The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from various external sources including CI logs, git commit diffs, and dependency manifest outputs to generate reports and suggested fixes. This creates an attack surface where a malicious actor could embed prompt injection instructions within a commit message or a simulated CI failure log to influence the agent's behavior during the QA process.
Ingestion points: Git commit history (git log), GitHub Action logs (gh run view), and dependency update metadata (npm/pip/cargo outdated).
Boundary markers: The instructions lack explicit technical delimiters or warnings for the agent to ignore instructions embedded within the logs, though they do mandate that all findings be grounded in concrete evidence.
Capability inventory: The skill has access to the Bash tool for shell execution, WebFetch for network access, and the ability to write files to the repository (reports and .gitignore).
Sanitization: No explicit sanitization or escaping of ingested log content is mentioned.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to execute various repository management and interrogation commands, including git, gh, npm, pip, pnpm, and cargo. While these are used for their intended QA purposes, they represent a significant execution capability within the repository environment.
[EXTERNAL_DOWNLOADS]: The skill is authorized to use WebFetch to retrieve external data, specifically identified for fetching upstream changelogs and dependency information to assess 'SDK drift'.

daily-qa