Tianshu (天枢) is a coding-agent runtime written in TypeScript, with a terminal TUI and a Tauri desktop app for macOS, Windows, and Linux that share one agent core. It targets a specific failure: a model drifting from the goal or declaring a task finished without proof. Tianshu places a cognitive execution layer it calls the CVM (cognitive virtual machine) between the model and real actions, and moves the goal, state, evidence, resources, permissions, and stop conditions out of the chat history into runtime-managed state.
A TaskContract pins the overall goal, and a delivery gate accepts a completion only when it carries runtime evidence such as tests, diffs, or verification commands, reporting a GREEN, YELLOW, or RED status with the changed files and an item-by-item completion audit. Convergence detection checks separately whether the agent is still making progress or stuck in a loop. The CVM runs 75 runtime hooks across five phases, and the project frames the problem as anchor collapse, where a salient local signal (a word in the prompt, an early conclusion, an instruction to comply) outweighs the global goal.
An A/B test on DeepSeek-V4-Flash across five tasks, reviewed by Claude Opus 4.7, measured the first CVM layer (a belief prompt) alone:
| Metric | Without belief prompt | With belief prompt |
|---|---|---|
| Tasks completed | 4/5 | 5/5 |
| Raised an objection unprompted | 0/5 | 3/5 |
| Asked about scope or impact | 0/5 | 1/5 |
| Intent over literal instruction | 1/5 | 4/5 |
A prefix-cache engine freezes the prompt prefix and appends increments, and reports steady-state cache hit rates of 95–99% in long sessions across models, with extra tuning for DeepSeek V4. Sixteen selectable domains (星域) bias the same model toward different working stances, such as weighing trade-offs, probing unknowns, or insisting a bug reproduces before calling it fixed. Supported providers include DeepSeek, Claude, GLM, Codex (GPT), Grok, MiniMax, MiMo, Kimi, and Gemini, plus any OpenAI-compatible endpoint such as Ollama or vLLM.
Features
- Evidence-gated delivery: a completion claim needs tests, diffs, or verification commands, checked by a delivery gate and a post-commit review
- CVM runtime hooks: 75 hooks across five phases sit between model output and real actions to detect and correct drift
- Prefix-cache engine: frozen prefix, incremental appendix, read deduplication, and cache inheritance on resume, with a cache diagnostics command
- Multi-agent orchestration: read-only scouting, parallel team work, multi-seat council review, and wave-based execution with typed work orders and model routing
- Domains: sixteen selectable stances that shift how the same model approaches a task
- Cockpit: panels for delivery verification, runtime advisories, cache hit rate and token cost, and risk level
- Plan and goal modes: an approve-then-execute plan workflow and a goal-driven mode that continues autonomously within a budget
- Handoff and rewind: structured session handoff into a new session, and rewind to any earlier point
- LSP integration: go-to-definition and find-references as agent tools for 20+ languages, with diagnostics fed back after each edit
- Permission modes: supervise, auto (the default), and full-auto with sandboxing and rollback
- MCP and skills: external tool servers and reusable workflow playbooks, loaded only for trusted projects
- Desktop extras: integrated terminal, theme studio, local voice input, phone-based remote approval, and a WSL-connected runtime on Windows
