hermes - 💡(How to fix) Fix [Feature]: Advisor Mode | Cross-endpoint guidance-only delegation for local-first setups [1 comments, 1 participants]

Problem or Use Case

Hi — wanted to float a small delegate_task extension before opening a PR, in case the design needs to change.

I run Hermes self-hosted on a single edge-class GPU. The main agent loop runs a 4B Gemma model (NVFP4-quantized) at ~60 tok/s. That's plenty for tool calls, terminal work, and most reasoning, but it bvisibly struggles on architecture decisions, ambiguous debugging, and "what am I missing?" judgment calls.

Two pressures push in opposite directions:

• Privacy / sovereignty. I work with code and data (legal-domain material, internal systems) I can't ship to third-party APIs. Self-hosting the main loop is non-negotiable.
• Quality on hard calls. Some decisions genuinely need a stronger model. Sending the entire agent transcript to a frontier API to get one architectural opinion exfiltrates more than the question needs.

What I'd find useful is a guidance-only escape hatch at the delegate_task layer: send a minimized packet to a stronger model running on a separate endpoint, get back recommendations only, then keep executing locally. No tools, no transcript replay.

Proposed Solution

The smallest change I think buys this:

1. New advisor config section in ~/.hermes/config.yaml:

   advisor:                                                                              
     model: ""          # empty = inherit parent          
     provider: ""       # empty = inherit parent                                         
     base_url: ""       # direct OpenAI-compatible endpoint
     api_key: ""                                                                         
     max_iterations: 3                                    
     allow_tools: false

2. New advisor: bool arg on the existing delegate_task tool.
3. When advisor=true: the child reads advisor.* (so it can route to a different endpoint than the main loop), runs with empty toolsets by default, and gets a "guidance-only, do not claim execution" system prompt.

That's the only part that genuinely can't live as a skill — the cross-endpoint routing requires harness-level config + provider resolution. Everything else (when to invoke it, the structured packet shape, the output contract) is just instructions to the model and seems like a textbook skill candidate. I'm planning a separate skill for that side, either bundled in optional-skills/ or on the Skills Hub.

Alternatives Considered

I drafted this as a top-level advisor section because the endpoint is independent of delegation's, but I'd happily nest it under delegation.advisor if that's cleaner.

I want to check whether "advisor" is redundant with something already in flight delegate_task or whether it's complementary.

Feature Type

Configuration option

Scope

Medium (few files, < 300 lines)

Contribution

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

Debug Report (optional)