Skip to content

feature: Programmatic Context Engineering plugin in AuthBridge #641

Description

@OsherElhadad

Feature Description

Add a new ContextGuru plugin to AuthBridge.

The plugin will be applied around agent LLM calls - before and after each LLM invocation - and will programmatically modify the agent context in order to reduce context size while preserving as much task accuracy as possible.

This is planned to be part of the end-of-July 0.7.0 release. A PR implementing the initial version will be opened soon.

Motivation

Agent contexts can grow quickly during long-running tasks, especially when they include conversation history, tool outputs, intermediate reasoning artifacts, retrieved documents, and repeated state. This increases:

  • LLM latency
  • token cost
  • risk of exceeding model context limits
  • risk of degraded answer quality due to noisy or irrelevant context

AuthBridge is a good integration point because the context engineering layer can be applied transparently around LLM calls, without requiring every agent implementation to manually manage context compression.

cc @huang195

Proposed Solution

Proposed Behavior

Introduce a new AuthBridge plugin, tentatively named context-guru, that can intercept LLM-bound requests and LLM responses.

Scope

Initial scope for 0.7.0:

  • Add a new AuthBridge plugin for programmatic context engineering.
  • Support applying the plugin before and after LLM calls.
  • Provide a configurable policy for context reduction.
  • Add configuration examples and documentation.
  • Add tests for basic context reduction behavior and safety invariants.

Want to contribute?

  • I would like to work on this issue.

Additional Context

No response

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

No type

Projects

Status
Done

Relationships

None yet

Development

No branches or pull requests

Issue actions