- Implementation plan with 9 tasks covering bug fixes and testing - Summary report documenting root causes, solution, and verification - Both documents provide comprehensive reference for future maintainers
43 KiB
Fix Position Tracking and P&L Calculation Bugs Implementation Plan
For Claude: REQUIRED SUB-SKILL: Use superpowers:executing-plans to implement this plan task-by-task.
Goal: Fix three critical bugs in position tracking: (1) cash reset to initial value each day, (2) positions lost over weekends, (3) incorrect profit calculations showing trades as losses.
Architecture: Remove redundant _write_results_to_db() method that creates corrupt position records with cash=0 and holdings=[], and fix profit calculation logic to compare same-day position values instead of previous-day portfolio value.
Tech Stack: Python 3.12, SQLite3, pytest
Root Cause Analysis:
-
Bug #1 & #2 (Cash reset + positions lost):
ModelDayExecutor._write_results_to_db()calls non-existent methods (get_positions(),get_last_trade(),get_current_prices())- These return empty values, creating corrupt records with
cash=0,holdings=[] get_current_position_from_db()then finds this corrupt record as "latest", causing reset
-
Bug #3 (Incorrect profit calculations):
- Current logic compares portfolio value to previous day's final value
- When you buy stocks, cash decreases and stock value increases equally → portfolio value unchanged → profit=0 shown
- Should compare to start of current day (after price changes) to show actual gains/losses from trading
Task 1: Write Failing Tests for Current Bugs
Files:
- Create:
tests/unit/test_position_tracking_bugs.py
Step 1: Write test demonstrating cash reset bug
Create tests/unit/test_position_tracking_bugs.py:
"""
Tests demonstrating position tracking bugs before fix.
These tests should FAIL before implementing fixes, and PASS after.
"""
import pytest
from datetime import datetime
from api.database import get_db_connection, initialize_database
from api.job_manager import JobManager
from agent_tools.tool_trade import _buy_impl
from tools.price_tools import add_no_trade_record_to_db
@pytest.fixture
def test_db_with_prices(tmp_path):
"""Create test database with price data."""
db_path = str(tmp_path / "test.db")
initialize_database(db_path)
# Insert price data for testing
conn = get_db_connection(db_path)
cursor = conn.cursor()
# 2025-10-06 prices
cursor.execute("""
INSERT INTO price_data (symbol, date, open, high, low, close, volume, created_at)
VALUES ('NVDA', '2025-10-06', 185.5, 190.0, 185.0, 188.0, 1000000, ?)
""", (datetime.utcnow().isoformat() + "Z",))
# 2025-10-07 prices (Monday after weekend)
cursor.execute("""
INSERT INTO price_data (symbol, date, open, high, low, close, volume, created_at)
VALUES ('NVDA', '2025-10-07', 186.23, 190.0, 186.0, 189.0, 1000000, ?)
""", (datetime.utcnow().isoformat() + "Z",))
conn.commit()
conn.close()
return db_path
@pytest.mark.unit
class TestPositionTrackingBugs:
"""Tests demonstrating the three critical bugs."""
def test_cash_not_reset_between_days(self, test_db_with_prices):
"""
Bug #1: Cash should carry over from previous day, not reset to initial value.
Scenario:
- Day 1: Start with $10,000, buy 5 NVDA @ $185.50 = $927.50, cash left = $9,072.50
- Day 2: Should start with $9,072.50 cash, not $10,000
"""
# Create job
manager = JobManager(db_path=test_db_with_prices)
job_id = manager.create_job(
config_path="configs/test.json",
date_range=["2025-10-06", "2025-10-07"],
models=["claude-sonnet-4.5"]
)
# Day 1: Initial position (action_id=0)
conn = get_db_connection(test_db_with_prices)
cursor = conn.cursor()
cursor.execute("""
INSERT INTO trading_sessions (job_id, date, model, started_at)
VALUES (?, ?, ?, ?)
""", (job_id, "2025-10-06", "claude-sonnet-4.5", datetime.utcnow().isoformat() + "Z"))
session_id_day1 = cursor.lastrowid
cursor.execute("""
INSERT INTO positions (
job_id, date, model, action_id, action_type,
cash, portfolio_value, session_id, created_at
)
VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
""", (
job_id, "2025-10-06", "claude-sonnet-4.5", 0, "no_trade",
10000.0, 10000.0, session_id_day1, datetime.utcnow().isoformat() + "Z"
))
conn.commit()
conn.close()
# Day 1: Buy 5 NVDA @ $185.50
result = _buy_impl(
symbol="NVDA",
amount=5,
signature="claude-sonnet-4.5",
today_date="2025-10-06",
job_id=job_id,
session_id=session_id_day1
)
assert "error" not in result
assert result["CASH"] == 9072.5 # 10000 - (5 * 185.5)
# Day 2: Create new session
conn = get_db_connection(test_db_with_prices)
cursor = conn.cursor()
cursor.execute("""
INSERT INTO trading_sessions (job_id, date, model, started_at)
VALUES (?, ?, ?, ?)
""", (job_id, "2025-10-07", "claude-sonnet-4.5", datetime.utcnow().isoformat() + "Z"))
session_id_day2 = cursor.lastrowid
conn.commit()
conn.close()
# Day 2: Check starting cash (should be $9,072.50, not $10,000)
from agent_tools.tool_trade import get_current_position_from_db
position, next_action_id = get_current_position_from_db(
job_id=job_id,
model="claude-sonnet-4.5",
date="2025-10-07"
)
# BUG: This will fail before fix - cash resets to $10,000 or $0
assert position["CASH"] == 9072.5, f"Expected cash $9,072.50 but got ${position['CASH']}"
assert position["NVDA"] == 5, f"Expected 5 NVDA shares but got {position.get('NVDA', 0)}"
def test_positions_persist_over_weekend(self, test_db_with_prices):
"""
Bug #2: Positions should persist over non-trading days (weekends).
Scenario:
- Friday 2025-10-06: Buy 5 NVDA
- Monday 2025-10-07: Should still have 5 NVDA
"""
# Create job
manager = JobManager(db_path=test_db_with_prices)
job_id = manager.create_job(
config_path="configs/test.json",
date_range=["2025-10-06", "2025-10-07"],
models=["claude-sonnet-4.5"]
)
# Friday: Initial position + buy
conn = get_db_connection(test_db_with_prices)
cursor = conn.cursor()
cursor.execute("""
INSERT INTO trading_sessions (job_id, date, model, started_at)
VALUES (?, ?, ?, ?)
""", (job_id, "2025-10-06", "claude-sonnet-4.5", datetime.utcnow().isoformat() + "Z"))
session_id = cursor.lastrowid
cursor.execute("""
INSERT INTO positions (
job_id, date, model, action_id, action_type,
cash, portfolio_value, session_id, created_at
)
VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?)
""", (
job_id, "2025-10-06", "claude-sonnet-4.5", 0, "no_trade",
10000.0, 10000.0, session_id, datetime.utcnow().isoformat() + "Z"
))
conn.commit()
conn.close()
_buy_impl(
symbol="NVDA",
amount=5,
signature="claude-sonnet-4.5",
today_date="2025-10-06",
job_id=job_id,
session_id=session_id
)
# Monday: Check positions persist
from agent_tools.tool_trade import get_current_position_from_db
position, _ = get_current_position_from_db(
job_id=job_id,
model="claude-sonnet-4.5",
date="2025-10-07"
)
# BUG: This will fail before fix - positions lost, holdings=[]
assert "NVDA" in position, "NVDA position should persist over weekend"
assert position["NVDA"] == 5, f"Expected 5 NVDA shares but got {position.get('NVDA', 0)}"
def test_profit_calculation_accuracy(self, test_db_with_prices):
"""
Bug #3: Profit should reflect actual gains/losses, not show trades as losses.
Scenario:
- Start with $10,000 cash, portfolio value = $10,000
- Buy 5 NVDA @ $185.50 = $927.50
- New position: cash = $9,072.50, 5 NVDA worth $927.50
- Portfolio value = $9,072.50 + $927.50 = $10,000 (unchanged)
- Expected profit = $0 (no price change yet, just traded)
Current bug: Shows profit = -$927.50 or similar (treating trade as loss)
"""
# Create job
manager = JobManager(db_path=test_db_with_prices)
job_id = manager.create_job(
config_path="configs/test.json",
date_range=["2025-10-06"],
models=["claude-sonnet-4.5"]
)
# Create session and initial position
conn = get_db_connection(test_db_with_prices)
cursor = conn.cursor()
cursor.execute("""
INSERT INTO trading_sessions (job_id, date, model, started_at)
VALUES (?, ?, ?, ?)
""", (job_id, "2025-10-06", "claude-sonnet-4.5", datetime.utcnow().isoformat() + "Z"))
session_id = cursor.lastrowid
cursor.execute("""
INSERT INTO positions (
job_id, date, model, action_id, action_type,
cash, portfolio_value, daily_profit, daily_return_pct,
session_id, created_at
)
VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
""", (
job_id, "2025-10-06", "claude-sonnet-4.5", 0, "no_trade",
10000.0, 10000.0, None, None,
session_id, datetime.utcnow().isoformat() + "Z"
))
conn.commit()
conn.close()
# Buy 5 NVDA @ $185.50
_buy_impl(
symbol="NVDA",
amount=5,
signature="claude-sonnet-4.5",
today_date="2025-10-06",
job_id=job_id,
session_id=session_id
)
# Check profit calculation
conn = get_db_connection(test_db_with_prices)
cursor = conn.cursor()
cursor.execute("""
SELECT portfolio_value, daily_profit, daily_return_pct
FROM positions
WHERE job_id = ? AND model = ? AND date = ? AND action_id = 1
""", (job_id, "claude-sonnet-4.5", "2025-10-06"))
row = cursor.fetchone()
conn.close()
portfolio_value = row[0]
daily_profit = row[1]
daily_return_pct = row[2]
# Portfolio value should be $10,000 (cash $9,072.50 + 5 NVDA @ $185.50)
assert abs(portfolio_value - 10000.0) < 0.01, \
f"Expected portfolio value $10,000 but got ${portfolio_value}"
# BUG: This will fail before fix - shows profit as negative or zero when should be zero
# Profit should be $0 (no price movement, just traded)
assert abs(daily_profit) < 0.01, \
f"Expected profit $0 (no price change) but got ${daily_profit}"
assert abs(daily_return_pct) < 0.01, \
f"Expected return 0% but got {daily_return_pct}%"
Step 2: Run tests to verify they fail
Run: ./venv/bin/python -m pytest tests/unit/test_position_tracking_bugs.py -v
Expected: All 3 tests FAIL demonstrating the bugs
Step 3: Commit failing tests
git add tests/unit/test_position_tracking_bugs.py
git commit -m "test: add failing tests demonstrating position tracking bugs"
Task 2: Remove Redundant _write_results_to_db() Method
Files:
- Modify:
api/model_day_executor.py:161-167(remove call) - Modify:
api/model_day_executor.py:435-558(remove entire method)
Step 1: Remove the call to _write_results_to_db()
In api/model_day_executor.py, find the execute_async() method around line 161-167:
# Commit and close connection before _write_results_to_db opens a new one
conn.commit()
conn.close()
conn = None # Mark as closed
# Store positions (pass session_id) - this opens its own connection
self._write_results_to_db(agent, session_id)
# Update status to completed
Replace with:
# Commit and close connection
conn.commit()
conn.close()
conn = None # Mark as closed
# Note: Positions are written by trade tools (buy/sell) or no_trade_record
# No need to write positions here - that was creating duplicate/corrupt records
# Update status to completed
Step 2: Delete the entire _write_results_to_db() method
In api/model_day_executor.py, delete lines 435-558 (the entire method):
# DELETE THIS ENTIRE METHOD (lines 435-558):
def _write_results_to_db(self, agent, session_id: int) -> None:
"""
Write execution results to SQLite.
...
"""
# ... entire method body ...
Also delete the helper method _calculate_portfolio_value() at lines 533-558:
# DELETE THIS TOO (lines 533-558):
def _calculate_portfolio_value(
self,
positions: Dict[str, float],
current_prices: Dict[str, float]
) -> float:
"""
Calculate total portfolio value.
...
"""
# ... entire method body ...
Step 3: Run unit tests to see what breaks
Run: ./venv/bin/python -m pytest tests/unit/test_model_day_executor.py -v
Expected: Some tests FAIL because they mock non-existent methods
Step 4: Commit the removal
git add api/model_day_executor.py
git commit -m "fix: remove redundant _write_results_to_db() creating corrupt position records"
Task 3: Fix Unit Tests That Mock Non-Existent Methods
Files:
- Modify:
tests/unit/test_model_day_executor.py:21-43 - Modify:
tests/unit/test_model_day_executor.py:185-295 - Modify:
tests/unit/test_model_day_executor_reasoning.py:240-266
Step 1: Update create_mock_agent() helper
In tests/unit/test_model_day_executor.py, find the create_mock_agent() function (lines 21-43):
def create_mock_agent(positions=None, last_trade=None, current_prices=None,
reasoning_steps=None, tool_usage=None, session_result=None,
conversation_history=None):
"""Helper to create properly mocked agent."""
mock_agent = Mock()
# Default values
mock_agent.get_positions.return_value = positions or {"CASH": 10000.0}
mock_agent.get_last_trade.return_value = last_trade
mock_agent.get_current_prices.return_value = current_prices or {}
# ...
Replace with (remove references to deleted methods):
def create_mock_agent(reasoning_steps=None, tool_usage=None, session_result=None,
conversation_history=None):
"""Helper to create properly mocked agent."""
mock_agent = Mock()
# Note: Removed get_positions, get_last_trade, get_current_prices
# These methods don't exist in BaseAgent and were only used by
# the now-deleted _write_results_to_db() method
mock_agent.get_reasoning_steps.return_value = reasoning_steps or []
mock_agent.get_tool_usage.return_value = tool_usage or {}
mock_agent.get_conversation_history.return_value = conversation_history or []
# Async methods - use AsyncMock
mock_agent.run_trading_session = AsyncMock(return_value=session_result or {"success": True})
mock_agent.generate_summary = AsyncMock(return_value="Mock summary")
mock_agent.summarize_message = AsyncMock(return_value="Mock message summary")
# Mock model for summary generation
mock_agent.model = Mock()
return mock_agent
Step 2: Update tests that verify position writes
In tests/unit/test_model_day_executor.py, find TestModelDayExecutorDataPersistence class (lines 182-345).
The tests test_writes_position_to_database and test_writes_holdings_to_database need to be updated because positions are now written by trade tools, not by the executor.
Replace these tests with:
@pytest.mark.unit
class TestModelDayExecutorDataPersistence:
"""Test result persistence to SQLite."""
def test_creates_initial_position(self, clean_db):
"""Should create initial position record (action_id=0) on first day."""
from api.model_day_executor import ModelDayExecutor
from api.job_manager import JobManager
from api.database import get_db_connection
# Create job
manager = JobManager(db_path=clean_db)
job_id = manager.create_job(
config_path="configs/test.json",
date_range=["2025-01-16"],
models=["gpt-5"]
)
# Mock successful execution (no trades)
mock_agent = create_mock_agent(
session_result={"success": True, "total_steps": 10}
)
with patch("api.model_day_executor.RuntimeConfigManager") as mock_runtime:
mock_instance = Mock()
mock_instance.create_runtime_config.return_value = "/tmp/runtime_test.json"
mock_runtime.return_value = mock_instance
executor = ModelDayExecutor(
job_id=job_id,
date="2025-01-16",
model_sig="gpt-5",
config_path="configs/test.json",
db_path=clean_db
)
with patch.object(executor, '_initialize_agent', return_value=mock_agent):
# Mock _handle_trading_result to avoid database writes
with patch.object(executor, '_handle_trading_result', new_callable=AsyncMock):
executor.execute()
# Verify initial position created (action_id=0)
conn = get_db_connection(clean_db)
cursor = conn.cursor()
cursor.execute("""
SELECT job_id, date, model, action_id, action_type, cash, portfolio_value
FROM positions
WHERE job_id = ? AND date = ? AND model = ?
""", (job_id, "2025-01-16", "gpt-5"))
row = cursor.fetchone()
assert row is not None, "Should create initial position record"
assert row[0] == job_id
assert row[1] == "2025-01-16"
assert row[2] == "gpt-5"
assert row[3] == 0, "Initial position should have action_id=0"
assert row[4] == "no_trade"
assert row[5] == 10000.0, "Initial cash should be $10,000"
assert row[6] == 10000.0, "Initial portfolio value should be $10,000"
conn.close()
def test_writes_reasoning_logs(self, clean_db):
"""Should write AI reasoning logs to SQLite."""
# This test remains the same as before (line 297-344)
# ... (keep existing test)
Step 3: Update test_model_day_executor_reasoning.py
In tests/unit/test_model_day_executor_reasoning.py, find the test that calls _write_results_to_db directly (around line 240-266).
Delete or skip this test since the method no longer exists:
@pytest.mark.skip(reason="Method _write_results_to_db() removed - positions written by trade tools")
def test_write_results_links_position_to_session(test_db):
"""DEPRECATED: This test verified _write_results_to_db() which has been removed."""
# Delete this entire test or mark as skipped
pass
Step 4: Run tests to verify fixes
Run: ./venv/bin/python -m pytest tests/unit/test_model_day_executor.py tests/unit/test_model_day_executor_reasoning.py -v
Expected: All tests PASS
Step 5: Commit test fixes
git add tests/unit/test_model_day_executor.py tests/unit/test_model_day_executor_reasoning.py
git commit -m "test: update tests after removing _write_results_to_db()"
Task 4: Fix Profit Calculation Logic (Bug #3)
Files:
- Modify:
agent_tools/tool_trade.py:144-157(buy function) - Modify:
agent_tools/tool_trade.py:287-300(sell function) - Modify:
tools/price_tools.py:417-430(no_trade function)
Background:
Current profit calculation compares portfolio value to previous day's final value. This is incorrect because:
- When you buy stocks, cash ↓ and stock value ↑ equally → portfolio unchanged → profit=0
- But we're comparing to previous day's final value, which makes trades look like losses
Correct approach:
Profit should be calculated by comparing to the start-of-day portfolio value (same day, action_id=0). This shows actual gains/losses from price movements and trading decisions.
Step 1: Fix profit calculation in buy function
In agent_tools/tool_trade.py, find the profit calculation in _buy_impl() (around lines 144-157):
# Get previous portfolio value for P&L calculation
cursor.execute("""
SELECT portfolio_value
FROM positions
WHERE job_id = ? AND model = ? AND date < ?
ORDER BY date DESC, action_id DESC
LIMIT 1
""", (job_id, signature, today_date))
row = cursor.fetchone()
previous_value = row[0] if row else 10000.0 # Default initial value
daily_profit = portfolio_value - previous_value
daily_return_pct = (daily_profit / previous_value * 100) if previous_value > 0 else 0
Replace with:
# Get start-of-day portfolio value (action_id=0 for today) for P&L calculation
cursor.execute("""
SELECT portfolio_value
FROM positions
WHERE job_id = ? AND model = ? AND date = ? AND action_id = 0
LIMIT 1
""", (job_id, signature, today_date))
row = cursor.fetchone()
if row:
# Compare to start of day (action_id=0)
start_of_day_value = row[0]
daily_profit = portfolio_value - start_of_day_value
daily_return_pct = (daily_profit / start_of_day_value * 100) if start_of_day_value > 0 else 0
else:
# First action of first day - no baseline yet
daily_profit = 0.0
daily_return_pct = 0.0
Step 2: Fix profit calculation in sell function
In agent_tools/tool_trade.py, find the profit calculation in _sell_impl() (around lines 287-300):
# Get previous portfolio value
cursor.execute("""
SELECT portfolio_value
FROM positions
WHERE job_id = ? AND model = ? AND date < ?
ORDER BY date DESC, action_id DESC
LIMIT 1
""", (job_id, signature, today_date))
row = cursor.fetchone()
previous_value = row[0] if row else 10000.0
daily_profit = portfolio_value - previous_value
daily_return_pct = (daily_profit / previous_value * 100) if previous_value > 0 else 0
Replace with:
# Get start-of-day portfolio value (action_id=0 for today) for P&L calculation
cursor.execute("""
SELECT portfolio_value
FROM positions
WHERE job_id = ? AND model = ? AND date = ? AND action_id = 0
LIMIT 1
""", (job_id, signature, today_date))
row = cursor.fetchone()
if row:
# Compare to start of day (action_id=0)
start_of_day_value = row[0]
daily_profit = portfolio_value - start_of_day_value
daily_return_pct = (daily_profit / start_of_day_value * 100) if start_of_day_value > 0 else 0
else:
# First action of first day - no baseline yet
daily_profit = 0.0
daily_return_pct = 0.0
Step 3: Fix profit calculation in no_trade function
In tools/price_tools.py, find the profit calculation in add_no_trade_record_to_db() (around lines 417-430):
# Get previous value for P&L
cursor.execute("""
SELECT portfolio_value
FROM positions
WHERE job_id = ? AND model = ? AND date < ?
ORDER BY date DESC, action_id DESC
LIMIT 1
""", (job_id, modelname, today_date))
row = cursor.fetchone()
previous_value = row[0] if row else 10000.0
daily_profit = portfolio_value - previous_value
daily_return_pct = (daily_profit / previous_value * 100) if previous_value > 0 else 0
Replace with:
# Get start-of-day portfolio value (action_id=0 for today) for P&L calculation
cursor.execute("""
SELECT portfolio_value
FROM positions
WHERE job_id = ? AND model = ? AND date = ? AND action_id = 0
LIMIT 1
""", (job_id, modelname, today_date))
row = cursor.fetchone()
if row:
# Compare to start of day (action_id=0)
start_of_day_value = row[0]
daily_profit = portfolio_value - start_of_day_value
daily_return_pct = (daily_profit / start_of_day_value * 100) if start_of_day_value > 0 else 0
else:
# First action of first day - no baseline yet
daily_profit = 0.0
daily_return_pct = 0.0
Step 4: Run bug tests to verify fix
Run: ./venv/bin/python -m pytest tests/unit/test_position_tracking_bugs.py::TestPositionTrackingBugs::test_profit_calculation_accuracy -v
Expected: Test PASSES
Step 5: Commit profit calculation fixes
git add agent_tools/tool_trade.py tools/price_tools.py
git commit -m "fix: correct profit calculation to compare against start-of-day value"
Task 5: Verify All Bug Tests Pass
Step 1: Run all bug tests
Run: ./venv/bin/python -m pytest tests/unit/test_position_tracking_bugs.py -v
Expected: All 3 tests PASS
Step 2: Run full test suite
Run: ./venv/bin/python -m pytest tests/ -v --cov=. --cov-report=term-missing --cov-report=html --tb=short
Expected: All tests PASS, coverage maintained or improved
Step 3: If any tests fail, debug and fix
If tests fail:
- Read the error message carefully
- Check which assertion failed
- Add debug prints to understand state
- Fix the issue
- Re-run tests
- Commit the fix
Step 4: Commit passing tests
git add -A
git commit -m "test: verify all position tracking bugs are fixed"
Task 6: Integration Test with Real Simulation
Files:
- Create:
tests/integration/test_position_tracking_e2e.py
Step 1: Write end-to-end integration test
Create tests/integration/test_position_tracking_e2e.py:
"""
End-to-end integration test for position tracking across multiple days.
Tests the complete flow: ModelDayExecutor → trade tools → database → position retrieval
"""
import pytest
from datetime import datetime
from api.database import get_db_connection, initialize_database
from api.job_manager import JobManager
from api.model_day_executor import ModelDayExecutor
from unittest.mock import patch, Mock, AsyncMock
def create_test_db_with_multi_day_prices(tmp_path):
"""Create test database with prices for multiple consecutive days."""
db_path = str(tmp_path / "test.db")
initialize_database(db_path)
conn = get_db_connection(db_path)
cursor = conn.cursor()
# Create price data for 5 consecutive trading days
dates = ["2025-10-06", "2025-10-07", "2025-10-08", "2025-10-09", "2025-10-10"]
base_price = 185.0
for i, date in enumerate(dates):
# Prices gradually increase each day
open_price = base_price + (i * 2.0)
close_price = base_price + (i * 2.0) + 1.5
for symbol in ["NVDA", "AAPL", "MSFT"]:
cursor.execute("""
INSERT INTO price_data (symbol, date, open, high, low, close, volume, created_at)
VALUES (?, ?, ?, ?, ?, ?, ?, ?)
""", (
symbol, date, open_price, open_price + 3, open_price - 1,
close_price, 1000000, datetime.utcnow().isoformat() + "Z"
))
conn.commit()
conn.close()
return db_path
@pytest.mark.integration
class TestPositionTrackingEndToEnd:
"""End-to-end tests for position tracking across multiple days."""
def test_multi_day_position_continuity(self, tmp_path):
"""
Test that positions correctly carry over across multiple trading days.
Scenario:
- Day 1: Start with $10,000, buy 5 NVDA @ $185
- Day 2: Cash should be $9,075, holdings should be 5 NVDA
- Day 3: Buy 3 AAPL @ $189
- Day 4: Should have 5 NVDA + 3 AAPL
- Day 5: Sell 2 NVDA @ $193
- Final: Should have 3 NVDA + 3 AAPL + increased cash
"""
db_path = create_test_db_with_multi_day_prices(tmp_path)
# Create job for 5-day simulation
manager = JobManager(db_path=db_path)
job_id = manager.create_job(
config_path="configs/test.json",
date_range=["2025-10-06", "2025-10-07", "2025-10-08", "2025-10-09", "2025-10-10"],
models=["test-model"]
)
# Mock agent that makes specific trades each day
def create_trading_agent(day, trades):
"""Create mock agent that executes specific trades."""
mock_agent = Mock()
mock_agent.get_conversation_history.return_value = []
# Mock async methods
async def run_session(date):
# Execute trades for this day
from agent_tools.tool_trade import _buy_impl, _sell_impl
from tools.general_tools import get_config_value
job_id = get_config_value("JOB_ID")
session_id = get_config_value("SESSION_ID")
for trade in trades:
if trade["action"] == "buy":
_buy_impl(
symbol=trade["symbol"],
amount=trade["amount"],
signature="test-model",
today_date=date,
job_id=job_id,
session_id=session_id
)
elif trade["action"] == "sell":
_sell_impl(
symbol=trade["symbol"],
amount=trade["amount"],
signature="test-model",
today_date=date,
job_id=job_id,
session_id=session_id
)
return {"success": True}
mock_agent.run_trading_session = AsyncMock(side_effect=run_session)
mock_agent.generate_summary = AsyncMock(return_value="Mock summary")
return mock_agent
# Day 1: Buy 5 NVDA
day1_trades = [{"action": "buy", "symbol": "NVDA", "amount": 5}]
# Day 2: No trades
day2_trades = []
# Day 3: Buy 3 AAPL
day3_trades = [{"action": "buy", "symbol": "AAPL", "amount": 3}]
# Day 4: No trades
day4_trades = []
# Day 5: Sell 2 NVDA
day5_trades = [{"action": "sell", "symbol": "NVDA", "amount": 2}]
all_trades = [day1_trades, day2_trades, day3_trades, day4_trades, day5_trades]
dates = ["2025-10-06", "2025-10-07", "2025-10-08", "2025-10-09", "2025-10-10"]
# Execute each day
for i, (date, trades) in enumerate(zip(dates, all_trades)):
with patch("api.model_day_executor.RuntimeConfigManager") as mock_runtime:
mock_instance = Mock()
mock_instance.create_runtime_config.return_value = f"/tmp/runtime_{i}.json"
mock_runtime.return_value = mock_instance
executor = ModelDayExecutor(
job_id=job_id,
date=date,
model_sig="test-model",
config_path="configs/test.json",
db_path=db_path
)
mock_agent = create_trading_agent(i, trades)
with patch.object(executor, '_initialize_agent', return_value=mock_agent):
result = executor.execute()
assert result["success"], f"Day {i+1} execution failed"
# Verify final positions
conn = get_db_connection(db_path)
cursor = conn.cursor()
# Get last position
cursor.execute("""
SELECT p.cash, p.portfolio_value
FROM positions p
WHERE p.job_id = ? AND p.model = ?
ORDER BY p.date DESC, p.action_id DESC
LIMIT 1
""", (job_id, "test-model"))
final_position = cursor.fetchone()
# Get final holdings
cursor.execute("""
SELECT h.symbol, h.quantity
FROM holdings h
JOIN positions p ON h.position_id = p.id
WHERE p.job_id = ? AND p.model = ?
ORDER BY p.date DESC, p.action_id DESC, h.symbol
LIMIT 10
""", (job_id, "test-model"))
holdings = {row[0]: row[1] for row in cursor.fetchall()}
conn.close()
# Verify final state
assert "NVDA" in holdings, "Should have NVDA position"
assert holdings["NVDA"] == 3, f"Expected 3 NVDA (bought 5, sold 2) but got {holdings['NVDA']}"
assert "AAPL" in holdings, "Should have AAPL position"
assert holdings["AAPL"] == 3, f"Expected 3 AAPL but got {holdings['AAPL']}"
# Cash should be less than initial $10,000 (we bought more than sold)
assert final_position[0] < 10000, f"Cash should be less than $10,000 but got ${final_position[0]}"
# Portfolio value should be roughly $10,000 (prices didn't change much)
# Allow some variation due to price movements
assert 9800 < final_position[1] < 10200, \
f"Portfolio value should be ~$10,000 but got ${final_position[1]}"
Step 2: Run integration test
Run: ./venv/bin/python -m pytest tests/integration/test_position_tracking_e2e.py -v
Expected: Test PASSES
Step 3: Commit integration test
git add tests/integration/test_position_tracking_e2e.py
git commit -m "test: add e2e integration test for multi-day position tracking"
Task 7: Update Documentation
Files:
- Modify:
CHANGELOG.md - Modify:
docs/developer/database-schema.md
Step 1: Update CHANGELOG.md
Add entry at the top of CHANGELOG.md:
## [Unreleased]
### Fixed
- **Critical:** Fixed position tracking bugs causing cash reset and positions lost over weekends
- Removed redundant `ModelDayExecutor._write_results_to_db()` that created corrupt records with cash=0 and holdings=[]
- Fixed profit calculation to compare against start-of-day portfolio value instead of previous day's final value
- Positions now correctly carry over between trading days and across weekends
- Profit/loss calculations now accurately reflect trading gains/losses without treating trades as losses
### Changed
- Position tracking now exclusively handled by trade tools (`buy()`, `sell()`) and `add_no_trade_record_to_db()`
- Daily profit calculation compares to start-of-day (action_id=0) portfolio value for accurate P&L tracking
Step 2: Update database schema documentation
In docs/developer/database-schema.md, find the section on the positions table and update the daily_profit and daily_return_pct field descriptions:
### positions
| Column | Type | Description |
|--------|------|-------------|
| ... | ... | ... |
| daily_profit | REAL | **Daily profit/loss compared to start-of-day portfolio value (action_id=0).** Calculated as: `current_portfolio_value - start_of_day_portfolio_value`. This shows the actual gain/loss from price movements and trading decisions, not affected by merely buying/selling stocks. |
| daily_return_pct | REAL | **Daily return percentage compared to start-of-day portfolio value.** Calculated as: `(daily_profit / start_of_day_portfolio_value) * 100` |
| ... | ... | ... |
**Important Notes:**
- **Position tracking flow:** Positions are written by trade tools (`buy()`, `sell()` in `agent_tools/tool_trade.py`) and no-trade records (`add_no_trade_record_to_db()` in `tools/price_tools.py`). Each trade creates a new position record.
- **Action ID sequence:**
- `action_id=0`: Start-of-day position (created by `ModelDayExecutor._initialize_starting_position()` on first day only)
- `action_id=1+`: Each trade or no-trade action increments the action_id
- **Profit calculation:** Daily profit is calculated by comparing current portfolio value to the **start-of-day** portfolio value (action_id=0 for the current date). This ensures that:
- Buying stocks doesn't show as a loss (cash ↓, stock value ↑ equally)
- Selling stocks doesn't show as a gain (cash ↑, stock value ↓ equally)
- Only actual price movements and strategic trading show as profit/loss
Step 3: Commit documentation updates
git add CHANGELOG.md docs/developer/database-schema.md
git commit -m "docs: update changelog and schema docs for position tracking fixes"
Task 8: Manual Testing with Real Simulation
Step 1: Create test configuration
Create configs/position_tracking_test.json:
{
"agent_type": "BaseAgent",
"date_range": {
"init_date": "2025-10-06",
"end_date": "2025-10-10"
},
"models": [
{
"name": "position-test",
"basemodel": "anthropic/claude-sonnet-4-20250514",
"signature": "position-tracking-test",
"enabled": true
}
],
"agent_config": {
"max_steps": 15,
"initial_cash": 10000.0,
"stock_symbols": ["NVDA", "AAPL", "MSFT", "GOOGL", "META"]
},
"log_config": {
"log_path": "./data/agent_data"
}
}
Step 2: Run simulation
# Make sure MCP services are running
cd agent_tools
python start_mcp_services.py &
cd ..
# Run simulation in DEV mode (no API costs)
DEPLOYMENT_MODE=DEV python main.py configs/position_tracking_test.json
Step 3: Verify results via API
# Start API server
uvicorn api.main:app --reload &
# Get job results
curl http://localhost:8000/api/v1/jobs | jq '.'
# Get positions for the job
JOB_ID="<job-id-from-above>"
curl "http://localhost:8000/api/v1/jobs/${JOB_ID}/positions?model=position-tracking-test" | jq '.'
Step 4: Verify position continuity
Check the output JSON:
- Cash continuity: Each day's starting cash should equal previous day's ending cash
- Holdings persistence: Stock positions should persist across days unless sold
- Profit accuracy: Profit should be 0 when buying/selling (no price change), and non-zero only when prices move
Step 5: Document test results
If all looks good, create a commit:
git add configs/position_tracking_test.json
git commit -m "test: add manual test config for position tracking verification"
Task 9: Final Verification and Cleanup
Step 1: Run complete test suite
./venv/bin/python -m pytest tests/ -v --cov=. --cov-report=term-missing --cov-report=html --tb=short
Expected: All tests PASS, coverage ≥ 90%
Step 2: Check for any remaining references to deleted methods
git grep -n "get_positions\|get_last_trade\|get_current_prices\|_write_results_to_db\|_calculate_portfolio_value" -- '*.py'
Expected: Only references in:
- Test files (mocking for legacy tests)
- Comments explaining the removal
- This plan document
If any production code references these, they need to be removed.
Step 3: Clean up any debug prints or temporary code
Review all modified files for:
- Debug
print()statements - Commented-out code
- TODO comments that should be addressed
Step 4: Final commit
git add -A
git commit -m "fix: complete position tracking bug fixes - all tests passing"
Step 5: Create summary report
Create docs/plans/2025-11-03-position-tracking-fixes-summary.md:
# Position Tracking Bug Fixes - Summary
**Date:** 2025-11-03
**Issue:** Three critical bugs in position tracking system:
1. Cash reset to initial value each trading day
2. Positions lost over non-continuous trading days (weekends)
3. Profit calculations showing trades as losses
## Root Causes
1. **Bugs #1 & #2:** `ModelDayExecutor._write_results_to_db()` called non-existent methods (`get_positions()`, `get_last_trade()`, `get_current_prices()`) on BaseAgent, resulting in corrupt position records with `cash=0` and `holdings=[]`. When `get_current_position_from_db()` retrieved positions for the next day, it found these corrupt records, causing cash resets and position losses.
2. **Bug #3:** Profit calculation compared portfolio value to **previous day's final value** instead of **start-of-day value**. Since buying/selling stocks doesn't change total portfolio value (cash ↓, stock value ↑ equally), this showed trades as having profit=0 or small rounding errors.
## Solution
1. **Removed redundant method:** Deleted `ModelDayExecutor._write_results_to_db()` and `_calculate_portfolio_value()` entirely. Position tracking is now exclusively handled by trade tools (`buy()`, `sell()`) and `add_no_trade_record_to_db()`.
2. **Fixed profit calculation:** Changed all profit calculations to compare against start-of-day portfolio value (action_id=0) instead of previous day's final value. This accurately reflects gains/losses from price movements and strategic trading.
## Files Modified
**Core Changes:**
- `api/model_day_executor.py`: Removed `_write_results_to_db()` call and method definitions
- `agent_tools/tool_trade.py`: Fixed profit calculation in `_buy_impl()` and `_sell_impl()`
- `tools/price_tools.py`: Fixed profit calculation in `add_no_trade_record_to_db()`
**Test Changes:**
- `tests/unit/test_position_tracking_bugs.py`: New tests demonstrating and verifying fixes
- `tests/unit/test_model_day_executor.py`: Updated mock helper and data persistence tests
- `tests/unit/test_model_day_executor_reasoning.py`: Skipped test for removed method
- `tests/integration/test_position_tracking_e2e.py`: New end-to-end integration test
**Documentation:**
- `CHANGELOG.md`: Added fix notes
- `docs/developer/database-schema.md`: Updated profit calculation documentation
## Verification
- All unit tests pass
- All integration tests pass
- Manual testing with 5-day simulation confirms position continuity
- Profit calculations accurate
## Impact
**Before:**
- Cash reset to $10,000 each trading day
- Positions lost after weekends
- Trades showed as $0 profit (misleading)
**After:**
- Cash carries over correctly between days
- Positions persist indefinitely until sold
- Profit accurately reflects price movements and trading strategy
Step 6: Final commit
git add docs/plans/2025-11-03-position-tracking-fixes-summary.md
git commit -m "docs: add summary report for position tracking bug fixes"
Success Criteria
All of the following must be true:
✅ All tests in tests/unit/test_position_tracking_bugs.py PASS
✅ All existing unit tests continue to PASS
✅ Integration test demonstrates position continuity across 5 days
✅ Manual testing shows correct cash carry-over
✅ Manual testing shows positions persist over weekends
✅ Manual testing shows profit=0 for trades without price changes
✅ Code coverage maintained at ≥ 90%
✅ No references to deleted methods in production code
✅ Documentation updated
✅ CHANGELOG updated
Rollback Plan
If issues are discovered after deployment:
-
Revert commits:
git revert <commit-hash-range> -
Re-run tests to verify rollback:
./venv/bin/python -m pytest tests/ -v -
Investigate root cause of rollback:
- Check logs
- Review test failures
- Identify what was missed
-
Create new fix with additional tests
Notes for Future Maintainers
Position Tracking Architecture:
-
Single source of truth: Trade tools (
buy(),sell()) andadd_no_trade_record_to_db()are the ONLY functions that write position records. -
ModelDayExecutor responsibilities:
- Create initial position (action_id=0) on first day via
_initialize_starting_position() - Manage trading sessions and reasoning logs
- Does NOT write position records directly
- Create initial position (action_id=0) on first day via
-
Action ID sequence:
action_id=0: Start-of-day baseline (created once, used for profit calculations)action_id=1+: Incremented for each trade or no-trade action
-
Profit calculation:
- Always compare to start-of-day (action_id=0) portfolio value
- Never compare to previous day's final value
- This ensures trades don't show false profits/losses
Testing:
test_position_tracking_bugs.pycontains regression tests for these specific bugs- If modifying position tracking, run these tests first
- Integration test (
test_position_tracking_e2e.py) verifies multi-day continuity