The Agent Skills Directory

[PROMPT_INJECTION]: The skill ingests untrusted code diffs and finding descriptions to perform verification. This represents an indirect prompt injection surface where malicious content in a file could attempt to manipulate the audit verdict. This is the primary intended function of the tool and is mitigated by design.
Ingestion points: Data is ingested via the finding_packet (referenced in references/verdict-prompt.md).
Boundary markers: The skill uses markdown code block delimiters to isolate untrusted content in the verification prompt.
Capability inventory: The verification process uses read-only workspace tools such as git, grep, and cat to inspect the project state.
Sanitization: The skill implements redaction rules to ensure secrets and sensitive tokens are not included in audit logs or verification prompts.
[COMMAND_EXECUTION]: The tool utilizes standard development commands (git status, git diff, grep, cat) to research findings. These operations are conducted within a read-only sandbox during the verification phase, preventing unauthorized modification of the codebase.

seek-verdict