The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill guides the agent to generate and execute "minimal reproducer scripts" in Step 8 of the logic-review-guide.md to verify findings. While the instructions include safety constraints—such as prohibiting network requests, file writes, and database operations—this involves the runtime execution of code generated by the agent based on potentially untrusted input.
[REMOTE_CODE_EXECUTION]: The Step 8 verification process involves the agent checking for local runtimes (e.g., python3, node, go) and executing scripts to confirm the presence of bugs and the effectiveness of remedies.
[EXTERNAL_DOWNLOADS]: The skill references an external repository (github.com/hyhmrright/logic-lens) as the upstream source for supporting materials and a grading script (grade-iteration.py). It instructs the agent to obtain required materials from this source if they are not present locally.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze untrusted code provided by the user.
Ingestion points: Code or snippets shared by the user for review.
Boundary markers: The skill provides a structured framework (Premises, Trace, Divergence) and instructions to treat the input as code to be analyzed.
Capability inventory: Step 8 allows for the execution of generated scripts using local runtimes (python3, node, go).
Sanitization: The skill lacks explicit sanitization or escaping of the user-provided code before it is potentially incorporated into the reproducer scripts used in the execution verification gate.

logic-review