matlab-debugging - SKILL.md Agent Skill

name: matlab-debugging description: Diagnose MATLAB errors and unexpected behavior. Breakpoints, workspace inspection, try-catch diagnostics, and common error patterns. Use when debugging functions, tracing errors, inspecting variables, or diagnosing runtime failures. license: MathWorks BSD-3-Clause metadata: author: MathWorks version: "2.0"

Investigating and Debugging MATLAB Code with MCP Tools

You have access to a live MATLAB session via MCP tools. Use them to actively investigate code — whether debugging errors, understanding behavior, or answering questions about how MATLAB code works. Don't just guess from code alone.

When to Use

User encounters a MATLAB error message or unexpected result
User wants to set breakpoints or inspect variable state
Tracing why a function produces wrong output
NaN/Inf values appearing unexpectedly
A MATLAB MCP tool returns an error or exception — including runtime errors, syntax errors, undefined function/variable errors, or failed test results
User asks "why is my MATLAB code not working", "help me debug", or shares a MATLAB stack trace

When NOT to Use

Code quality review without a runtime problem — use matlab-reviewing-code instead
Performance profiling — use performance optimization workflows
Writing tests for correctness — use matlab-testing instead
Understanding MATLAB APIs or language features without a specific bug

Static Analysis vs Runtime Debugging

Not every issue needs the live MATLAB session. Choose the right approach:

Static analysis is enough when: syntax errors, unused variables, obvious logic mistakes, or issues visible from reading the source code alone. Use the Read tool and check_matlab_code.
Runtime debugging is needed when:
- The error depends on actual data values, types, or dimensions
- The output is wrong but the code looks correct
- The user says "it doesn't work" but the code looks fine — check actual data
- You need to know what variables contain at a specific point in execution

When in doubt, start with static analysis. Escalate to runtime debugging when you can't determine the root cause from source alone.

Auto-Trigger on MATLAB Errors

When a MATLAB MCP tool returns an error (runtime error, syntax error, undefined function/variable, dimension mismatch, failed assertion, etc.), do not silently move on or guess at a fix. Instead:

Recognize the error — Look for patterns like Error using ..., Undefined function or variable, Index exceeds ..., Error in ..., MATLAB stack traces, or failed test results in MCP tool output.
Ask the user for permission — Before launching into investigation, offer:

"I noticed a MATLAB error: <brief error summary>. I can use the matlab-debugging skill to dig into this — inspect variables, trace the call stack, and identify the root cause. Want me to investigate?"
Proceed only after confirmation — Once the user agrees, follow the investigation workflow below.

This applies whether the error came from the user running code, or from you running code on the user's behalf (e.g., verifying a fix, running tests).

Available Tools

Tool	Use For
`mcp__matlab__run_matlab_file`	Run .m scripts — prefer this for executing user scripts and verifying fixes
`mcp__matlab__evaluate_matlab_code`	Quick diagnostics: inspect variables, evaluate expressions, test small snippets
`mcp__matlab__check_matlab_code`	Static analysis of .m files (warnings, unused vars, potential issues)
`mcp__matlab__run_matlab_test_file`	Run a MATLAB test file
`mcp__matlab__detect_matlab_toolboxes`	List installed toolboxes — use when "Undefined function" may be a missing toolbox

Prefer run_matlab_file over evaluate_matlab_code for running scripts. Only use evaluate_matlab_code for short diagnostic commands (checking a variable, testing an expression, etc.) — not for re-running entire scripts inline.

Workflow

1. Understand the Goal

Determine what the user needs:

Debugging — runtime error, wrong output, unexpected behavior
Understanding — how does this code work, what does this function do
Investigating — why does this variable have this value, where does this data come from
Exploring — what functions are available, how is this codebase structured

2. Gather Information via MATLAB

Use mcp__matlab__evaluate_matlab_code to run diagnostic commands.

Read source code: Use the Read tool for .m files on disk — not MATLAB's type or dbtype. Only use MATLAB's which to locate files you haven't found yet, and which -all to check for shadowing.

Preview large data: Use varName(1:min(5,end),:) or head(T) to preview slices instead of dumping entire variables.

Runtime debugging — check desktop mode first:

Before using breakpoints, check if MATLAB has a desktop:

desktop('-inuse')  % true = desktop mode, false = no-desktop

Desktop mode (desktop('-inuse') returns true):

Use the full breakpoint workflow:

Set a breakpoint before running:

dbstop if error          % Pause on any error
dbstop if caught error   % Pause on error inside try-catch (silent failures)
dbstop if warning        % Pause when a warning is issued
dbstop if naninf         % Pause on NaN or Inf
dbstop in file at line   % Pause at a specific line

Run the code via run_matlab_file — MATLAB pauses at the breakpoint.

Inspect the call stack:

dbstack                  % See full call stack with file names and line numbers

whos to see what's in scope — check variable names, sizes, and types before inspecting any values. For large arrays/tables, preview a slice (varName(1:5,:), head(T)) instead of displaying the whole thing.
Navigate frames and inspect variables in each scope:
```
dbup                     % Move up one frame (toward caller)
dbdown                   % Move back down (toward callee)
```
After dbup/dbdown, variable inspection commands operate in that frame's local scope — use this to check inputs/outputs at each level.

Resume or exit:

dbcont                   % Continue execution to next breakpoint or end
dbquit                   % Exit debug mode entirely
dbclear all              % Remove all breakpoints when done

Note: Interactive stepping (dbstep) is unreliable via MCP — each evaluate_matlab_code call is a separate command, so step state may not persist.

No-desktop mode (desktop('-inuse') returns false):

Do NOT use dbstop if error, dbstop if naninf, dbstop if warning, or any breakpoint that pauses execution. In no-desktop mode, pausing breakpoints cause the MCP eval to hang indefinitely.

Use these strategies instead:

try-catch wrappers — Wrap suspect code to capture the error and inspect state after failure:

try
    result = suspectFunction(data);
catch ME
    fprintf('Error: %s\n', ME.message);
    fprintf('In: %s line %d\n', ME.stack(1).name, ME.stack(1).line);
    whos  % Show variables in scope at failure
end

Tracer conditional breakpoints — Use dbstop with a condition that always returns false so MATLAB never actually pauses, but prints the value as a side effect. Define a tracer function:
```
function out = tracer(val)
    disp(val);
    out = false;
end
```
Then set a conditional breakpoint that calls it:
```
dbstop in myScript at 42 if tracer(myVar)
```
When line 42 executes, MATLAB evaluates tracer(myVar), which prints the value and returns false — execution continues without pausing. This probes variables at specific lines without modifying the source file. Remove tracer breakpoints when done: dbclear all
Run then inspect — For scripts, run via run_matlab_file, then use evaluate_matlab_code to inspect workspace variables after execution. The base workspace persists across MCP calls within a session.

3. Investigate Iteratively

This is the core of debugging — use your judgment:

Read the relevant source code to understand intent
Inspect variables that appear on or near the failing line
Evaluate sub-expressions to isolate which part fails
Test hypotheses by running small snippets in MATLAB
Be selective — don't dump the entire workspace. Only inspect what is relevant to the problem. If there are 7 variables in scope but only 2 are used on the failing line, inspect those 2.

When the error occurs inside a MathWorks built-in function, the bug is almost always in the user's code that called it. Walk up the stack to user code.

4. Question the Diagnosis

The user's description of the problem may not match the actual problem. Before fixing what they say is broken, verify it yourself:

Test the claimed failure directly — If the user says ismember or strcmp doesn't work, run it yourself with their actual data. Often the operation works fine and the real issue is the logic around it.
Check data before blaming code — When operations on data "don't work," inspect the data: types (class), actual values (sample a few), whitespace (strtrim), hidden characters (double(str)), and dimensions (size).
Verify what the code does vs what the user describes — Read the code and confirm which variables and columns are actually being used. The user may describe their intent but the code may do something different.
Consider whether the approach itself is wrong — Sometimes the code has no bug per se, but the approach is unnecessarily complex or fragile. Suggest idiomatic MATLAB alternatives: table joins instead of manual loops, vectorized operations instead of element-wise comparisons, built-in functions instead of hand-rolled logic.

Common Errors — What to Check First

Error Pattern	Diagnostic Steps
`Undefined function or variable 'X'`	`which X`, `exist('X','file')`, `exist('X','var')`, check `path`, use `detect_matlab_toolboxes` to verify toolbox is installed
`Index exceeds array dimensions` / `Index exceeds the number of array elements`	`size(arr)` and inspect the index expression — often off-by-one or empty array
`Not enough input arguments`	Check how the function is called at the call site, `nargin` inside the function, compare with function signature
`Too many input arguments`	Same as above — caller passing extra args, or calling a script as if it were a function
`Matrix dimensions must agree` / `Dimension mismatch`	`size(A)`, `size(B)` for both operands — often one is row and the other column
`Subscript indices must either be real positive integers or logicals`	Check index variable: `class(idx)`, `min(idx)`, look for 0 or negative values, NaN, or floating-point indices
`Dot indexing is not supported for variables of this type`	`class(var)` — usually accessing a struct field on a non-struct (cell, array, table)
`Unable to perform assignment`	Check `class` and `size` of both sides of the assignment
`Out of memory`	`whos` to find large variables, check for accidental array growth in loops
`Maximum recursion limit`	Check for missing base case or infinite mutual recursion — `dbstack` at error point
NaN propagation / wrong branch taken	In desktop mode, use `dbstop if naninf` to catch where NaN is first created. In no-desktop mode, use `dbstop in file at line if tracer(suspect)` to probe values without pausing. NaN comparisons (`>`, `<`, `==`) are always `false`, so `if` takes the wrong branch silently. Trace upstream: check for `0/0`, `Inf-Inf`, or NaN in input data. Use `any(isnan(var))` to test.

Gotchas

Pausing breakpoints hang in no-desktop mode. When desktop('-inuse') returns false, dbstop if error, dbstop if naninf, dbstop if warning, and unconditional line breakpoints cause MCP eval to hang indefinitely. Always check desktop mode first. In no-desktop mode, use tracer conditional breakpoints or try-catch wrappers instead (see no-desktop strategies above).
dbstep does not work reliably via MCP. Each evaluate_matlab_code call is a separate command, so step state may not persist between calls. Use breakpoints (dbstop in file at line) to pause at specific locations instead of trying to step through code. For tracing execution flow, insert temporary fprintf() statements or use try-catch blocks.
Do not use type or dbtype to read source code. Use the Read tool instead — it's faster and doesn't consume MATLAB session output bandwidth. Only use which to locate files you haven't found yet.
Always dbclear all when done debugging. Leftover breakpoints from a previous session will cause unexpected pauses in later runs.
dbup/dbdown scope is per-call. Each evaluate_matlab_code invocation resets to the current frame. Chain dbup with variable inspection in the same evaluate_matlab_code call:
```
dbup; whos; myVar(1:min(5,end),:)
```
A separate evaluate_matlab_code call after dbup will be back in the original frame.
Never run restoredefaultpath. It removes the MCP server's own packages from the MATLAB path, breaking the eval channel (mcpEval not found). If this happens, the only recovery is to restart MATLAB entirely.
No toolboxes required. This skill uses only base MATLAB debugging commands.

Key Principles

Use the tools — You have a live MATLAB session. Run commands, inspect state, test hypotheses. Don't debug from code reading alone.
Be selective — Request only what's relevant. Don't dump all variables or read every file in the stack.
Check size before fetching data — Always know the dimensions before displaying a variable. Never blindly disp an unknown variable.
Focus on user code — Errors surfacing in built-in functions are almost always caused by bad inputs from user code.
Read before fixing — Always read the actual source code before suggesting changes. Never guess what code looks like.
Verify fixes — When practical, run the corrected code in MATLAB to confirm.