The Agent Skills Directory

[COMMAND_EXECUTION]: The scripts/update-preview.py script executes multiple system commands using subprocess.run, including git, daytona, and pnpm. These are used to manage the preview lifecycle and environment within the sandbox.
[DYNAMIC_EXECUTION]: The update-preview.py script dynamically constructs a Python program as a string and executes it inside a remote Daytona sandbox using python3 -c.
Evidence: The script uses remote.replace("REF", repr(args.ref)).replace("WEB", repr(web)) to inject arguments into a code template before execution.
Mitigation: The script applies strict regex validation to input arguments (--ref must be a 40-character hex string) and uses shlex.quote and repr() to safely handle injected values.
[REMOTE_CODE_EXECUTION]: The skill is designed to download and execute code from remote repository branches/commits to generate previews.
Evidence: git fetch origin ref followed by pnpm install and pnpm build within the sandbox.
Mitigation: The skill enforces the use of full 40-character commit SHAs to prevent branch hijacking and executes the code within isolated Daytona sandboxes.
[INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied data such as commit SHAs and stage names which are used to control the build process.
Ingestion points: Command-line arguments (--ref, --stage) in scripts/update-preview.py.
Boundary markers (absent): No specific delimiters are used in the prompt instructions to isolate these inputs, though the helper script enforces strict formatting.
Capability inventory: File writing (/tmp/preview-update.log), network operations (git fetch), and process management (os.kill) are performed by the script.
Sanitization: Inputs are strictly validated against regex patterns ([0-9a-f]{40}) before being used in commands or scripts.

preview-my-work