The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to process potentially untrusted external data, which creates a vulnerability surface where malicious instructions could be embedded in the data being analyzed.
Ingestion points: SKILL.md explicitly mentions processing user-uploaded files (e.g., 'Unzip the uploaded my-project.zip') and delegating tasks from GitHub issues or bug reports.
Capability inventory: The agent possesses the capability to read, write, and execute code within its workspace environment.
Boundary markers: The instructions in DESIGN.md and IMPLEMENT.md attempt to mitigate risk by advising the orchestrator to define explicit scope boundaries and clear acceptance criteria.
Sanitization: No explicit content sanitization is mentioned, though the skill relies on an isolated container as a primary security control.
[DYNAMIC_EXECUTION]: The skill's primary purpose is the autonomous generation and execution of code.
Evidence: SKILL.md and IMPLEMENT.md describe a workflow where the agent explores the codebase, implements solutions, and runs test suites to verify its own work.
Context: This behavior is the intended primary purpose of the skill and is constrained to a dedicated, isolated container session as described in the execution environment section.
[COMMAND_EXECUTION]: The agent is instructed to run shell commands and development tools to perform its tasks.
Evidence: IMPLEMENT.md and REVIEW.md contain instructions for the agent to run test suites (e.g., pytest, mypy) and report the results.
Context: Command execution is scoped to the agent's workspace and is used for implementation verification and codebase exploration.

code-agent