The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the GitHub CLI (gh) to interact with repository data, including viewing issues and creating pull requests. It also executes standard unix utilities (jq, awk, sed, git) to manage internal state and verify progress.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from GitHub Issue titles and bodies. It mitigates injection risks by delegating processing to specialized worker skills and maintaining a multi-stage verification pipeline (Inspect and Review phases) where machine-checkable evidence and adversarial peer reviews are used to validate outcomes.
[DYNAMIC_EXECUTION]: The orchestrator supports an 'Evaluate' phase that launches applications using user-defined commands from configuration (e.g., npm run dev). It also includes a self-improvement loop in the 'Retrospective' phase, which allows for automated proposals to update reference files. This is protected by a 'Tier' system that restricts automatic merging to non-executable reference files and enforces human review for changes to logic, scripts, or core skill definitions.
[REMOTE_CODE_EXECUTION]: The conductor delegates implementation and testing tasks to other agent skills. While these skills may perform code operations, the orchestrator acts as a supervisor, enforcing 'fail-closed' recovery contracts and verifying all outputs against predeclared expectations and cryptographic hashes.

spec-orchestrate