Gen Assembly

pc:gen-assembly

Generate a PartCAD assembly: an .assy file that composes parts (and sub-assemblies) into one product. The text after the command ($ARGUMENTS) is the assembly description. You author both the ASSY and its component parts and prove the whole thing passes pc test. There is no legacy ai assembly pipeline — this capability is new.

1. Decompose

$ARGUMENTS is the description. Break it into components and the spatial/mechanical relationships between them: which parts, how many of each, and how they attach to or move relative to one another. Clarify only load-bearing ambiguities.

2. Inventory or generate the parts

shCopy code
pc --no-ansi list parts          # what already exists to reuse

Reuse existing parts; generate any missing component with the /pc:gen-part flow. Each part must pass pc test on its own before you compose it.

3. Author the ASSY

Scaffold (this creates the empty file and the assemblies: entry), then write <name>.assy:

shCopy code
pc --no-ansi add assembly assy <name>.assy

An ASSY file is a tree of nodes under links: (reference: PartCAD "Assembly YAML" docs, docs/source/assy.rst).

Container — the top level, and any grouping node:

yamlCopy code
name: <optional>
description: <the description>
location: <optional [[x,y,z], [ax,ay,az], angle_deg]>
links:
  - <node>
  - <node>

Part node — places a part; use exactly one placement method:

yamlCopy code
# explicit placement (translation, then rotation of angle_deg about the axis):
- part: <part-in-this-package  or  //package:part>
  name: <optional instance name>
  location: [[x, y, z], [ax, ay, az], angle_deg]

yamlCopy code
# OR connect by ports (no interface mating):
- part: <...>
  connectPorts:
    with: <this part's port, if it has more than one>
    name: <target instance already in the assembly>
    to: <target port, if the target has more than one>

yamlCopy code
# OR connect by interfaces (universal mating):
- part: <...>
  connect:
    with: <this part's interface, if more than one>
    name: <target instance already in the assembly>
    to: <target interface, if more than one compatible>
    # optional disambiguation: withInstance / withPort / toInstance / toPort

location, connectPorts, and connect are mutually exclusive — pick one per node. with names the interface/port on the part being added; to names it on the part already in the assembly.

Both connectPorts and connect also take two optional sections that describe the connection rather than place it:

yamlCopy code
- part: <...>
  connect:
    name: <target instance already in the assembly>
    comment: <free form context for a human or an LLM; never parsed>
    how:
      stage: <label; consecutive nodes sharing it are connected at the same time>
      pushForceMax: <force in N, default 5>
      pushDistance: <staging distance in mm; default 1.5x the part's length along the interface Z>
      turnDirection: <cw (default) or ccw>
      turnTorqueMax: <torque in N*m, default 0>
      threadStep: <lead in mm per full turn, default 0.00>
      holdWith: <interface(s) to hold the part being added by>
      holdWithInstance: <instance(s) of holdWith>
      holdWithForceMin/Max: <force in N to hold it with, default 3/7; holdWithForce sets both>
      holdTo: <interface(s) to hold the target by>
      holdToInstance: <instance(s) of holdTo>
      holdToForceMin/Max: <force in N to hold the target with, default 3/7; holdToForce sets both>

pushDirection is not a field: it is deduced from the interfaces (an interface's +Z points into the object it belongs to, which is the way an incoming part travels) and reported with the rest.

Everything that is required to perform the assembly must be codified in how and the other fields — never only in comment, which no tool reads. The hold* fields default to the connect: section of the part or assembly definition in partcad.yaml:

yamlCopy code
parts:
  <name>:
    type: step
    connect: # (optional) what this part contributes to every connection
      hold: <interface(s) to hold it by>
      holdInstance: <instance(s) of hold>
      holdForceMin/Max: <force in N; holdForce sets both>

threadStep defaults to the thread of the interfaces being connected, declared once on the interface that introduces it:

yamlCopy code
interfaces:
  m3:
    threadStep: 0.5   # inherited by every interface that inherits m3
    selfScrew: false  # true when it cuts its own thread instead of matching one

Two connected interfaces must agree on threadStep unless one declares selfScrew. Run pc test -a <name> to check that: it fails an assembly whose instructions contradict themselves — a mismatched thread, or a *Min above its *Max — and passes one that takes every default.

Sub-assembly node — identical to a part node but with assembly: instead of part:, and it may carry its own nested links:.

Prefer connect/connectPorts when the parts define interfaces/ports; otherwise use explicit location. Work in millimeters and degrees.

Do not guess the names with:, to:, withInstance:, toInstance:, withPort: and toPort: take. Ask each part what it has, and read the exact spelling out of the log:

shCopy code
mkdir -p /tmp/pc-render                                  # -O needs the directory to exist
pc render -t png -O /tmp/pc-render --with-all <part>

That draws every port of the part with its name, and every interface instance with a line out to the ports that belong to it — and lists all of them in the log, under the N port(s) drawn on the projection: line, which is the list to copy from. pc info <part> reports the same names as text.

Then mark the assembly not manufacturable in partcad.yaml (it is generated, not a catalog item) so pc test passes:

yamlCopy code
assemblies:
  <name>:
    type: assy
    path: <name>.assy
    manufacturable: false

Do not add a manufacturing: section: assy is the only method an assembly has and it comes with the type. (additive/subtractive/forming/sheet_metal are ways of making a part and mean nothing here.) Marking it manufacturable instead makes pc test require that every part in it can be bought or made from a declared supplier — only do that when that is true.

4. Group what repeats into sub-assemblies

An assembly of any size repeats itself, and a flat file hides that: a reader cannot see that four of those brackets are the same bracket, PartCAD builds each one from nothing, and the instruction book walks through the same six steps four times.

So once the parts are in the right places, look at what recurs and give each recurring group a name of its own. Name them <product>/<piece> — an object whose name has a / in it is written into that sub-directory, so the pieces sit together beside the thing they belong to.

yamlCopy code
assemblies:
  gearbox/idler-stage:
    type: assy
    desc: One idler on its shaft, with the bearing either side

Three things follow from it, and they are the reason to do it rather than a tidiness argument:

• It is built once. Every further use is a cache hit, and a change inside one piece rebuilds that piece rather than the product.
• It is an object in its own right, so it can be rendered, tested and inspected alone — which matters most when the whole product is too big or too slow to render at all.
• The instruction book documents it once, and says how many to make.

Judge the split by what the design actually repeats, not by carving the model into even parts. A group that occurs once is not a piece; a group that occurs eight times is, even if it is three parts.

Check the arithmetic: the parts in each piece, multiplied by how many times the piece is used, plus whatever is left loose, has to come to what you started with. It is the cheapest test there is and it catches a piece that quietly gained or lost a part.

5. Replace the coordinates with connections

A part placed by location: is a part nothing holds up. The file says where it ended up, not what puts it there — so nothing can check it, an instruction book can only say "at (144, 9.6, 0)", and moving anything means recomputing everything downstream by hand. A part joined through its ports says the thing that is true: this feature of it meets that feature of what it sits on, and PartCAD works out the coordinates.

Convert in this order, and verify after each step (§6).

Find the pairs by where the ports are

Two ports are a joint when they are at the same point. That is the whole test, and it is the same test whatever the parts are:

• Compute every port of the part in the pose it is already in.
• Compute every port of the parts already placed.
• A pair at the same point, within a tolerance that reflects how the coordinates were written, is a joint.

Do not derive port names from a part's name or its nominal grid. A part whose name says it is 2 by 2 may carry four ports on one face and one on the other; a part with a hollow middle has ports round its rim and none inside. Naming conventions are a convenience, not a contract — ask the part where its ports are (pc info <part>, or read implements: out of its configuration) and match on position.

Grow the order; do not follow the construction order

A joint may only name a node placed before it, so the order the nodes are written in decides how many of them can be joined at all. Written in the order the thing is physically built, a whole first layer has nothing beneath it and all of it stays on coordinates.

It does not have to be written that way. A part joins to what is under it or to what is over it, and the second breaks the deadlock: place one part of the first layer, add the part above that bridges it to its neighbour, and the neighbour then joins upward into that bridge. So grow the order — take whatever can be joined to what is already down, and only when nothing can, place another part by coordinates and carry on. Break ties by the original order, so a piece that needs no help still reads the way it was written.

The first node needs no location at all

Neither coordinates nor a joint: it is the assembly's origin, and where that goes is for whatever places the assembly to decide. Saying it twice is how the two come to disagree.

That moves the assembly's frame onto that node, so whatever places the assembly has to take up the difference — by the offset the node used to sit at, turned the way the assembly is turned. An assembly whose parts are all placed against one shared grid is the exception: leaving its first node's coordinates out moves that node and nothing else.

A turn is a difference, not an orientation

When a part is joined to one that lies a different way, the joint carries the difference between the two — not how either of them lies on its own. A part turned a quarter turn, sitting on another turned the same quarter turn, is not turned relative to what it sits on and the joint needs no turn at all.

Getting this backwards is the most expensive mistake in the conversion, because the result still builds: the part lands one feature away, which looks like a near miss rather than a wrong rule.

Where a joint does need a turn, prefer a port pair that does not: parts usually meet over several features, and a pair that lies square avoids the question entirely. Take the turned pair only when it is the only one.

Let sub-assemblies join each other

A port inside a sub-assembly is the sub-assembly's own business until it says otherwise. map: is how it says otherwise — a name of its choosing against the node inside it, the interface that node implements, and the instance:

yamlCopy code
assemblies:
  gearbox/idler-stage:
    type: assy
    map:
      shaft-out: [bearing-b, //pub/std/bearing:bore, outer]

The interface is not renamed — it is a contract — but the instance name is the piece's to pick. The parent then connects to shaft-out like any other port. The map does not reach inside a sub-assembly of a sub-assembly: that one externalizes what it wants seen, and this one maps that. A boundary is crossed one object at a time, which is what keeps a piece free to be rearranged inside without breaking what connects to it.

Note that a sub-assembly is joined through a port of a part inside it, and that part may lie differently from the sub-assembly itself — in which case the joint turns the whole piece by the difference. The rule above applies unchanged; it is just easier to miss one level up.

When a part has no port where the joint is

Leave it on coordinates and say why. A joint through an interface a part has not got is not a joint, and one through a port in the wrong place is a joint that lies — it will place something silently wrong, which is worse than coordinates that are merely silent.

Then fix it where it belongs: a part missing a mating feature it really has is a gap in the package that serves the part, not something to work around in the assembly. Report it or fix it there, and the assembly gets shorter for free.

A part that is a standard part under a name of its own (an enrich, such as a leg cut from a post) has the standard part's ports. Where the joint is near one of them rather than on it, give it a name of the leg's own with map: instead of declaring a new port by coordinates: {port: <its port>, moveX: ..., turnZ: ...}, in the frame of that port, written in terms of the enrich's with values where it depends on them ("%length * 25.4%"). See "New names for what an object has" in the configuration docs.

Do not predict a mate; try it

Where a joint comes out wrong and the geometry says it should not, stop reasoning about the kernel's mate arithmetic and ask it. Build a throwaway assembly holding every hypothesis — each candidate port of the part, against each candidate port of the target, at each candidate turn — beside the same part placed by the coordinates you are trying to reproduce. Instantiate it once and read back where everything landed. Whatever matches the reference is the answer.

It costs one instantiation and it is not a guess. Two of the conversions this guidance came from were settled this way after several rounds of plausible reasoning had each produced a different wrong answer.

6. Validate, render it from several angles, iterate

shCopy code
pc --no-ansi test -a <name>          # build gate: geometry instantiates

Prove each step changed nothing

Grouping into pieces and replacing coordinates with joints are both supposed to leave the product exactly as it was. Neither announces it when they do not: the assembly still builds, still renders, and a part one feature out of place looks like the design.

So compare, rather than look. Instantiate the assembly before and after, walk it down to the parts, and check every part is the same part at the same place. No CLI command prints that — pc info -a gives the configuration, the ports and a hash of the source, which changes when the file is rewritten however little the geometry moved — so walk the tree yourself:

pythonCopy code
import asyncio, json, sys
import partcad as pc

ctx = pc.init(".")
assembly = ctx.get_assembly("//<package>:<name>")
asyncio.run(assembly.do_instantiate())

out = []
def walk(node):
    for child in node.children:
        item = child.item
        asyncio.run(item.do_instantiate())          # a sub-assembly is lazy
        if getattr(item, "children", None):
            walk(item)                              # compose the transform here
        else:
            out.append([item.name, child.location])
json.dump(out, open(sys.argv[1], "w"), default=str)

Run it against the tree before the change and after it, and diff the two.

A render is not the check. Two drawings of an assembly with one part moved are nearly identical pictures, and anti-aliasing alone will differ between two runs that are geometrically the same — so a pixel diff answers a question you did not ask. Compare the placements.

Two cautions from doing this badly:

• Compare in world coordinates once a frame has moved. Dropping the first node's location moves the assembly's own frame, so every part inside it "differs" while the product is untouched. Compose the transforms down through the tree — rotations included — and compare where each part ends up in the end.
• Compare as a multiset, not pairwise. Sorted lists of (part, position) shift wholesale when one entry changes, so a single misplaced part reads as dozens of differences and buries what actually moved.

Expect zero. A conversion that moves one part has a bug, not a rounding difference.

That gate passes for an assembly whose parts are all in the wrong place, so the design is decided by looking at it — and one picture cannot decide it. An assembly is judged on where its components are, and position along the viewing direction is precisely what a projection collapses: a part offset into the screen, mirrored, or turned a quarter turn sits perfectly in the one view that hides the error. Render four — front, top and right to place things in all three axes, iso to see the product whole. A rendered file is named after the object, so give each view a directory of its own or they overwrite each other:

shCopy code
for view in front top right iso; do
  mkdir -p /tmp/pc-render/$view                     # -O expects it to exist
  pc --no-ansi render -a -t png --view $view -O /tmp/pc-render/$view <name>
done

Each view directory gets <name>.png plus one file per sub-assembly, since a sub-assembly is an object in its own right. View all four of <name>.png — and the sub-assemblies' own views when a nesting is what went wrong — and check against the decomposition from §1:

• every component is there, as many times as it should be, and nothing else is;
• each one is where it belongs in all three axes — a placement is only confirmed by two views that share no viewing direction;
• each one is oriented as it belongs, including where a symmetric part hides it: a flipped bracket reads the same from the front and differs from the top;
• nothing interpenetrates that should not, and nothing floats where it should seat — a gap and a contact look alike from the front and separate from the side.

Add back, left or bottom when a component stays hidden behind another in all four, and --viewport-origin X,Y,Z / --viewport-up X,Y,Z to look along a joint no named view shows. /pc:render covers these options in full.

Where a component is in the wrong place and the ASSY uses connect: rather than location:, the views tell you that it is wrong; the next section tells you why.

When a connection comes out wrong

A part in the wrong place tells you that a connect: is wrong, not why — ports and interfaces are not geometry, so nothing about them is on the plain render. Draw them:

shCopy code
mkdir -p /tmp/pc-render/ports
# Every port of every part. Add --with-interfaces for each interface joined to
# its ports, or --with-all for both.
pc --no-ansi render -a -t png --view iso -O /tmp/pc-render/ports --with-ports <name>

These write <name>.png like any other render, so an overlay needs a directory of its own or it lands on top of the plain views from above — and --no-ansi because the port names are reported on stderr, which is half of what makes the picture readable.

On an assembly these walk everything inside it and place each child's ports where the assembly put the child, so both ends of a connection are on one picture. Each port is a coordinate frame whose long arrow is +Z — the direction a part travels along when it is connected through that port — and each is named <part-instance>:<port>.

Being a frame rather than geometry does not exempt it from what §4 is about: one projection still collapses the axis a port is offset along, so two frames a millimetre apart along the line of sight are drawn one on top of the other. When the iso view leaves that ambiguous, render the overlay from the same four views the geometry was checked from.

Read the picture:

• Connected as intended: the two frames coincide and their +Z arrows point in opposite directions.
• A fixed gap between the two frames: the connection resolved, but one of the parts places that port wrong. Fix the part's implements:, not the ASSY — and use /pc:add-interfaces for that.
• Frames coincide but the part is rotated: the ports met and the roll (their X axes) is what disagrees, or the wrong instance of a symmetric interface was picked. Pin it with withInstance:/toInstance:.
• The part did not move at all: the connect: named something that does not exist or is ambiguous. The log lists every port drawn, with its exact name — compare it against what the ASSY says.
• Two parts connected through the wrong pair: --with-interfaces shows which ports each interface instance owns, which is what a connect: selects by. If a bolt pattern shows up as four separate instances rather than one with four ports, the interface is declared wrong.

Iterate on the --with-ports render, not the plain one, for as long as a connection is the thing being fixed. A part whose own port is in the wrong place is fixed in the part, not in the ASSY — render that part on its own with --with-all (/pc:add-interfaces covers reading it) before touching the assembly that uses it.

Then adjust the placements or mates, re-test, and re-render. Iterate until every view matches. pc ide view -a <name> gives an interactive view.

7. Finalize

Summarize the structure — parts, sub-assemblies, key placements — and how to view it (pc ide view -a <name>, or pc render -a -t png --with-all <name> for a picture with the connection metadata on it).

Say what is still placed by coordinates and why, one reason per case. "The rest use location:" tells the next reader nothing; "these four sit where nothing below them carries a port, and that part's package declares none on that face" tells them where to look and what would remove it. An assembly that is honest about its remaining coordinates is one somebody can finish.

For an assembly whose connections are worth keeping a picture of, declare the drawing as a file type so pc render keeps it up to date along with everything else:

yamlCopy code
assemblies:
  <name>:
    render:
      svg-with-ports:
        package: //builtin/render
        path: render_svg.py
        extension: ports.svg
        with_ports: true

examples/feature_interface does this for a part and for the assembly it belongs to.