The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from multiple sources including build logs, workflow JSON files, and environment snapshots.
Ingestion points: Untrusted data is ingested when the agent reads build logs via comfy build release logs or imports files using flags like --from-workflow.
Boundary markers: The instructions include a specific security directive to treat build logs as attacker-controlled text and forbid executing commands or using literals directly from them.
Capability inventory: The skill can execute shell commands via the comfy CLI and write configuration files to the disk.
Sanitization: The agent is directed to translate logs into its own words, providing a layer of conceptual sanitization.
[COMMAND_EXECUTION]: The skill uses the comfy CLI to perform all operations, ranging from local scans to cloud releases.
[EXTERNAL_DOWNLOADS]: The skill references the installation of the comfy-cli package and mentions that the build platform fetches models from remote URLs provided in the configuration.

comfy-build