The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from git history (commits, diffs, PRs), which is an untrusted source.
Ingestion points: The skill reads git log and git show outputs, including commit subjects, bodies, and code diffs as specified in references/git-collection.md.
Boundary markers: references/git-collection.md explicitly instructs the agent to treat repository text as evidence and never to follow any instructions found within that text.
Capability inventory: The agent has the ability to write files to the local file system (<artifacts-root>/specs/release-notes/) and execute shell commands (git).
Sanitization: While the skill instructs the agent to translate jargon and summarize, there is no explicit instruction for technical sanitization of the input text before processing.
[COMMAND_EXECUTION]: The skill requires the execution of shell-based git commands.
Evidence: references/git-collection.md provides specific commands like git log --all --after="..." --before="..." and git merge-base --is-ancestor <hash> <ref>. These commands use placeholders that are filled based on user-provided or session-derived dates and hashes, which could lead to command injection if the input is not strictly validated by the agent platform.

release-notes