The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes Bash commands to perform LaTeX compilation, directory management, and file auditing. Permitted tools are restricted in the frontmatter to a specific set including latexmk, xelatex, pdflatex, biber, bibtex, mkdir, ls, wc, and cp. These commands are invoked in Phase 1 (pre-flight) and Phase 2 (compile-fix loop) to manage build artifacts and process documents.
[UNVERIFIABLE_DEPENDENCIES_AND_REMOTE_CODE_EXECUTION]: In Phase 6.1, the skill runs a local Python script using 'uv run python' to verify output consistency. The script, 'verify_outputs.py', is located within a shared utility directory ('_shared') relative to the skill root, representing internal framework functionality rather than an external or untrusted dependency.
[DYNAMIC_EXECUTION]: The skill implements an autonomous compile-fix loop that iterates up to 5 times. It parses compilation logs for known error patterns defined in 'references/known-errors.md' and dynamically modifies the LaTeX source to resolve issues, such as missing packages or font conflicts, before retrying the build.
[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted content from user-provided LaTeX and bibliography files during its citation audit (Phase 5) and source pathology lint (Phase 4). While these ingestion points exist, the skill's capabilities are focused on structural and bibliographic analysis, with no instructions detected that would lead to agent bypass or unsafe operations.
[DATA_EXPOSURE_AND_EXFILTRATION]: No exfiltration patterns were detected. The skill's operations are confined to the local project directory and standard output paths. Network tools like curl or wget are not included in the 'allowed-tools' configuration.

latex