The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Fetches reference source code from official GitHub repositories of organizations including OpenAI, Google Gemini, Microsoft, and Zed Industries. These clones provide a 'ground truth' evidence base for validating platform behavior claims and are used for reading rather than execution.
[COMMAND_EXECUTION]: Utilizes the GitHub CLI (gh) for issue management, pull request reviews, and release automation. It also executes standard build and test commands using npm, npx, and git to verify code quality and manage versioning manifests.
[PROMPT_INJECTION]: The instructions employ a highly structured 'Engineering Manager' persona and assertive language (e.g., 'ABSOLUTE', 'NON-NEGOTIABLE') to enforce strict adherence to the project's verification and testing protocols. This serves to guide subagent behavior towards reliable outcomes.
[PROMPT_INJECTION]: The skill processes external data from GitHub issues and pull requests, creating an indirect prompt injection surface. This is mitigated by a mandatory 'Claim Verification' gate that requires hard evidence (reproducible tests or official documentation) before any code changes are implemented. The analysis flow involves:
Ingestion points: External data is ingested through gh issue view and gh pr view commands in triage-issue.md and review-pr.md.
Boundary markers: Explicit instructions and the 'Claim Verification' gate instruct the agent to ignore unverified claims and focus on technical proof.
Capability inventory: The skill has capabilities for executing tests via npx vitest, performing GitHub operations via gh, and modifying version manifests.
Sanitization: All behavioral claims are cross-referenced against freshly-cloned official source code in the refs/ directory and verified through reproduction scripts.

context-mode-ops