The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the Bash tool to run arbitrary commands defined in the project's config.yaml file (keys: verify_commands, verify.test_commands, verify.build_command, verify.type_check_command). This is a high-risk capability if the project configuration is controlled by an untrusted party.\n- [REMOTE_CODE_EXECUTION]: The skill automatically detects and executes scripts from project-controlled files like package.json (scripts.test), Makefile, and pytest.ini. This leads to the execution of untrusted code defined in the repository.\n- [PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection because it processes content from external artifacts (specs, designs, and tasks) to generate reports and calculate a pass/fail verdict.\n
Ingestion points: Step 1 reads task, spec, and design artifacts via mem_search and mem_get_observation (SKILL.md).\n
Boundary markers: Absent; instructions do not specify the use of delimiters for the ingested data.\n
Capability inventory: Access to the Bash tool (Steps 6 and 7) and data persistence via mem_save (Step 10).\n
Sanitization: Absent; content is analyzed directly for correctness and completeness checks.\n- [DATA_EXFILTRATION]: The ability to run arbitrary shell commands defined in project-local configuration files provides a vector for exfiltrating sensitive environment data or local files using standard CLI tools if the project configuration is malicious.

sdd-verify