v8.7.3

Released: 2026-09-28

English

Fixes and reliability

• Stop advertising six V3 preview roles during install/update onboarding: agentroles.ccb_checker, agentroles.ccb_clarification_broker, agentroles.ccb_plan_reviewer, agentroles.ccb_round_checker, agentroles.ccb_worker and agentroles.su_ccb. Explicit installation and updates for installed roles still work, and missing-source diagnostics remain visible. This is separate from the previously released V1/V2 config-panel filtering.

• Codex input protection no longer depends on a GPT model-name prefix or its capitalization. Recognize supported composer footers by layout and rendering evidence so switching model labels does not strand queued messages. This supersedes the capitalization-only fix in [#361].

• Add regression coverage for Claude's changing suggestions, Tab-accepted drafts, deferred sending and return-before-ask ordering. Claude and OMP already use model-independent observation; this release does not replace their production adapters.
• Preserve the existing current-turn gate, chronological ask/back FIFO and fixed 180-second draft wait. Unknown layouts remain blocked on protected paths.

Verification and upgrade

• 322 focused composer, FIFO and Claude execution checks passed. Real Claude qualification completed 19 jobs with one attempt each, covering consecutive result chains, human-initiated busy turns, early draft clearing, the real 180-second wait and modal exit recovery.
• 188 update, RolePack and config-panel regression checks passed, including the six-role recommendation filter and preserved missing-source diagnostics.
• Update CCB, then stop and start affected projects when their work can safely be interrupted to load the new daemon code. Keep .ccb and provider conversation history. Inspect outstanding jobs before restarting; do not resubmit accepted requests blindly.
• Android metadata follows the existing coordinated release pipeline; this release adds no mobile UI feature. Platform-specific support status and prerequisites remain unchanged.

Known limitations

• Custom single-field Codex status bars without recognized hints or a dim field separator, and wrapped or additional-row footer layouts, remain unqualified. The change does not prove every healthy-pane delivery stall in [#356] fixed; reporter verification remains open.
• Claude native plugin read/fill APIs were tested only in an isolated experiment and are not enabled by CCB. Missed edit events and empty reads during modals or without a bound composer prevent treating that experiment as production-ready.
• Final observation, clearing and terminal paste/Enter are not atomic. Existing limitations around attachments, custom editor modes and clear primitives remain.

中文

修复与稳定性

• 安装/更新引导不再推荐六个 V3 预览角色：agentroles.ccb_checker、agentroles.ccb_clarification_broker、agentroles.ccb_plan_reviewer、agentroles.ccb_round_checker、agentroles.ccb_worker 和 agentroles.su_ccb。仍可显式安装、更新已安装角色，缺失来源诊断也继续显示。这与此前已发布的 V1/V2 配置面板筛选是两项不同修改。

• Codex 输入保护不再依赖 GPT 模型名前缀及其大小写，改用布局和渲染证据识别受支持的输入框底栏，避免切换模型名称后消息卡在队列中。本修复取代 [#361] 中仅兼容大小写的处理。

• 补充 Claude 动态建议、Tab 接受后的草稿、延迟发送和先返回后 ask 的顺序回归测试。Claude 与 OMP 原本已采用不依赖模型名的判断；本版没有替换它们的正式适配器。
• 保留当前处理回合结束门禁、按时间排序的 ask/back FIFO，以及固定 180 秒草稿等待。受保护路径遇到未知布局时仍阻止投递。

验证与升级

• 322 项输入框、FIFO 和 Claude 执行相关检查通过。真实 Claude 验证完成 19 个任务，每个任务仅一次 attempt，覆盖连续结果链、用户主动启动的忙碌回合、提前清空草稿、真实 180 秒等待和退出弹窗后恢复。
• 188 项更新、RolePack 和配置面板回归检查通过，覆盖六个角色的推荐过滤以及保留缺失来源诊断。
• 更新 CCB 后，在受影响项目的工作可安全中断时停止并重新启动项目，使新 daemon 代码生效。保留 .ccb 和 provider 对话历史；重启前检查未完成任务，不要盲目重发已接受的请求。
• Android 元数据沿用现有同步发布流程，本版没有新增手机界面功能。各平台支持状态和外部依赖要求保持不变。

已知限制

• 不带已识别提示或淡色字段分隔符的 Codex 自定义单字段状态栏，以及折行或多行底栏，仍未获得支持验证。本次变更不能证明 [#356] 中所有健康 pane 投递卡住问题均已解决，仍需报告者复测。
• Claude 原生插件 read/fill 接口只进行了独立实验，CCB 不会启用。编辑事件遗漏，以及弹窗或未绑定输入框时返回空文本等问题，仍阻碍其正式接入。
• 最后观察、清空及终端粘贴/Enter 不是原子操作；附件、自定义编辑模式和清空方式的既有限制仍然存在。