kernelCAD

Use this when the user wants to start from their own STEP or STL file (a part they designed elsewhere, a vendor model). Give exactly one source: url (a public https direct-download link), upload (a file the user attached in ChatGPT), or content_base64 (raw bytes, small files only). Max 50 MB. The format is detected from the bytes (or pass format: 'step' | 'stl'). Writes the file and a starter <name>.kcad.ts whose base body is lib.fromSTEP / lib.fromSTL into output_dir (default: a fresh temp dir), evaluates it, and returns { ok, format, model_path, script_path, code, bbox }. Pass script_path as { file } to evaluate_script, add_feature, dfm_review, export and the other tools, or edit code to build on the body. STL is a mesh: booleans, bbox and export work, but faces stay planar facets (use mesh_to_features for an editable feature tree). A non-watertight STL is refused unless allow_open is true.

Use this when you need to read facts about a model. One reader, selected by `of`: - 'assembly' — physical assembly inventory (parts, bboxes, connectors, mates, disconnected solids). - 'robot' — URDF/SDFormat export preview (links, joints, planning groups, end-effectors, issues). - 'step' — inspect an imported STEP file: solid tree, exact bbox/volume, cylindrical holes (breached bores flagged partial; holeDetection: 'exact' | 'heuristic'). - 'shape' — volume / surfaceArea / bbox for one feature ({ feature_id? }). - 'mass' — mass, centre of mass, centroidal inertia tensor (inertia6 + 3x3 inertiaMatrix), principalMoments/principalAxes, symmetry flags, and optionally the radius of gyration about an arbitrary axis ({ feature_id?, density?, gyration_axis? }); density in kg/m^3, defaults to 1000 (water). - 'features' — features captured by the script (kind, id, params, transforms, suppression). - 'assemblies' — assembly intent (assemblies, parts, connectors, joints). - 'topology' — canonical face names + edge count for a feature ({ feature_id? }). - 'edges' — edges of a shape with optional EdgeQuery ({ feature_id?, query? }); returns @kc[...] refs. - 'face-edges' — boundary edges of a named canonical face ({ feature_id?, face_name }). - 'faces' — faces of a shape with optional FaceQuery ({ feature_id?, query? }); returns @kc[...] refs. - 'face-labels' — user-applied labels visible in the script. - 'mates' — mates captured by the script. - 'constraints' — sketch constraints captured by the script. - 'part-stats' — bundled parts-catalog statistics. - 'bend-table' — sheet-metal bend table for a flattened pattern. - 'params' — declared model parameters. - 'part-categories' — top-level part-catalog categories available in the bundled (and configured remote) catalog. - 'part-families' — part families within a category ({ category? }); count + exemplar ids per family. - 'bom' — bill of materials ({ assembly? }): one row per distinct part (grouped by geometry/catalog identity, no…

Evaluate supplied kernelCAD code or a file to capture an authored NURBS curve (nurbsCurve/spline3d). Returns degree, distinct knots and multiplicities, poles, weights, rational and uniformWeights flags, hasNonUnitWeights, and parameter range. Defaults to the last captured curve in that script; pass feature_id to choose one. The source may load referenced assets. Reports authored curve data, not exact geometry-derived B-rep edge/face NURBS; surface extraction and blending are not supported. Does not update the active MCP session or write a project.

Use this when you need to list the kernelCAD script-runtime surface: global functions (box, path, selectEdges, helix, etc), Shape methods (fillet, sweep, lower, etc), Sketch methods (extrude, revolve, sweep), PathBuilder methods, EdgeQuery/FaceQuery key sets, and featureKindFaceLabels (which globals accept opts.faceLabels and valid value shapes). Use this to discover what is callable from a .kcad.ts script. Call this BEFORE concluding kernelCAD lacks a capability — its NURBS freeform surfacing (loft, sweep, boundary-fill, G2 blend) is easy to miss from tool names alone.

Return the kernelcad-authoring skill — conventions for writing .kcad.ts scripts (API surface, parameters, evaluation contract, common pitfalls) — in sections. Call with NO arguments first: you get the quick-start (units, axes, the return rule, the rules that prevent most failures, the intent → API routing table, the "Before you finish" checklist) and an index of every section with a one-line summary (under 16k chars). Then pull what you need: `{ section: "<id>" }` returns one section (e.g. "top-level-functions", "materials", "dfm-gates-print-readiness"); `{ query: "fillet after subtract" }` returns the best-matching sections. `{ section: "all" }` returns the full ~140k-char skill (SKILL.md plus its reference files) — only if your client accepts large responses. Clients that list resources can also read `kernelcad://skills/authoring` (the SKILL.md core). OUTPUT: { uri, mimeType, mode, text, sections? } — `mode` is index | section | search | all | not-found.

Use this when you need a canonical pattern snippet for a CAD task. Search the kernelCAD cookbook for canonical pattern snippets. Returns top-k snippets matching the natural-language query, ranked by BM25 over title/tags/keywords/trigger. Use when you need a canonical pattern for fillet-after-subtract, non-overlapping booleans, sketch-to-extrude flows, etc. Returns empty if no snippet scores above the relevance floor — proceed without cookbook help in that case.

Use this when you need the kernelCAD 26-code diagnostic catalogue with hint templates. Tiny one-shot call; useful for an agent that wants to pre-populate retry strategies. Hints are also inline on every emitted diagnostic — this tool just gives you the canonical list up front.

Use this when the user wants to union, subtract, or intersect two imported triangle mesh solids (STL, OBJ, or 3MF). Supports union, subtract (A minus B) and intersect. Pass each operand as { file } on the MCP server or { data } with base64 mesh bytes; formats are detected or may be specified. Returns volume in mm³, triangle count and a binary STL as data_base64; pass that payload to import_step to continue modeling. Inputs must be closed, consistently oriented 2-manifold meshes. Open, non-manifold, or invalid meshes return an error. This computes mesh geometry and does not create an editable B-rep feature tree.

Mesh a kernelCAD .kcad.ts source server-side and return a COMPACT geometry summary — overall bounds plus, per feature, its id, kind, triangle count, and bounding box. Use this to INSPECT a model's geometry without a viewer: confirm a part is the size/shape you expect, see how many triangles each feature contributes, or check that every feature produced geometry. This runs the full server-side OCCT pipeline (the same one the Studio renderer uses), so it evaluates modern sources (assembly, path, .material, …) that the legacy client worker cannot. INPUT: `source` (required) the .kcad.ts script text; `fileName` (optional) a label for diagnostics; `params` (optional) a map of parameter-name → number overrides applied before meshing (stateless slider recompute). OUTPUT: { ok, bounds, featureCount, features: [{ id, kind, triangleCount, bbox: { min:[x,y,z], max:[x,y,z] } }], failedFeatureIds, diagnostics }. `ok` is true when every feature meshed; `failedFeatureIds` lists features that failed to compile (and `ok` is then false). Raw vertex/index/normal arrays are NEVER returned — this is a summary only. To SEE the rendered model, call open_in_studio (includes PNG preview + viewer); use get_latest_render only for alternate views.

Use this when you are handed an STL, OBJ or 3MF of a mostly prismatic mechanical part (plate, bracket, spacer, flange, housing block) and need an EDITABLE kernelCAD model of it rather than a faceted lib.fromSTL import. Deterministic, measured, self-verifying: it welds and checks the mesh, segments planes and cylinders, picks the extrusion axis, slices each band and fits exact lines / arcs / circles, snaps near-round values (each snap recorded), then emits a readable .kcad.ts with named param()s — a revolve for concentric round stacks, extruded profiles otherwise, .hole()/.holes() for through, blind and counterbored bores (axial and side-drilled), .cutout() for pockets, .fillet() for constant-radius edge blends (radius measured on the sharp edge, edges grouped by radius and picked with the shortest exact edge query), boolean subtractions for what no drilling feature can reach. It then EVALUATES that script and compares it with the mesh: volume IoU (column ray casting) and symmetric surface deviation (max + RMS), over up to 4 refinement passes. Returns { script, ledger, fidelity: { maxDeviationMm, rmsMm, volumeIoU, verdict: faithful | approximate | failed, thresholds }, unmatchedRegions, features, passes }. A fillet or sharp reading is kept by which measures better; variable-radius blends and chamfers are reported, not forced. The verdict is computed from the numbers — faithful needs IoU >= minIoU AND max deviation <= maxDeviationMm AND a watertight mesh AND no unmatched region. Freeform surfaces, tilted planes and side bosses are listed in unmatchedRegions (reference.mesh.freeform-region-unmatched), never silently dropped. The ledger uses fact ids equal to param names, so resolve_assumptions on the written <out>.ledger.json yields paramOverrides for set_param. Pass { out } to write the script and ledger; the mesh itself is never modified. Hosted: give the file as an https:// URL, a data: URI (base64, max 25 MB) or "asset:<id>" from request_upload (when the file is o…

Save the model as a kernelCAD project AND show it: persists a project revision, returns the interactive Studio viewer (MCP Apps / ChatGPT outputTemplate), and includes a PNG preview in the SAME tool result (image content + previewUrl when available). ONE VIEWER PER PROJECT PER RESPONSE: every call renders a new viewer card. Do all validation with evaluate_script first, then call open_in_studio exactly once, after validation succeeds; reuse the returned URL, revision and preview, and never call it again after a successful result in the same response. Re-calling with unchanged code for the same project returns that revision instead of writing a new one. A heavy model can answer status:"running" (code tool.still-running) before the client times out; that is NOT a success: call open_in_studio again with exactly the same arguments, which joins the running save and returns its result. CONTINUE, DO NOT DUPLICATE: to change a model you already saved, pass `project: '<slug>'` from the last result (or from list_my_projects) — this adds a revision to that project. Without `project`, the save goes to the project this chat already saved, or to your project with the same title (case, punctuation and v2/rev3/(1) ignored) from the last 24 h. Views, poses, colours and options of one design are revisions of ONE project — keep the same title and slug. Pass `new_project: true` only when the user wants a separate design. A different cookbook (execute_cookbook, or open_in_studio of that cookbook's unchanged source) starts its own project; pass `project` to add a revision, or `continue_design: true` to keep revising the design this chat is already saving. The result starts with the slug to pass next time, a /p/<slug>?version= link for this revision, and that revision's title. Use this when the user wants to SEE or share the model — do NOT call get_latest_render afterwards for the happy path; the preview is already here when previewDelivered is true. Pass the full `.kcad` source as `code`…

Order the part made: creates the production request and returns a pay_url for the user. Propose the manufacturing method from the user's words and the part, pass it as method with a one-line reason. The price is the real maker price; never invent or send one. Tell the user the card is only held, and a person confirms the final price before charging. If the result is status "pricing", the price is still being worked out: call again in a few seconds.

Start or read the real price for making the part. Cheap. Returns status "ready" with estimate_cents and summary, or "pricing" while the makers are asked (call again in a few seconds). Creates nothing and charges nothing.

Use this when you need to wrap a 2D closed curve onto a 3D face. Insert a `<shape>.projectCurve({ source, face, scaleMode? })` chained call into a kernelCAD script. The `source` is the structured `{ kind: "sketchCommands", commands: [...] }` wire format the runtime API accepts. Wraps the curve onto the face along the face normal; pair with `.extrude(d)` / `.cut(...)` for raised or engraved logos on curved bodies. Open-wire projection (`asEdge: true`) is not implemented and is rejected at edit time. Side-effect-free; returns modified code plus diagnostics.

Use this when you need to resolve or inspect topology against a script's lowered geometry. Pass the script as `code` (inline; works on a hosted server) or `file`. Selected by `mode` (default 'evaluate'): - 'evaluate' — list the faces a Query matches ({ query }). Pass expect:'unique' to assert exactly-one. - 'resolve' — resolve one ref to one face ({ ref }). - 'lineage' — the creating/modifying history of a face ({ ref }; feature_id defaults to the last feature). Query forms (faces only; edge/vertex kinds return query.unsupported-entity-type): - a bare face label: 'top', 'bottom', 'left', 'right', 'front', 'back', or a hole/cutout slot like 'wall'. - @kc[<feature-id>/face/<label>] — feature ids are the ones refs print, e.g. '@kc[box_1/face/top]', '@kc[hole_1/face/wall]'. - @kcq[face(<filter>, ...)] — filters AND together: createdBy("box_1"), withLabel("top"), geometryType(PLANE|CYLINDER|CONE|SPHERE|TORUS|BSPLINE_SURFACE), closestTo([x,y,z]) or closestTo([x,y,z], k), containsPoint([x,y,z]); combine with union(a, b), intersection(a, b), subtraction(a, b). Example: '@kcq[face(geometryType(CYLINDER), closestTo([0,0,10]))]'. Each matched face comes back with its canonical @kc ref, centroid, normal and area.

Use this when you need the faces a ray hits in a local Studio model, nearest first, with points, distances, and normals. Inputs are model units (mm); direction is normalized. Requires the local Studio URL and session token for the same script, reads the current model generation, and returns a set_selection argument object with every hit. feature_id must identify a displayed terminal shape; intermediates have no current viewer shapeIndex. Intersections use the viewer per-face tessellation, so curved-surface points and normals are mesh approximations.

Use this when you need to remove a feature line from a script. Remove a single line from a kernelCAD script identified by a substring match. Returns the modified code plus diagnostics from re-evaluating. Refuses to remove the line containing the return statement. Side-effect-free.

Use this when you need to LOOK at a kernelCAD model — render its script to deterministic PNG views for visual self-check (the visual half of the evaluate → render → inspect → fix loop), with NO studio or dev server required. Pass { code } (inline source) or { file } (a .kcad.ts path), exactly one. Renders the canonical engineering views (front, right, top, iso — pass { views } for a subset, e.g. ["iso"] for fastest iteration) plus an optional { pose: "<az>,<el>" } arbitrary camera angle (degrees; az=0,el=0 is front, +az rotates CCW around +Z, +el lifts the camera). NO STUDIO / DEV-SERVER REQUIRED: a prebuilt static player (dist/headless-player) is served from an ephemeral local port automatically; a running studio dev server is used as fallback, and { base_url } forces one. The only environment dependency is playwright chromium (npx playwright install chromium). Pass { focus } or { hide } (arrays of feature ids or assembly part names, mutually exclusive) to isolate parts — same semantics as `kernelcad render --focus/--hide`. Pass { section: { axis, position, flip? } } to cut a cross-section and inspect INTERIOR geometry (wall thickness, internal pockets, whether a bore runs through) rather than only the outer shell. Pass { explode: { factor, mode? } } to pull a multi-part assembly apart (mode: "mate-axis" default, or "radial") using the same mesher as `kernelcad render --explode` — requires assembly.model()/solvedModel(). PNGs are written to { out_dir } (default: a fresh temp session directory) and returned as absolute paths with per-view camera descriptions (kernelCAD is Z-up). Mechanism truth runs first, same protocol as `kernelcad render`: a broken mechanism still renders but every tile is watermarked MECHANISM BROKEN (KERNELCAD_RENDER_STRICT=1 refuses instead); read { mechanism, mechanism_failure_codes }. The probe runs full BREP interference sweeps and can dominate latency on large assemblies — pass { no_mechanism_check: true } for fast iteration (the preview…

Use this when evaluate_script reported an error and you want the fix applied rather than described. Takes the candidates why_did_this_fail derives for a diagnostic, applies them one at a time, re-evaluates after each, and keeps the first that clears the diagnostic without introducing new errors. Never edits outside the repair region (failing feature statement + its input statements + the param() lines it reads) — an out-of-region patch is refused with tool.repair.out-of-region. Returns the repaired source in `new_code` (the caller persists it), a unified `diff`, and before/after health maps. Pass { file? | code?, diagnostic?: '<id>'|'first-error', strategy?: 'apply-first'|'try-all'|'dry-run', max_attempts?: number }.

Use this when the user needed something kernelCAD does not have, so the gap is recorded instead of papered over. kind: asset (search_assets returned coverage none or poor: the catalog does not model it), capability (a tool cannot do what was needed), output (a tool answered, but not with what was needed), docs (a tool description did not say how to use it). One short sentence in `need`; name the tool in `tool` when one fell short. Stores no personal data beyond your account id when signed in. Then tell the user plainly what is missing.

Submit a manufacturing quote request after the user has reviewed and explicitly approved the checklist and every displayed warning. KernelCAD staff will confirm the final itemized price. This sends the request for human review and does not charge the customer or start checkout.

Use this when the user has a photo, PDF drawing, STEP, STL, DXF, SVG or 3MF file on their device and you need it in a kernelCAD tool (trace_from_image, drawing_to_cad, mesh_to_features, inspect of:"step", or lib.fromSTEP / fromSTL / fromDXF / fromSVG / referenceImage in a script). Do not ask the user to paste the file as text or base64, and do not type coordinates from an image by hand. Returns a one-time `uploadUrl` (valid 30 minutes) and an `asset` reference. Give the user the link; after they say the upload is done, pass the `asset` value ("asset:<id>") as the file argument, for example trace_from_image({ imageUrl: "asset:<id>" }) or lib.fromSTEP("asset:<id>"). If the file is already at a public https:// URL, pass that URL instead. On clients that support MCP URL elicitation the upload page opens for the user directly and the call returns once the file is stored (status "uploaded"); then use the asset right away without asking the user anything.

Use this when you need to confirm or override the open facts in an assumption ledger from trace_from_image (missing scale, inferred/assumed values) before committing geometry built from a reference photo. Reads the persisted `<model>.ledger.json` at `ledgerPath`, applies each resolution — `{ id, confirm: true }` to accept a fact as-is, or `{ id, value }` to override it — rewrites the ledger file, and returns the updated ledger plus `paramOverrides` (factId -> value) to feed straight into `set_param`. Pair with the `kernelcad-from-reference` skill.

Run a deterministic mechanism-fitness review on supplied kernelCAD code or a file. It performs a full evaluation, replaces the active MCP session model on success, and clears that session if evaluation fails. The script may load referenced assets. The review validates the assembly/mate graph and connector placement, samples declared mate limits, optionally checks interference at sampled poses, and reports connector workspace bounds, fitness findings, and a repair mode (none, local-fix, parameter-tune, or topology-redesign). Runs within timeBudgetMs (default 90 s); stages that do not fit are listed in skippedStages and were not checked. Hosted: lib.fromSTEP / fromSTL / meshFromSTL / fromDXF / fromSVG and referenceImage take a direct string literal or a const bound to one (computed expressions are unsupported). They accept "asset:<id>" (from request_upload), an https:// URL or a data: URI as the path.

Return the newest brush-painted review packet from a Studio session. After sharing a /p/<slug> link, the user can open it in the browser and paint marks over the 3D viewport to give visual feedback. Call this tool with the `slug` from that link to see the strokes — screenshot + mask + struck part names plus an optional one-line note and intent tags (e.g. "too thick", "missing", "wrong angle") describing WHAT is wrong — and act on the feedback. The slug is the capability: no OAuth required when passing `slug`; private projects require the owner to be signed in. Omit `slug` to fetch your own latest packet from your signed-in account (requires OAuth). By default returns short-lived signed Storage URLs for the screenshot + mask + meta.json plus the struck part names — small and context-friendly. Pass `paths_only: false` to also base64-inline the PNGs for clients that cannot fetch the signed URLs over HTTP.

Use this when you need to know whether a part will hold a load. Runs a linear-static structural study and returns evidence — peak von Mises stress (MPa), peak displacement (mm), the minimum safety factor against the material yield (taken away from the clamp-singular edges of the fixed faces; the raw clamp-edge peak is reported separately as peakAtSupportMPa), per-region hot spots named by @kc[...] face ref, mesh trust flags, an equilibrium residual, and (CalculiX only) stress-heatmap PNG paths. The study is EITHER declared in the script with `shape.feaStudy({ material, fixed, loads, meshSize?, minSafetyFactor? })` OR passed on the call against the shape the script returns: { code: 'return box(100, 20, 10);', material: 'aluminum-6061', fixed: { atX: 0 }, loads: [{ faces: { atX: 100 }, force: [0, 0, -200] }] }. Faces are a FaceQuery ({ atX | atY | atZ, byNormal: 'Z', ofSurfaceType, near: [x,y,z], ... }; with near it names the ONE matching face nearest the point), a face label ('top', 'bottom', 'left', 'right', 'front', 'back'; it must name ONE face — after a boolean 'top' can be several, which is an error listing them) or an @kc[...] ref. Force is the TOTAL [Fx, Fy, Fz] in N on those faces (a scalar force + direction is accepted). Materials: mild-steel, aluminum-6061, pla, petg, abs, nylon (aliases steel, aluminum), or { E, nu, yield } in MPa. Solver (`solver`, default 'auto'): CalculiX + gmsh quadratic tets when that toolchain is installed; otherwise the built-in in-process voxel solver (8-node incompatible-mode bricks), which needs nothing installed and is marked summary.solver 'in-process-voxel' with trust.meshTrusted false — an estimate (accurate for prismatic bending, approximate at curved faces and notches), never a verified pass. solver: 'calculix' fails up front with fea.solver.unavailable and the install command when the toolchain is absent. Also: `study` (a declared study name; default the last), `output_dir` (keeps the deck / summary JSON), `mesh_size` (mm…

Search the public kernelCAD asset catalogue of published models. This is not find_part (screws, bearings, NEMA). Read-only. Anonymous callers are allowed. Do not send an API key. Use source:"library" for the free built-in model library. Search its titles, tags or kit IDs, and download editable .kcad.ts from downloads.source. Library slugs are kit/id. For community projects use source:"community". With no source, an empty or unavailable community search falls back to the library for supported filters. Library supports q, category (kit or manufacturing), tags, license, triangle bounds, kit, limit and cursor. Its default order is kit/id; use sort:"triangles" for descending complexity. Rig, IK, latest and remixed are community metadata. Each community row has slug, title, license (SPDX, always present), category, tags, kit, stats, has_rig, has_ik, ik_type, remix_count and thumbnail_url. UNITS: stats.size_m is {x, y, z} in metres. Y is up. stats.origin_y is the model origin height in metres. stats.triangles is the triangle count. RIG: has_rig means the asset has named moving parts. Pass rig:true to keep only those. get_asset returns rig as { <name>: { type: "rotate" or "slide", axis: "x"|"y"|"z" (local), range: [min, max] (degrees for rotate, metres for slide), node } }. has_ik / ik_type mark an inverse-kinematics target. ik is "any", "arm" or "legs". FILTERS, all optional: q (title, tags and description), category, tags (every listed tag must match), license (SPDX id), min_triangles, max_triangles, rig (true = moving parts only), ik ("any"|"arm"|"legs"), kit, sort ("latest"|"triangles"|"remixed"), cursor (next_cursor from the previous page), limit (1-100, default 24). Only public, gallery-listed assets with a redistributable licence are returned. Use get_asset(slug) for params, presets and the rig. Use get_asset_pack(slug) for a short-lived signed zip URL.

Use this when you need to upload a .gcode file (e.g. written by export with target: "model", format: "gcode") to a real network printer and, by default, start the print. protocol: 'octoprint' (POST /api/files/local with an X-Api-Key), 'moonraker' (Klipper's POST /server/files/upload), or 'bambu-lan' (Bambu Lab LAN-mode: FTPS implicit-TLS upload on port 990 as user 'bblp' with the printer's LAN access code, then an MQTT print-start command on port 8883 — requires access_code and, unless start_print is false, serial). bambu-lan uploads a .gcode.3mf print project (the printer prints 3MF projects, not bare G-code): the G-code goes in as Metadata/plate_1.gcode, inside the model 3MF from export format '3mf' when model_3mf_path is given (use options { arrange: 'plate', slicer: 'bambu' } so the printer shows the same named, coloured parts), else inside a minimal 3MF shell; the file name always ends in .3mf ('part.gcode' -> 'part.gcode.3mf'). Optional printer (a bundled profile id, the same ids as the export 3mf/gcode options.printer) is checked against protocol before anything is read or sent: an unknown id is refused with the list of valid ids, 'bambu-lan' needs a Bambu Lab profile and a Bambu Lab profile needs 'bambu-lan'. Pass { dry_run: true } to validate connectivity/authentication only, without uploading or starting a print. Never logs or echoes api_key/access_code.

Use this when you need to edit a param() default value in a kernelCAD script. Returns the modified code as text plus diagnostics from re-evaluating the result. Caller persists the new code via standard file-write tools (this tool has no side effects). With preview: true it builds the edit in a scratch session instead and returns the diff_geometry summary plus the requirements-ledger and S1 binding diffs (which requirements flip, which margins shrink, which selectors moved or re-bound), without touching the active session, source, revision or caches. A preview whose revised build fails returns ok:false and no diff. Inspect after.requirementsOk to see whether the edit preserves declared intent; preview.ok only says the preview could be computed.

Use this when you need to set how the script returns its assembly. Replace the final top-level return statement with `return <assembly>.model();` or `return <assembly>.solvedModel(poses, options?);`. Use solvedModel for mate-authored mechanisms so FK and validation run. Returns modified source plus diagnostics from re-evaluation.

Use this when you need to replace the shared local Studio face/edge selection, or clear it with selection:null. The viewer highlights the pick on its next sync (300 ms). Does not edit geometry. Pass model_generation from get_selection; writes from an older generation are rejected. Use only current shape indices. Requires a local Studio session.

Use this when you need to solve the mate graph and get part poses. Run the v0.6 mate-graph solver on the active assembly. Returns { status, poses, iterations? } where each pose is a serialized Transform ({ translation, rotateAxis, rotateDeg }). Optional poses overrides mate pose values by mate name.

Use this when you need to solve a 2D sketch constraint set. Solve a 2D sketch constraint set. Side-effect-free: pass { entities, constraints } and receive solved entities plus the original constraints. Entities are POINT, LINE, and CIRCLE records; constraints use the kernelCAD constraint vocabulary.

Use this when you need to know where and why a design breaks across a tolerance/dimension range, not just whether it builds at one nominal value. Declares one or more param() names with a { values: [...] } list or a { min, max, steps } range (number params are bisected; choice/string params are evaluated as discrete points), re-evaluates the script once per cartesian-product combination (capped at 64 combos — exceeding it truncates to the first 64 and emits kinematic.sweep-tolerance.combo-cap-exceeded), and measures every requirement in the S2 ledger (script require(), dfmSpec, gated FEA, mechanical intent) plus the standard gates on each combo: interference, mounting-hole diameter agreement, and joint-axis binding (all three default on); reachability only when gates.reachable names a tip_link + target. After the coarse grid it bisects each numeric pass/fail transition down to boundaryTolerance and reports the boundary value, the failing side, and the binding requirement/gate names (e.g. "wall fails below 1.840 mm; binding: wall"). Returns per-combo verdicts, firstFailure per gate, firstRequirementFailure per requirement, boundaries[], and a coverage statement that says "checked at N points plus boundaries", never "proven".

Use this when you need to trace a reference image into normalized waypoints. Trace pixel-space features from a reference photo into normalized [0..1] waypoints the agent can map to mm via a known scale anchor and feed to path().spline / path().nurbsSegment. Three backends are dispatched behind the scenes: `opencv` (deterministic; uniform-bg silhouette only), `vision-llm` (vision model; named points/cluttered backgrounds; needs a configured vision endpoint), and `hybrid` (opencv silhouette + LLM-labeled named points). Default backend is `auto`. Per-feature confidence is reported. Pair with the `kernelcad-trace-from-image` skill for the conversion-to-mm pipeline. Hosted: give the file as an https:// URL, a data: URI (base64, max 25 MB) or "asset:<id>" from request_upload (when the file is on the user's device).

Use this when you need to check a design against a rule set. One verifier, selected by `check`: - 'assembly' — static assembly review ({ file | code }, or the session evaluate_script left): mate validator, default-pose interference, joint topology, mechanical plausibility. ok:false with `blocking[]` exactly when review_cad's blocking rules fire on these findings; review_cad also sweeps joint ranges and runs physics. - 'urdf' — structural validity of a .urdf file ({ urdf_path }). - 'dfm' — print-readiness gates declared by dfmSpec() ({ file | code }). - 'dfm-preflight' — sheet-metal flat pattern vs a job-shop's ordering rules ({ vendor, material, thicknessIn|thicknessMm, ... }). Returns the vendor's catalog source pages as `sources`; when `disclosure` is present, show it with those links. - 'assembly-path' — continuous straight-line removal of each assembly part along ±X/±Y/±Z and optional world-space directions. Returns valid directions, blocking parts and one of three states (removable / blocked / unknown); sequence peels the directional blocking graph and reverses it into a build sequence with a model stamp. Fastened-mate and fixed-connect clusters become subassemblies: each is planned recursively inside and installed as one unit at the top level (`Step.subassembly`). Remaining valid orders use deterministic heuristics (subassemblies first, outside-in, fewest reorientations, fewest tool changes); stuck sets get a budgeted blocking-reduction partition and a sampling fallback whose paths are labelled found-by-search, never proven. A fastener with a fastened-mate `fastener` declaration (axis, pitch, optional tool envelope) is tested along its helical path and reported `motion: helical`; a press/snap `fit` declaration becomes a `requires-force` step, never simulated. A modelled driver/socket/hex-key envelope swept along the approach axis adds `toolBlocked` { fastener, blockers } and a failing S2 ledger entry; unmodelled tools add nothing. Threaded fasteners without m…

Use this when you need to trace why a feature failed. Walk the upstream chain of a failing feature. Returns the diagnostics of the requested feature plus the diagnostics of every upstream feature in topological order (the requested feature is the last entry). Per-code hints are inline on every diagnostic — call lookup_diagnostics for the full catalogue. Pass { file?, code?, feature_id? }.