Skip to content

PROMPT.md

Latest commit

 

History

History
22 lines (18 loc) · 1.85 KB

PROMPT.md

File metadata and controls

22 lines (18 loc) · 1.85 KB

Open Council: Architecture & Coding Standards

Project Overview

Open Council is a CLI-first, open-source multi-agent LLM orchestration tool designed to replicate Perplexity's "Council of Models." It utilizes a "Bring Your Own Key" (BYOK) hybrid model, mixing local offline models with cloud-based endpoints.

Core Tech Stack

  • Language: Python 3.11+
  • Orchestration: LangGraph (State management, cyclic/acyclic routing)
  • LLM Routing: LiteLLM (Universal API translation)
  • Concurrency: asyncio
  • CLI UI: Rich (Live displays, progress spinners, markdown rendering)

Strict Coding Standards

  1. Async Everything: All network calls, file I/O, and LLM generations MUST be asynchronous (async def, await). Never use blocking standard library calls (like requests). Use aiohttp or httpx for standard web requests.
  2. The Semaphore Rule: Never make raw API calls directly from a LangGraph node. All LLM calls and Search calls MUST be routed through our core/throttle.py which wraps them in an asyncio.Semaphore to prevent 429 Rate Limits.
  3. Graceful Fallbacks: The system must never crash on an LLM API failure. Use LiteLLM's built-in fallback array (Primary -> Secondary -> Local Ollama).
  4. No Raw Prints: Standard print() statements will corrupt the async terminal UI. All CLI output must be routed through the Rich console components defined in src/open_council/ui/.
  5. Type Hinting: Enforce strict Python typing (typing.Dict, typing.List, typing.Annotated). LangGraph state relies entirely on accurate TypedDict and Pydantic schemas.

Graph Design Philosophy

  • Nodes should be pure functions that take a State dict, perform an async action, and return a dictionary containing ONLY the keys they intend to update.
  • Do not mutate the state inside the node directly; let LangGraph handle the state merging.