The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is configured to automatically run all commands listed in the steps array of the state.json file during the validation phase. This provides a direct path for arbitrary command execution if the content of the state file is manipulated or contains unsafe shell instructions.
[INDIRECT_PROMPT_INJECTION]: The skill ingests task goals, patterns, and implementation details from state.json and PROGRESS.md. These files serve as an ingestion point for untrusted data that could potentially contain instructions designed to override the agent's behavior or safety guidelines during the development process.
[DYNAMIC_EXECUTION]: The skill delegates git operations and state updates to a background agent utilizing a specific external model (openrouter/z-ai/glm-4.7-flash). This creates a chain of autonomous actions that are performed without direct user visibility or review of the specific instructions sent to the secondary model.
[PROMPT_INJECTION]: The instructions explicitly direct the agent to skip user confirmation at multiple stages, such as proceeding immediately to Step 1 after preconditions are met and starting the next task immediately after a commit. These directives suppress the human-in-the-loop safety mechanism, allowing the agent to perform multiple file modifications and shell executions autonomously.

tracer-dev