Language: English | 简体中文

Simulink Automation Suite

[!IMPORTANT] This project is paused (since June 2026). MathWorks has released an official, actively-maintained agentic toolkit for Simulink — for production use, please switch to the official Simulink Agentic Toolkit.

The differentiators explored here (the guarded write-safety model: dry-run / rollback / read-back verification; and post-simulation performance analysis), together with a new motor-control-focused direction, will continue in a separate, dedicated project. This repository remains available as-is for reference.

Simulink Automation Suite is a Claude Code plugin for Simulink automation workflows through MATLAB Engine for Python.

• Canonical plugin name: simulink-automation-suite
• Read Analysis — the simulink-analyzer agent autonomously explores model topology, traces connections, audits parameters, and returns structured findings without polluting conversation context.
• Post-Simulation Analysis — the sim-analyst agent extracts signals from sl_sim_result, evaluates dynamic performance (rise time, overshoot, settling time), and returns structured conclusions without exposing raw waveform data to the main conversation.
• Write Automation — the simulink-automation skill guides safe parameter modification with dry-run preview, precondition guards, and rollback support.
• Runtime Python module path: simulink_cli (unified CLI entrypoint)

Positioning

Simulink Automation Suite is built to make Simulink analysis agent-native in Claude Code:

• It exposes Simulink context as deterministic, machine-readable tool outputs.
• It lets AI reason on real model topology/parameters instead of screenshots.
• It keeps workflows real-time and token-efficient with clipping/projection controls.

In short: the plugin helps AI understand first, then assist.

Why This Plugin Exists

Common AI+Simulink workflows are often one of these:

1. Screenshot-based discussion: fast but shallow, visual-only understanding.
2. Export-and-parse flow: richer context but heavy, delayed, and token-expensive.

This plugin provides a third path: direct, structured, runtime model analysis for agents.

How It Works

1. Claude Code invokes the simulink-automation skill for write/meta tasks, or dispatches the simulink-analyzer agent for read analysis.
2. The skill resolves MATLAB session context (session list/use/current/clear) with exact-name matching, using either an explicit --session or a previously selected active session.
3. It executes one of the 18 available actions: schema, list_opened, scan, connections, inspect, find, highlight, set_param, model_new, model_open, model_save, model_close, model_update, block_add, line_add, line_delete, block_delete, simulate, or session.
4. Results are returned as a single machine-readable JSON payload on stdout; warnings never spill raw text into stdout, and stderr is reserved for maintainer-facing diagnostics.
5. Failures use stable error codes for reliable agent recovery.
6. Write operations (set_param) use dry-run preview (default), machine-executable apply_payload, rollback payloads, guarded execute via expected_current_value, and read-back verification.

Prerequisites

Before using session-bound actions (list_opened, scan, connections, inspect, find, highlight):

1. Install and activate MATLAB on your machine.
2. Install MATLAB Engine for Python in the same Python interpreter that runs this plugin.
3. In MATLAB Command Window, run:

matlab.engine.shareEngine

Troubleshooting:

• engine_unavailable: MATLAB Engine for Python is unavailable in the active Python environment. Fix interpreter/environment installation.
• no_session: MATLAB Engine is available, but no shared MATLAB session is visible. Run matlab.engine.shareEngine in MATLAB, then retry.

Quick Start

1. Add the marketplace source

/plugin marketplace add Mistakey/simulink-automation-suite

2. Install the plugin from marketplace

/plugin install simulink-automation-suite@simulink-automation-marketplace

3. Invoke the namespaced skill

/simulink-automation-suite:simulink-automation Scan gmp_pmsm_sensored_sil_mdl recursively and focus on controller subsystems.

4. Verify plugin registration (optional)

/plugin list simulink-automation-suite@simulink-automation-marketplace

Scenario Examples

For end-to-end Claude Code prompts and screenshots (single bilingual page), see:

• docs/examples/claude-code-scenarios.md

Core Actions

Output Controls

Use output clipping/projected fields when you need compact payloads:

python -m simulink_cli scan --model "my_model" --max-blocks 200 --fields "name,type"
python -m simulink_cli inspect --model "my_model" --target "my_model/Gain" --param "All" --max-params 50 --fields "target,values"
python -m simulink_cli connections --target "my_model/Gain" --detail ports --max-edges 50 --fields "target,edges,total_edges,truncated"
python -m simulink_cli find --model "my_model" --name "PID" --max-results 50 --fields "path,type"

JSON Request Mode

--json is a first-class entrypoint and is mutually exclusive with flag-based action arguments. schema returns structured metadata for each action field (type, required/default/enum, description). JSON mode is the canonical contract surface for complex strings and newlines; use it whenever values need escaping or contain embedded line breaks.

python -m simulink_cli --json "{\"action\":\"schema\"}"
python -m simulink_cli --json "{\"action\":\"list_opened\",\"session\":\"MATLAB_12345\"}"
python -m simulink_cli --json "{\"action\":\"scan\",\"model\":\"my_model\",\"recursive\":true,\"session\":\"MATLAB_12345\"}"
python -m simulink_cli --json "{\"action\":\"inspect\",\"model\":\"my_model\",\"target\":\"my_model/Gain\",\"param\":\"Description\",\"summary\":true}"
python -m simulink_cli --json '{"action":"connections","target":"my_model/Gain","direction":"both","depth":1,"detail":"summary","max_edges":50,"fields":["target","upstream_blocks","downstream_blocks"]}'
python -m simulink_cli --json '{"action":"find","model":"my_model","name":"PID","max_results":50,"fields":["path","type"]}'
python -m simulink_cli --json '{"action":"set_param","target":"my_model/Gain1","param":"Gain","value":"2.0"}'
python -m simulink_cli --json '{"action":"model_new","name":"my_model"}'
python -m simulink_cli --json '{"action":"model_open","path":"C:/models/my_model.slx"}'
python -m simulink_cli --json '{"action":"model_save","model":"my_model"}'
python -m simulink_cli --json '{"action":"block_add","source":"simulink/Math Operations/Gain","destination":"my_model/Gain1"}'
python -m simulink_cli --json '{"action":"line_add","model":"my_model","src_block":"Sine","src_port":1,"dst_block":"Gain","dst_port":1}'
python -m simulink_cli --json '{"action":"line_delete","model":"my_model","src_block":"Sine","src_port":1,"dst_block":"Gain","dst_port":1}'
python -m simulink_cli --json '{"action":"block_delete","destination":"my_model/Gain1"}'
python -m simulink_cli --json '{"action":"simulate","model":"my_model"}'
python -m simulink_cli --json '{"action":"model_update","model":"my_model"}'
python -m simulink_cli --json '{"action":"model_close","model":"my_model"}'

Safety Model (Write Operations)

• dry_run defaults to true and returns both rollback and machine-executable apply_payload
• apply_payload carries expected_current_value, so execute can reject stale previews instead of mutating blindly
• Replay the returned apply_payload exactly; do not manually reconstruct the guarded execute payload
• Stale preview replay returns precondition_failed without mutating the model
• Execute mode reads back the value to verify the write
• If read-back does not confirm the requested value, the action returns verification_failed and preserves rollback/write-state recovery metadata
• Every response includes a rollback payload for one-command undo, preserving an explicit session override when one was used
• The value field is always a string and may legitimately include literal percent signs, for example "%.3f"
• One parameter per invocation (no batch operations)

Guarded Edit Loop

The standard single-parameter agent loop is:

1. inspect the current parameter state.
2. Run set_param with dry_run=true.
3. Replay the returned apply_payload.
4. inspect again to confirm the new value.
5. Replay rollback if you need to restore the original value.

Preview response excerpt:

{
  "action": "set_param",
  "dry_run": true,
  "current_value": "1.5",
  "proposed_value": "2.0",
  "apply_payload": {
    "action": "set_param",
    "target": "my_model/Gain1",
    "param": "Gain",
    "value": "2.0",
    "dry_run": false,
    "expected_current_value": "1.5"
  },
  "rollback": {
    "action": "set_param",
    "target": "my_model/Gain1",
    "param": "Gain",
    "value": "1.5",
    "dry_run": false
  }
}

If the target changes between preview and execute, replaying that saved apply_payload returns precondition_failed. If the write runs but read-back does not confirm it, the action returns verification_failed.

Strict Defaults and Error Contract

• Session matching is exact-only (no fuzzy matching).
• If multiple MATLAB shared sessions exist, either select one via session use <name> or pass --session explicitly for MATLAB-bound actions.
• If no opened model can be resolved to an active root, scan and find return model_not_found.
• unknown_parameter means the caller supplied a request field or flag that is not part of the contract.
• param_not_found means the target block does not expose the requested runtime parameter.
• Invalid JSON or wrong JSON field types return invalid_json.

Error envelope:

{
  "error": "<stable_code>",
  "message": "<human_readable_message>",
  "details": {},
  "suggested_fix": "<optional_next_step>"
}

Common error codes:

• invalid_input
• invalid_json
• unknown_parameter
• json_conflict
• engine_unavailable
• no_session
• session_required
• session_not_found
• state_write_failed
• state_clear_failed
• model_required
• model_not_found
• subsystem_not_found
• invalid_subsystem_type
• block_not_found
• param_not_found
• precondition_failed
• set_param_failed
• verification_failed
• model_already_loaded
• model_save_failed
• source_not_found
• block_already_exists
• model_dirty
• port_not_found
• line_already_exists
• line_not_found
• update_failed
• simulation_failed
• inactive_parameter
• runtime_error

Session management commands may return state_write_failed or state_clear_failed when the local plugin state file is not writable.

If no MATLAB shared session exists, run matlab.engine.shareEngine in MATLAB and retry.

What's Inside

simulink_cli/           # Unified CLI package (single entrypoint)
├── __main__.py         # python -m simulink_cli
├── core.py             # Action registry, JSON/flag parsing, schema, routing
├── errors.py           # Error envelope builder
├── json_io.py          # JSON I/O utilities
├── validation.py       # Input hardening
├── session.py          # MATLAB session management
├── model_helpers.py    # Path resolution helpers
└── actions/            # One module per action
    ├── scan.py
    ├── inspect_block.py
    ├── connections.py
    ├── find.py
    ├── highlight.py
    ├── list_opened.py
    ├── set_param.py
    ├── model_new.py
    ├── model_open.py
    ├── model_save.py
    ├── model_close.py
    ├── model_update.py
    ├── block_cmd.py
    ├── line_add.py
    ├── block_delete.py
    ├── line_delete.py
    ├── simulate_cmd.py
    └── session_cmd.py
agents/                 # Published agent definitions
├── simulink-analyzer.md  # Read-analysis agent (topology, search, connections, inspection)
└── sim-analyst.md        # Post-simulation analysis agent (signals, waveforms, performance)
skills/                 # Plugin skill definitions (docs only, no Python code)
└── simulink_automation/  # Write automation + meta-query skill
    ├── SKILL.md
    └── reference.md
tests/                  # Test suite

Verification

python -m unittest discover -s tests -p "test_*.py" -v
claude plugin validate .