[roadmap] Q4 2026 Agent: reliable execution and recoverable collaboration / 可靠执行与可恢复协作 #261

Parent roadmap: #107. Proposed delivery window: November–December 2026. Priorities and dates are planning targets; owners and capacity require confirmation.

Goal and scope / 目标与范围

Make conversational delegation deliver actual editable project content. Users and Agents must alternate work under the same authorization, object versions and locks. Prioritize a reliable single-Agent creation loop; multi-Agent delivery is not a prerequisite for the November film.

让对话委托产生真实、可编辑的项目内容，用户与 Agent 在相同授权、版本和锁规则下交错工作。优先稳定单 Agent 创作闭环，多 Agent 不作为11月底成片的前置条件。

Work and milestones / 工作与里程碑

• By November 30 · P0: resolve tool discovery, invalid parameters, expired editing sessions and result-delivery failures needed by the reference-video workflow. Cover project/capability discovery, assets, references, generation, adoption, required editing, timeline arrangement and export through shared services, CLI and tools. Provide actionable help and structured results.

• First half of December · P0: correlate conversations, Agent runs, tool calls, jobs and receipts. Recover from disconnection, restart and partial failure; query unknown submissions and reuse completed assets. Distinguish stopping a reply, stopping orchestration and canceling accepted jobs.

• By December 31 · P0/P1: inspect actual outputs, identify local creative defects and rerun affected work only. Compare DSV4.1 and GPT-6 on the same task, assets, permissions and resource constraints.

• December · P1, evidence-gated: compare single-Agent batch orchestration against multi-Agent delegation for independent shots or asset checks. Productize only if measured quality, elapsed time, human intervention and cost justify it. A finding of no benefit is valid.

• Coordinate supported timeout/cancellation and late-result contracts with vLLM-Omni; application operations remain OpenDirector's responsibility. Do not introduce real-time duplex as an implicit dependency.

• **11月底 · P0：**修复参考视频成片所需的工具发现、错误参数、编辑会话过期和产物交付问题；通过共用服务补齐查询、素材、引用、生成、采用、必要编辑、时间线和导出入口，提供结构化结果与可操作帮助。

• **12月上半月 · P0：**关联会话、Agent run、工具、任务和回执；断线、重启及部分失败后可接续，未知提交查询原任务，复用已完成素材；区分停止回复、停止编排和取消任务。

• **12月底 · P0/P1：**检查真实产物、修复局部创作问题并仅重做受影响工作；按同任务、素材、权限和资源约束对比 DSV4.1 与 GPT-6。

• **12月 · P1，按证据决定：**比较单 Agent 批量编排与多 Agent 委派独立镜头/素材检查的收益，再决定是否产品化；无收益也是有效结论。

• 与 vLLM-Omni 对齐已支持的超时、取消和迟到结果语义，应用侧业务仍由 OpenDirector 执行，不默认引入实时双工。

Implementation and dependencies / 实现影响与依赖

Reuse shared project services and preserve runtime boundaries. #166 remains an architecture proposal, not a required runtime rewrite. Coordinate November essentials with the canvas/timeline, provider and workflow roadmaps. Existing implementation tasks: #198, #137, #132, #131, #246, #247, #15, #85, #166. Upstream coordination: vllm-project/vllm-omni#8264.

复用共享业务服务和现有 runtime 边界；#166 作为架构方案评估，不要求重写 runtime。与画布/时间线、provider、工作流协调11月必需能力，已有具体 Issue 继续承接实现。

Acceptance / 验收标准

• UI and Agent can alternate the same project without losing edits or bypassing locks; delivered results remain editable. / 双入口交错工作不丢内容、不绕过锁，结果可编辑。
• Exercise real disconnection, long jobs, restart and partial recovery without duplicate submission or incorrect adoption. / 验证断线、长任务、重启和局部恢复，无重复提交或错误采用。
• Report technical success, Agent quality assessment and user approval separately; state browser participation in export explicitly. / 区分技术成功、Agent质量判断与用户认可，明确导出是否依赖浏览器。
• Preserve revision-bound output-quality, timing, intervention, repeated-work and resource evidence for model and Agent-strategy comparisons. / 模型与编排对比保留绑定版本的质量、耗时、人工介入、重复工作及资源证据。