我在构建本地优先的 AI 辅助软件开发工具，方向包括多代理编排、Agent Skills、DeepSeek Harness 插件，以及更容易观察和验证的开发者工具。
我关心的不是“看起来很智能”的系统，而是在真实工程压力下仍然可靠的系统：权限边界清晰、工具调用可追踪、完成状态有证据、失败时也能被检查。
- 面向编码和软件交付的多代理编排。
- 面向 API、工具、MCP、发布安全的 Agent Skills。
- 本地优先、可观察、可验证的 AI 开发者工具。
- 小而完整的展示型项目，用来表达交互、视觉和想法。
一个证据驱动的多模型工程代理，包含明确的权限边界、工具证据、验证关卡、CLI 和 WebUI。
Windows 桌面宠物，用本地 hooks 和本地 HTTP bridge 展示 AI coding agent 正在做什么。
DeepSeek Harness 的独立第三方插件和工具集合，包含检查、诊断和日常操作辅助能力。
用于 DeepSeek Harness 的 provider 连通性和延迟手动检查工具。
面向 coding agents 的完成状态审计 sidecar，用证据判断任务是 completed 还是 blocked。
Profile README 中的访客计数展示组件，用来给主页增加一点可见的访问反馈。
面向 API 契约演进的 Agent Skills：breaking changes、compatibility、deprecation、idempotency、webhooks 和 OpenAPI 工作流。
面向工具和 MCP 契约的 Agent Skills：设计、权限、人类确认、breaking-change review 和评估。
面向 AI 辅助编码之后交付阶段的 Agent Skills：secrets、repros、flaky-test triage、feature flags 和 environment drift。
- 对外展示：
Wanbinyu - 主要工具：
multi-agent-orchestrator、pet-tty、wanbinyu-harness-toolbox - DeepSeek Harness 插件：
dsh-* - Skills 集合：
api-platform-skills、ai-surface-skills、ship-guard-skills - 创意实验：
qixi-* - 历史镜像：
skill-*
独立的 skill-* 仓库是已归档的历史镜像；新的工作会进入上面的三个 skills 集合。
I build local-first tools for AI-assisted software development, including multi-agent orchestration, Agent Skills, DeepSeek Harness plugins, and developer tools that are easier to observe and verify.
I care less about systems that merely look intelligent and more about systems that remain useful under real engineering pressure: clear permission boundaries, traceable tool use, evidence-backed completion, and inspectable failure modes.
- Multi-agent orchestration for coding and software delivery.
- Agent Skills for API, tool, MCP, and release-safety workflows.
- Local-first AI developer tools that are observable and verifiable.
- Small showcase projects for interaction, visuals, and ideas.
An evidence-driven multi-model engineering agent with explicit permission boundaries, tool evidence, verification gates, CLI, and WebUI.
A Windows desktop pet that shows what an AI coding agent is doing through local hooks and a local HTTP bridge.
Independent third-party plugins and tools for DeepSeek Harness, including inspection, diagnostics, and operational helpers.
Manual provider connectivity and latency checks for DeepSeek Harness.
A sidecar completion audit for coding agents where evidence decides completed vs blocked.
The visitor counter displayed in this profile README, used as a small visible feedback element for the showcase page.
Agent skills for API contract evolution: breaking changes, compatibility, deprecation, idempotency, webhooks, and OpenAPI workflows.
Agent skills for tool and MCP contracts: design, permissions, human approval, breaking-change review, and evaluation.
Agent skills for shipping after AI-assisted coding: secrets, repros, flaky-test triage, feature flags, and environment drift.
- Showcase:
Wanbinyu - Main tools:
multi-agent-orchestrator,pet-tty,wanbinyu-harness-toolbox - DeepSeek Harness plugins:
dsh-* - Skill collections:
api-platform-skills,ai-surface-skills,ship-guard-skills - Creative experiments:
qixi-* - Historical mirrors:
skill-*
Standalone skill-* repositories are archived historical mirrors; new work lives in the three skill collections above.