monitor

Monitor - Autonomous Learning Loop

Closes the autonomous loop: check trades, close hits, evaluate predictions, generate summaries.

All agents use subagent_type: general-purpose with explicit model to ensure MCP tool access. Include "Do NOT use the Edit tool" in every prompt.

Workflow

Step 1: Get Current Prices

Delegate using Task with subagent_type: general-purpose, model: haiku:

"You are the market-monitor agent. Get current prices for ALL symbols that have open trades. Call get_portfolio_state() from crypto-learning-db first to see which symbols have open positions. Then use get_exchange_prices(symbol=...) from crypto-exchange MCP to get live prices for each symbol. Return a JSON object like: {\"BTC/USDT\": 98500, \"ETH/USDT\": 3200} Do NOT use the Edit tool."

Step 2: Check Open Trades Against SL/TP

Using the prices from Step 1 and the open trades from get_portfolio_state():

For each open trade, check:

• Long trade: Did price drop to or below stop_loss? Did price rise to or above take_profit?
• Short trade: Did price rise to or above stop_loss? Did price drop to or below take_profit?

If SL or TP was hit, delegate using Task with subagent_type: general-purpose, model: opus:

"You are the portfolio-manager agent. Close trade {trade_id}. Current price is ${price}. The {SL/TP} at ${level} was hit. Call close_trade(trade_id='{trade_id}', exit_price={price}, close_reason='{SL/TP} hit at ${level}') from crypto-learning-db. Do NOT use the Edit tool."

Step 2b: Trailing Stop Adjustment

For open trades that were NOT closed (still active), check if the trade is profitable:

• Long: current price > entry_price
• Short: current price < entry_price

If profitable AND the current stop_loss hasn't been optimally trailed, delegate using Task with subagent_type: general-purpose, model: sonnet:

"You are the risk-specialist agent. Read agents/risk-specialist.md for context. Trade {trade_id} ({symbol}, {side}) is profitable. Entry: ${entry}, Current: ${price}, SL: ${stop_loss}, TP: ${take_profit}. Analyze whether to trail the stop-loss. Use calculate_volatility(symbol=...) and get_support_resistance(symbol=...) from crypto-technical MCP. If you recommend adjusting, call update_trade(trade_id='{trade_id}', stop_loss={new_sl}, notes='your reasoning') from crypto-learning-db. Rules: only trail in profitable direction, never widen the stop, leave room for normal volatility. Do NOT use the Edit tool."

Step 3: Post-Mortem for Closed Trades

For each trade closed in Step 2, delegate using Task with subagent_type: general-purpose, model: opus:

"You are the learning-agent. Read agents/learning-agent.md for your analysis framework. Trade {trade_id} was just closed ({result}, PnL: {pnl}). Do a full post-mortem:

1. Call query_predictions(trade_id='{trade_id}') from crypto-learning-db to find all predictions
2. For each pending prediction, evaluate: read the original, compare to what happened, write a detailed NL evaluation
3. Call validate_prediction() with your evaluation for each one
4. Call upsert_pattern() if you identify a named pattern
5. Write a post-mortem report to data/reports/

Do NOT use the Edit tool."

Step 4: Evaluate Expired Predictions

Delegate using Task with subagent_type: general-purpose, model: opus:

"You are the learning-agent. Call find_expired_predictions(current_prices='{prices_json}') from crypto-learning-db using the current prices. For each expired prediction:

1. Read the original prediction text
2. Compare to the current price / market state
3. Write a natural language evaluation explaining how close it was and why
4. Call validate_prediction() with your evaluation

Do NOT use the Edit tool."

Step 5: Monthly Summary (conditional)

Check if today is the last day of the month (or within 2 days of month end). If yes, delegate using Task with subagent_type: general-purpose, model: opus:

"You are the learning-agent. Generate a monthly summary. Call generate_summary(summary_type='monthly') from crypto-learning-db. Also check if it's end of quarter — if so, call generate_summary(summary_type='quarterly') too. Do NOT use the Edit tool."

Step 6: Report

Present a summary of what happened:

## Monitor Report

**Open Trades:** X active
**Trades Closed This Run:** X (list with PnL)
**Predictions Evaluated:** X expired validated
**Summary Generated:** Yes/No

### Next Actions
- [Any trades approaching SL/TP levels]
- [Any concerning patterns noticed]