The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill extracts shell commands from the Verify and Final Verification fields of the blueprint.md file and executes them using subprocess.run(shell=True). This allows the agent to run arbitrary validation logic defined within the repository's blueprint artifacts.
[DATA_EXFILTRATION]: During operation, the skill captures detailed repository metadata, including file existence, sizes, and SHA-256 hashes of all implementation targets, as well as the output of Git status commands. This state information is persisted in a construction.json bundle located in the external run directory.
[PROMPT_INJECTION]: The skill is primarily driven by instructions ingested from blueprint.md and request.md. These files contain natural language goals, approaches, and behavioral slices that the agent must interpret and execute, bypassing direct user prompting for individual steps.
[INDIRECT_PROMPT_INJECTION]: The skill's architecture relies on processing instructions embedded in repository files.
Ingestion points: blueprint.md and request.md files located in the blueprints/<task>/ directory.
Boundary markers: The skill validates the 'sealed' state of the blueprint using blueprint-validation HMAC-like receipts that bind the document to a specific request hash and repository HEAD. It also enforces a specific blueprint-contract version.
Capability inventory: The skill uses subprocess.run for shell execution, performs Git operations, and has write access to both the target repository and an external state directory.
Sanitization: The skill performs basic checks on extracted commands to ensure they are not empty and do not contain template placeholders (e.g., <fill-me>) before execution.

construct