The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes vendor-specific tools listCodeStepTools, describeCodeStepTools, and generateRevision. These tools enable the agent to inspect build environments and initiate the code revision lifecycle within the duvoai platform infrastructure.
[PROMPT_INJECTION]: The SKILL.md file contains instructions for the agent to optimize payloads for the generateRevision tool by omitting redundant SDK references. While flagged as potential concealment by automated scanning, these instructions are functional optimizations for an internal toolchain and are paired with a mandatory requirement to present a diff to the user for explicit approval before any changes are finalized.
[INDIRECT_PROMPT_INJECTION]: The skill processes user input to generate code, creating a surface where malicious instructions could potentially be passed to the code generation engine.
Ingestion points: User queries about the SDK and requests for program changes in SKILL.md.
Boundary markers: Explicit markers are absent in the prompt interpolation for generateRevision.
Capability inventory: listCodeStepTools, describeCodeStepTools, and generateRevision (used to write code to the platform's build system).
Sanitization: No explicit sanitization of user intent is described.
Mitigation: The risk is mitigated by a mandatory diff review process where the agent must "present the diff, and accept or decline on the user's word" before any code is committed.

code-step-author