0.15 Alpha: routing authority, continuous interaction and release tracking / 路由权限、连续交互与发布跟踪

[Lucas1479]

Purpose and current status

Track the product goals, cross-PR dependencies, design decisions, and release acceptance for 0.15 Alpha. Implementation is already underway in Draft PRs. This issue adds a shared discussion and acceptance entry point; implementation details and test records remain in the individual PRs and linked documents.

The release must achieve three goals together: correctly understand who the user wants to do what, preserve existing task context, and keep character responses timely. Better routing accuracy must preserve task continuity and character responsiveness.

User-visible behavior and authority boundaries

• Questions or progress requests addressed to an existing executor remain messages. They must not automatically add requirements or start another execution.
• Amendments retain the existing Work identity and accepted requirements. An independent deliverable creates a separate Draft.
• Ordinary Chat, Provider messages, Work changes, and application actions remain distinct, with the correct source and target.
• The character may begin responding before professional planning finishes. Models own interpretation and expression; the Host owns identity, permissions, acceptance, and execution. A character response does not itself authorize execution.
• Stop, recovery, and replay preserve task identity and permission constraints without duplicate execution.

Implementation scope and dependencies

PR	Responsibility	Current relationship
#74	Routing authority, cooperative addressing and continuation, professional Work planner, recovery, and related AUIP boundaries	Core change targeting main; Draft
#75	Compact Companion panel, VN-style VOICE animation, and caption restoration on reopen	Independently based on main; Draft; existing product discussion in #61
#76	Optional, experimental ACP Work Providers	Stacked on #74; Draft; no real production-use evidence yet
#77	Complete staged Slice export review and associated backend permission checks	Stacked on #74; Draft

For #76 and #77, merge the base first, then retarget to main and verify the resulting diff and checks. Review each PR independently. Before release, explicitly include or defer each companion feature; sharing the 0.15 version does not establish acceptance.

Current design and compatibility

#74 currently enables cooperative routing and the professional planner by default. Disable only the planner to use basic cooperative routing. Disable COOPERATIVE_CHAT_ENABLED and restart to return to the original Chat routing implementation. These are alternative configurations, not three serial model passes. A failed professional plan does not automatically switch strategies.

Rollback retains shared identity, Session, and recovery fixes; it does not promise to restore every historical implementation detail. Basic cooperative routing still has a documented limitation around explicit Provider constraints. Preserve compatibility with public installation profiles, the CPU/model-less path, Linux AEC, macOS wallpaper and keyboard behavior, and optional character retrieval.

Release acceptance checklist

These checkboxes track acceptance of the final release candidate. Existing focused test evidence does not automatically complete an entire item.

• Routing and continuity: cover ordinary Chat, questions to an existing executor, amendments to existing Work, independent deliverables, and source/target identification. Confirm no unintended task creation or loss of accepted requirements.
• Authority and recovery: verify Host acceptance, permission denial, stop, recovery, and replay without unauthorized or duplicate execution.
• Character responsiveness: with the professional planner enabled, verify warm ordinary Chat and ordinary Chat after a professional task. Under the existing definition, typed submission to the first non-empty audio-device write should be ≤1.5 seconds. Record the environment, model, cache conditions, and samples; report cold startup and task-response latency separately.
• Configuration and rollback: verify the default professional cooperative, basic cooperative, and original Chat paths. Document configuration changes, known limitations, and any necessary migration steps.
• Integration and compatibility: run the relevant Python/Electron checks, builds, and public source-release checks against the final candidate, and complete the applicable packaged desktop, device, and platform interaction acceptance. Distinguish tested from untested scope.
• Release scope: explicitly include or defer feat: 0.15 Alpha compact VN-style companion panel #75, feat(acp): experimental ACP Work Providers for 0.15 Alpha #76, and feat(slice): review complete staged exports in 0.15 Alpha #77. Complete acceptance for each included feature and retain ACP's optional and experimental status.
• Release notes: synchronize final defaults, compatibility, rollback, known issues, and unverified scope, and complete reviews of all included PRs.

Existing evidence and limits

#74 records historical same-model routing comparisons, deterministic regression, configuration rollback, and combined-branch checks. The two physical warm ordinary-chat first-audio samples were 1.125 seconds and 1.078 seconds; the second followed a completed professional-planner task. These are two historical samples under stated conditions, not a full final-release journey pass or a long-run latency percentile guarantee. Historical routing scores, contract tests, and physical audio evidence support different conclusions and cannot replace one another.

Detailed records:

• 0.15 Alpha scope and E2E definition
• Impact and validation scope
• Historical routing baseline

Discussion and closure

Feedback is welcome on concrete interactions: whether the current default strategy matches expected use, whether task continuity or authority boundaries have gaps, and which companion features should ship in this Alpha.

Close this issue once release scope is settled, included PRs are merged, and final acceptance and release notes are complete. Merging the core PR does not automatically close this tracker. TTS engine replacement, MPS concurrency investigation, and additional texture performance work remain separate discussions and are not added as release prerequisites by this issue.

---

目的与当前状态

集中跟踪 0.15 Alpha 的产品目标、跨 PR 依赖、设计决策和发布验收。实现已在 Draft PR 中展开；本 Issue 补充统一的讨论与验收入口，具体实现和测试明细继续维护在各 PR 及关联文档中。

核心目标是同时做到：正确理解用户想让谁做什么、连续使用已有任务上下文、及时给出角色回应。不能以牺牲任务连续性或角色响应速度来换取路由准确率。

用户可见的行为与权限边界

• 向已有执行者询问进展或提出问题，保持为消息，不自动增加任务要求或发起新的执行。
• 修改已有任务时，保留 Work 身份和已接受的要求；独立交付需求创建单独的 Draft。
• 普通 Chat、Provider 消息、Work 修改和应用操作保持明确区分，正确识别来源与目标。
• 角色可以在专业规划完成前开始回应；模型负责理解和表达，Host 负责身份、权限、接受与执行。角色回应本身不构成执行授权。
• 停止、恢复和重放保持同一任务的权限与身份约束，不能重复执行。

实现范围与依赖

PR	负责范围	当前关系
#74	路由权限、协同寻址与连续交互、专业 Work planner、恢复及相关 AUIP 边界	核心变更，目标分支为 main，当前 Draft
#75	紧凑 Companion 面板、VN 风格 VOICE 动画及重开字幕恢复	独立基于 main，当前 Draft；已有产品讨论见 #61
#76	可选、实验性的 ACP Work Providers	依赖 #74，当前 Draft；尚无真实生产使用证据
#77	Slice 完整暂存导出审阅及配套后端权限检查	依赖 #74，当前 Draft

#76、#77 应先完成基础分支合并，再调整到 main 并核实差异及检查。各 PR 保持独立审查；发布前明确每个附属功能纳入还是延后，不能因同属 0.15 就自动视为已验收。

当前设计与兼容性

#74 当前默认启用协同路由和专业 planner。仅关闭 planner 可使用基础协同路由；关闭 COOPERATIVE_CHAT_ENABLED 并重启可切回原 Chat 路由实现。三种策略是可选配置，不是三个串行模型步骤；专业规划失败后不自动切换策略。

回退保留共用的身份、Session 和恢复修复，不承诺恢复所有历史实现细节。基础协同路由对显式 Provider 约束仍有已记录的限制。公共安装配置、CPU/model-less 路径、Linux AEC、macOS 壁纸与键盘行为以及可选角色检索都需要保留兼容性。

发布验收清单

以下复选框用于最终候选版本验收；已有局部测试证据不自动代表整项完成。

• 路由与连续交互：覆盖普通聊天、询问已有执行者、修改已有 Work、独立交付、来源与目标识别，确认不会误建任务或丢失已接受要求。
• 权限与恢复：验证 Host 接受、权限拒绝、停止、恢复和重放，不产生未授权执行或重复执行。
• 角色响应：专业 planner 开启时，验证预热普通聊天及专业任务后的普通聊天；按已有定义，文本提交至首次非空音频设备写入应 ≤1.5 秒。记录环境、模型、缓存和样本，冷启动及任务回应耗时另列。
• 配置与回退：验证默认专业协同、基础协同和原 Chat 路径，说明配置变化、已知限制及必要迁移方式。
• 集成与兼容：针对最终候选运行相关 Python/Electron 检查、构建及公共源码发布检查，并完成涉及的打包桌面、设备和平台交互验收；区分已测与未测范围。
• 发布范围：确定 feat: 0.15 Alpha compact VN-style companion panel #75、feat(acp): experimental ACP Work Providers for 0.15 Alpha #76、feat(slice): review complete staged exports in 0.15 Alpha #77 各自纳入或延后；若纳入，完成对应验收，并保留 ACP 的可选和实验性标记。
• 发布说明：同步最终默认行为、兼容性、回退方式、已知问题和未验证范围，完成所纳入 PR 的审查。

已有证据与限制

#74 已整理历史同模型路由比较、确定性回归、配置回退和组合分支检查。实机预热普通聊天首声样本为 1.125 秒和 1.078 秒，其中第二个样本位于专业规划任务完成后。这是两个有条件的历史样本，不是最终发布版本全流程通过或长期延迟分位数保证。历史路由评分、契约测试和实机音频分别支持不同结论，不能互相替代。

详细记录：

• 0.15 Alpha 范围与 E2E 定义
• 影响与验证范围
• 历史路由基线

待讨论与关闭条件

欢迎围绕具体交互提出反馈：当前默认策略是否符合使用预期，任务延续或权限边界是否有遗漏，以及哪些附属功能应纳入本次 Alpha。

在发布范围明确、所纳入 PR 合并且最终验收与发布说明完成后关闭本 Issue。核心 PR 合并本身不自动关闭总跟踪。TTS 引擎替换、MPS 并发调查和额外纹理性能优化保持独立讨论，不作为本 Issue 新增的发布前置条件。