The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to execute shell commands to inspect changes and verify implementation integrity.
Evidence: The 'First Action' requires the agent to run git diff to inspect code changes directly.
Evidence: The 'Mechanical Checks' section instructs the agent to run the project's canonical test suite and lint/static analysis commands.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from the code implementation, task descriptions, and implementer status reports, which could contain malicious instructions intended to manipulate the review outcome or the commands executed.
Ingestion points: The skill ingests tasks.md, requirements, design documents, implementer status reports, and the actual code diff.
Boundary markers: The instructions include defensive prompts such as "Do not trust the implementer report as source of truth" and "Read the spec yourself," which serve as boundary markers to mitigate accidental obedience to embedded instructions.
Capability inventory: The skill can execute git diff and arbitrary validation commands discovered in the project environment.
Sanitization: No explicit sanitization or filtering is defined for the commands or the content of the status reports before processing.
[DYNAMIC_EXECUTION]: The skill executes validation and linting commands that are determined dynamically based on the project environment and user inputs.
Evidence: The 'Regression Safety' check relies on 'validation commands discovered by the controller' to run the test suite.

kiro-review