The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill utilizes npx to download and execute the claude-flow package from the npm registry. While this is a standard mechanism for running developer tools, it involves runtime code execution from an external registry.
[COMMAND_EXECUTION]: Extensive use of shell commands is documented for verification workflows, including git for repository rollbacks, jq for JSON result processing, and curl for metric exportation. The skill also includes a command to launch a local web dashboard for monitoring verification activities.
[PROMPT_INJECTION]: Indirect prompt injection surface identified. The skill ingests untrusted data from agent outputs to determine quality scores and trigger automated actions.
Ingestion points: Processes agent task outputs and source code files during verify check and truth scoring operations as described in SKILL.md.
Boundary markers: No specific delimiters or instructions are documented to isolate the verification logic from potentially malicious instructions embedded within the data being analyzed.
Capability inventory: Possesses significant system capabilities, including command execution via npx, filesystem modifications through the --auto-fix feature, and repository manipulation via rollback commands.
Sanitization: Lacks documented validation or sanitization of agent outputs before they are processed by truth-scoring algorithms or used to drive automated environment changes.
[DATA_EXFILTRATION]: Documented features include sending verification metrics to external monitoring endpoints such as DataDog and Prometheus via curl POST requests. While intended for observability, this establishes a pattern for outbound data transfer.

verification-quality