Work Harness v0.2.2

Work Harness v0.2.2 turns the single-instance framework into a production-configurable release while keeping real Provider credentials under operator control.

Highlights

• Added dual Provider credential channels: environment-variable injection or authenticated UI entry with encrypted persistence.
• Added production first-run configuration mode with separate liveness and Provider readiness checks.
• Added production Compose templates, configuration validation, resource limits, hardened Caddy headers, and credential-key-aware backups.
• Updated the workbench to a low-saturation gray/white square group-chat interface.
• Kept independent Provider and model routing for the main coordinator group and individual child agents.

Compatibility

• Custom OpenAI-compatible and DeepSeek-dialect endpoints remain supported.
• Both Chat Completions and Responses API families remain supported.
• This release remains a single-instance SQLite deployment; it does not add multi-tenant identity or horizontal scaling.

Validation

• 71 backend tests and 15 frontend tests passed.
• Python 3.11 and 3.12 CI, frontend production build, wheel/sdist verification, Compose validation, and container build passed.
• The release automation publishes Windows x64 executable assets and multi-architecture GHCR images.

Work Harness v0.2.1

后端性能与群聊体验更新 / Backend performance and group-chat UX update

• 普通低/中风险任务采用自适应综合复核，默认双产物流程由 11 次模型调用降至 7 次；高风险、研究与决策任务继续使用三维独立复核。
• 增加完整运行与逐模型服务并发背压、幂等工具调用缓存、请求体/速率保护，以及确定性 Trace 采样。
• SSE 改为由 LangGraph 节点更新驱动，减少空闲轮询和重复 Checkpoint 读取。
• 前端改为类似 QQ/微信的三栏群聊，加入左右消息气泡、群成员栏、移动端适配和中文口语化 Agent 功能标签。
• 保持自定义 OpenAI 协议、DeepSeek 方言、Chat Completions 和 Responses 兼容，不包含任何真实模型密钥。
• 新增生产运行配置、后端基准脚本和两份架构分析报告对应的实施说明。

验证结果：66 项后端测试、14 项前端测试、10/10 Mock 契约评估、类型检查、生产构建、Provider/Skill 兼容检查和 Python wheel/sdist 校验全部通过。

Release workflows will attach the Python distributions, Windows x64 executable, checksums, attestations, and multi-architecture GHCR image after GitHub Actions completes.

Work Harness v0.2.0

Windows 单文件封装 / Windows single-file packaging

• 提供可双击启动的 Windows x64 单文件 EXE，内置 FastAPI、Vue 工作台、配置、Prompt、Schema、评估用例和 Skill。
• Checkpoint、运行记录和 Provider Profile 持久化到 %LOCALAPPDATA%\WorkHarness，升级 EXE 不丢数据。
• 支持自定义 OpenAI 协议与 DeepSeek 方言，以及 Chat Completions / Responses；不包含任何模型密钥。
• 主 Agent 和每个子 Agent 继续支持独立 Provider/模型路由。
• Release 自动生成并证明 wheel、sdist、Windows EXE、SHA-256 和多架构 GHCR 镜像。

Download work-harness.exe and work-harness.exe.sha256, verify the checksum, then double-click the executable. See docs/windows-executable.md for details.

Work Harness v0.1.0

Work Harness v0.1.0 is the first public alpha release of the controlled, observable, and testable multi-agent workflow harness.

Work Harness v0.1.0 是首个公开 Alpha 版本，提供受控、可观测、可测试的多 Agent 工作流框架。

Highlights / 主要功能

• Bounded LangGraph planning, parallel execution, independent review, repair, human approval, and synthesis.
• OpenAI-compatible and DeepSeek protocol behavior for Chat Completions and Responses.
• Persistent Provider Profiles with separate main-agent and child-agent model routing.
• Modular Skills, Tool adapters, capability parity checks, and deterministic Mock evaluation.
• Vue workbench, SSE updates, SQLite restart recovery, and one-command Docker Compose deployment.
• English and Simplified Chinese README documentation.

Distribution / 发布产物

The release workflow attaches the Python wheel and source distribution, and publishes the multi-platform container image as ghcr.io/1ces1amese/work-harness:0.1.0.

发布工作流会附加 Python wheel 与源码包，并发布 ghcr.io/1ces1amese/work-harness:0.1.0 多架构容器镜像。

Important / 注意

This alpha release is designed for local development, contract testing, manual real-model evaluation, and single-instance deployment. It contains no model credentials. SQLite deployments must remain single-instance.

该 Alpha 版本适用于本地开发、协议测试、手动真实模型验收和单实例部署，不包含任何模型密钥；SQLite 部署不得横向扩容。