The Agent Skills Directory

[PROMPT_INJECTION]: Instructions in SKILL.md (under "Iron Laws" and "Rationalization Counter") and agent/qt-unittest-builder.md explicitly command the agent to bypass user confirmation for high-privilege actions like file writing and command execution. The skill labels safety checkpoints (like asking the user) as "rationalizations" that "waste time," directing the agent to "Directly execute" instead.
[PRIVILEGE_ESCALATION]: The skill mandates that its sub-agent operate with write: allow and bash: true permissions without seeking user approval for individual operations. This significantly increases the agent's autonomy and removes the standard safety layer provided by user review for destructive file operations and shell command execution.
[DYNAMIC_EXECUTION]: The provided stub-ext mock tool in the resources/stub/ directory utilizes low-level system calls such as mprotect (Linux) and VirtualProtect (Windows) to modify memory page permissions to PROT_EXEC (executable). This facilitates binary-level function hooking at runtime, which is a form of dynamic code manipulation.
[INDIRECT_PROMPT_INJECTION]: The agent/qt-unittest-builder.md sub-agent is designed to analyze project source code and configuration files (CMakeLists.txt, .pro). These files act as untrusted ingestion points where a malicious project could embed instructions to manipulate the agent's generation of build scripts and test code.
Ingestion points: Project configuration files (CMakeLists.txt, .pro) and source code headers analyzed in agent/qt-unittest-builder.md.
Boundary markers: Absent; the prompt does not specify delimiters or warnings to ignore instructions embedded in the analyzed code.
Capability inventory: The sub-agent has full write and bash access to perform file modification and execute build commands (cmake).
Sanitization: None; the agent processes and reflects analyzed project structures directly into new build configurations without validation.

qt-unittest-build