The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to locate and execute a local binary or source executable named harness or .harness/bin/harness. It specifies flags such as --steps and --plan for execution. This is a core part of the skill's intended functionality for running automated code reviews.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data in the form of git diffs (merge-base(base, head)..head) and structured output from the harness tool. This creates a potential surface for instructions embedded within the code changes to influence the agent's behavior during the review and remediation process. The skill attempts to mitigate this by instructing the agent to focus on material findings and evidence-backed dispositions.
Ingestion points: Git diffs and reviewer tool outputs (SKILL.md).
Boundary markers: The skill references "structured reviewer outputs" which implies a format for tool responses (SKILL.md).
Capability inventory: The agent has the ability to execute the harness command and write file corrections (SKILL.md).
Sanitization: No specific prompt-level sanitization or escaping of diff content is described.

change-review-workflow