Skip to main content

​
Overview

Zeus uses the DeepAgents framework’s middleware mechanism to inject additional capabilities during Agent execution. Middleware executes in order, each handling specific concerns.

​
Middleware List

​
SummarizationMiddleware

Automatically summarizes long conversation histories to avoid exceeding the model’s context window.
  • Model Profiles: The system includes built-in context window configurations for common models (e.g., gpt-4-turbo: 128K input / 4K output, gpt-4o: 128K input / 16K output, claude-3-5-sonnet: 200K input / 8K output). Default configuration is used when a model is not in the predefined list.
  • Trigger Condition: Summarization is triggered when token count reaches 85% of max_input_tokens.
  • Retention Policy: Always retains the most recent 5 messages; older messages are compressed into summaries.

​
FilesystemMiddleware

Provides filesystem operation capabilities, supporting configuration of working directories and allowed file extensions.

​
AnthropicPromptCachingMiddleware

Enables prompt caching (ephemeral mode) for Anthropic models, reducing repeated System Prompt token billing across multi-turn conversations.

​
PatchToolCallsMiddleware

Automatically repairs malformed JSON tool calls, supporting up to 3 automatic retries.