The before-manifest was taken pre-bake while save mints *D geometry on its clone, so every new dimension failed verification 1-vs-7. Compute before on a bake-simulated clone: baked-vs-baked proves roundtrip stability (bake correctness itself is pinned by the explode geometry tests). Pinned by save_verified_accepts_new_dimension_with_baked_block.
The DXF reader keeps marker identities on the block record but never materializes the entities, so block_entity_handle/block_end_handle dangle after every DXF load (reopened baked *D dimension blocks hit this every time). Re-create missing sentinels at load finalize, reusing the file handle when free and minting otherwise. Idempotent and DWG-safe: loads whose markers already resolve are untouched. Pinned by dimlinear_survives_save_reload_with_baked_geometry.
resources/* vary as sessions open and snapshots land, so an hour of advertised freshness was optimistic. protocol_result now takes the TTL value: tools/list and discovery keep 3600000, resources/list/read/templates get 60000, on every protocol version. No tool surface change; digest untouched.
The GUI requires request_id and document_id on every non-query op, but the bridge sent neither on capture: every ocs_capture failed with request_id_required, then document_required. request() now always mints the GUI request id for capture (a client-supplied id wins and is echoed in metadata), and capture carries the active document. The op=capture read path answers guidance pointing at ocs_capture instead of a missing-path failure. Live-proven against the open window: plain and annotated captures return image content.
The cached PID snapshot predates a GUI spawn, so the newborn pid missed from it read dead and the bridge deleted its just-written descriptor inside the launching call itself: every fast launch reported still-starting forever while a healthy window sat undiscoverable. Two guards: the snapshot cache is invalidated right after spawning, and deletion now waits out a 120s file-age grace (unknown age counts as newborn). Slow starters that write late were the only ones surviving before, which is why this was intermittent. Covered by stale_snapshot_may_delete predicate tests.
Runs the official MCP Inspector against a fresh release build on every push and PR: initialize purity, four titled tools, headless resources/list, both resource templates. Headless-safe methods only; drawing-dependent coverage stays in unit tests and the pinned digest. Server addressed via config file (positional CLI targets break on Windows shells); TTL assertions stay at unit level (the CLI normalizes result envelopes on output).
A dispatched rebuild checks out the tag's commit but runs main's
workflow; v2026.40 has no stage-bundled.sh, so the macOS bundle step
failed. Skip the plugin with a warning in that case only.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The wasm check broke on the MCP `tools` op, the block library's folder
picker and DWG thumbnails, and the PDF export-window culling helpers.
Gate each one like the code around it; `read_handle` needs no gate.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Everything was written and only clipped at view time, so a window over a
corner of a large drawing produced a near-empty page of several MB.
Wires and hatches whose sheet bounds miss the window are now skipped.
Fixes#1568
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A dimension draws from its *D block by name, but loading an xref skipped
anonymous blocks and kept the name verbatim, so its dimensions fell back
to a regeneration (giant text at the origin) or a host block of the same
name. The load path now imports them prefixed like other blocks and
renames the dimension's block; BIND is unchanged.
Fixes#1578
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Splitting a boundary object erased it and added the pieces, so the hatch
lost that source and rebuilt from what was left, often the whole circle.
The pieces now take over the old object's association before it is
erased.
Fixes#1572
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A point carried a key vertex, so Endpoint (which outranks Node) claimed
it, and an absolute-size glyph's strokes were indexed for Nearest and
Intersection. Points now offer only their Node, in the indexed and the
unindexed snap paths alike.
Fixes#1570
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Each band segment was a rectangle ending square at its vertex, so the
outer corner of every joint showed a notch. Every segment end now carries
the tangent of half its turn, and the shader slides that end's corners
onto the joint's bisector so neighbours meet in a miter. Turns sharper
than the limit keep butt ends. Constant-width runs inside a polyline split
around its arcs no longer count as tapered, which had kept them unmitred.
Fixes#1565
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The codec writes R2018 AuxHeader maintenance fields as RL, so AutoCAD no
longer reads a 2018 save as student-made. The kernel drops zero-area
spikes before triangulating, so a hatch whose boundary folds back on
itself fills on screen instead of vanishing.
Fixes#1564Fixes#1569
Co-Authored-By: Claude Opus 5.5 <[email protected]>
rustc 1.99 overflows its default 8 MB stack compiling the crate in
release on Windows and macOS. Raise RUST_MIN_STACK to 64 MB in the cargo
config and the release workflow (the latter also covers rebuilding older
commits). The Windows Build step now runs in bash so a failed cargo build
stops the job instead of surfacing later as a missing exe at signing.
workflow_dispatch gains tag/commit inputs to rebuild a published release.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Short interactive waits returned a bogus 10-minute timeout on first expiry, killing the re-poll pattern and blocking task creation. The dismiss branch now fires only when the requested wait actually reaches the ceiling; shorter waits return running as before. Predicate covered by interactive_timeout asserts.
tasks/cancel marks the task so polls report terminal cancelled with neither result nor error; late completions are discarded. Task envelope errors move from grandfathered -32000 to -32603. Agent schema sync on bridge start is now skippable via OCS_SKIP_SCHEMA_SYNC and logs one line when it writes.
resources/templates/list serves the two genuine parameterized URIs (pyramid tiles, snapshots) with TTL hints instead of -32601. Tool titles move top-level per the 2026-07-28 Tool shape (dropped from annotations); annotation keys are allowlist-tested. resources/list ids are sorted so client and prompt caches hit. Tool schema digest rolls to fbba987a; pin updated deliberately after reviewing the titles-only surface diff.
How to iterate the bridge: fast builds, digest pin, session model, discovery, launch serialization, cancellation, error routing, safety model, rebuild swaps, conformance gates. No diary entries, no harness names, no build-specific data.
run() reads stdin on a feeder thread into a channel (CancelPump); wait loops harvest lines in 50ms quanta instead of sleeping blind. A notifications/cancelled naming the in-flight MCP id dismisses the GUI prompt with a second cancel exchange, drops the late result, and sends nothing back. Cancels arriving between polls resolve via the inflight map; re-polls for dismissed ops answer cancelled without touching the GUI. Interactive picks (user_select/getpoint) get a 10-minute single-call ceiling ending in isError timeout text; other ops keep the 60s clamp. Ping-during-block stays unanswered while one exchange occupies the loop (bounded residual).
Unknown tools become -32602 Protocol Errors carrying the four valid names (SEP-1303 keeps them there; the isError shape hid them from strict clients). Bridge/GUI infrastructure failures (no sessions, dead TCP, unwindable startup) become -32603 via the CallError split; tool-domain validation stays isError so the model keeps self-correcting. Verified observable in a live client session: bogus session surfaces readable -32603 naming ocs_sessions, model recovers.
resources/read reports -32602 on the modern path (2026-07-28 forbids emitting -32002) and keeps -32002 on legacy, where it was the code. Non-object top-level input (batches included - no batch semantics defined) gets -32600 Invalid Request instead of silence, so senders never hang. Id-less objects stay silent (notifications). Ping arm deliberately untouched: the method was removed only on the modern path, and legacy health checks may rely on it.
serverInfo carries only name/title/version so strict clients never reject the handshake. Build extras (build_rev, build_profile, tool_schema) live canonically in bridge_identity, surfaced via capabilities.experimental and the in-band bridge stamp - no copies to drift. Adds conformance guards: spec-pure initialize with unknown-version fallback, and JSON-RPC errors for unknown methods, missing tool names and missing resource URIs.
Global MCP instructions don't reliably reach agents (fresh sessions never announced the build). The directive now lives in the ocs_sessions description every agent reads first, pointing at the bridge object. Tool schema digest rolls to ede58332; pin updated deliberately after reviewing the description-only diff.
List results (tools/list, resources/list, resources/read) carry ttlMs and cacheScope on the legacy 2025-11-25 path too; legacy clients ignore unknown fields per the SEP's backward-compat rule. The modern envelope (resultType, _meta.serverInfo) stays negotiated. Legacy-shape pin updated deliberately to assert presence.
Discovery skips pid-verified dead sessions and deletes stale descriptors; probes are bounded so dead sessions delay discovery by at most one timeout. The bridge exits after the in-flight request once it detects a rebuild, and error results omit structuredContent so strict clients surface real errors. Concurrent launches are serialized with automation/starting.lock: the first caller spawns the GUI, later callers wait for the same descriptor instead of opening another window. Session states and hello/capabilities responses now carry the bridge build identity (version, build_rev, tool_schema), and the agent instructions point the build announcement at it.
_LINE / .LINE reached show_command with the prefix and matched nothing;
strip it the way the Start-tab gate already does. The before/after
screenshots only illustrate the PR and stay on its branch.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
F12, the drafting dialog and restored settings each re-derived DYNMODE and
lost the mode on the way: F12 off set 0, so on came back as 3 even after
DYNMODE 1. set_dyn_input now makes off negative (temporarily off) and on
restore the stored mode, and the bare query prints the stored value.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
ensure_layer had started trimming the name, so an entity on " A" got a
table entry "A" it never resolves to. Register the name as given again
and drop the unused return value. The five copies of the dirty-flag check
become sync_registered_layers.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Fix joined words (livelloiselezionati, livelloeliminato), typos (Gstione,
formatazione), lowercase prompt starts, leftover English (object(s),
selected), and meanings that drifted from the source: SLICE plane miss,
named range not found, worksheets, extractable objects, data link name,
count column, pointer input, solid, Stretch verb form, and the locked
layer message mixing layer/livello.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A full-strength fill on a large tab header read louder than the selected
tab's outline. A faint wash of the same blue still points at the tab
without looking selected.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
A command started from the command line, an alias, a shortcut or Repeat
now lights the same ribbon button a click on it would, so users learn
where the tools they type live. A button inside a dropdown lights the
dropdown, a slide-out tool lights its group title, a tool hidden in a
collapsed panel lights the panel's opener, and a button on another tab
lights that tab's header. Always on, in the existing active-button blue.
Buttons are matched by the command they dispatch, so ids that differ
from their command still resolve; the tab on show wins when a command
appears twice, and a clicked button keeps its highlight. Transparent
commands leave the running command lit. The #355 rule that a highlight
lasts only while something runs moves from the ribbon click handler into
dispatch_command, so typed commands are cleared the same way.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
With objects already selected, OFFSET opens at the side step with the
preview following the cursor instead of asking for a distance first
(#422). At the side step a typed distance places the offset at once on
the cursor's side, and a negative one on the other side; Enter at the
start repeats the last distance (#418). Erase and Layer, which lived on
the skipped distance step, are offered there too. With nothing selected
the distance-first flow is unchanged.
The opposite side is found by mirroring the cursor through its nearest
point on the outline, capped at half the offset distance so it cannot
cross a narrow gap back to the cursor's own side. Two new prompts are
added to the catalog and every locale.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
#1590 cloned the whole document on every plot to evaluate plot-time
fields, even in drawings without a single FIELD; a large drawing paid a
full copy per plot. The copy is made only when the drawing holds fields.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
The PR pinned codec and kernel to its fork through [patch]; codec #77 and
kernel #73 are merged upstream, so the patch sections are replaced by
the upstream pins (codec 8d23408, kernel d22a270, graph 9b5ae39) in this
same commit: the kernel now applies the body transform during the lift,
and an OCS without this PR's change would apply it twice. The new codec
also brings AssocCurveValue, registered with the plugin API's type
registry, and EmbeddedEntity::Body, named as a polyline profile.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
