The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the repository's branch diff and GitHub pull request metadata, creating a surface where embedded instructions could potentially influence the reviewer subagents. 1. Ingestion points: The skill uses git diff and gh pr view to collect data that is interpolated into the prompts for six parallel reviewers. 2. Boundary markers: The shared context block uses headers to delimit sections, but does not provide explicit escape sequences to prevent the subagents from interpreting instructions contained within the diffs or PR comments. 3. Capability inventory: Reviewer subagents are configured as general-purpose, granting them the ability to read arbitrary files and execute search tools in the local environment. 4. Sanitization: The skill does not sanitize or filter the content of the diffs or PR metadata before processing. However, the subagent prompts include strict analytical directives that mitigate the risk of following adversarial instructions.
[DYNAMIC_CONTEXT_INJECTION]: The skill uses the ! syntax to execute shell commands at load time for context gathering. Evidence: The SKILL.md file contains several ! commands including git rev-parse, git diff, git ls-files, and gh pr view. Context: These commands are benign developer utilities used to characterize the state of the repository for the review task and do not incorporate external or unvalidated input into the command strings.

squad-review