The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted content from GitHub event payloads, which may contain malicious instructions.
Ingestion points: Event payloads (specifically the JSON string in the payload field) from githubarchive tables processed during query execution (SKILL.md).
Boundary markers: The skill includes explicit instructions for the agent to treat these payloads as data, ignore embedded instructions, and report them as evidence (SKILL.md).
Capability inventory: The skill uses libexec/raptor-bq-query to execute SQL and can write query files to the local file system (SKILL.md).
Sanitization: The raptor-bq-query wrapper is read-only and runs in a network-pinned sandbox (SKILL.md).
[COMMAND_EXECUTION]: The skill involves writing SQL queries to local files and executing a command-line wrapper (libexec/raptor-bq-query) to interact with BigQuery (SKILL.md).
[EXTERNAL_DOWNLOADS]: The skill instructions specify installing the official google-cloud-bigquery and google-auth packages from standard Python registries (SKILL.md).

github-archive