The editor

magnet gui opens your project on a live canvas in the browser. You place and wire blocks, run the simulation and read the traces, while every edit is written into your .mag source. The text is the single source of truth and the canvas is a view of it.

$ magnet gui
     Serving /home/you/hello_pid at http://127.0.0.1:7317/?token=…
Ctrl-C to stop

magnet gui is magnet serve --open. It starts a local kernel in the project directory that serves the editor and a WebSocket on one localhost port, guarded by the token in the URL.

There is no account, cloud or separate backend. The token is also written to .magnet/kernel.json for other clients.

Hand-written text on the canvas

The canvas shows a hand-written model faithfully:

• The step’s inputs and outputs are ports.
• Each probed signal has a magnifier on its output port.
• Every hand-written statement, such as a model call, a fby or arithmetic, is a custom block named for what it binds, wired by name and read-only. Double-click one to read its statement, and Open in source jumps to its line in the file.

Recognizing arbitrary hand-written magscript as palette blocks is a deliberate non-goal. The full block vocabulary belongs to what the canvas authors itself.

A walkthrough

1. Edit the text. Change kp: 1.0 to 2.0 in your editor and save. The canvas updates in about 300 ms.

2. Add a block. Drag a Gain, Saturation or Sine Wave from the palette and wire it in. A Gain is an expression and needs no field. A stateful block arrives as:

   • a field
   • a step line
   • a params entry

3. Edit parameters. Double-click a block. Only the changed value moves in the text, and your comments and formatting stay. The Variable name field renames the block and every use of it in one edit.

4. Pin a type, or let it follow. Every block with an input, such as a Gain, a Saturation or a Unit Delay, takes the type of what is wired into it. Its Type reads Inherited, with the type it resolves to, and its text states none (let clamp = limit(u, -1.0, 1.0);, d: UnitDelay). A block with no input, like a Sine Wave, has a type of its own.

   Pick a type to pin it (let clamp: f32 = …, d: UnitDelay<f32>). A wire never changes a pinned type, and wiring in something else is a type error on the block, not a silent rewrite. Pick Inherited to let it follow again.

5. Probe a signal. Hover an output port and click the magnifier, press P, or right-click and pick Probe this signal. The step gains the argument in @probe(...). Click again to unprobe.

6. Add ports. Drag an Input or Output from the palette. A port added to a module is added to every call of that module in the package in the same edit.

7. Group into a module. Select two or more catalog blocks and press Ctrl+G. Their statements, fields and params entries move verbatim into a new model and step pair. Ctrl+Z restores.

8. Run. The simulation settings set dt and the end time, which are the manifest’s [simulate] table. Run simulates the current text. The first run pays a cargo build, and reruns of an unchanged model start instantly.

9. Iterate. Change a value and Run again. The viewer flags a trace as stale whenever the text changed after its run.

In Simulate mode a declared input reads 0.0, so drive a loop from a Step source. Replay mode drives the inputs from a .csv in the project instead, one row per tick.

A module instance has two names. Rename instance… renames the field, such as rotor_1. Rename module… renames the file, the model and every use path, field type and params entry that reaches it.

Layouts

Three layouts on the top bar arrange the panes for an activity:

• Modelling puts the Canvas Editor beside the Output Viewer.
• Simulation puts the Canvas Editor beside the Simulation Viewer.
• Deployment puts the Text Editor beside the Deployment Settings: the [deploy] table, Flash and Watch.

In Deployment, the Model row lists every model the package can ship, and the Firmware row every crate that builds a binary with the chip its runner specifies. Picking either writes the manifest. A firmware walks through it.

Traces

The Simulation layout’s viewer shows one row per probed signal, grouped by the model that produced it. An array probe is one expandable row. Stop ends a runaway run, and a stale flag marks a trace whose text changed since.

Its two sections, Custom Plots and Probes, fold to their headers, and dragging a header’s grip swaps their order. Every plot folds to its title row, and Full screen opens it across most of the window, its settings with it. Remove probe on a probe’s row takes the probe off the model once you Apply; its trace stays until the next run.

To navigate a plot:

• Drag across it to zoom in on that span of time.
• Ctrl and the scroll wheel zoom about the cursor. In full screen, the wheel alone does.
• Shift turns that zoom to the value axis.
• A middle-button drag pans.
• A double-click returns to the whole run.

A trace stays on screen until the next run replaces it, because a run costs a build. A new tab does not replay a run for a model that has changed since.

Model instances and feedback

Double-click an instance to view its contents. Its inports and outports are the module’s ports on the parent canvas. Instances are created on the canvas, but their interiors are edited in the text, so the drilled-in view is read-only.

Model calls schedule as full feedthrough (see Causality, delays, and feedback). Wiring an instance’s output back to its input reports an algebraic loop. Place a Unit Delay on the feedback path to break it.

A stateful catalog block such as an Integrator or PID is a module too, and a loop through it needs the same Unit Delay. Stateless blocks such as Gain, Sign and Switch are expressions and never block a loop.

The Output Viewer

The Output Viewer shows the generated #![no_std] crate live. Hover a block to highlight its generated lines. Diagnostics outline the offending blocks in red, with the full error in the panel. View › Toggle Output Left/Right moves it to the other side.

Version and updates

Help › Magnetite Version opens a dialog that shows:

• which magnet serves the canvas
• its project root and kernel address
• a link to these docs

Update and Uninstall run magnet self-update and magnet uninstall and print the result.

Uninstalling keeps your license and settings, so a reinstall picks up the same seat. It shuts the kernel down when it finishes.