The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted source code files from the local repository, which could contain adversarial instructions or deceptive comments (e.g., docstrings or logic summaries) intended to influence the agent's behavior during the cleaning process.\n
Ingestion points: Local source files identified via git diff commands as described in SKILL.md.\n
Boundary markers: The skill mandates an interactive workflow where every proposed change is presented to the user for explicit approval (y/n) before any filesystem modifications occur, as detailed in references/review-workflow.md.\n
Capability inventory: The agent performs file system read/write operations, displays diffs for user review, and executes git commands.\n
Sanitization: The skill relies on manual human verification of suggested removals rather than automated content filtering to mitigate potential prompt injection within the code.\n- [COMMAND_EXECUTION]: The skill instructs the agent to execute shell commands (git diff --cached --name-only and git diff HEAD~1 --name-only) to determine the scope of files to be scanned. This is a standard and expected administrative task within the skill's intended developer-centric workflow.

deslop