DeepSeek Releases Cordis, a Dynamic Composition Theory
Key point
DeepSeek-AI and Peking University have released Cordis, a formal theory-based meta-framework for runtime plugin removal and dependency management.
Details
Filling the Theoretical Gap in Dynamic Composition
Existing software has a rich formal theory for static composition fixed at compile time, but lacks a theoretical foundation for dynamic composition, which involves loading/unloading components at runtime. Researchers from Peking University and DeepSeek-AI proposed the meta-framework Cordis to fill this gap by dividing the requirements of dynamic composition into temporal composition (safe state restoration) and spatial composition (structural dependency management), and formalizing them.
Limitations of Existing Plugin Systems
The paper points out issues with existing systems using the VSCode extension system as an example.
- Temporal Limitations: Among the top 100 extensions, 87 contain executable code, making a host restart mandatory upon removal. Restarting initializes caches and connection states, resulting in availability loss.
- Spatial Limitations: Among the top 100, only 7 explicitly declare dependencies between non-builtin extensions. The lack of typed
anyreturn values makes verifiable interface-based dependency management impossible.
Theoretical Foundation: Effect and Coeffect Systems
Cordis combines two formal theories.
- Effect System: Tracks side effects that a program has on the world at the type level. Through monadic effects and algebraic effects, it structurally pairs operations with their inverses (undo) to guarantee safe state restoration.
- Coeffect System: Tracks resources or permissions that a program requires from the environment as context. It is used to statically analyze and verifiably resolve dependencies between components.
Integration with DeepSeek Harness
Cordis is introduced as a plugin framework included in a vendor manner in the official documentation of DeepSeek Harness. The agent preset configuration filename agent.cordis.yml reflects this relationship. This research also serves as documentation explaining what formal theory the DSH plugin ecosystem, already utilized by the community, is built upon.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.