* claim: #5822
* Batch: consecutive parameter edits can rebuild the design once (#5822)
An agent's batch of parameter edits, one parameter per step, evaluated the
design once per edit: an edit early in the timeline rebuilds nearly all of
it, and 21 such edits ran for minutes.
engine.batch and the MCP batch tool take coalesce_parameters (default
false). With it, each run of consecutive plain parameter edits
(parameters.change with only a name and an expression, or with changes) is
set as one parameters.change {changes}: one rebuild for the run. Any other
step ends a run, so later steps see the design up to date. Every step
still gets its own result (coalesced: the run's length); the run is one
undo step. A run whose combined edit is refused (a new parameter, which
changes can't make; a bad expression) runs step by step, so a refused edit
is still named by its own step. The app's off-thread evaluation computes
the run the same way. The grouping is solidcraft_engine::batch_params,
shared by both runners.
21 parameter edits on a plate with 21 extruded bosses (in-process MCP,
ci profile): 40.1 s step by step, 3.5 s coalesced, the same volume.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Batch coalescing: say a run skips the states between its edits; test it against step by step (#5822)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* claim: #4099
* Measure: a face or edge picked by its point alone (#4099)
inspect.measure (and the MCP measure tool) refused a face or edge item
without an index ("missing field `index`"). Face and edge items are now
read as body + point with an optional index: the point finds the face
(the same lookup face picking uses) or the nearest edge within the pick
tolerance, before the selection is built. A missing body or a point on
no face or edge is a clear error. Explicit indices still work.
Rebased onto main as one commit (was 120d26f7, df488f9a, db9d1073 plus
merges of main).
- A request the app reports as still running (#1618) is waited for on the
same connection through `control.pending`; past the wait the error says it
was not cancelled and will still apply (or only write its file), later calls
are not sent behind it, and the new read-only `pending` tool returns its
result once it lands.
- A failed batch says what really happened: `rolled_back` only when the
design is back where the batch started (as #3288 / #3242 on main: the app's
`rolled_back` is believed; for an older app, every completed step undone;
only a rollback asked for is reported as one), plus `applied`, `not_run` and
`failed.still_running`.
- A batch step may name a tool (set_parameter, list_commands,
describe_command, execute, measure, inspect_design, undo, redo, export, save,
open, new_design, crash_reports); screenshot, body_topology, batch and
pending are refused before anything is sent.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* docs(mcp): Windows setup for Claude Desktop (paths, two servers, Store config file)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs(mcp): say the Store build can keep the config elsewhere, not that the other path fails
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The app opens its localhost control channel on port 27878 (or the next free
one) unless Preferences > Assistant turns it off; --control and
SOLIDCRAFT_CONTROL_PORT still win. A current.port discovery file lets
solidcraft-cli mcp find the running app with no arguments.
Co-authored-by: Claude Sonnet 5.5 <noreply@anthropic.com>
* claim: #4966
* Agent chat: serve the bundled skills' reference through an MCP tool (#4966)
The bundled skills start with "read reference.md", but SolidCraft's own
chat turns the agent's file tools off, so its agent could not follow them.
The MCP server gains a read-only `skill_reference` tool that returns a
skill's documents from copies built into the binary. It takes a skill
name and a file name from a fixed list, never a path, and refuses an
unknown skill or file with the available ones. The skills point to it,
and docs/mcp.md and skills/README.md describe it.
* claim: #3654
* Control channel: control.running lists what the app is running, across connections (#3654)
control.pending answers only for the connection that asks, so a second client,
the MCP bridge after a reconnect or a script saw {"pending": null} while the app
was busy. control.running answers for the whole app, the same on every
connection and at once (the connection's own thread answers it, also while it
is busy with a timed-out request): every request the app has been sent and not
answered yet, oldest first, with its connection, id, method, command id (or a
batch's step count), seconds since it arrived and whether it still waits its
turn. A timed-out request stays listed until the app answers it, also after its
connection closed. Parameters are never shown.
The window marks a request started when it takes it (ControlRequest.started),
which tells queued from running. The MCP server gets a read-only `running` tool.
---------
Co-authored-by: MateuszSychXpl <151755914+MateuszSychXpl@users.noreply.github.com>
* joint.inspect: what each joint is attached to, for agents
A read-only `joint.inspect {joint?}` gives each joint's definition: for each
side the occurrence, its component and path, the snap as saved, and what
each persistent face/edge name refers to (the feature that made it, and the
sketch and curve a swept side came from), whether the origin resolves and
why not; plus kind, offset, angle, flip, values with limits, expressions,
the motion links it is in, and its solve errors and conflicts. Its id ends
in `.inspect`, so agent gates treat it as reading only.
MCP `inspect_design` includes the definitions when the design has joints
(`joints: false` leaves them out). joint.list and joint.inspect share the
motion-link JSON.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* i18n: catalog entries for Inspect Joint (en, ja, zh)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Spike: typed parameter contracts and JSON Schemas for solid.box and solid.extrude
The schema RFC's spike (plan/schema-spike.md). solid.box and solid.extrude get
typed contracts with generated JSON Schemas (2020-12), checked in front of the
unchanged handlers by a per-session level: off (the default, which changes
nothing), warn (the call runs and gains param_warnings) or strict (the call is
refused and the design left alone). Also engine.describe and the MCP tool
describe_command, `solidcraft-cli mcp --param-check off|warn|strict`, a
FeatureResult type for feature replies, and an env-var audit recorder
(SOLIDCRAFT_PARAM_AUDIT) used to check the contracts against every test.
Local spike branch; not for upstream as is (see the results report).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* Parameter contracts: put the audit recorder behind a feature; drop two ignored test keys
The audit recorder is now on only in a build with the engine's `param-audit` feature
and `SOLIDCRAFT_PARAM_AUDIT` set. Its code is still compiled, linted and tested without
the feature. `scripts/param_audit.py` summarises a test run's audit logs, and the schema
module's docs give the steps for converting a further command.
Two tests passed keys their command ignores, which the audit found: `position` on
`solid.box` (`hole_face_tests.rs`) and `symmetric` on `solid.extrude` (`face_tests.rs`).
Both are removed, so the tests build what they always did. Comments no longer call
the contracts a spike.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
Measure now reads the picked item's exact shape (cylindrical faces, circular and arc edges,
sketch circles, arcs and lines, profiles, planes, points):
- A circle, hole or arc shows Ø, R and its length (circumference, or arc length and angle).
- A slot (a sketch profile, or a half-cylinder end of a slot in a solid) shows its width,
centre-to-centre length and overall length.
- Two items where one is round measure from the centre or axis by default (axis to axis,
axis to plane, edge or point), with the closest and farthest distances listed too; the
panel switches between centre, min and max, and the canvas line follows.
- inspect.measure and MCP `measure` (new `items` argument) add diameter, radius,
center_distance, min and max; existing fields are unchanged.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Fixes#1167. While a sketch is edited, a body that a rolled-off, unsuppressed Remove deletes is not shown (unless the sketch is built on it); the history stays rolled back, only the view changes. Drawing, picking, screenshots, thumbnails and the browser's Bodies folder agree; document.inspect reports removed_later. Checked on main + this branch: cargo check --workspace, the engine sketch_edit_removed/occurrences tests and the UI viewport/browser tests.
solidcraft-cli snapshot takes --camera JSON and --view with a saved view name (through view.apply, as the MCP screenshot does), plus --no-grid and --no-axes; the MCP screenshot and ui.render take grid/axes, the headless render skips a hidden origin axis, and the app's ui.render follows the Grid and Origin switches. Closes#1245.
file.insert_step and doc.open on a STEP file take `up` (turn a Y-up file so its up is +Z) and `names` (names for the imported bodies). The MCP tool and the app's control requests pass them through; an import without them works as before. Tests in the kernel's STEP reader, the engine, MCP and the file-load UI; docs/step-import.md and docs/mcp.md describe both.
A solidcraft-mechanisms agent skill with a worked example that a test replays through the MCP server (the actuator follows the hinge with the stated stroke and the drive stops at the rod's limit); the skills index and docs/mcp.md list it.
The bridged MCP server starts when no app is listening, says so in its instructions and in each tool error, and connects on the first call after the app starts. A batch that could not be sent says the app is not running. docs/mcp.md describes it.
A design note for SolidCraft in the browser: tiered storage, a SharedWorker job broker, GPU dispatch and in-browser agent control through the same engine commands, with the generic pieces in craft-libs. Linked from the MCP and web getting-started docs.
Lands the Render workspace (R3), bounded preview-image jobs with cancellation and atomic PNG export (R4a), and opaque PBR looks as one change.
- Render workspace: an overlay-free product preview on the installed GPU device. Returning to Design restores its camera. render.scene and render.settings.apply are added.
- Image jobs: render.preflight/start/status/cancel/export run bounded CPU preview jobs. Export captures the job, overwrite choice and document, and checks the caller's file scope before writing.
- Opaque PBR: looks can carry metal fraction, roughness and dielectric IOR. The product viewport and preview PNG use them; geometry and mass are unchanged. The v1 catalog export refuses PBR rather than dropping it.
- Render setups and PBR looks are saved under design format 21. Labels are in en/de/ja/ko/zh/ru.
Supersedes #706 and #800. Refs #620, #673, #756, #1004.
Contributed by @0xlillith.
Co-authored-by: lillith <35337584+0xlillith@users.noreply.github.com>
A design carries metadata (format 21): project, part number, revision, author, organization, description and license, plus other tools' data under reverse-DNS keys, kept as written. Edited in the Design Properties window (Document Settings), set by document.metadata.set, reported by inspect_design, and written into STEP, IGES and 3MF where the format has a place for them. Translated, with tests.
PathScope::check reports a missing folder inside the scope as NotFound with the nearest existing folder, instead of calling it outside the allowed folders. Paths outside the scope and dangling links are still refused. Fixes#1077.
doc.open and the STEP/mesh imports run on the file loader's worker with the request's path scope, and a font named by file path in a design is checked against the scope before it is read.
On Windows the temporary folder often has an 8.3 short name, which the path scope refuses. vfs::long_names spells folders SolidCraft derives from the environment with long names, and the MCP scratch folder and the recovery folder use it, so screenshot without a path and the recovery commands work there again.
docs/mcp.md gains a "Camera and close-up recipe" section: the orbit camera's conventions (radians, eye direction for yaw and pitch, orthographic half-height, which side inspect.section keeps) and a seven-step recipe for a close-up of one region of a large STEP model. Documentation only.
A design with components and joints exports as a URDF robot description (file.export to .urdf, the MCP export tool, solidcraft-cli run --out, File > Export URDF...). It writes one binary STL per component in metres, links with mass, centre of mass and inertia from material density, and maps rigid, revolute, slider, cylindrical, pin-slot, planar and ball joints to URDF joints or short chains. Loops, broken joints and loose parts are reported as warnings. Options: mesh_prefix, effort, velocity. Docs in docs/urdf.md.
The app's control protocol gets engine.batch: the steps run in one request (so one frame), each as engine.execute runs it, stopping at the first failure and rolling back to the checkpoint by default. The MCP batch tool sends it to a running app (with the control token), is never resent after a lost reply, and falls back to step-by-step when the app refuses (agent cursor shown, or a step loads a file). A batch may take up to an hour to answer.
solidcraft --control PORT writes a random per-launch token to a user-only file (<data>/control/PORT.token; 0700/0600 on Unix, icacls on Windows; SOLIDCRAFT_CONTROL_TOKEN overrides) and refuses any connection whose first request lacks it. mcp --connect, the demo and packaging smoke clients send it. MCP execute/batch/open/save/export/screenshot are annotated destructive and open-world, new_design destructive. Docs: control-protocol.md, mcp.md, AGENTS.md, pending release note in releasing.md.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Agent tool calls (MCP, and bridged control requests carrying allow_paths) reach files only in the open design's folder (strictly inside home or the temporary folder, not in a hidden folder, and never its hidden files), the app's recovery folder, the server's private scratch folder and each --allow-path; --allow-any-path lifts the limit. Paths are resolved first, so .. and symlinks that lead out (including a save through a link, #845) are refused. On Windows, UNC, device, NT, verbatim non-drive, drive-relative, alternate-stream, reserved-device and 8.3 short-name paths are refused before any file system call, and the resolved path must be on a local drive. doc.recovery_discard and doc.recover {discard} delete only real recovery entries. Docs: mcp.md, control-protocol.md, README.
MCP execute accepts verbose: false too, using the same compact reply as batch (#599), and the MCP docs gain a section on compact replies.
Co-authored-by: lillith <35337584+0xlillith@users.noreply.github.com>
New view.apply command restores a named view, a standard view or any camera (view state, not an undo step); the desktop viewport follows it and the browser's Restore uses it. The MCP screenshot tool takes a named view or a camera. Un-ignores the #528 regression test. Fixes#528.
Co-authored-by: jonaswre <jonas.wrede@beyondit.gmbh>
MCP batch takes verbose: false for compact replies: feature steps leave out the full body list and sketch steps the ids they repeat outside created. Feature replies now also list changed_bodies and removed_bodies next to created_bodies, so a compact reply still shows every effect of a step. Un-ignores the #530 regression test. Fixes#530.
Co-authored-by: jonaswre <jonas.wrede@beyondit.gmbh>
Fixes#374. Feature replies gain created_bodies: the bodies the new feature made, apart from the ones it changed or that were already there. Documented in the execute tool description and docs/mcp.md.
Every panic (in a command, in the kernel, anywhere else) and every unclean
exit (abort, stack overflow, kill) leaves a report in a local crash folder:
message, file:line, a trimmed backtrace, version, commit, system and renderer,
the command and recent commands, and when possible the design and replay.json
that reproduce it (solvecraft-cli crashes replay <id>).
- Cost: nothing until the app installs the hook (tests, CLI and web never
do); then one small journal line per top-level command of the design on
screen, taken back with a truncate when the command changed nothing.
- Unclean exits: each run holds a lock on its journal; a journal nobody holds
at the next start belongs to an app that died. Autosaves restart the
journal from the autosaved design.
- Bounded: a crash's signature (place + message, numbers ignored) is counted,
not rewritten; at most 5 new reports per run, 20 reports, 50 MB, 90 days;
recovered kernel panics are pruned first.
- Agents: diagnostics.crash_reports / crash_report / crash_reports.clear
commands, the MCP crash_reports tool, solvecraft-cli crashes, solvecraft
--crash-log; diagnostics.crash_test (with SOLVECRAFT_CRASH_TEST) checks it.
- People: the Crash Reports window opens after a crash (Help > Crash
Reports…) with a reviewed issue draft per report; nothing is sent.
- Release builds keep line tables for file:line in backtraces.
Tests: child processes that really panic in a command, in the kernel, on a
thread, and abort; dedup counts, limits, hostile folders, journal parsing,
the UI window and draft, the MCP tool. docs/crash-reports.md.
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: bherbruck <brennen.herbruck@gmail.com>
Params docs name other commands ("usable in sketch.delete"), so filtering by
an id returned 56 commands. The filter now looks only at id, label, tab and
panel; the tool description and docs/mcp.md say so.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RjGw4gA5ZvKrU8MQP6sBgT
DuBearteandClaude Opus 5.5
raykholo
bherbruckandClaude Opus 5.5
SS-ENandClaude Opus 5.5
rbousigues88andClaude Opus 5.5
Mateusz SychandMateuszSychXpl
Irakli Kourtis
Lasse Lian
NOWN-DEV
William Dunn
selune
Igor Churmeev
Kevin Ravensberg
Jonas Wrede
CypherNaugh_0x