The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection as it ingests and processes data from external, potentially untrusted sources to derive its verification requirements.
Ingestion points: Pull request bodies, branch descriptions, and external task or specification files are used to resolve implementation intent during Phase 0.
Boundary markers: The intent is distilled into a numbered requirement list within the plan.md artifact to provide a structured contract for validation.
Capability inventory: The skill possesses the ability to execute shell commands (baseline gates and fix waves) and write to the local file system (Phase 3 fix loop).
Sanitization: The instructions do not specify any explicit sanitization or escaping of the external content before it is used to influence agent behavior.
[COMMAND_EXECUTION]: The skill executes local development commands to validate implementations against their specifications.
Evidence: As defined in references/stack-adapters.md and references/fix-spec-template.md, the agent runs shell commands for testing (e.g., npm test, pytest), linting (e.g., eslint, ruff), and type checking (e.g., tsc, mypy).
[EXTERNAL_DOWNLOADS]: The skill may perform external downloads to resolve environment issues during preflight.
Evidence: The 'unblock protocol' described in references/preflight.md allows the agent to install missing dependencies or binaries (with a limit of two attempts) to enable the execution of gate commands.

nuke-verify