Spaces:

---

saptarashmi-lab

/

Multi-Agent-Review-Board

like 0

Sleeping

Changes:

• Added a Hugging Face token resolution flow with clear precedence:
  HF_TOKEN (Space secret) > HUGGINGFACEHUB_API_TOKEN > HF_OAUTH_ACCESS_TOKEN.
  Updated UI messaging to recommend setting HF_TOKEN to avoid OAuth access issues.
• Wrapped InferenceClient.chat_completion calls with retries and exponential backoff
  (with jitter) to handle cold starts, rate limits (429), and transient 5xx/timeouts.
• Updated default models to only include chat-capable, Inference-API-compatible models.
  Kept meta-llama/Llama-3.1-8B-Instruct and Qwen/Qwen2.5-Coder-32B-Instruct.
  Added Qwen/Qwen2.5-7B-Instruct and meta-llama/Llama-3.2-3B-Instruct.
  Removed google/gemma-2-2b-it, deepseek-ai/DeepSeek-R1, and mistralai/Mistral-7B-Instruct-v0.3.
• Improved custom model validation by checking repo existence and issuing a lightweight
  chat_completion ping to verify chat compatibility and gated access early.
• Fixed peer-review attribution by including stable agent labels and IDs
  (Agent N (id=X)) to prevent agents from confusing ownership of responses.
• Updated review logic so all agents receive all other agents’ latest responses
  in rounds 2+, rather than restricting peer context to a single agent.
• Changed agent IDs to start at 1 instead of 0 and ensured IDs remain stable
  when agents are added or removed.