The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/multi_model_review.py executes several local command-line interface (CLI) tools, including gemini, codex, and claude. It utilizes asyncio.create_subprocess_exec to invoke these binaries, passing large prompts containing user-sourced code and specifications as command-line arguments.
[DATA_EXFILTRATION]: The skill reads source code and specification files from the local filesystem (via _collect_specs, _collect_programs, and _collect_tests) and transmits this data to external AI service providers, such as api.openai.com. While this is required for its function as a multi-model reviewer, it involves the outbound transmission of potentially sensitive project data to third-party servers.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection. It ingests untrusted data from local project files and interpolates it directly into system prompts for multiple LLMs. This creates a surface where a malicious actor could embed instructions within code comments or documentation to manipulate the review output.
Ingestion points: Files read from --spec-dir, --program-dir, and --test-dir via SpecProgramTestCollector in scripts/multi_model_review.py.
Boundary markers: Uses markdown code blocks (```yaml) and headers (## SPECIFICATION) to delimit content within the prompt.
Capability inventory: Performs network requests via httpx to OpenAI and local Ollama endpoints, and executes local binaries via subprocess.
Sanitization: None. The content is read directly from the filesystem and truncated for length without escaping or filtering for embedded instructions.

multi-model-reviewer