The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted financial statements (XML) which provides a potential injection surface for instructions embedded in data fields.
Ingestion points: External XML statements are read and parsed by the scripts/adapters/ibkr.py script.
Boundary markers: The SKILL.md file contains explicit instructions for the agent to treat all statement fields and retrieved documents strictly as data and not as instructions.
Capability inventory: The skill possesses the ability to execute sibling skill scripts via the CLI and write local report files using scripts/analyze_portfolio.py.
Sanitization: A safe() function in scripts/analyze_portfolio.py escapes Markdown control characters and HTML tags (e.g., |, <, >) before interpolating parsed data into workpapers and reports.
[DYNAMIC_EXECUTION]: The XML parsing logic in scripts/adapters/ibkr.py utilizes the standard library's xml.etree.ElementTree module.
Mitigation: The script implements a manual check for <!DOCTYPE and <!ENTITY declarations within the raw XML content to prevent XML External Entity (XXE) attacks.
[DATA_EXFILTRATION]: The skill is designed to handle sensitive financial data from broker exports.
Mitigation: Sensitive account identifiers are replaced with run-local pseudonyms (e.g., account-1, account-2) during normalization.
Mitigation: The instructions explicitly direct the agent to keep raw exports and balances local and out of version control, emphasizing that external research should only utilize anonymous instrument identifiers like ISINs or tickers.
[COMMAND_EXECUTION]: The skill provides a CLI for local parsing and instructs the agent to coordinate research by executing scripts from sibling skills (financial-reports-scraper and fundamental-analysis). These operations are scoped to the local environment to preserve data privacy and are consistent with the skill's stated purpose.

portfolio-analysis