The Agent Skills Directory

[DYNAMIC_EXECUTION]: The scripts/job.py script uses the importlib module to dynamically load and execute Python code from files in the jobs/ directory. These job scripts are intended to be authored by the AI agent. Although the skill employs a static linter (lint_source) to detect banned Playwright calls and hardcoded credentials, executing locally generated code remains a high-privilege capability that could be abused if the agent is compromised or the linter is bypassed.\n- [DATA_EXFILTRATION]: The scripts/pagestate.py utility communicates with an external service at https://api.typesafe.ai/v1/systemone to diagnose failed browsing runs. The script sends redacted page titles and visible text excerpts (masked for URLs, emails, and digits) to this third-party API. While redaction and a required API key are used as safeguards, this process involves transmitting content from authenticated browsing sessions to an external domain not on the trusted list.\n- [INDIRECT_PROMPT_INJECTION]: The skill interacts with external, untrusted websites and captures their state for error reporting. Malicious instructions embedded in these pages could potentially influence the agent during the task authoring or failure repair phases.\n
Ingestion points: External website content captured into failure.json and processed by pagestate.py.\n
Boundary markers: No explicit instruction wrappers or delimiters were identified to isolate untrusted web content from agent instructions.\n
Capability inventory: Dynamic execution of Python scripts via job.py and full browser automation via Playwright.\n
Sanitization: The redact function in pagestate.py masks URLs, email addresses, and digit sequences of four or more characters.\n- [COMMAND_EXECUTION]: The skill provides numerous CLI commands for scaffolding, linting, running, and diagnosing jobs, which involve shell-level operations and environment variable manipulation.

playwright-stealth-job