huiliyi37/Tianshu-harness

What it solves

Programming agents often suffer from "cognitive anchor collapse," where they lose track of global goals, make unverified assumptions, or give up when faced with minor obstacles. This project provides a runtime environment that prevents these behaviors, ensuring that agents actually complete tasks with verifiable evidence rather than just claiming to have finished.

How it works

It introduces a Cognitive Virtual Machine (CVM) that acts as a supervisory layer between the AI model and the real world. Instead of relying solely on the model's internal memory, the runtime manages a "Task Contract" that tracks goals, constraints, and success criteria. It uses a pipeline of 75 runtime hooks to monitor the agent's progress across five stages of each turn. Crucially, it enforces a "delivery gate"—the agent cannot mark a task as complete unless it provides evidence, such as passing tests or generated diffs.

Who it’s for

Developers who want to build or use more reliable, autonomous coding agents that can handle complex, multi-step engineering tasks without drifting from their original objectives.

Highlights

• Evidence-based delivery: Prevents agents from falsely claiming task completion by requiring verification steps.
• Cognitive Virtual Machine (CVM): A deterministic supervisory layer that observes, measures, and corrects agent behavior in real-time.
• Efficient Context Management: Features a prefix-caching engine that significantly reduces API costs and improves performance in long sessions.
• Unified Interface: Provides a consistent experience across a terminal-based TUI, a desktop application, and VS Code integration.
• Flexible Agent Personalities: Uses "Star Maps" (星域) to adjust the agent's behavioral tendencies without needing to modify the underlying model weights.