The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from git diffs and external reviewer responses. It mitigates injection risks by requiring a host-side adapter to enforce a closed structural schema, stripping instructions or tool calls from results. Ingestion points: review_target (SKILL.md) and delegated_result_record (review-workflow.md). Boundary markers: inert evidence framing and closed-structural schema. Capability inventory: git operations and file writes to docs/decisions/. Sanitization: secret-redaction overlay (review-workflow.md).
[COMMAND_EXECUTION]: The skill possesses the capability to execute git commands (e.g., commit, rebase, push) for repository management. These are restricted by mandatory gates (Commit-Selection Gate and History-Mutation Gate) requiring recorded selection sources and human confirmation for destructive or state-changing operations.
[DATA_EXFILTRATION]: The workflow manages exfiltration risks by enforcing a Secret Hygiene policy across all output boundaries, redacting sensitive content before it is persisted or displayed to the user.
[CREDENTIALS_UNSAFE]: The skill proactively manages credential safety by redacting API keys, JWTs, and private keys from all code snippets and records used during the review process.
[PROMPT_INJECTION]: The skill utilizes meta-level directives (shared-contract blocks) to define instruction precedence across multiple files. These are used to enforce stricter safety constraints rather than to bypass platform protections.

vibe-review