BrainForge LLM Integration

[NeonCharlie-24]

Description

Integrates BrainForge as a backend that dynamically discovers and registers its expert personas as advisors. Personas are fetched from the BrainForge API, merged with static YAML personas, and periodically re-synced. Includes auth management, an OpenAI-compatible LLM client, orchestrator safety fixes, and unit tests.

Issues

Closes #49

Other Notes

• BrainForge credentials are read from env vars (.env). Will need to add:
  • BRAINFORGE_USERNAME=guest
  • BRAINFORGE_PASSWORD=password
    to .env to run.
• NucleotidingsLLM returns a 500 error. All other personas/models have been tested with curl and return responses as expected.
• The periodic re-sync of personas introduced async edge cases that have been addressed, including ensuring the orchestrator uses the active provider's LLM (Gemini, vLLM) for routing decisions rather than pulling from an arbitrary persona, which could inadvertently select a BrainForge model.

[NeonDaniel]

Few implementation notes; main issue to address is the quality of responses from BrainForge models. I left a couple suggestions inline below, based mostly on this particular response:

NeonAI
12:02 PM
Thought
Exploring the core team at Neon AI, we find a group of highly skilled individuals dedicated to advancing AI technology.

What to do
Exploring the core team at Neon AI, we find a group of highly skilled individuals dedicated to advancing…
The team includes experts in machine learning, natural language processing, and computer vision, all committed to making AI…
Notable members such as Richard Leeds, Elon Gasper, and Clary Gasper bring a wealth of experience and innovation…
Next step
Exploring the core team at Neon AI, we find a group of highly skilled individuals dedicated to advancing…

[NeonCharlie-24]

Ah ok I passed extra_body.structured_outputs and got much better results! I also implemented some manual formatting specifically for the Brainforge model outputs and bypassed _ensure_compact_shape, which I think was causing the truncation.

[NeonCharlie-24]

The logic to skip _ensure_compact_shape was brittle (just a hardcoded string search) and not aligned with the other clients' response processing. Instead, I added maxLength constraints to the JSON schema to keep the field lengths within _ensure_compact_shape's truncation limits. This way BrainForge responses flow through the same post-processing path as every other provider without being clipped.

I also doubled max_tokens for BrainForge requests. The original token budget was causing some responses to be truncated before the JSON could close, resulting in malformed outputs.

I think this mostly solves the BrainForge response issue while maintaining parity with the other providers!

[NeonDaniel]

Couple nits; otherwise changes look good and worked locally for me