[Feature proposal] Edit the latest user message and regenerate in the same session

[Myoontyee]

An idle Desktop conversation could expose an Edit action on its latest ordinary user message. Saving would regenerate in the same Session instead of requiring a new conversation.

The reference implementation keeps attachments and response annotations, retains failed editing drafts, and checks for active or queued work and stale message sequences. It uses the existing append-only surface replacement mechanism; original events remain in the audit log while superseded prompt/reply rows are hidden in Chat. Completed file edits and other tool effects are not undone. A prompt already removed from the active context by compaction is refused rather than reconstructing context speculatively.

Validation includes real Agent-loop regeneration, durable revision records, and real Windows Electron editing and reload tests. The new model request excludes the old prompt and reply. This also preserves annotations described in #8750.

Reference implementation: 9d9544d6fc. Based on upstream dsh-v0.2.1-alpha.1. This is a feature proposal and reference implementation, not a request to merge an external PR.

[Myoontyee]

Implementation update: fixed a cold-restart corruption bug in my custom same-session editing prototype. This is a defect in the customization, not an upstream bug report.

The earlier implementation appended an empty developer/message surface-replacement marker using the previous turn/step after those scopes had already ended. The live view could still work, but a cold read correctly rejected the saved log with developer/message does not match an open turn and step. A renderer reload did not exercise this path.

The fix persists the edit range with the queued user prompt, then writes the standard replacement only inside the newly admitted open step. Rejected admission preserves the original model history. Released V4 validators and historical format generations are unchanged; the new optional queue field has a same-version compatibility record. Existing corrupt files are not silently rewritten by this source change.

Validation:

• Reproduced the exact failure with the previous implementation using the real compressed persistence backend.
• 11 focused tests pass across message editing, revision projection and the editor, including completed and queued edits saved to disk and reopened in a new context, continued input, and rejected admission.
• A real Electron smoke edits, fully quits, reopens the same persisted Session, edits again, then fully quits and reopens again. All three desktop processes exited; the same Session reopened and superseded prompts were absent from model input.

Source commit · Desktop restart regression.

The local Windows build retains the earlier Chinese IME, Markdown/TeX selection-copy, deep-link and other customization changes. No personal conversation data is included in these tests or this post.

Ctrl+Enter follow-up: the inline editor now saves and regenerates with Ctrl+Enter, while Enter inserts a newline. Composition, repeated keydown and disabled/busy states do not submit. Six focused editor tests pass; the real Desktop smoke uses Ctrl+Enter, regenerates, then closes and reopens the same Session successfully. Shortcut change.

[Myoontyee]

Update on alpha.2: the inline message editor now offers slash suggestions for current user-invocable skills and Host commands. A skill pick inserts the invocation into the edited prompt and Save/Regenerate preserves its annotations and attachments. A leading Host command changes the action to Execute command, uses the command controller, and does not simultaneously rewrite the previous prompt. Ctrl+Enter follows the displayed action; IME composition remains guarded.

The implementation explicitly injects the required Remote namespaces. During real Desktop testing we caught and fixed a prototype namespace collision (an installation endpoint named install conflicted with an internal Remote method); the corrected endpoint is installSkill.

Validation: actual Electron application, installation of a new skill while the conversation was already open, edited invocation reaching the model, command execution, and closing/reopening the same stored conversation; focused editor/composition tests and typecheck passed.

Source: https://github.com/Myoontyee/deepseek-harness/commit/0d6e8af4925d42704be2377013a5082ae4658f72

[Myoontyee]

定制分支更新：从“编辑最后一条”扩展为“从较早消息重新继续”

当前源码 143b420 已将编辑范围扩展到较早的人工用户消息：最新消息仍按已有编辑路径处理；选择较早消息时，后续可见对话尾部被撤回，再按修改后的输入继续。原始事件保留在日志中，已发生的文件修改、命令和外部操作不会因此撤销。运行中或存在排队任务时仍不可编辑。

源码与 Host/Client 测试改动：
https://github.com/Myoontyee/deepseek-harness/commit/143b42096fb4ea5212c1890b0a6af9f8edc325c5

建议将后续产品契约明确为两个独立动作：

• 编辑并从这里继续：预览受影响的后续对话范围，说明执行副作用不会回滚；提交时由 Host 再次校验当前日志版本、源消息和任务状态。
• 从这里分叉：保留原会话，使用原生 fork 能力创建另一条会话；不能将它伪装成原地编辑。

附件、批注和用户正文作为同一次编辑的快照；未保存草稿、并发新消息、已压缩或不支持重写的来源需要明确的拒绝/恢复行为。远端编辑在所属 Host 上执行，失联后先查提交结果，不盲目重复生成。

独立验收：较早/最新消息、运行和排队保护、版本冲突、批注附件保留、重启后的历史投影、旧尾部不再次进入当前模型请求，以及已执行文件操作仍存在的清楚提示。

这是当前实现补报及后续验收建议，不表示本轮已重新完成全部平台或真实模型测试。