Agent can now manage simulations via conversation:
- 'start' (run simulation on token)
- 'stop' (stop running simulation)
- 'status' (check if running, show progress)
- 'results' (get current/latest simulation results)
Example conversations:
- 'Run a simulation on GRIAN' → starts simulation
- 'How's the simulation going?' → shows status
- 'Stop the simulation' → stops and shows results
- 'Show me the results' → shows simulation results
Agent can now run backtests via conversation:
- User can say 'backtest this strategy' or 'test on [token]'
- Agent extracts token address (from conversation or asks user)
- Agent runs backtest and presents results in chat with:
- Total Return
- Final Balance
- Win Rate
- Max Drawdown
- Sharpe Ratio
- Agent suggests strategy adjustments based on results
Prototype implementation that allows:
1. Normal conversation with the AI
2. Tool-calling to update trading strategies
Created new ConversationalAgent that uses CrewAI with tools:
- get_current_strategy: Check current bot strategy
- update_trading_strategy: Update bot's trading configuration
The agent can now respond to questions like 'What is this?' without
forcing JSON output, and can update strategies when user provides
specific parameters.
Refs #51
- Use CrewAI's LLM class directly with api_base parameter instead of custom subclass
- Remove broken MiniMaxLLM inheritance from LLM
- Update agent creation to use LLM(model, api_key, api_base) pattern
The issue was that inheriting from CrewAI's LLM class caused the api_base
to be set to None. Now we use CrewAI's LLM directly with the correct parameters.
Fixes#43
Changes:
1. Updated API endpoint from api.minimax.chat to api.minimax.io
2. Changed default model from MiniMax-Text-01 to MiniMax-M2.7
(MiniMax-Text-01 is not available for all API key plans)
3. Updated .env.example with correct default model
MiniMax API docs: https://platform.minimax.io/docs/api-reference/text-anthropic-apiFixes#43
- Fix relative import path in crew.py (from ..core to ...core)
- Update __init__.py exports to match actual class names
- Remove incorrect CrewAgent and LLMConnector exports
LLM was outputting nested params structure but engines expect flat fields.
This caused backtesting and simulation to never trigger any trades.
Changes:
- llm_connector.py: Update prompt to output flat condition structure
- crew.py: Update StrategyValidator to validate flat structure
- crew.py: Update StrategyExplainer to read flat structure
Fixes#25
