The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill implements a self-update mechanism in scripts/self_update.py and a companion manager in scripts/companion_manager.py. These scripts can download ZIP packages from a remote server (as specified in API responses), verify their SHA-512 hashes, and extract them directly into the skill directory, replacing existing executable scripts. This pattern allows for arbitrary code execution if the update server or the communication channel is compromised.
[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection. It ingests data from external sources, including search result snippets from the Bocha API (via scripts/event_study_local.py) and financial datasets from the Quant Buddy API. This external content is interpolated into the agent's context for analysis and report generation without the use of delimiters or explicit instructions for the model to ignore embedded commands. This risk is significant given the skill's capabilities to execute shell commands and modify files.
Ingestion points: scripts/event_study_local.py (Bocha search results), scripts/fetch_fastquery_csv.py (external CSV data).
Boundary markers: Absent in the instructions and scripts.
Capability inventory: subprocess.run in scripts/call.py and scripts/companion_manager.py.
Sanitization: No sanitization logic found for external text snippets.
[DYNAMIC_EXECUTION]: The script scripts/research_contract.py uses the __import__ function to dynamically load the math module at runtime to perform validation checks.
[COMMAND_EXECUTION]: The skill's primary entry point, scripts/call.py, uses the subprocess module to execute other internal scripts such as executor.py and self_update.py. Furthermore, scripts/companion_manager.py invokes powershell.exe to manage directory junctions on Windows, demonstrating a clear capability to spawn and interact with system processes.

quant-buddy-skill