The Agent Skills Directory

[PROMPT_INJECTION]: The skill utilizes an indirect prompt injection surface by instructing the agent to gather information from external web searches and incorporate it into the skill's knowledge base and validation logic.
Ingestion points: Web search results gathered during Phase 2 as defined in SKILL.md and references/update-checklist.md.
Boundary markers: Absent; there are no specific instructions to the agent to delimit or ignore potentially malicious instructions embedded in the search snippets.
Capability inventory: The agent has the capability to read and write files within the repository (Read, Glob, Write) and perform web searches (WebSearch).
Sanitization: No explicit sanitization or validation steps are defined for the content retrieved from the web before it is used to update internal files.
[COMMAND_EXECUTION]: The skill involves the dynamic modification of executable validation hooks based on information retrieved from the web.
Evidence: Phase 7b and 7c in SKILL.md and references/update-checklist.md instruct the agent to modify validate-udonsharp.sh (shell script) and validate-udonsharp.ps1 (PowerShell script).
Context: These scripts are described as "PostToolUse validation hooks" in references/skill-structure.md, implying they are executed by the environment. Modifying these scripts based on external SDK changes is a core part of the renovation process but introduces a pathway for logic changes in executable files.

unity-vrc-skills-renovator