The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses subprocess.run in assets/forge.py and assets/verify_evidence.py to interact with the project's Git repository for operations like status checks, diffing, and metadata retrieval. These actions are essential for its primary function of capturing evidence linked to specific commits.
[DYNAMIC_EXECUTION]: The assets/verify_evidence.py script dynamically executes shell commands provided via the --cmd argument during proof recording tasks. This capability is used to run formal verification or property-based testing tools within a clean checkout of the project. Evidence: Commands are executed using subprocess.run after being parsed by shlex.split. Mitigation: The skill implements a runner_matches validation function that restricts execution to an allowlist of recognized testing runners (e.g., verus, kani, pytest, cargo test, vitest, go test).
[INDIRECT_PROMPT_INJECTION]: The skill processes data from the target repository, creating an attack surface for indirect prompt injection. Ingestion points: assets/forge.py reads and parses features/*.md and .manifold/*.json files from the project directory. Boundary markers: The skill does not use explicit boundary markers to delimit untrusted data in these files, though it enforces specific structural headers. Capability inventory: The skill possesses file-writing capabilities (scaffolding and recording) and command-execution capabilities (git and validated test runners). Sanitization: Data is parsed using standard JSON/Markdown logic, and commands are verified against a predefined list of valid runners.

verification-skill-forge