The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill dynamically identifies the path to its execution scripts using a find command that searches the local directory structure (*/figures/agents/figure-qa-scripts). It then executes these discovered scripts using uv run. This dynamic path resolution allows for the execution of arbitrary scripts if a user is working within a malicious project structure that mimics the expected plugin layout.
[COMMAND_EXECUTION]: The skill uses uv run to execute Python scripts for figure analysis. It also employs shell utilities including find, grep, ls, and test for file type detection and path discovery. Furthermore, it executes a Python one-liner to extract code from Jupyter Notebooks before analysis.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it reads untrusted data into the agent's context without sanitization.
Ingestion points: SVG source code, Python plot scripts, and Jupyter Notebook cells are read in references/figure-qa-procedure.md.
Boundary markers: Absent; content is read directly for VLM evaluation.
Capability inventory: Access to shell execution (uv run), file system searches (find), and file reads.
Sanitization: None; the raw content is passed to the VLM rubric pass. Maliciously crafted figures could contain instructions to force a 'ship' status or bypass journal compliance checks.

figure-qa