printpipe

printpipe make "I want a stand for my headphones"

Six screens, one shell: new → generate → model gate → slice settings → toolpath gate → print monitor. A sentence goes in one end; a file sits on the printer at the other. Generation, mesh validation, bed placement, slicing and upload are automatic. It stops exactly twice, both times to let a person look at something.

What it removes is everything after modelling: export, open the slicer, pull the SD card, plug it into the laptop, copy, plug it back, scroll the file list.

1 · A sentence

The new-project screen: a dark card with a large heading, a prompt box containing a sentence, a Generate Model button, and four example chips below it — “Describe the part you want. The pipeline does the rest.” The chips are past jobs — a tolerance test coupon, a 53→58 mm filter adapter, a Raspberry Pi wall mount.

The model comes back as OpenSCAD rather than a finished mesh. That one choice is what makes the rest possible: the geometry stays readable, stays parameterised, and re-renders in about a second.

2 · The model gate

Dragging a slider re-renders through OpenSCAD’s -D in about a second and never touches the model. Changing the sentence goes back to the LLM and takes minutes. The interface keeps those apart because they cost orders of magnitude apart, and blurring them would train the wrong habit.

The slider labels are not authored anywhere. They are the comments already sitting next to each parameter in the .scad file — the model wrote them, and nothing had been reading them out.

The parameter panel for the headphone stand: nine sliders, each labelled with its OpenSCAD variable name and the comment written beside it in the source — Every row is a variable in model.scad. “gusset_ang — angle of the support under the arm, above 45° so it needs no supports” is the comment, not a label someone wrote for the UI.

3 · Slice settings

Printer profile and filament profile are separate layers; the panel on the right shows this job overrides five values.

Printer and filament are two profile layers rather than one file, because PLA’s 100 % fan on PETG looks fine until the part splits along a layer under load, and that mistake should not be one edit away from the printer definition.

4 · The toolpath gate

The gcode toolpath viewer showing a sliced part in blue on a dark grid, a layer slider at 915 of 1251, and a footer with estimated time, filament mass, layer count, maximum Z and temperatures — Real gcode, not a preview of the model — every extrusion segment, coloured by its type, scrubbable layer by layer.

Each segment is drawn with a round cross-section, because a million same-coloured line segments stacked on each other is a smear rather than a drawing.

5 · The printer

The print monitor: a large progress dial at 13.5 percent with hours remaining, live nozzle and bed temperatures, a confirm-to-start panel, and a timeline of every pipeline step with timestamps — The job timeline keeps the whole run in one place — generate, validate, gate 1, slice, gate 2, upload, print.

Slicing and uploading are separate steps, so a file can be prepared while the machine is busy. Starting a print requires typing yes in full, in the browser exactly as in the terminal, and the server checks again before it does anything. The progress bar only counts the job if the machine is genuinely running this job’s file; otherwise it shows the queue position and how long the part ahead still needs.

Four orthographic views of a tall tapered headphone stand rendered in blue on a dark grid, with dimensions and face count printed above — And out the other end: 150 × 120 × 240 mm, 700 faces, from that one sentence.

The library problem

A folder of STL files cannot answer the two questions that come up daily: which of these are the same mesh, and which are versions of the same part. Filenames answer neither. station_main_part.stl and Weihao Ipad Holder (1) v5 v2.stl may be byte-identical; hotshoe v5.stl and hotshoe_updated.stl may not be.

So identity is the sha256 of the mesh bytes, and filenames are demoted to sightings — the same mesh may have been seen at several paths.

Grouping into families is deliberately stupid: strip the tokens that are obviously version noise (v4, new, 0.05, (1)) and key on what is left. It over-splits, and that is the failure direction I picked on purpose. ls does substring search, so a family split too finely is still findable; a family merged too eagerly hides parts silently.

One rule turned out to matter more than the rest: integers are identity, decimals are versions. The 58 in “53–58 mm adapter” names the part and dropping it merges that adapter into unrelated ones; 3.6 mm, 0.05, −0.01 are almost always a tolerance being re-cut. Getting that backwards splits the adapter family in half, which is how the rule got found.

Where it runs

A CLI and the web flow call the same functions, not two implementations; the only difference is where the two gates get answered. Docker for the toolchain, OctoPrint for the printer, Ender 3 S1 profiles, PETG or PLA.

The screens above are captures of it running — the parameter panel is this job’s actual model.scad, the slice figures are its real 1251 layers and 212.8 g.