| name | bfd-matlab-mcd |
| description | Use when building a closed STM32-to-Matlab workflow for J-Link HSS data capture, system identification, PID/LQR/MPC tuning, Kalman parameter estimation, or Matlab/Simulink code-generation feedback loops. |
BFD Matlab MCD Workflow
Self-Improvement Loop (Required)
- When this skill encounters a build, flash, debug, script-usage, Matlab, or code-generation workflow problem, record the resolved issue and lesson in
BFD-Kit/.learnings/ERRORS.md and/or BFD-Kit/.learnings/LEARNINGS.md.
- Unresolved capability gaps go to
BFD-Kit/.learnings/FEATURE_REQUESTS.md.
- Promote reusable fixes into the affected BFD-Kit asset in the same task when feasible.
- When a learning is promoted into a skill or script, append a short entry to
BFD-Kit/.learnings/CHANGELOG.md.
Purpose
Use this skill for the full loop:
- capture synchronized MCU scalar variables with J-Link HSS;
- store the run as an experiment dataset with manifest and compact summaries;
- run Matlab analysis without sending full CSV files into the AI context;
- use Matlab results for system identification, PID/LQR/MPC candidates, Kalman noise parameters, or MCD codegen checks;
- feed hardware evidence back into the next Matlab or firmware iteration.
Precheck
python3 ./.claude/skills/bfd-project-init/scripts/bootstrap.py --project-root . --mode check
/home/xuan/matlab/bin/matlab -batch "disp(version)"
If Matlab is installed elsewhere, pass --matlab-bin <path> or set MATLAB_BIN.
Capture A Dataset
python3 BFD-Kit/scripts/bfd_experiment.py capture-hss \
--elf "${STM32_ELF}" \
--symbol motor_cmd \
--symbol motor_speed \
--duration 2.0 \
--period-us 1000 \
--output-dir logs/experiments \
--experiment-name motor_step_001 \
--stimulus "step command"
The command writes capture.csv, capture.csv.meta.json, manifest.json, summary.json, summary.md, and Matlab templates under matlab/. Terminal output is compact by default.
Summarize Existing Data
python3 BFD-Kit/scripts/bfd_experiment.py summarize \
--dataset-dir logs/experiments/motor_step_001
Use summaries for AI review. Do not paste full CSV files unless the task explicitly requires raw samples.
Run Matlab Analysis
python3 BFD-Kit/scripts/bfd_experiment.py matlab-run \
--dataset-dir logs/experiments/motor_step_001 \
--analysis system-id \
--matlab-bin /home/xuan/matlab/bin/matlab
matlab-run defaults to --matlab-backend auto. In auto mode, BFD-Kit first tries the official MathWorks matlab-mcp-core-server; if the MCP server is unavailable or fails before a Matlab script runs, it records the MCP error under matlab_out/matlab_mcp_error.log and falls back to matlab -batch.
Use explicit MCP mode when validating an agentic Matlab/Simulink session:
python3 BFD-Kit/scripts/bfd_experiment.py matlab-run \
--dataset-dir logs/experiments/motor_step_001 \
--analysis mcd-check \
--matlab-backend mcp \
--mcp-server /path/to/matlab-mcp-core-server \
--mcp-matlab-root /home/xuan/matlab \
--satk-root /path/to/simulink-agentic-toolkit
The MCP path uses detect_matlab_toolboxes, check_matlab_code, and run_matlab_file when exposed by the server. If --satk-root is provided, it also runs satk_initialize through evaluate_matlab_code before the dataset analysis file, enabling Simulink Agentic Toolkit model-query and MBD workflows in the same session.
Supported --analysis values:
system-id: build iddata from classified input/output columns and estimate a conservative first-order model.control: produce baseline PID, LQR, and MPC candidates from an identified model.kalman: estimate measurement R and process Q covariance candidates from selected columns.mcd-check: record Matlab/Simulink codegen toolbox availability and MCU integration rules.
Hard Rules
- Use J-Link HSS only for fixed-address scalar globals/statics.
- Do not run HSS, RTT, GDB, or RAM polling concurrently on the same probe.
- Do not assume input/output columns. If
manifest.json lacks classification, update classification before trusting system identification output.
- Prefer the MathWorks Matlab MCP Core Server and Simulink Agentic Toolkit for agent-driven Matlab/Simulink work; use
matlab -batch only as an explicit or automatic fallback.
- Generated C/C++ must enter business-layer wrappers such as
USER/Modules or USER/APP, not CubeMX-managed directories.
- Treat Matlab-generated PID/LQR/MPC/Kalman parameters as candidates until validated by firmware build, flash, and hardware capture evidence.
References
