Overview

This dev log focuses on backend stabilization work for trading-agent. We added year fallback and PBR calculation logic to the DART financial data client, fixed a bug where current_price was missing from the market scanner pipeline, resolved an async/sync mixing issue in FastMCP middleware, and on the frontend improved SignalPanel with date-based grouping and upgraded the DAG workflow visualization.

Previous post: #4

DART Client Improvements

Background

Fetching financial data from the DART (Electronic Disclosure System) API had two problems:

1. Missing year data: When the latest year’s financial statements hadn’t been disclosed yet, the API returned an empty response. Without fallback logic to try a prior year, the analysis agent had to make decisions with no financial data.
2. PBR not calculated: PER is provided directly by the API, but PBR (Price-to-Book Ratio) is not. Even though market cap and net asset data was available, PBR wasn’t being calculated.
3. Industry-specific field differences: Financial statement line item names differ between the financial sector and general companies, causing parse errors for certain industries.

Implementation

Improvements made to backend/app/services/dart_client.py:

Year Fallback Logic:

# Try from current year, find a year with available data
for year in range(current_year, current_year - 3, -1):
    result = await self._fetch_financial_data(corp_code, year)
    if result and result.get("list"):
        break

Auto PBR Calculation:

# Calculate PBR from net equity (total capital) and market cap
if total_equity and total_equity > 0:
    pbr = market_cap / total_equity

Industry-specific Field Mapping:

Financial sector companies (banks, insurance, securities) use Operating Revenue instead of Operating Income, and reference Interest Income etc. instead of Revenue — branching logic was added for this.

Market Scanner Pipeline Fix

Background

After the market scanner scans stocks and passes them to each specialist agent (technical analysis, fundamental analysis, etc.), the current_price field was missing. The scanner fetches price data but wasn’t passing it along when calling downstream experts.

Implementation

flowchart LR
    A["Market Scanner"] -->|"Stock list +<br/>current_price"| B["Expert Agents"]
    B -->|"Analysis results"| C["Chief Analyst"]
    C -->|"Final signal"| D["Signal Queue"]

Updated backend/app/agents/market_scanner.py to explicitly pass current_price when calling experts:

# Before: price info missing from expert call
expert_result = await expert.analyze(stock_code, stock_name)

# After: current_price passed through entire pipeline
expert_result = await expert.analyze(stock_code, stock_name, current_price=price)

Also simplified the chief analyst’s debate logic in market_scanner_experts.py. The old approach had all expert opinions debating sequentially — reducing unnecessary rounds improved response time.

FastMCP Middleware async/sync Bug Fix

Problem

A method accessing context.state in MCP server middleware was a synchronous function being called with await:

# Bug: await on sync function
state = await ctx.get_state("trading_mode")  # TypeError!

FastMCP’s context state methods are synchronous functions. await-ing them either causes a TypeError when called on a non-coroutine, or in some Python versions silently returns None.

Fix

Removed await from open-trading-api/MCP/Kis Trading MCP/module/middleware.py and tools/base.py:

# Fix: sync method called directly without await
state = ctx.get_state("trading_mode")

Scheduled Tasks Activated

Updated scheduler-related settings in backend/app/models/database.py:

• Activated scheduled tasks that were previously disabled
• Adjusted cron timings to match Korean market hours (pre-market scan, intraday monitoring, post-market report)

Frontend Improvements

SignalPanel Date Grouping

Added date-based collapsible sections to frontend/src/components/dashboard/SignalPanel.tsx. Previously, all signals were listed chronologically, making it difficult to find signals from a specific date.

flowchart TD
    A["SignalPanel"] --> B{"Date grouping"}
    B --> C["2026-03-19<br/>5 signals"]
    B --> D["2026-03-18<br/>3 signals"]
    B --> E["2026-03-17<br/>8 signals"]
    C --> F["Collapsible<br/>expand/collapse"]

Daily Chart Data Extended

Extended the daily chart data fetch period in backend/app/services/market_service.py from 30 days to 90 days. This was needed to have sufficient data for moving average calculations (60-day, 90-day) in technical analysis.

DAG Workflow Styling

Updated the agent pipeline DAG visualization layout and expert chip styling in frontend/src/components/AgentWorkflow.css and AgentWorkflow.tsx. Adjusted node spacing, connector label positions, and overall container alignment for improved readability.

Commit Log

Takeaways

• async/sync mixing creates silent bugs. In Python, await-ing a sync function can return None instead of raising an error in some runtimes. When using libraries like FastMCP where sync and async coexist, you must verify each method’s signature.
• Missing data in pipelines is a common mistake. The scanner fetching price but not passing it to experts happened because each stage was tested independently. A reminder of the need for end-to-end tests.
• Financial data APIs require industry-specific handling. Financial sector financial statements are fundamentally structured differently from general companies. If you don’t pre-map these variations when wrapping the DART API, KeyError will hit you at runtime.
• Chart data range and analysis indicators must be designed together. Calculating a 90-day moving average requires at least 90 days of data — we were only fetching 30 days. Whenever adding technical analysis indicators, the data source’s range needs to be checked at the same time.