The Agent Skills Directory

[SAFE]: The skill implements a bridge between the AI agent and an Overleaf browser session to automate LaTeX compilation and diagnostic extraction. Actual behavior matches the stated purpose.
[COMMAND_EXECUTION]: The bash wrapper script executes local commands (python3, awk, sed, grep) and invokes the cdp.mjs tool from the chrome-cdp skill to communicate with the browser. These operations are within the scope of the skill's functionality.
[EXTERNAL_DOWNLOADS]: The skill does not perform external downloads or fetch remote code at runtime. It relies on local scripts (page-helpers.js, parse_log.py) and pre-installed local dependencies.
[DATA_EXFILTRATION]: The skill accesses browser session data (CSRF tokens and cookies) to trigger compilations. These credentials are only used to authenticate requests to the Overleaf domain (/project/:id/compile). No exfiltration to unauthorized third-party domains was detected.
[OBFUSCATION]: A static analysis hint flagged directional Unicode control characters (U+202A, U+202E, etc.) in parse_log.py. Technical review confirmed these characters are part of a regex (_BIDI) explicitly documented as stripping bidirectional markers from raw Overleaf logs to ensure clean parsing. This is a false positive and represents a sanitization best practice.
[INDIRECT_PROMPT_INJECTION]: The skill has an ingestion surface for untrusted data (LaTeX and BibTeX log files). While these logs could contain malicious content designed to influence the agent, the skill processes them as structured data. The analysis identifies this as a standard risk for any skill processing external logs, mitigated by the agent's internal safety guardrails.

overleaf-compile