The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructions direct the agent to clone an external GitHub repository (https://github.com/neteroster/CodexCont.git) and run the application using 'uv run python run.py'. Executing code from unverified third-party repositories allows for arbitrary code execution within the agent's environment.
[COMMAND_EXECUTION]: The setup and operation process requires executing shell commands such as 'git clone', 'uv sync', and 'uv run', which facilitate the deployment and execution of external, unvetted software.
[EXTERNAL_DOWNLOADS]: The skill fetches a complete codebase and its dependencies from a non-vendor GitHub account without integrity checks or source verification.
[INDIRECT_PROMPT_INJECTION]: The middleware ingests untrusted streaming data from upstream APIs and automatically injects instructions back into the reasoning flow.
Ingestion points: External streaming responses from OpenAI-compatible APIs (specified in SKILL.md configuration).
Boundary markers: Absent; the skill does not employ delimiters to separate proxied data from the agent's instruction context, nor does it warn the agent to ignore instructions within the stream.
Capability inventory: Network proxying, response folding, and automated prompt interpolation (CONTINUE_MARKER).
Sanitization: No sanitization or validation is performed on the reasoning tokens received from the upstream API before they are processed and potentially used to trigger further reasoning rounds.

codexcont-middleware