The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from repository files, issue tickets, and tool receipts to evaluate readiness. While this creates a surface for indirect prompt injection, the skill mitigates risks through explicit identity re-verification of the head and base refs before approval, and mandatory sanitization of evidence packs to prevent exfiltration of sensitive data like credentials.
Ingestion points: Reads repository source files, linked issue tickets, changelogs, and design critique snapshots.
Boundary markers: Utilizes strict status words (e.g., 'verified', 'attested') and requires a fresh re-read of the working surface immediately before accepting approval.
Capability inventory: Executes bundled shell scripts for repository analysis and dispatches companion skills (e.g., 'checking-simplicity') for specialized reviews.
Sanitization: Step 7 explicitly instructs the agent to sanitize the evidence pack by summarizing intent and omitting credentials or unnecessary personal data.
[COMMAND_EXECUTION]: The skill executes local, bundled bash scripts (surface-report.sh, changelog-union.sh, evidence-freshness.sh) to perform technical audits of the git environment. These scripts use standard POSIX tools and git commands, and they explicitly avoid network operations or external dependencies.
[SAFE]: The skill maintains a read-only posture regarding the host repository, delegating all write operations (staging, committing, pushing, or opening PRs) to separate finishing skills. Sensitive identity and surface state are captured and re-validated at the decision point to prevent state-change bypasses.

checking-pr-readiness