AnnoClaw 工作流 工作流

在一个控制台里继续处理 AnnoClaw 任务。

你可以在这里创建 AnnoClaw workflow、查看状态、跳转人工复核，并继续训练、导出和结果交付。

请先完成以下配置： 1. 在“待标注图片提交”区域填写 Dataset URI

OpenClaw hosted entrypoints and resources

Current site base URL
https://www.tjmakebot.com
托管入口地址
https://www.tjmakebot.com/api/openclaw-gateway/workflows
Compatibility templates and debug entry (expand only for engineering)
这些模板继续保留给节点级调试、旧用户迁移和最短验证链路，不再作为默认入口。
Advanced: view upstream API base
加载中...
推荐路径：OpenClaw 负责发起与编排，TjMakeBot Workbench 负责托管 Agent 工作流，主站编辑器负责人工精修。
工作流动作的套餐边界
当前套餐
Free Studio
Workflow execution access
创建 workflow、让 Agent 继续执行、训练、导出，以及兼容性任务都要求 Pro Workflow 或更高套餐。
Collab binding access
绑定协作项目和直接完成 review 动作要求 Team Review Cloud 或 Enterprise。
请先登录，再使用付费 workflow 套餐运行托管工作流动作。

Hosted workflow setup

左侧只负责配置本轮 Hosted workflow。真正的运行状态、当前阶段和下一步动作，都以右侧 Workflow lane 为准。

如果你不知道怎么操作：先只改 3 个字段。Project ID、Dataset 或图片 URI、标注对象。创建之后，不再靠猜按钮，直接看右侧 Workflow lane。
On-screen hint
当前是第 1 步：Create the hosted workflow first

Save the left-side setup as one workflowId first. Without it, review, training, export, and handoff cannot target the right session.

Area to focus on: Only fill in the first left-side section right now. Training and export settings can stay at their defaults.
Step 1 / Start here

Basic setup: fill these first

可编辑

For a first run, filling out only this section is enough. Leave training and export settings at their defaults.

Images to annotate

Tell the system which images to label first

这里支持两种方式：1）直接给一个 dataset zip URI；2）逐行提交图片 URI 列表。当前自动化工作台还不支持浏览器本地文件直传。

If your images are still only on your local computer, first move them into the main-site editor 做导入/整理，再把云端可访问 URI 填回这里。
Annotation label design

Then tell the system which objects to annotate

类别就是你要 AI 自动标出来的对象名称，例如 person、car、traffic-light。先用下面的快捷按钮或手动添加即可。

Quick add common labels
These labels will be annotated
行人汽车红绿灯
Part 2: Training and export settings (defaults usually work)
普通用户不需要一开始就展开这里。只有在进入训练或导出步骤时，再检查对应分组即可。
Step 3

Training settings

Only review this section when you are ready to start training. The defaults are good for a first demo.

Step 5

Export settings

You only need this section after training finishes. Ignore it during earlier steps.

Advanced mode: manual takeover / compatibility (most users can ignore this)
这里只有恢复和工程调试动作。按钮会严格按照当前 workflow 的 allowedActions 过滤，只显示现在真的允许的手动步骤。

Manual takeover actions available now

没有出现在这里的动作，说明当前 workflow 并不允许。普通情况下请继续走主流程。

The current workflow does not need manual takeover right now. Waiting for live sync or following the current stage lane is the safer choice.

Compatibility validation and low-level task debugging

保留给 OpenClaw 工程联调、最短 smoke test 和节点级 API 调试。普通用户不需要这里。

Workflow and task status messages appear here.
Current workflow lane

第 1 步：Create the hosted workflow first

Save the left-side setup as one workflowId first. Without it, review, training, export, and handoff cannot target the right session.

未开始
自动化工作台
Next action
-
Live sync
-
工作流 ID
-
Overall progress
0%
完成条件
After creation, the lane switches to "Open human review" and shows the workflowId, next action, and allowed actions.
避免这样做
请先完成以下配置： 1. 在“待标注图片提交”区域填写 Dataset URI

Workflow map: who owns each stage

这条 lane 只为一件事存在：让你一眼看清当前阶段、负责人和下一动作。主站负责前台页面和人工交接，AnnoClaw 负责 workflow 状态机。

1
Create hosted workflow
Hosted frontend · Do this now
The left-side configuration is saved as one workflowId. Human review, training, export, and handoff all continue from that workflow.
2
Hosted human review
人工复核 · 等待 workflow
The actual human review happens in the main-site editor. Open the review page, finish the review there, and return control to the agent.
3
OpenClaw agent resumes training
OpenClaw agent · Waiting for review
After review, the OpenClaw workflow advances training using allowedActions and resumeAction. Manual takeover is only needed for recovery.
4
Agent exports the model
Export stage · Waiting for training
After training completes, the OpenClaw agent triggers export and writes the artifact URI and version back to this workflow.
5
Review artifacts and handoff links
Hosted handoff · Waiting for export
When the loop completes, results stay on the hosted frontend, including the artifact URI, training handoff page, and editor deep-links.

Live snapshot

空闲
工作流 ID
-
项目 ID
-
协作项目 ID
-
Submitted assets
-
Training epoch
-
数据版本 ID
-
Release ID
-
成果物 URI
-
Artifact file
-
Next action: -
Agent resume action: -
人工任务: No pending human task
项目: -
协作项目: -
数据版本: -
Release ID: -
Annotation labels: -
Recommended plan: -
Download entry: -
Overall progress: 0%

Hosted pages and result handoff

所有面向人工的页面都保留在 TjMakeBot 主站里，包括复核页、训练恢复页和结果交接页；AnnoClaw 提供 workflow 状态和动作权限。

已关联协作项目待处理
-
这个 workflow 还没有映射到协作项目，所以后续复核和交付深链暂时不可用。
数据版本已同步待处理
-
它通常会在复核通过后、交付继续推进时创建。
已创建 Release待处理
-
它通常会在训练/导出之后创建，好让交付和训练血缘可以被追踪。
复核 URL:: -
训练 URL:: -

标注任务

Shared annotation entry for the agent workflow and low-level API tests

空闲
任务 ID: -
创建时间: -

训练任务

Can be restored into the training view or polled directly

空闲
任务 ID: -
创建时间: -

导出任务

Artifact URI, format, and callback are surfaced here

空闲
任务 ID: -
创建时间: -