The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill provides an installation command referencing a remote repository at https://github.com/zozo123/ariflow-swfactory. The repository name contains a typo ('ariflow' instead of 'airflow') compared to the skill's name and documentation, which is a suspicious naming pattern despite the author 'zozo123' matching the skill metadata.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to process GitHub issues as 'work orders', which are then used to drive agenticSpec and Plan stages. This creates an indirect prompt injection vulnerability surface.
Ingestion points: External GitHub issue bodies are ingested and preserved as 'intent' text in docs/factory/<issue>/intent.md.
Boundary markers: The skill documentation (SKILL.md, references/sandboxes.md) emphasizes the use of sandbox providers like E2B, Daytona, and Docker to isolate agent cells, though it requires 'intent' text to be preserved verbatim.
Capability inventory: The skill possesses the ability to execute subprocesses (uv run swfactory), perform file system writes, and manage git commits.
Sanitization: The skill documents terminal escape sequence sanitization for operator interfaces in references/operator-interface.md but lacks specific sanitization or filtering for the 'intent' text before it is used in agent prompts.
[COMMAND_EXECUTION]: The skill relies on various CLI tools and scripts including uv run swfactory, a native binary named swf, and helper scripts like scripts/snapshot_diff.py. These tools execute commands and logic derived from the repository environment.

airflow-software-factory