The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill designates projects/<name>/qa/README.md as the 'protocol SSOT' and instructs the agent to use lens briefs from this file 'verbatim' in sub-agent prompts. This pattern allows instructions stored in project documentation to control or override the behavior of the spawned QA agents.
[INDIRECT_PROMPT_INJECTION]: The skill interacts with the 'live product' including UI elements, API responses, and database stores. Content from these external sources could contain malicious text designed to manipulate the agents during the QA sweep.
Ingestion points: projects/<name>/qa/README.md, ${CLAUDE_PLUGIN_ROOT}/reference/qa-PROTOCOL-TEMPLATE.md, and the live product interface/API.
Boundary markers: None specified; instructions are pulled 'verbatim'.
Capability inventory: Spawning sub-agents, network operations via curl, browser automation via Playwright, and local file writes to the projects/ directory.
Sanitization: No evidence of sanitization or escaping of external content before interpolation into prompts.
[COMMAND_EXECUTION]: The skill relies on curl for API testing ('contract' and 'honesty' lenses) and Playwright MCP servers for browser automation. These tools provide network and automation capabilities that could be abused if the agent's instructions are compromised via injection.
[DYNAMIC_EXECUTION]: The skill manages a complex multi-agent workflow, spawning and coordinating multiple 'qa' agents with distinct roles and prompts generated at runtime based on file contents.
[EXTERNAL_DOWNLOADS]: The skill references a protocol template at ${CLAUDE_PLUGIN_ROOT}/reference/qa-PROTOCOL-TEMPLATE.md within the environment's kit repository, which it attempts to copy and use if the local project configuration is missing.

qa-sweep