The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to treat external plan files and adversarial review findings as authoritative instructions, which creates a surface for indirect prompt injection attacks.
Ingestion points: The skill reads plan.html, plan.md, and the output of 'fresh independent agents' during the adversarial review phase.
Boundary markers: The instructions emphasize following the plan exactly ('Treat the plan as the contract') and do not include specific delimiters or instructions to ignore embedded malicious prompts within the ingested data.
Capability inventory: The skill has the capability to execute shell commands for validation (tests, linters, builds), read and write repository files, and delegate work to other agents.
Sanitization: There is no evidence of sanitization or safety-filtering for the content of the plans or the reviewer findings before they are integrated into the execution flow.
[COMMAND_EXECUTION]: The skill is instructed to run repository-specific validation tools, including tests, linters, type checks, and build commands, which involves executing subprocesses within the host environment.

qstack-loop-no-nonsense