CCB v8.5.2

Released: 2026-07-30

English

A replacement pane now remains in a 90-second probation window and must
receive a new healthy observation before recovery succeeds. Queued work is
held while the runtime is probing.
Consecutive unstable recoveries back off for 30 seconds, 60 seconds,
120 seconds, 5 minutes, 10 minutes, and 30 minutes. After six attempts CCB
opens a recovery circuit and stops further respawns, crash captures, and
dispatch until an explicit restart or remount.
Each Provider runtime retains only the newest 50 pane-crash logs and matching
reason records. Pane history is cleared before respawn, and unchanged helper
manifests are no longer rewritten.
A stale CCB-managed Claude --continue binding can be removed without
changing Provider authentication. A missing managed Codex app server now
fails closed with an actionable restart/remount diagnostic.

Stable reply and cancellation guidance now lives once in managed project
memory instead of being repeated in every ask prompt.
Normal asks no longer inject the default reply block or require the Agent to
poll a cancellation file before every work step. Compact and silent asks add
only their requested mode metadata, while runtime interruption remains the
primary cancellation mechanism.

CCB Rich WezTerm launches now use a new process session with stdin, stdout,
and stderr detached from the invoking shell. GUI diagnostics and job-control
signals no longer pollute or stop the parent terminal job.
On Linux Wayland, the workbench provisions a private XCursor overlay for the
missing hand cursor while preserving the selected cursor theme and leaving
user and system cursor files unchanged.

No configuration migration is required.
Restart already-running CCB projects after upgrading so they load the new
recovery and managed-memory behavior.
Existing Rich users should run ccb update rich once to regenerate the
detached launcher and Wayland cursor compatibility overlay.
If an Agent reaches recovery-circuit-open, repair its Provider/session
state and then run ccb restart <agent> or remount the project.
npm-managed installations can run
npm install -g @seemseam/ccb@8.5.2.
Android users can install the signed CCB Mobile v8.5.2 APK from this release.

替换后的 pane 现在会进入 90 秒观察期，并且必须获得一次新的健康观测后
才能确认恢复成功；runtime 处于探测状态时，队列任务会保持等待。
连续不稳定恢复依次退避 30 秒、60 秒、120 秒、5 分钟、10 分钟和
30 分钟。第六次仍不稳定时，CCB 会打开恢复熔断，并停止继续 respawn、
捕获崩溃和派发任务，直到显式 restart 或 remount。
每个 Provider runtime 只保留最新 50 份 pane crash 日志及对应原因记录。
respawn 前会清空 pane history，内容未变化的 helper manifest 也不再重复
写盘。
CCB 可以移除失效的 Claude --continue 绑定，同时不改动 Provider
登录状态。托管 Codex app server 缺失时则会安全停止，并给出明确的
restart/remount 诊断。

稳定的回复和取消规则现在只写入一次托管项目记忆，不再在每次 ask
prompt 中重复注入。
普通 ask 不再附加默认回复段落，也不再要求 Agent 每一步都轮询取消文件。
compact 和 silent ask 只添加所需的模式元数据，runtime interruption
继续作为主要取消机制。

CCB Rich WezTerm 现在使用独立进程会话启动，并断开与调用 shell 的
stdin、stdout 和 stderr。GUI 诊断及 job-control 信号不会再污染或暂停
父终端任务。
在 Linux Wayland 下，workbench 会提供私有 XCursor overlay，补齐缺失
的 hand cursor，同时保留用户选择的 cursor theme，且不会修改用户或
系统 cursor 文件。

无需迁移现有配置。
升级后请重启已经运行的 CCB 项目，使其加载新的恢复和托管记忆行为。
已启用 Rich 的用户应执行一次 ccb update rich，重新生成隔离 launcher
和 Wayland cursor 兼容 overlay。
如果 Agent 进入 recovery-circuit-open，请先修复其 Provider/session
状态，再执行 ccb restart <agent> 或重新挂载项目。
npm 安装可执行：
npm install -g @seemseam/ccb@8.5.2。
Android 用户可从本 Release 安装签名后的 CCB Mobile v8.5.2 APK。