The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill (SKILL.md and references/blast-radius-procedure.md) explicitly instructs the agent to generate and execute local proof scripts or tests, typically within the repository's .tmp/ directory, to verify the safety of code changes.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data such as code diffs, pull requests, and proposed changes. Because the agent is prompted to generate and run code based on this input, it creates a potential vulnerability where malicious instructions within the diff could influence the generated script or test logic.
Ingestion points: Ingests code diffs, pull requests (PRs), and developer-supplied proposed changes as primary input in SKILL.md.
Boundary markers: The instructions mention honoring authorized scope and report-only limits, but the skill lacks strict data/instruction delimiters to separate untrusted input from agent instructions.
Capability inventory: The skill has the capability to write files (proof scripts) and execute local shell commands for script running and testing (referenced in SKILL.md and blast-radius-procedure.md).
Sanitization: The procedure includes a general instruction to "strip anything private before it goes anywhere public," but lacks technical sanitization or validation of the ingested diff data before it influences code generation.
[COMMAND_EXECUTION]: The skill workflow involves triggering the execution of local scripts and tests to prove "safety facts" about the code under review.

blast-radius