The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted data from the project's source code and feature files to identify drift and update documentation.
Ingestion points: The 'Source wave' phase (described in references/maintain-verification-skill-procedure.md, Step 2) involves launching subagents to read user-facing feature files and source code.
Boundary markers: The instructions explicitly state that 'Children never drive the app and never edit files' and that the subagents are bound by the 'authorized scope'.
Capability inventory: The skill has the ability to modify the verification directory (SKILL.md, features/, and harness scripts), execute harness scripts, start/stop local application instances, and open Pull Requests.
Sanitization: There is no mention of specific sanitization or filtering of the project source code before it is interpreted by the agent or used to update the feature map.
[DYNAMIC_EXECUTION]: The skill's maintenance workflow involves the creation, modification, and execution of harness scripts.
Evidence: Step 5 of the maintenance procedure (references/maintain-verification-skill-procedure.md) specifies fixing 'harness gaps' by ensuring scripts are executable and their invocation is documented. These scripts are then executed during the 'Live pass'.
[COMMAND_EXECUTION]: The skill is designed to manage the lifecycle of local application instances, which involves executing shell commands.
Evidence: SKILL.md (External Side Effects) and the 'Live pass' section in the procedure documentation describe starting and tearing down local instances for servers, UIs, or CLIs depending on the skill's launch model.

maintain-verification-skill