The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes the GitHub CLI (gh), unzip, and python3 to automate repository inspection. These commands are executed using node:child_process.execFile with sanitized arguments, which effectively prevents shell injection vulnerabilities.- [PROMPT_INJECTION]: The skill processes untrusted data from GitHub, such as pull request comments and Action logs, creating a surface for indirect prompt injection. Findings: 1. Ingestion points: Data enters via pr-snapshot.ts, pr-threads.ts, and ci-failures.ts when reading PR metadata or downloading job logs. 2. Boundary markers: The instructions explicitly tell the agent to summarize output and cite file paths rather than echoing raw content; additionally, the sanitizeForTerminal function is used to strip control characters. 3. Capability inventory: The skill can execute gh commands and local utilities like unzip and python3 for log processing. 4. Sanitization: The skill uses sanitizeForTerminal to remove C0/C1 control characters from external strings and employs a slug function to ensure filenames are safe.- [CREDENTIALS_UNSAFE]: The skill accesses the user's GitHub token via gh auth token for API authentication. It includes specific logic in scripts/lib.ts and REFERENCE.md to prevent this token from being logged or leaked to external blob storage during log retrieval.

gh