The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes notebook cell contents and kernel state, which serve as ingestion points for untrusted data if a notebook file is sourced externally.
Ingestion points: The skill reads cell code and variables from the active marimo kernel namespace as described in reference/code-mode.md.
Boundary markers: Execution results are streamed via a structured SSE (Server-Sent Events) interface in scripts/execute-code.sh, which separates stdout, stderr, and metadata.
Capability inventory: The skill has the capability to execute arbitrary Python code in the notebook kernel using scripts/execute-code.sh.
Sanitization: The skill does not perform sanitization of notebook content, as its purpose is to provide direct access to the kernel environment for pair-programming tasks.
[COMMAND_EXECUTION]: The skill utilizes shell scripts to perform discovery and execution tasks.
The scripts/discover-servers.sh and scripts/execute-code.sh scripts use standard system utilities like curl, jq, awk, and grep to interact with the local filesystem and the marimo HTTP API.
Shell variables are handled safely; for instance, Python code is injected into JSON payloads using jq --arg to prevent command injection.
[REMOTE_CODE_EXECUTION]: The skill facilitates the execution of Python code on a notebook server, which is the primary functionality of a pair-programming agent.
Execution is performed by sending HTTP POST requests containing code to the /api/kernel/execute endpoint of a marimo server.
scripts/execute-code.sh includes a security warning if the target URL host is not a local address (e.g., localhost or 127.0.0.1).
[EXTERNAL_DOWNLOADS]: The skill documentation describes the use of external JavaScript libraries for rendering rich UI widgets.
reference/rich-representations.md mentions importing libraries like D3.js and Flechette from the well-known CDN https://esm.sh for browser-side visualization.

marimo-pair