Tool count 131 -> 154; 474 -> 1242 collected tests.
The release exists for two measured reasons. A 149-tool catalog costs a client
~40,300 tokens before it has asked for anything, and a drafter searching the way
drafters think -- FILLET, BPOLY, QSELECT, WBLOCK -- scored df=0 against the stock
index, because those words were simply not in the data.
Discovery and batching
- DISCOVERY_MODE=search replaces the catalog with search_tools + call_tool,
ranked over an AutoCAD command and synonym corpus (discovery/aliases.py) on
top of fastmcp's BM25 index. Measured: 40,305 -> 356 tokens advertised.
- cad_batch runs a step list in one round trip; fields= projects result-heavy
responses. The `core` profile is gone -- discovery replaced it.
Honesty debt (Tier 0)
- OCS/ECS: `extrusion` appeared nowhere in this codebase, so mirrored and
rotated geometry reported coordinates that were quietly wrong. backends/ocs.py
translates at the backend boundary; every tool coordinate is WCS on both
engines, and a tilted plane refuses rather than answering with a number that
looks fine.
- drawing_save_as(fmt="dwg") wrote DXF bytes to a .dwg path. drawing_audit never
repaired. analysis_measure_area shoelaced points the caller typed in.
- ezdxf calls gained a per-call timeout, so one hung call can no longer wedge a
server whose document lock is a single asyncio.Lock.
Architecture
- The 1854-line backend ABC is 19 contract modules; @capability(key, reason=)
makes optional methods declarative and a gate fails on undeclared holes.
- 8 layout/viewport tools, CHSPACE, CAD_PROGID escape hatch.
- Wave A: selection filters, boundary tracing, hatch depth, annotation objects,
measurement -- 13 tools of a planned 39. The REGION and 2D-boolean family was
cut on evidence: add_region() yields zero ACIS bytes and the greiner_hormann
substitute loses 28.2% of the area on a square with one semicircular edge.
Verified against a live AutoCAD 2026, which found four COM defects no amount of
reading would have caught -- AcadPViewport has no ViewCenter, and the line that
set it had been swallowed by a bare except since v1.4.
Benchmark v3 (15 tasks) and what it caught
- v2 scored this server 10/10, which carried no information: every task in it
exercised something the server was built around. Five tasks were added because
they can fail, and three did while being written -- a hatch could not be
measured though the refusal listed HATCH as measurable, the boundary tools and
the measure tool disagreed about the same shape (139.2177 vs 139.2699), and a
circle stored as two bulged vertices measured 0.0.
- Competitor reports are NOT re-scored. They are pinned v1.4 runs against v2 and
were never asked these five questions; the chart shows those rows `not run`.
- Correctness A/B: 21/24 -> 24/24, zero regressions.
Performance, stated plainly
The published v1.4 perf report was recorded on CPython 3.14 and the new one on
3.11.15, so the apparent 2-6x speedup is the interpreter, not this release.
Re-measured on one interpreter, v1.5.0 is 28% slower creating entities and 22%
slower on the 10k roundtrip. Fully attributed: EZDXF_CALL_TIMEOUT=0 restores
v1.4.0's exact numbers. That is the deadlock fix being paid for, the knob is
documented, and winning it back is on the 1.6 roadmap.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Leaderboard restyled to the LLM-benchmark look: dark background, rank-numbered labels, per-rank color palette, 'NN / 100 - grade' annotations. The live-run, task-matrix and performance charts share the same dark theme for a consistent README identity. Charts regenerated deterministically from their data files.
Co-authored-by: Cursor <cursoragent@cursor.com>
source_review.json extended with three projects reviewed 2026-07-23 (LokmenoWer/best-cad-mcp B/74, daobataotie/CAD-MCP B/35, zh19980811/Easy-MCP-AutoCad C/30) and our entry updated to the released v1.4.0 (95: CI matrix, PyPI packaging, A/B gate, paper space, ISO 286, opt-in solids). Chart regenerated via render_chart; README benchmark section now opens with the leaderboard + 9-row table, source-review lane merged into it.
Co-authored-by: Cursor <cursoragent@cursor.com>
Ran benchmarks/compare_versions.py against the v1.3.0 tag: 21/21 on both versions, 0 fixed / 0 regressed - the v1.4.0 feature+infra work introduced no correctness regressions. Main README gains lane 4 (version A/B, with the historical v1.0.0 -> v1.1.0 contrast); benchmarks/README records the release-gate result and drops the stale 'performance coming' stub.
Co-authored-by: Cursor <cursoragent@cursor.com>
ruff format was never run repo-wide; CI now enforces 'ruff format --check', so normalize every module once. No behavior changes.
Co-authored-by: Cursor <cursoragent@cursor.com>
- benchmarks/correctness_suite.py: 21 deterministic headless (ezdxf) checks
across core ops, dimensions, arrays, geometry, query/parity, render, the
quality gate, and engineering.
- benchmarks/compare_versions.py: one-command A/B runner — checks out an older
ref into a throwaway worktree and runs the same suite against both, each check
in its own subprocess so a crash is a miss, not a run-killer.
- benchmarks/README.md: methodology, results table, honesty caveats.
- README: Benchmarks section now reports the measured result —
v1.0.0 8/21 (38.1%) -> v1.1.0 21/21 (100%); 13 fixed, 0 regressions.
