Design → test → fix, before printing

The user prints on a Bambu Lab printer (default P1S, 256 mm cube, 0.4 mm nozzle, PLA unless they say otherwise). Never hand over a design that phyx3d says needs-changes without telling them exactly why.

1. Understand the job

Before modelling, pin down (ask only if you can't infer a sensible default):

• what the part does, how it is held (screwed to a wall, sits on a desk, clamped…) and what load it takes (e.g. "2 kg shelf" ≈ 20 N; add ×2 for safety),
• key dimensions / what it must fit,
• material (default PLA; PETG for outdoor/impact/heat > 50 °C; TPU for flexible).

2. Model it (build123d MCP)

• Design with the build123d tools in small steps and look at their previews.
• Millimetres, Z up. A single part: put the intended print face at z = 0. Several parts that fit together (an enclosure, a lid, modules): model each one in its assembled position so fit checks compare real geometry, and rotate it into print orientation only when exporting.
• Printable-by-design rules: walls ≥ 1.2 mm (≥ 2 mm under load); 45° chamfers (not fillets) on downward edges; teardrop holes in vertical walls; fillet r ≥ 2 mm at inside corners that take load; avoid floating parts.
• Export an STL to an absolute path (e.g. ./out/<name>.stl).

3. Test it (phyx3d MCP) — always in this order

1. analyze_model {path, material} → read verdict, checks[].findings, and todo. Look at the returned image.
2. If anything about orientation is off (supports, stability, warp): suggest_orientation {path} and re-run analyze_model with the best rotate.
3. If the part carries load: stress_test with realistic fixed + loads in the design's own axes. rel regions are the easiest: {rel:{min:[0.85,0,0],max:[1,1,1]}} = the far 15 % along +X. Pass the chosen print rotate too — layer direction changes the strength. Aim for safety factor ≥ 2 (≥ 3 for impact/repeated loads). Snap-fits are different: a snap moves by a fixed distance (barb reach + clearance), not under a known force. Test it with displacements: [{region: <the barb>, move: [0,0,<travel>]}] instead of loads, with a fine mesh (resolution ≥ 60000 for thin arms). You get the safety factor at that travel and pushForces (N, how hard it is to press). Force-guessed snap FEA can be wrong by large factors either way. Lay snap springs flat on the bed where possible, so they bend along the layers.
4. If it must stand, stack, or survive knocks: simulate_physics with tilt, push, stack or drop.
5. Optional, if Bambu Studio is installed: slice_bambu for exact time and filament.

3b. Things that move — robots, vehicles, arms, grippers, linkages

Use simulate_mechanism whenever parts move relative to each other.

1. Model each moving part as its own STL in assembly position (all parts share one coordinate system, Z up, floor at z = 0). Leave ≥ 0.3 mm clearance between separate parts.
2. Write a <name>.mech.json next to the STLs:
   • parts: {id, file} (or shape box/cylinder/sphere for quick tests), fixed: true for a base screwed down, payloads: [{mass, at}] for batteries/objects carried, friction: 0.9 for TPU tyres/feet.
   • joints: {id, type: revolute|prismatic|fixed|ball, parent (id or "world"), child, anchor [x,y,z], axis, limits} — anchor = the pin/shaft centre.
   • motor: pick a real part via preset (SG90, MG90S, MG996R, DS3218, STS3215 servos in degrees; TT, N20, JGA25, NEMA17 motors in rpm; LINEAR in mm) and a target: number, keyframes, or expression of t (e.g. "20*sin(2*pi*1.2*t + pi)"). Expressions may read yaw, pitch, roll, x, y, z, speed of the tracked part for simple feedback (e.g. steer a walker with (22 + 0.1*yaw)).
3. Run it and read: distance/speed/heading drift/fell (mobile robots), joint ranges (machines), each motor's p95 and peak vs rated torque, saturated share, tracking error, collisions between parts, overlap at start. Look at the filmstrip.
4. Fix: a motor saturated or p95 > 0.7 × rated → stronger preset, shorter lever, lighter part, slower motion. Falls over → wider stance, lower body, slower gait. Veers → symmetric design or feedback. Parts collide → clearance or phase change. Then also run stress_test on the most loaded part using the motor's peak torque as the load.
5. A walking gait that works as a starting point: trot, diagonal legs in phase, hip A*sin(ωt+φ), knee B*max(0,-cos(ωt+φ)) (lifts the foot while the leg swings forward).

3c. Parts that lock together — dovetails, T-slots, bayonets, snaps, detents, threads, press fits

Run check_interlock whenever two printed parts join (lids, doors, rails, caps, clips). Model every part in its assembled position, then give spec: {parts:[{id, file}], interlocks:[{type, moving, against, axis}]} + base_dir (or a .interlock.json path).

• axis = the direction the moving part travels to go in. Twist types also need depth (push-in before turning), angle (sign = direction) and center (a point on the axis); screws need pitch + turns. Unusual motions: insert: [{move}|{rotate, axis, about}|{screw, pitch, axis, about}].
• Read: fit (clamped = zero clearance on opposite sides → will bind), insertion (jams where?), escapes (free play per direction; an unexpected free direction → add a stop), engagement.net (catch height − free play; needs ≥ 0.2 mm detent / ≥ 0.6 mm lock, in whole layers), press-fit interference (0.02–0.25 mm per side), wrongWays ("assembles" → add a key so it only fits one way).
• Then check the catch's strength: stress_test with displacements = the engagement height (needed) at the barb.

4. Fix and repeat

Apply the todo fixes in the CAD, re-export, re-test. Iterate until the verdict is ready, or printable-with-care with every remaining warning explained. Usually 2–4 rounds.

For enclosures and other multi-part assemblies, also:

• Fit-check against a model of what goes inside (board, battery, device): model it with its real dimensions and check clearances in the assembled position, not by eye.
• Test locks and snaps with check_interlock: it measures the real free play and whether the catch still holds after it, and finds paths that jam.
• Watch for slicer settings that break the design: elephant's foot on the first layer, seam placement on mating faces, and hole shrinkage — say which settings (or compensation) the design assumes.

5. Report to the user

Keep it short and concrete:

• final STL path and how to print it: orientation (rotation values), supports yes/no, material, walls/infill you tested with,
• the numbers that matter: safety factor under their load, tip angle, filament/time estimate,
• anything you could not verify (strength is a rough guide; warp is a heuristic),
• tip: they can open phyx3d serve (http://localhost:5217) to see every check you ran, live, in 3D.