current-session live-probe 的两个明确 DENY。若动作仍落盘,不要继续循环重装、Trust、重启;切换到已经通过双 DENY 的宿主,当前推荐 Claude Code。