continuum-cursor

Continuum + Cursor Skill

Cursor reads .cursor/rules/*.mdc files. The Continuum repo ships .cursor/rules/continuum.mdc — copy it into any downstream project.

Minimal MDC file for a Continuum project

Create .cursor/rules/continuum.mdc:

---
description: Continuum agent framework conventions for this project
globs: ["src/**/*.py", "tests/**/*.py", "playground/**/*.py"]
alwaysApply: true
---

# Continuum Project Rules

This project uses **Continuum** (import: `continuum`, package: `shyftlabs-continuum`).

## Setup invariants

- Python 3.13. Venv: `python3.13 -m venv .venv`
- All public APIs are **async**. Entrypoints: `asyncio.run(main())`
- Always `load_dotenv()` before importing continuum settings
- `OPENAI_API_KEY` required even when using Anthropic/Gemini (mem0 embedder)
- Infra via docker compose: Redis :6380, Milvus :19530 (or Qdrant :6333)

## Core imports

\`\`\`python
from continuum.agent import BaseAgent, AgentRunner
from continuum.agent.config import AgentConfig, AgentMemoryConfig, RunnerConfig
from continuum.agent.types import Handoff, MemoryScope, RunContext, EventType
from continuum.llm import LLMClient, LLMConfig
from continuum.memory import MemoryClient
from continuum.session import SessionClient
\`\`\`

## Provider routing

Model string prefix determines provider (when gateway is not set):
- `gpt-*`, `openai/*` → OpenAI SDK
- `claude-*`, `anthropic/*` → Anthropic SDK  
- `gemini/*`, `google/*` → Gemini via OpenAI-compat

When `SMART_GATEWAY_URL` is set, all models route through GatewayProvider automatically.
Use `gateway_mode="strict"|"modest"|"quality"` on BaseAgent to control routing tier.

## Critical don'ts

- No `litellm` — removed entirely
- No `response_format` on Anthropic — instruct via system prompt and parse manually
- No JSON mode + tools on Gemini simultaneously — pick one
- MemoryScope enum: `from continuum.agent.types import MemoryScope` (not from `continuum.memory`)
- SessionClient.add_message takes `ChatMessage(...)` not `role=` kwargs