The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill dynamically resolves and loads other project-specific skills at runtime. The activation contract and hard rules specify that the orchestrator must 'Resolve matching project skills before starting' and pass these paths to external actors such as jd-judge-a, jd-judge-b, and jd-fix-agent. The output format also includes a 'skill_resolution' field that tracks whether paths were injected or pulled from a fallback registry. Loading instructions from computed paths is a medium-risk pattern as it can lead to the execution of unintended skills if the resolution logic is influenced by the environment or target content.
[INDIRECT_PROMPT_INJECTION]: The skill presents an attack surface for indirect prompt injection because it ingests untrusted data and possesses file-writing capabilities. 1. Ingestion points: The workflow in SKILL.md and the Judge Prompt in prompts-and-formats.md describe reading an 'immutable target' consisting of exact file paths provided at runtime. 2. Boundary markers: While judges are described as 'read-only', there are no explicit delimiters or instructions to ignore malicious content embedded within the target source code. 3. Capability inventory: The jd-fix-agent is authorized to apply 'atomic work units', which involve modifying the source files to resolve confirmed findings. 4. Sanitization: No sanitization or validation of the target content is documented before processing, meaning malicious instructions in the target could potentially influence the judges' findings or the fix actor's modifications.
[COMMAND_EXECUTION]: The skill orchestrates the execution of several sub-components and tools, including jd-judge-a, jd-judge-b, and jd-fix-agent. It also references the use of the gentle-ai CLI for capturing results and schema validation. The reliance on these external actors to perform sensitive operations like file modification increases the impact of potential compromises in the orchestration logic.

judgment-day