hermes - 💡(How to fix) Fix [Feature]: Message interceptor hooks

Problem or Use Case

Currently, when the Hermes agent is invoked—particularly within sandboxed environments or integrated into messaging platforms (Telegram, Discord, etc.)—it triggers a full orchestration cycle by default. This involves loading a comprehensive System Prompt and extensive context, often resulting in an overhead of +10k input tokens per request.

For simple interactions or sensitive tasks, this overhead is excessive ("overkill"). On local LLMs or resource-constrained systems, this leads to:

• High Latency: Significant delays while the model processes the massive prompt prefix before generating a response.
• Resource Exhaustion: Unnecessary memory and compute consumption for tasks that do not require complex agentic reasoning.
• Privacy & Data Sovereignty: The inability to easily route specific prompts to local scripts or "private-only" models without passing the data through the main, context-heavy Hermes pipeline.

Proposed Solution

I propose the implementation of Message Interceptor Hooks directly within the Hermes agent communication layer. This mechanism would allow for user-defined triggers (prefixes or regex) to intercept a message and bypass the standard agent logic and its massive system prompt entirely.

Key Functionality:

• Custom Trigger Mechanism: Users can define specific identifiers (e.g., a $ prefix) and secondary regex patterns (e.g., matching keywords like mail, weather, or private).
• Selective Redirection: If a match is found, the System Prompt and context loading are skipped. The payload is instead routed to a local handler, which could be a script or a smaller, more private local model.
• Example Workflow: * User Input: $ Get my last mail
  • Detection: The interceptor identifies the $ prefix, and a regex detects the keyword mail.
  • Action: Instead of sending the request to the main LLM with a 10k token prefix, Hermes executes a local command: python get_mail_private.py --prompt "Get my last mail".
  • Result: The user gets their information instantly and privately. This also allows sensitive data to be redirected to a local model with fewer resources, ensuring the information never leaves the local environment.

This creates a "Manual Routing" system where privacy-sensitive or deterministic tasks are handled locally and instantly, while complex, non-sensitive queries still utilize the full agentic pipeline.

Alternatives Considered

• Separate "Light" Agents: Creating a secondary, lightweight Hermes instance for simple tasks. However, this still involves the overhead of the agent framework and does not offer the same raw execution speed as a direct CLI/script hook.
• Manual System Prompt Stripping: Dynamically reducing the System Prompt based on intent. This is inefficient, prone to breaking the agent's reasoning capabilities for complex tasks, and requires constant configuration switching.
• Standard Model Routing: Using existing routers to switch between models. While useful, this usually doesn't solve the problem of bypassing the "System Prompt" prefix entirely to save input tokens and latency on a single local instance.

Feature Type

Gateway / messaging improvement

Scope

None

Contribution

• I'd like to implement this myself and submit a PR

Debug Report (optional)