The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to discover and execute arbitrary startup commands found within the repository's documentation and configuration files, such as Makefiles or package scripts (e.g., npm start, make). This allows the agent to run potentially complex and unreviewed scripts defined by the repository owner.
[DYNAMIC_EXECUTION]: The skill dynamically generates an executable control CLI (scripts/control) using Python or the repository's native language. It then executes this generated script to interact with and drive the application during the verification process.
[PERSISTENCE]: The skill establishes a permanent, project-local skill by creating a dedicated directory (.agents/skills/verify-APP/) and relative symlinks in platform-specific skill directories like .claude/skills/. This effectively installs a new capability that persists across agent sessions within the repository.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection as it is designed to ingest and process potentially untrusted data from the application's runtime output.
Ingestion points: The agent reads terminal transcripts, logs, and HTTP response bodies during the verification pass (SKILL.md, Step 3 and 5).
Boundary markers: The skill suggests using JSON for CLI output to provide structure, but does not specify delimiters for natural language evidence.
Capability inventory: The skill can execute shell commands, create files, and establish symlinks.
Sanitization: Instructions include masking sensitive credentials (e.g., auth headers, tokens) with ***, but no sanitization is specified for executable instructions embedded in the ingested application output.

verify