matlab-runtime-skill

name	matlab-runtime-skill
description	Use when a computational math repository or task uses MATLAB files, MATLAB README commands, MATLAB toolboxes, or MATLAB MCP execution tools.

MATLAB Runtime Skill

MATLAB is a runtime backend, not the workflow driver. The workflow remains Skill-first and conversation-first: computational_math_reproduction_workflow_skill writes plans, records evidence under outputs/{run_id}/, and pauses for human approval before consequential execution.

This Skill defines how to inspect, plan, and optionally execute MATLAB code in computational math reproduction tasks. The local helper scripts/matlab_runtime.py can summarize MATLAB files, README commands, likely toolboxes, entrypoint candidates, and local matlab/octave executable availability.

When To Use

• Source contains .m, .mlx, .mat, MATLAB project files, or MATLAB README commands.
• A repository depends on MATLAB toolboxes.
• The user wants MATLAB execution, tests, static analysis, or toolbox detection.
• MATLAB MCP tools are available or need to be checked.

When Not To Use

• The task is a pure Python, Julia, C++, or R repository with no MATLAB artifacts.
• The user wants MATLAB Agentic Toolkit installation itself. In that case, use the official MATLAB Agentic Toolkit setup guidance as a reference and ask for approval before changing global Codex configuration.
• MATLAB, Octave, or MATLAB MCP is not configured yet. Use matlab_environment_setup_skill first.

Execution Boundary

• Do static file inspection without approval.
• Before executing MATLAB code, write or update the run plan and ask for approval.
• Use MATLAB MCP tools only when available in the current agent session.
• If MCP tools are unavailable but local matlab or octave is available, generate a CLI run plan and ask for run-plan approval before execution.
• If neither MCP nor CLI execution is available, produce a MATLAB runtime plan and verification instructions instead of pretending execution happened.
• Save external execution logs to outputs/{run_id}/logs/run.log.
• Do not install MATLAB, install toolboxes, or change global MCP configuration without explicit approval.

Preferred MCP Tools

Capability	Use
detect_matlab_toolboxes	Confirm MATLAB version and installed toolboxes.
check_matlab_code	Static Code Analyzer checks for .m files.
run_matlab_file	Run scripts or programs from files. Prefer this over long inline code.
run_matlab_test_file	Run MATLAB unit tests with structured results.
evaluate_matlab_code	Short diagnostics and quick variable/toolbox checks only.

Local Helper

Use this helper for static preflight and plan evidence:

conda run -n ai4math python -m skills.matlab_runtime_skill.scripts.matlab_runtime --source <repo> --out outputs/{run_id}

The helper writes matlab_runtime_summary.json when --out is provided. It does not execute MATLAB code. repo_analyzer.py and run_planner.py use the same helper logic to attach MATLAB runtime summaries and generate approved matlab -batch or octave --eval candidate plans when a local executable is available.

Workflow

1. Inspect MATLAB files, README commands, project files, and toolbox references.
2. Read references/INDEX.md.
3. Determine whether MATLAB MCP tools, local matlab, or local octave are available.
4. If unavailable, route to matlab_environment_setup_skill for agent-neutral setup.
5. If execution is available, include the exact MATLAB/MCP/Octave action in plan.md and wait for approval.
6. Execute only the approved action, capture logs, and summarize evidence.
7. Route failures to failure_diagnosis_skill.

Codex MCP Setup Reference

The official MATLAB Agentic Toolkit recommends registering MATLAB MCP with Codex using:

codex mcp add matlab -- "<MCP_SERVER_PATH>" --matlab-root "<MATLAB_ROOT>" --matlab-display-mode "<DISPLAY_MODE>"

After registration, the MATLAB MCP server should have a longer timeout such as tool_timeout_sec = 600 in the Codex config. Treat setup as a user-approved environment change, not an automatic reproduction step.