Processing Layer - Pluggable Workers

[illeatmyhat]

The name "plugin" is too crowded in the AI space. We're still workshopping the name.

Objective

While it is possible for developers to consume kaizen directly using the KaizenClient, it should be possible for developers to add features to kaizen without contributing to kaizen directly.

Methodology

Contributors want to affect various API surfaces. They should be able to write a single component which contains all of the changes they want to integrate. That component should be installable as a python package. Python has a number of features to support this, such as Protocols.

API Surfaces

• Contributors should be able to take agent trajectories, process them, and store them in the DB asynchronously.
• Contributors should be able to provide MCP Tools (e.g. a get_guidelines() function) so that agents have an easy to understand method to retrieve plugin-specific entities.

Sample Implementations

• Semantic "Fact" Extraction & Retrieval
• Guideline Extraction & Retrieval

Maintenance Layer

This work implies a need to identify and change certain functionality to run independently of functions

• Conflict Resolution
• Scoring?
• Clustering?

Plugin Conflicts

• Plugins should warn or error when being loaded, and anything created by other plugins was overwritten.

Configuration

• There should be a programmatic way to load plugins via the KaizenClient, and a method using config files when running an MCP or HTTP server