The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions direct the agent to use Bash to perform environment detection, run test diagnostics, and potentially apply fixes to the codebase. This is consistent with the skill's stated purpose of auditing and improving test suites.
[DATA_EXFILTRATION]: The skill includes instructions to redirect the 'full findings' of the audit to an external command or handler named /atlas-report instead of displaying them in the standard output. While this moves potentially sensitive information about the test suite's health (such as coverage gaps or flaky test details) to a different reporting format (HTML), it appears to be a design choice for creating structured reports rather than an attempt to leak data to an unauthorized third party, especially considering the vendor context (tonone-ai).
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data by reading and analyzing the user's local test files and configurations. While this presents a surface for indirect prompt injection (where malicious instructions inside a test file could attempt to influence the agent), the skill does not have capabilities that would make such an attack high-impact in this context.
Ingestion points: Reads local test files, CI configurations, and coverage reports (SKILL.md).
Boundary markers: None explicitly defined in the instructions.
Capability inventory: Uses Bash, Read, and WebFetch (SKILL.md).
Sanitization: No specific sanitization or validation of the ingested test code is mentioned.

proof-audit