huiliyi37
huiliyi37 is an individual/solo open-source developer and creator of Tianshu Harness (天枢), an AI coding-agent runtime. Tianshu is intended to make coding agents behave more like reliable engineering partners by managing goals, state, permissions and evidence outside the model, and by requiring verification before delivery.
At a Glance
- Individual developers
- Open-source contributors and ecosystem developers
- Teams working on real software repositories
- Developers using DeepSeek, Claude, GLM, OpenAI-compatible or other supported model providers
- +1 more
AI Tools by huiliyi37
(1)
Tianshu Harness
Open Source AI Coding Agent
Discussions
No discussions yet
Be the first to start a discussion about huiliyi37
Latest News
Tianshu Harness v3.36.1 released: sub-agent delegation hotfixes and tool-call protocol hardening.
v3.36.0 prerelease published.
v3.35.0 released with native table preview, attachment path handles and a full-height TUI viewport.
v3.34.1 and runtime-v3.34.1 released, covering Skill Python runtime/environment fixes and a self-contained CLI runtime.
Products & Services
Open-source TypeScript coding-agent runtime, distributed as a terminal CLI/npm package, desktop application and VS Code extension sharing one agent kernel. It uses a Cognitive Virtual Machine with task contracts, evidence-gated delivery, runtime hooks, Sensorium state awareness, multi-agent orchestration, memory, handoff and rewind.
An interactive terminal TUI plugin for the official DeepSeek Harness. It provides a lightweight ANSI rendering engine, streaming Markdown/tool cards, 16+ themes, slash commands and selectors, input history, local preference persistence, LSP diagnostics, image/vision bridging, code search and cross-session memory.
Market Position
Tianshu positions itself as an open-source coding-agent runtime rather than a chat-only coding assistant. Its differentiator is the CVM/runtime-supervision approach—task contracts, evidence gates, verification loops, permission controls and stateful orchestration—combined with a shared terminal/desktop/editor kernel. It competes broadly with other coding-agent CLIs, IDE agents and DeepSeek Harness interfaces, while emphasizing reliable, auditable delivery and long-session cost control.
Leadership
Founders
huiliyi37
The GitHub account owner and primary creator/maintainer of Tianshu-harness and dsh-tianshu-tui. The public profile identifies the account through these repositories and describes the work as bringing Tianshu Harness to developers; no prior employers or other professional background were published in the reviewed sources.
Executive Team
huiliyi37
Creator and primary maintainer
Owns the GitHub account and repositories that publish Tianshu-harness and dsh-tianshu-tui; the public profile presents Tianshu Harness as this creator's work.
Founding Story
The project was started to address a recurring problem in coding agents: models may be persuaded to agree, follow a local instruction too literally, loop on tools, or claim completion without evidence. Tianshu's initial vision was to place a Cognitive Virtual Machine (CVM)—a deterministic runtime supervision layer—between the model and the real world so goals, state, evidence, permissions and termination conditions remain under runtime control.
Business Model
Revenue Model
Open-source Apache-2.0 runtime and documentation are available for free; the official site also sells optional Pro access and a limited lifetime membership. The site says payment is currently handled through Alipay with manual activation.
Pricing Tiers
Download the open-source version without login; includes the runtime, documentation and basic installation entry.
Spark long-session channel, Computer Use GUI automation, multi-perspective orchestration/unattended automation and up to two activated devices.
The same Pro capabilities; the site presents this as approximately ¥17/month and says it saves ¥77 versus monthly billing.
Includes the Pro authorization path, lifetime membership benefits and a sponsor identity marker; limited to 200 memberships on the pricing page.
Target Markets
- Individual developers
- Open-source contributors and ecosystem developers
- Teams working on real software repositories
- Developers using DeepSeek, Claude, GLM, OpenAI-compatible or other supported model providers
- Users needing terminal, desktop or editor-integrated coding-agent workflows
- Real-world software development in local repositories
- Reading codebases, making multi-file changes and running tests or verification commands
- Long-running coding sessions where context, memory, cost and handoff matter
- Multi-agent planning, review and parallel implementation
- CI/headless automation through prompt, stream-json and goal modes
- Supervised desktop/browser automation and remote approval (Pro features)