The Agent Skills Directory

[COMMAND_EXECUTION]: Instructions in SKILL.md direct the agent to run shell commands including ls, grep, and cat to inspect the local file system and source code content.
[REMOTE_CODE_EXECUTION]: The skill requires the execution of a local shell script ./qa-playwright-capture.sh to capture visual evidence. The content of this script is not provided or verified within the skill package, representing an unverifiable execution of external logic.
[PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection (Category 8) by ingesting untrusted data from local source files and providing the agent with command execution capabilities.
Ingestion points: The agent reads content from resources/views/, *.html, and public/qa-screenshots/test-results.json as part of its mandatory process.
Boundary markers: No delimiters or instructions are provided to the agent to disregard potential instructions embedded within the analyzed files.
Capability inventory: The agent has access to shell execution (ls, grep, cat, and local script execution) which could be abused if malicious instructions are processed.
Sanitization: No filtering or validation of the ingested file content is performed before processing.

agency-evidence-collector