autocad-mcp-pro/CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

[1.6.0] — 2026-09-25

The v1.6 roadmap, shipped as five tracks: 154 → 247 registered tools in 28 groups, the correctness suite at 48 / 48 against v1.5.1's 26 / 48 with zero regressions, the 21-task benchmark matrix at 21 / 21, and a live AutoCAD 2026 run of every track's COM paths. Two roadmap criteria are not green, and Not in 1.6 below says which and why: track D (Plant 3D) was dropped, and headless creation throughput is still below v1.4.0's. Roadmap: docs/superpowers/specs/2026-09-15-v1.6-roadmap-design.md.

Track A of 1.6: the server becomes a P&ID expert and a P&ID reader. Nine pid_* tools in a new SECTION 17, two generic contracts every engine implements, six critique focuses that join drawing_finalize, and a TOOL_PACKS gate so a mechanical-only client never pays for the P&ID surface. Spec: docs/superpowers/specs/2026-09-15-v1.6-pid-design.md.

Track E of 1.6: the environment a drawing lives in. 38 tools in three new sections — styles (18), page setup and templates (19), environment (20) — the drawing_settings facade grown, a system-variable catalog, five bundled templates, several open documents headlessly, portable layer states, named views and UCS, and the live-only group (launch, preferences, document properties, operator prompts) refused honestly headlessly. Spec: docs/superpowers/specs/2026-09-16-v1.6-settings-design.md.

Tracks B and G of 1.6: the complete mechanical 2D part drawer and the sheet standard. 25 tools in four new sections — mechanical parts (21), standard parts (22), mechanical annotation (23) and sheet and delivery (24) — one part model and one view engine instead of a zoo of per-part generators, a new mech pack, six mech_* critique focuses and four new contract members (xref_attach, xref_manage, image_attach, drawing_export_dwg). Spec: docs/superpowers/specs/2026-09-22-v1.6-mech-design.md.

Track F of 1.6: architectural 2D, the plan model. 12 tools in a new SECTION 25 — a wall network with hosted doors and windows drawn by one engine (L / T / X junctions resolved, openings cut, poché by material), rooms computed as the faces the walls enclose and read back from any plan, schedules read off the drawing, a furniture and sanitary catalogue, the structural grid and plan symbols, exterior dimension chains and the whole plan in one transaction — a new arch pack, three arch_* critique focuses, and English or Turkish labels. No new contract member. Spec: docs/superpowers/specs/2026-09-23-v1.6-arch-design.md.

Track H of 1.6: understanding drawings somebody else made. Six tools in two new sections — understanding and QA (26: a one-call report, a scale check between a P&ID and its layout, a revision diff, topology checks) and plant takeoffs (27: pipe and cable quantities whose topology comes from the P&ID and whose lengths come from the layout) — one DXF snapshot per drawing (a live document through Document.Export, never renamed), a vocabulary in five languages, three topo_* critique focuses, a new plant pack and the office extra. Read-only by default; nothing invented. No new contract member. Spec: docs/superpowers/specs/2026-09-24-v1.6-understand-design.md.

Added

• Styles — 10 tools, pack core. dimstyle_list/create/modify/set_current, textstyle_list/create/set_current, mleaderstyle_list/create, drawing_apply_standard. ISO-25 and ANSI (metric) presets are authored data pinned by tests/test_standards_presets.py against ISO 129-1 / ASME Y14.2 (DIMTXT 2.5 / 3.0, DIMASZ 2.5 / 3.0, DIMEXO 0.625 / 1.5, DIMDSEP , / ., DIMTAD 1 / 0 …). dimstyle_modify accepts the 33-name whitelist only (the 17 preset variables plus 16 more — DIM_VARIABLE_WHITELIST) and refuses an out-of-range value by name before writing; changed lists what moved. A font neither bundled nor on the engine's font path is written and reported font_resolved: false, never refused. mleaderstyle_create works on both engines: ActiveX exposes no MLeaderStyle collection, but the ACAD_MLEADERSTYLE dictionary holds full IAcadMLeaderStyle objects, so the live engine creates the style through AddObject and the property writes — no mleaderstyle capability key exists.

• Page setup & templates — 6 tools, pack core. page_setup_list/apply (PAPER_SIZES ISO A0–A4 / ANSI A–E, CTB_CATALOG, SCALES), plot_style_list (catalog + installed), batch_plot — mediabox_mm parsed from each PDF's /MediaBox, so ISO_A3 landscape is 420 × 297 because the file says so — drawing_template_list, drawing_template_save (.dwt refused headlessly, capability dwt_write). Five bundled templates built by scripts/build_templates.py through the server's own methods (--check proves the DXFs byte-reproducible); their .dwt twins were built once on AutoCAD 2026 by the smoke. drawing_new(template="iso_a3_mech") resolves the bundled file and reports {name, source}.

• Environment — 22 tools, pack settings. document_list/activate/close on both engines (the headless backend keeps a document registry; drawing_close is the active-document case); layer_state_save/restore/ list/delete as portable XRECORD snapshots — restore applies only the requested properties, reports missing_layers (skipped, never created) and new_layers (untouched), and the state survives save/reopen; view_named_save/restore/list; ucs_list/set/restore (non-orthogonal axes refused with the measured angle; tool coordinates stay WCS); system_variable_describe over a 90-entry catalog (known: false + nearest names on a miss); drawing_properties_get/set (custom properties on both engines, summary fields COM-only: dwgprops); system_launch (live_application), system_preferences_get/set (whitelisted keys with ranges, read-only keys refused by name, old/new reported: preferences), user_pick_point/select and system_prompt_message (interactive_prompt; a cancelled pick is {cancelled: true}, a COM_CALL_TIMEOUT overrun is timed_out).

• drawing_settings grows by limits, grid/grid_spacing, snap/snap_spacing, ortho, polar/polar_angle, psltscale, annotation_scale, linear_units, angular_units, dimstyle, textstyle; system_set_variable refuses an out-of-range value for a catalogued name before writing.

• Benchmark evidence. benchmarks/tasks_v4.py adds page_setup_truth; the correctness suite grows to 32 with settings_dimstyle_iso25_values, settings_layer_state_roundtrip, settings_pdf_mediabox_a3. A/B against v1.5.1: 26 / 32 → 32 / 32, six miss → pass, zero regressed. scripts/smoke_settings_com.py runs the track once against the live seat (record below).

• Benchmark evidence for tracks B and G. benchmarks/tasks_v5.py (18 tasks, the default matrix; v4 / v3 / v2 stay addressable) adds mech_assembly: a flange-coupling A3 sheet built through the code the tools call — frame and title block, a hub with a full section and chain dimensions, a flange with its bolt-circle end view, ISO 4014 bolts and ISO 4032 nuts, the parts list read off the drawing and linked balloons — gated on drawing_critique(focus=None) returning zero issues and a finalize score of at least 90 (it scores 100); a test drops one balloon and watches the gate fail by name. The correctness suite grows to 39 with mech_part_roundtrip, mech_section_hatch_area (1200 mm² read back out of the drawing), mech_iso286_on_dimension, mech_thread_unrepresented_is_caught, std_part_iso4014_m12, sheet_frame_iso5457_a3 and bom_balloon_link. A/B against v1.5.1: 26 / 39 → 39 / 39, thirteen miss → pass, zero regressed. The token suite measures a TOOL_PACKS=core,mech row: 193 tools / 61,673 tokens against 224 / 72,163 by default.

• Live COM smoke. scripts/smoke_mech_com.py executed once on AutoCAD 2026 (25.1s (LMS Tech)): the A3 frame and ISO 7200 title block, a two-step shaft with a full section and 12 chain dimensions, an ISO 4014 M12x60 block carrying 1 ACADMCP_MECH chunk, an ISO 7573 parts list read off the drawing with its linked balloon, an xref attached and detached, and a real 31687-byte .dwg written by ActiveX SaveAs — drawing_critique(focus=None) returned 0 issues (record below).

• Architecture — 12 tools, pack arch. arch_wall, arch_opening, arch_stair, arch_room, arch_rooms_detect, arch_schedule, arch_grid, arch_symbol, arch_catalogue_list, arch_catalogue_insert, arch_dimension_chains, arch_plan_from_spec. The plan travels as ACADMCP_ARCH XDATA through the mechanical codec, generalised to take an application id; draw_prims takes a role map. The arch layer set is ISO 13567-style with ISO 128 lineweights; wall poché follows engineering/arch/materials.py, whose rows each name their source. Areas are measured, never typed. Lean gains four and reaches 64.

• Benchmark evidence for track F. benchmarks/tasks_v6.py (19 tasks, the default matrix; v5 to v2 stay addressable) adds arch_roundtrip: the example plan drawn by arch_plan_from_spec and read back by arch_rooms_detect within 0.1 % of each label and of the hand-computed net floor, the schedules listing the drawn tags, gated on a zero-issue drawing_critique(focus=None) and a finalize score of at least 90; a test removes one room label and watches the gate fail by name. The correctness suite grows to 43 with arch_junction_l_t_x, arch_room_area_net, arch_opening_cuts_wall and arch_rooms_detect_foreign. A/B against v1.5.1: 26 / 43 → 43 / 43, seventeen miss → pass, zero regressed. The token suite measures a TOOL_PACKS=core,arch row: 191 tools / 60,568 tokens against 236 / 78,772 by default.

• Live COM smoke. scripts/smoke_arch_com.py executed once on AutoCAD 2026 (25.1s (LMS Tech)): the example plan drawn in one transaction, both rooms read back within 0.1 % of their labels and of the hand-computed net floor, and drawing_critique(focus=None) returned 0 issues (record below).

• Understanding foreign drawings — 4 tools, pack core; plant takeoffs — 2 tools, pack plant. drawing_understand, drawing_scale_check, drawing_diff, drawing_topology_check, pipe_takeoff, cable_takeoff. Every one reads a snapshot of frozen WCS records (engineering/understand/snapshot.py); labels are read by one parser (%%c, \U+2205, the U+2300 sign, ø and MTEXT format runs decoded, Cyrillic look-alike tags folded to Latin) and layer names and texts by one EN / TR / RU / NL / DE vocabulary. The takeoff workbook is written as CSV always and as XLSX with the new office extra (openpyxl, also in full). Lean gains drawing_understand and reaches 65.

• Benchmark evidence for track H. benchmarks/tasks_v7.py (21 tasks, the default matrix; v6 to v2 stay addressable) adds takeoff_roundtrip and understand_foreign, both on the synthetic plant pair of tests/fixtures/plant_pair.py, whose truth is computed from the positions it placed; each gate is shown to fail on a broken input. The correctness suite grows to 48 with scale_check_detects_schematic, pipe_takeoff_rmst_exact, cable_takeoff_roundup, diff_detects_known_edits and topology_known_defects. A/B against v1.5.1: 26 / 48 → 48 / 48, twenty-two miss → pass, zero regressed.

• scripts/field_test_takeoff.py runs both takeoffs on a user's own P&ID and layout and writes the report next to them; it refuses an output folder inside the repository and prints no quantity unless --show is given.

• Live COM smoke. scripts/smoke_understand_com.py executed once on AutoCAD 2026 (25.1s (LMS Tech)): the live snapshots matched their files type for type in model space, the DWG route read the same P&ID, both takeoffs matched the truth and the diff reported exactly the one entity moved (record below).

• P&ID — 9 tools, pack pid.

  • pid_symbol_list / pid_symbol_insert: a 43-symbol ISO 10628-2 / ISA-5.1 catalogue authored in code (11 valve bodies × 6 actuators, 6 rotating machines, 4 heat exchangers, 6 parametric vessels with nozzle ports, 7 inline items, the instrument bubble in 4 types × 5 locations, 5 line markers, 3 connectors), placed as real blocks with a TAG ATTRIB, named ports, and an ACADMCP_PID XDATA payload. Ports are transformed through the INSERT's real rotation, scale and mirror; a stretched symbol (|x_scale| ≠ |y_scale|) has no circle for a radial port and is refused rather than answered with a wrong point.
  • pid_line_draw: port-to-port orthogonal routing (auto picks from seven candidates by fewest bends, then length; direct; or waypoints), eight ISA-5.1 line classes mapped to layer / linetype / marker blocks, and line numbers filled from a format (default {size}-{service}-{seq}-{spec}-{insulation}; an empty field drops out with its separator). A route that cannot honour the stub on both ports is refused by name (route((0,0), 0°, (3,8), 180°, stub=5) raises) rather than drawn short; a non-axis port direction, a NaN coordinate and an unfillable format token are refused before anything is written. Lines it crosses are counted and reported (crossings), never silently drawn over — and never refused.
  • pid_tag_parse: the ISA-5.1 identification-letter grammar (Table 4.1, pinned by tests/data/isa51_tags.json), with the equipment-prefix map.
  • pid_graph: the reader. Every symbol INSERT becomes a node with a source and a confidence — catalogue name 1.0, ACADMCP_PID payload 0.95, TAG-style attribute or keyword block name 0.6 with inferred ports, an INSERT a plain line merely touches 0.3. Confidence is the minimum of its inputs, never raised by agreement between heuristics. Each line end resolves to a port, a junction on another line, or dangling with the nearest connectable port as a hint, always within its own space (scope="all" never joins two layouts). Edge length is the bulge-aware length the engine reports, not a chord walk (12.5 % short on a semicircular jump).
  • pid_instrument_index / pid_line_list / pid_equipment_list: derived from the graph only, optionally written as CSV. An instrument tapping a pipe reports the pipe's line number, never its own signal wire.
  • pid_from_spec: a whole sheet from one JSON spec in one transaction — rolled back entirely on the first refusal — returning the handles, the graph and the critique; dry_run=True routes in memory and touches nothing.

• block_define(name, entities, attdefs) on both engines: a block from typed primitive specs (line, circle, arc, polyline, text, solid) and ATTDEFs, no source entities needed; overwrite=True replaces the contents so existing INSERTs keep the name. Validation runs before any write: a DXF-illegal layer name, a duplicate or trailing-newline ATTDEF tag, a multi-line or trailing-caret string are TypeError/ValueError naming entities[i] / attdefs[i], and nothing is created.

• entity_get_xdata / entity_set_xdata on both engines, typed by DXF group code. The two limits AutoCAD enforces silently are enforced loudly and before the write: a string over 255 characters and an entity budget of 16 KB across every application (three 15 KB apps on one entity used to pass headlessly and would have failed SetXData live). LF/CR in a string is refused — it split a DXF tag, the saved file failed to reopen and, because every undo snapshot is a DXF save, transaction_rollback broke after one such write. nan/inf are refused rather than written as literal text that reads back as null.

• Six critique focuses — pid_dangling_line, pid_duplicate_tag, pid_incompatible_connection, pid_untagged_instrument, pid_illegal_tag, pid_unconnected_equipment — run under focus=None and inside drawing_finalize, over one cached graph. They score evidence only: a drawing without a recognised P&ID node returns [], a 0.3 touch guess can neither create nor suppress a finding, and a line ending on a sheet frame (an INSERT the drawing does not vouch for) is dangling. drawing_refine hints name only ports pid_line_draw accepts.

• Two resources: autocad://pid/symbols (the catalogue with ports and parameter schemas) and autocad://standards/isa51 (the letter tables and the grammar's disambiguation rules).

• Benchmark evidence. benchmarks/tasks_v4.py adds pid_roundtrip (the example sheet drawn through pid_from_spec, read back by a graph builder that never sees the spec); the correctness suite grows to 29 with three P&ID checks. A/B against v1.5.1: 26 / 29 → 29 / 29, three miss → pass, zero regressed (ab-v1.5.1-vs-v1.6.0-dev.json).

• scripts/smoke_pid_com.py runs the track once against the live AutoCAD on the machine, in a new document; scripts/render_pid_catalog.py and scripts/render_readme_pid.py rebuild the two README images.

Fixed

• Every tool call raised RecursionError after a few hundred in-process sessions. fastmcp 3's enable() / disable() append a Visibility transform per call, and a tool lookup walks the whole transform stack as nested calls. _apply_tool_profile runs in the lifespan, once per client connection, so each session stacked two more layers. The first Linux CI run of the 1.6 tree found it: on CPython 3.11, which counts C frames against the recursion limit, system_status itself failed from about 90 % of the suite on. The profile's transforms now replace the previous ones instead of stacking on them, the idiom _apply_discovery_mode already used, and a test switches the profile 600 times and then calls a tool.

• The same CI run failed 21 fake-ActiveX tests off Windows (they build a real pywin32 VARIANT and now skip without pywin32, like the suite's other VARIANT-dependent COM tests; the Windows lane runs them), and a test that spelled a path through a directory it never created, which only Windows normalises away.

• Found by the 1.6.0 README renders (each with a test that fails on the old code):

  • The exterior dimension chains of arch_dimension_chains and arch_plan_from_spec were drawn at the dimension style's DIMSCALE 1, so a 1:50 chain's 2.5 text was 2.5 drawing units high — 0.05 mm on the sheet — and the plan render showed empty boxes where the numbers belong. Every other plan annotation already followed the plot scale, and now the chains do too: DIMSCALE is the plan's scale while they are drawn (reported as dimscale, replacing the old advice to set it by hand), and the drawing's own DIMSCALE is put back afterwards. Headlessly each chain dimension carries the scale as its override. The live engine reaches the same DIMSCALE through SetVariable; that combination has not been re-run on a live seat since this change.
  • A headless TABLE drew its title as the first column's cell, so a title such as DOOR SCHEDULE wrapped inside that column and ran over the header row. The title row is now merged across the table, as AutoCAD's TABLE draws it.
  • A headless TABLE's cell margin was a flat 1.0 drawing unit — 0.02 mm on a 1:50 schedule, where the text sat on the cell line and LEFT read as FT. The margin now scales with the lettering (0.4 × text height, at most 0.15 × row height), which leaves every table at the default 2.5 text as it was.
  • drawing_new left the Standard text style on ezdxf's default txt, which has no superscript two: a room label's m² read m? when AutoCAD opened the DXF and rendered as a box headlessly. Standard is now arial.ttf, as AutoCAD 2026's acadiso.dwt ships it, so both engines begin from the same Standard; the five bundled templates are rebuilt on it and scripts/build_templates.py --check reproduces them.

• Hatch islands on a live seat. hatch_add_boundary was refused on COM (hatch_edge_paths), so a transverse section of any hollow part - a bushing, a flange bore - could not be drawn in AutoCAD. Each typed edge now becomes a temporary LINE, true ARC or ELLIPSE appended with AppendInnerLoop and deleted after Evaluate; a chain of lines is one polyline. Measured on AutoCAD 2026: a two-arc island leaves exactly 10000 - 100*pi of a 100 x 100 hatch, the OD40/ID20 sleeve's section B-B hatches 942.2864 mm2 (the view engine's declared 180-gon), and an open loop is refused with nothing left behind.

• block_find_references found nothing on a live seat. ActiveX names a block reference AcDbBlockReference, so entity_list(type_filter="INSERT") matched nothing. The COM type filter now lets the DXF name find it (INSERT, LWPOLYLINE, and every *DIMENSION class for DIMENSION); the reported type is unchanged.

• Found by the tracks B + G evidence (each with a test that fails on the old code):

  • dimension_diameter put its text beyond the second chord point headlessly and beyond the first on AutoCAD (AddDimDiametric's ChordPoint, measured: a ⌀50 from (x, 170) to (x, 220) annotated at (x, 160)). The two engines annotated opposite sides of the same part, and headlessly every front-view diameter of mech_part_draw stood on its own chain dimension — dim_overlap fired on any single-step disk. The headless engine now follows ActiveX.
  • On a live seat mech_part_draw died after drawing every view: its anchor POINT goes on MECH, and ActiveX refuses entity.Layer on a missing layer (Key not found) where ezdxf creates it. Every mechanical draw path now ensures its target layers before the first entity.
  • bom_extract returned nothing live: ActiveX names a block reference BLOCKREFERENCE, and the parts list filtered on INSERT only. Both names are read.
  • The live _entity_info reported no text_position for a dimension (the entity arrives narrowed to IAcadEntity, which has no TextPosition), so dim_overlap compared bounding-box centres — a diameter and its bore read "0.00 mm apart". It is read through the concrete class now.
  • pid_graph raised KeyError: 'start' inside drawing_critique(focus=None) on a LINE whose point read AutoCAD had refused while busy; such a line is left out instead.
  • The validator's fake_dimension_text read a standard-part designation (ISO 4014 - M12x60) as a hand-typed dimension, so every parts list drawn as text cost its sheet a warning. The designation is exempt; a measurement beside it still flags.

• Track A hardening (tests/test_track_a_hardening.py; the remaining review findings are filed with a disposition each in docs/analysis/track-a-review-backlog.md):

  • entity_create_block_ref of an undefined block name is refused before writing (it returned a handle to a dangling INSERT).
  • block_explode keeps ATTRIB text: each ATTRIB becomes a TEXT entity on both engines instead of being dropped.
  • pid_from_spec refuses a NaN coordinate before the transaction opens and rolls back on BaseException, not only Exception.
  • A signal line no longer receives a fabricated line number and label; the reader reports line_number: null for it.
  • A bubble's exit point honours the first waypoint instead of the port direction.
  • A foreign or hand-edited ACADMCP_PID payload is read tolerantly, and an existing polyline with a bulge is tested as its arc, not its chord, so a line through a crossing jump counts as a crossing.
  • block_insert attribute values are validated as strings; an int or None is a TypeError naming the tag, not str()-coerced.
  • block_define refuses a NaN base point and a layer that does not exist in the drawing (it wrote both and drawing_audit repaired them later).

• Headless block_insert(attributes=…) lost its ATTRIBs. The ezdxf path used add_auto_blockref, which wraps the INSERT in an anonymous *U block: the handle it returned reported block_name *U1, and block_get_attributes on it returned {}. Reproduce on v1.5.1 with block_create_from_entities of a block with one ATTDEF, then block_insert(name, attributes={"TAG": "P-101"}) and block_get_attributes(handle). It now inserts the named block with add_auto_attribs, so the INSERT carries the name and the ATTRIBs, on both engines alike. block_insert of an undefined name is refused before writing (it used to return a handle whose dangling INSERT drawing_audit later deleted).

• Live LWPOLYLINEs report as POLYLINE through ActiveX (AcDbPolyline), which the P&ID line reader's type filter excluded: on AutoCAD every auto-built line number restarted at seq 1 and crossings was always 0. Both spellings are accepted; the COM _entity_info polyline branch now returns points (and per-vertex bulges) for a live LWPOLYLINE.

• Polyline bulges now leave both engines in WCS: a mirrored polyline's arcs were handed out in the entity frame next to WCS points, so a reader tested them on the wrong side of the chord.

• INSERT properties gain geometry_bbox (drawn geometry only, ATTRIBs excluded) and mirrored on both engines. The COM box is exact at right angles and measured off them; a member ActiveX cannot measure is carried by its corners and the box says approximate: true.

• entity_list lost every type-specific property on a live seat with a makepy cache. ModelSpace.Item (and SelectionSet.Item) is declared IAcadEntity*, so once win32com.client.gencache had run against AutoCAD every listed entity arrived as the gen_py IAcadEntity wrapper: measured on AutoCAD 2026, item.StartPoint on an AcDbLine raised AttributeError, and a LINE row kept only its bounding box (no start / end, no polyline points, no text). The room reader therefore saw 0 wall segments and arch_plan_from_spec refused its first room label live, while every headless test passed. _entity_info now re-dispatches exactly that wrapper through its runtime class (_com_concrete_entity); a concrete wrapper from HandleToObject is left alone. The same narrowing strips start from every listed LINE, so it would equally explain the pid_graph KeyError: 'start' recorded above as a busy-refused point read (not re-measured).

• Track H, found after the groups landed. Six defects in the readers, each with a test that failed before its fix:

  • A truncated DXF (an interrupted save, a partial copy) made ezdxf raise StopIteration, which asyncio.to_thread cannot hand back to an awaiting caller, so every track H tool reading one would have hung. The snapshot now refuses it as "not a readable DXF file".
  • take_snapshot of a DWG the operator already had open opened it again read-only and closed "it" unsaved - possibly their copy, with their unsaved work. An open document is now looked up first, exported as it stands and left open; only a document the snapshot opened is closed.
  • drawing_diff: a LINE drawn back to front and moved was reported as reshaped; an INSERT carrying one attribute tag twice could lose an edit of the first value (a repeated tag is now compared as the list of its values); the cluster list is capped like the other lists (truncated["clusters"], the markup still clouds every cluster); and a whole-drawing move keeps only the 32 nearest candidates per entity.
  • A line number was read as an equipment tag: 100-P-001 yielded the pump P-001. A token that is the tail of a word beginning with a digit is no longer a tag.
  • The pipe network merged ends and dropped stubs within tol = 1.0 drawing units - a metre on a metre P&ID. The tolerance now defaults to 1 mm in the P&ID's inferred unit, and a dropped stub is counted and said.
  • A wiring callout was counted as a place its panel is written: Wiring to CP-2 made CP-2 ambiguous in drawing_scale_check and drawing_understand, and Wiring to CP1 a phantom tag CP1. On the synthetic pair the check matched 10 tags and 45 pairs (13.3 % within ±10 %) where the generator placed 11 and 55 (21.8 %). A callout is no longer a tag occurrence - the rule the line-network reader already kept.

• Track H, found by the field test (the takeoffs run on a real P&ID pair that stays outside the repository). Each fix has a test that failed first:

  • The pipe network took every piping or electrical layer, so valve, pump and annotation layers (sizes, materials, numbers) became pipe runs. vocab.is_network_layer is now the one rule for the network, the topology check and the topo_* focuses: piping or electrical at 0.9, and naming neither equipment nor annotation.
  • Measured on the P&ID, a run still needed equipment tags at both ends, so a field P&ID whose pipes rarely end on a tagged outline gave an empty table. On the P&ID each piece now goes to the room it lies in, with its diameter kept piece by piece. A layout takeoff that cannot route every run now says how many, why, and what the P&ID alternative is.
  • TEXT caret notation (^J, a line break) was left in place, so a room label's number read as 6^J.... plain() now decodes it.
  • A label that only mentions a room was read as a room: a control cabinet naming the room it serves, or a note pointing into a room. An unnumbered label naming equipment is no room. Without room faces, when the drawing numbers its rooms, only numbered labels take pieces.
  • The bare Russian "to" before a panel (к CP1, к щиту CP-2) is a wiring callout.

• Track H, what the field test asked for (each with a synthetic test):

  • Rooms drawn as frames. A room label no wall face holds takes the smallest closed, straight-sided polyline around it that holds no other room number and is at least ten label heights a side (label+frame, confidence 0.8). This is how a P&ID draws its rooms, so P&ID lengths are cut at the room frames.
  • The nearest panel, flagged. cable_takeoff(panel_rule="nearest"), the default: a load that no callout places, whose tag is on the layout, takes the nearest panel there. Only names with a panel prefix (is_panel_name: CP, JB, DB, MCC ...) or names a callout gives are candidates. The row is flagged panel_inferred, and an inferred panel never makes a package. panel_rule="stated" leaves such a load open. The benchmark and the correctness check read stated panels only.
  • Possible packages and detail copies. pipe_takeoff reports detail_copies: compact groups of tags drawn twice (repeated), and compact groups of one equipment's own parts, such as M150B, M150C and M150D of M150 (family). A pasteurizer drawn in detail, whose internal pipes its vendor supplies, shows up this way. Nothing is removed until exclude_regions=["D2"] names it. The excluded runs go to the control rows as excluded_region and are totalled in excluded_m.

Changed

• Release evidence re-measured for 1.6.0. The A/B gate against v1.5.1 (ab-v1.5.1-vs-v1.6.0.json: 26 / 48 → 48 / 48, 22 fixed, 0 regressed); the reference adapter on matrix v7 (autocad-mcp-pro.json: 21 / 21, replacing the v3 run of 1.5); the 100-point rubric re-scored per category with the evidence for every point withheld (93; the competitor rows are their July snapshots, not re-reviewed); the perf lane as medians of five alternating runs, with the v1.4.0 comparison published beside it (perf-v1.4.0-vs-v1.6.0.json). The four charts are regenerated from those files, and the README is rewritten around them, with the mechanical and architectural sheets rendered by scripts/render_readme_mech.py and scripts/render_readme_arch.py.
• ruff skips docs/superpowers: ruff 0.16 formats Markdown code blocks, and reflowing a plan's quoted code would change what the document says was written.
• MAX_DXF_BYTES defaults to 512 MB (was 50 MB): a real plant layout read for track H was a 188 MB DXF. The snapshot's refusal names the file's size, the limit and the variable; 0 still disables the check.
• Three tool groups appear in system_about and docs/tool-inventory.json: styles (10), page_setup (6), environment (22) — 204 tools in 23 groups; LEAN_TOOL_NAMES 55; capability keys dwgprops, dwt_write, live_application, preferences, interactive_prompt, registry_sysvar and the documents, named_views, ucs and layer_states (mode: "xrecord") features in both maps.
• TOOL_PACKS=all|core,pid,settings (default all) advertises only the vertical packs a client uses; core is always on and unknown names are ignored with a warning. It composes with TOOL_PROFILE=lean (which now carries pid_symbol_insert, pid_line_draw and pid_graph, and the five settings essentials: 55 tools, 52 with TOOL_PACKS=core) and ENABLE_3D; system_about reports tool_packs. Measured idle cost: 199 tools / 59,315 tokens by default, 190 / 55,593 with TOOL_PACKS=core,settings, 168 / 48,825 with TOOL_PACKS=core, 55 / 16,360 lean, 2 / 356 in discovery mode.
• PID_LAYERS (drawing_apply_iso_layers("pid")) gains PROCESS-LINE-TEXT; drawing_plan / drawing_preflight report LAYER_SET_INTENT_MISMATCH when a P&ID intent is planned onto a non-pid layer set.
• The prompt_pid_diagram prompt template is rewritten onto the tools (pid_from_spec, pid_symbol_insert, pid_line_draw, pid_graph) instead of describing primitives.
• CritiqueFocus grows by the six pid_* values; benchmark matrix v4 (--matrix v3 / v2 reproduce the earlier sets); correctness suite 29.
• README release snapshot: 204 tools · 8 resources · 5 prompts · 3284 collected tests (this machine; it was 166 · 2423 after track A). CI has not run on this tree yet, and the pre-existing 1369-vs-1363 drift between the README and Linux collection predates both tracks and is superseded by the new figure.

Not in 1.6

• Track D — Plant 3D, read-only — is dropped, and roadmap criterion 5 is withdrawn, not met. The maintainer does not use Plant 3D, and a reader tested against nobody's real projects would be a guess. TOOL_PACKS keeps no plant3d pack; the plant pack is track H's takeoffs.
• Roadmap criterion 7 is not met: headless creation is slower than v1.4.0. Medians of five alternating runs, one CPython 3.11.15 and one ezdxf over each tree's own code: 2,000 line creates 244.5 → 306.1 ms (1.25×), the 10,000-line build / export / reopen 1,847.9 → 2,553.5 ms (1.38×); with EZDXF_CALL_TIMEOUT=0 still 1.17× and 1.20×. The premium quality pass went the other way, 150.6 → 65.3 ms. A cProfile of the event-loop thread over the 2,000 creates counts 491,071 function calls on 1.6.0 against 319,038 on v1.4.0. The timeout is part of the cost, not all of it, and earlier notes that switching it off restored v1.4.0's numbers were wrong. Carried to 1.7.
• Four mechanical feature tables ship without rows: DIN 471/472, DIN 509, DIN 332 and ISO 3601-2. Their drawing code and provenance tests are in; every size is refused by name until rows verified against the standard's own table are transcribed.

Live COM smoke

scripts/smoke_pid_com.py, AutoCAD 2026 (25.1s (LMS Tech)), Windows 11, 2026-09-16, in a new document (Drawing2.dwg) — block_define → pid_symbol_insert → entity_set_xdata → pid_line_draw → pid_graph → six focuses → entity_get_xdata, all through the COM engine, exit 0:

{
  "autocad": "25.1s (LMS Tech)",
  "active_document": "Drawing2.dwg",
  "handles": {"P-101": "98", "V-201": "A8", "FCV-101": "B8", "FIC-101": "C5", "OP-1": "CF"},
  "graph_stats": {
    "nodes_by_kind": {"connector": 1, "equipment": 2, "instrument": 1, "valve": 1},
    "edges_by_class": {"electric": 1, "process_major": 3},
    "dangling": 0,
    "unclassified": 0,
    "confidence_min": 1.0,
    "entities_scanned": 16
  },
  "crossings_total": 0,
  "critique_issues": 0,
  "critique": [],
  "xdata_chunks": 2
}

The headless engine reports the identical graph_stats, crossings_total, critique_issues and xdata_chunks on the same spec. The first attempt exited 2 on this machine because of a stale win32com gen_py cache for the AutoCAD type library (a package directory holding only a __pycache__) — an environment fault, not a backend one; deleting %TEMP%\gen_py\3.11 cleared it and no backend change was needed.

Live COM smoke — settings (track E)

scripts/smoke_settings_com.py --build-dwt, AutoCAD 2026 (25.1s (LMS Tech)), Windows 11, 2026-09-22, in a new document (Drawing27.dwg) — styles → page setup → batch_plot (the PDF's own /MediaBox) → the five .dwt twins built and iso_a3_mech.dwt reopened → layer state round trip → named view → UCS → preferences (one reversible set) → SummaryInfo → prompt line, all through the COM engine, exit 0:

{
  "autocad": "25.1s (LMS Tech)",
  "system_launch": {
    "launched": false,
    "attached": true,
    "version": "25.1s (LMS Tech)",
    "document": "Drawing9.dwg",
    "visible": true,
    "progid": "AutoCAD.Application",
    "backend": "com"
  },
  "dwt_built": [
    {
      "name": "iso_a3_mech",
      "path": "templates\\iso_a3_mech.dwt",
      "bytes": 18240,
      "magic": "AC1032",
      "is_dwg_container": true,
      "format": "dwt"
    },
    {
      "name": "iso_a1_arch",
      "path": "templates\\iso_a1_arch.dwt",
      "bytes": 17281,
      "magic": "AC1032",
      "is_dwg_container": true,
      "format": "dwt"
    },
    {
      "name": "iso_a3_pid",
      "path": "templates\\iso_a3_pid.dwt",
      "bytes": 18528,
      "magic": "AC1032",
      "is_dwg_container": true,
      "format": "dwt"
    },
    {
      "name": "ansi_b_mech",
      "path": "templates\\ansi_b_mech.dwt",
      "bytes": 17250,
      "magic": "AC1032",
      "is_dwg_container": true,
      "format": "dwt"
    },
    {
      "name": "ansi_d_arch",
      "path": "templates\\ansi_d_arch.dwt",
      "bytes": 17278,
      "magic": "AC1032",
      "is_dwg_container": true,
      "format": "dwt"
    }
  ],
  "document": "Drawing27.dwg",
  "styles": {
    "dimstyle": {
      "ok": true,
      "name": "ISO-25-SMOKE",
      "values": {
        "DIMTXT": 2.5,
        "DIMASZ": 2.5,
        "DIMEXE": 1.25,
        "DIMEXO": 0.625,
        "DIMGAP": 0.625,
        "DIMTAD": 1,
        "DIMTIH": 0,
        "DIMTOH": 0,
        "DIMDEC": 2,
        "DIMDSEP": ",",
        "DIMLUNIT": 2,
        "DIMZIN": 8,
        "DIMBLK": "",
        "DIMTXSTY": "ISOCP",
        "DIMLWD": -2,
        "DIMLWE": -2,
        "DIMSCALE": 1.0
      },
      "written": [
        "DIMASZ",
        "DIMBLK",
        "DIMDEC",
        "DIMDSEP",
        "DIMEXE",
        "DIMEXO",
        "DIMGAP",
        "DIMLUNIT",
        "DIMLWD",
        "DIMLWE",
        "DIMSCALE",
        "DIMTAD",
        "DIMTIH",
        "DIMTOH",
        "DIMTXSTY",
        "DIMTXT",
        "DIMZIN"
      ],
      "current": true,
      "textstyle_created": true
    },
    "textstyle": {
      "ok": true,
      "name": "ISOCP-SMOKE",
      "font": "isocp.shx",
      "font_resolved": true,
      "current": false
    },
    "mleaderstyle": {
      "ok": true,
      "name": "SMOKE",
      "values": {
        "arrow_size": 2.5,
        "landing_gap": 1.0,
        "text_style": "ISOCP",
        "text_height": 2.5
      },
      "textstyle_created": false
    },
    "mleaderstyles_on_seat": [
      "SMOKE",
      "Standard"
    ]
  },
  "page_setup": {
    "applied": {
      "paper": "ISO_A3",
      "orientation": "landscape",
      "size_mm": [
        420.0,
        297.0
      ],
      "canonical_media_name": "ISO_A3_(420.00_x_297.00_MM)",
      "plot_style": "monochrome.ctb",
      "plot_style_known": true,
      "scale": "1:1",
      "scale_factor": 1.0,
      "scale_ratio": [
        1.0,
        1.0
      ],
      "dxf_standard_scale_type": 16,
      "activex_standard_scale": 16,
      "plot_area": "layout",
      "plot_type": 5,
      "device": "DWG To PDF.pc3",
      "margins_mm": null,
      "center": null,
      "paper_units": "mm"
    },
    "changed": {
      "paper": [
        "",
        "ISO_A3"
      ],
      "canonical_media_name": [
        "",
        "ISO_A3_(420.00_x_297.00_MM)"
      ],
      "size_mm": [
        [
          0.0,
          0.0
        ],
        [
          420.0,
          297.0
        ]
      ],
      "plot_style": [
        "",
        "monochrome.ctb"
      ],
      "device": [
        "Microsoft Print to PDF",
        "DWG To PDF.pc3"
      ],
      "margins_mm": [
        [
          0.0,
          0.0,
          0.0,
          0.0
        ],
        [
          17.0,
          17.0,
          5.0,
          5.0
        ]
      ],
      "paper_units": [
        "inches",
        "mm"
      ]
    },
    "pdf": {
      "layout": "Layout1",
      "ok": true,
      "path": "C:\\Users\\VECTOR\\AppData\\Local\\Temp\\acadmcp_settings_smoke_sapms54v\\Drawing27-Layout1.pdf",
      "bytes": 2774,
      "mediabox_mm": [
        420.158,
        297.039
      ],
      "paper": "ISO_A3"
    },
    "installed_plot_styles": [
      "Autodesk-Color.stb",
      "Autodesk-MONO.stb",
      "DWF Virtual Pens.ctb",
      "Fill Patterns.ctb",
      "Grayscale.ctb",
      "Screening 100%.ctb",
      "Screening 25%.ctb",
      "Screening 50%.ctb",
      "Screening 75%.ctb",
      "acad.ctb",
      "acad.stb",
      "monochrome.ctb",
      "monochrome.stb"
    ]
  },
  "template": {
    "path": "C:\\Users\\VECTOR\\Documents\\GitHub\\Autocad-MCP\\templates\\iso_a3_mech.dwt",
    "source": "bundled",
    "document": "Drawing28.dwg",
    "dimstyle_current": [
      "ISO-25"
    ],
    "papers": [
      "ISO_A3"
    ]
  },
  "documents": [
    {
      "name": "Drawing9.dwg",
      "path": null,
      "active": false,
      "saved": false,
      "entity_count": 0
    },
    {
      "name": "Drawing27.dwg",
      "path": null,
      "active": true,
      "saved": false,
      "entity_count": 0
    }
  ],
  "layer_state": {
    "saved": {
      "ok": true,
      "name": "SMOKE",
      "layer_count": 3,
      "replaced": false,
      "chunks": 2,
      "backend": "com"
    },
    "restored": {
      "ok": true,
      "name": "SMOKE",
      "applied": {
        "layers": 3,
        "properties": [
          "on",
          "frozen",
          "locked",
          "color",
          "linetype",
          "lineweight",
          "plot",
          "current"
        ],
        "current_layer": "0"
      },
      "missing_layers": [],
      "new_layers": [],
      "backend": "com"
    },
    "deleted": {
      "ok": true,
      "deleted": "SMOKE",
      "backend": "com"
    }
  },
  "view": {
    "saved": {
      "ok": true,
      "name": "SMOKE_VIEW",
      "center": [
        100.0,
        50.0
      ],
      "height": 80.0,
      "width": 160.9964412811388,
      "replaced": false,
      "backend": "com"
    },
    "restored": {
      "ok": true,
      "name": "SMOKE_VIEW",
      "center": [
        100.0,
        50.0
      ],
      "height": 80.0,
      "width": 160.9964412811388,
      "applied": "zoom_window",
      "backend": "com",
      "viewctr": [
        100.0,
        50.0
      ],
      "viewsize": 80.0
    }
  },
  "ucs": {
    "set": {
      "ok": true,
      "name": "SMOKE_UCS",
      "origin": [
        10.0,
        20.0,
        0.0
      ],
      "x_axis": [
        1.0,
        0.0,
        0.0
      ],
      "y_axis": [
        0.0,
        1.0,
        0.0
      ],
      "replaced": false,
      "current": true,
      "backend": "com"
    },
    "world": {
      "ok": true,
      "name": "world",
      "current": true,
      "backend": "com"
    }
  },
  "preferences": {
    "cursor_size": 5,
    "read_only": [
      "Files.SupportPath",
      "Files.TemplateDwgPath",
      "Files.PrinterStyleSheetPath",
      "Files.PrinterConfigPath"
    ],
    "template_dwg_path": "C:\\Users\\VECTOR\\AppData\\Local\\Autodesk\\AutoCAD 2026\\R25.1\\enu\\Template"
  },
  "properties": {
    "set": {
      "ok": true,
      "summary_written": [
        "author",
        "title"
      ],
      "custom_written": [
        "ACADMCP_SMOKE"
      ],
      "custom_deleted": [],
      "backend": "com"
    },
    "get": {
      "summary": {
        "title": "ACADMCP settings smoke",
        "subject": "",
        "author": "smoke_settings_com.py",
        "keywords": "",
        "comments": ""
      },
      "summary_available": true,
      "custom": {
        "ACADMCP_SMOKE": "1"
      },
      "backend": "com"
    }
  },
  "prompt_message": {
    "ok": true,
    "text": "ACADMCP settings smoke: done",
    "backend": "com"
  },
  "failures": []
}

The headless engine reports the identical mediabox_mm (to 1e-9: matplotlib writes an exact 420 × 297; AutoCAD's PDF driver writes 420.158 × 297.039, inside spec §9's ±0.5 mm), layer_state.restored and properties.get.custom on the same steps; summary_available is false headlessly and the five summary fields are null (capability dwgprops).

What the run refused before the green one, each fixed in the backend and pinned by a fake-ActiveX test:

• Three runs died with RPC_E_CALL_REJECTED (-2147418111) on the read that follows a document switch — Documents.Count straight after Documents.Open, after Close(False) inside document_close, and in _ensure_document_state — AutoCAD's own message filter refusing the call while it still switches documents. pywin32 has no CoRegisterMessageFilter, so the backend now waits that window out itself (_wait_out_rejected_call, read-only callables only, 0.25 s → 1 s pauses inside a 20 s budget) at those sites; any other error is not retried.
• layer_list on the reopened template raised AttributeError: 'IAcadLayer' object has no attribute 'Color'. Did you mean: 'color'?. The type library spells IAcadEntity / IAcadLayer color and Lineweight; late-bound dispatch resolved .Color / .LineWeight case-insensitively, but once a makepy cache exists (this machine's was regenerated on 2026-09-17) every object a method returns is a generated wrapper and the lookup is exact. The backend now spells both members as the type library does (works on both wrappers) and the fakes model the measured names.
• layer_state_restore raised 'IAcadObject' object has no attribute 'Count': Dictionaries.Item and IAcadDictionary.Item are declared IAcadObject* (the narrowing _com_unnarrow already handled for MLEADER styles); the save before it went through the typed Dictionaries.Add, so only the reopen paths failed. Both now un-narrow.
• A state saved with layer 0 current came back with two warnings and 0 unapplied: the Freeze put on layer 0 raises 'Invalid layer' whatever the value, even False while thawed (measured on a scratch document; color / Linetype / Lineweight / Plottable / LayerOn / Lock / ActiveLayer all fine on it). Freeze is now written only when it would change the layer.
• Two of the smoke's own expectations predated the merged tree and were corrected to it: mleaderstyle_create is implemented on COM through the ACAD_MLEADERSTYLE dictionary (spec §4: no mleaderstyle key exists), so the smoke now asserts the created style is listed with arrow_size 2.5; and plot_style_list marks catalogue rows with installed: true (the source: "installed" rows are files the catalogue does not know), so the monochrome.ctb check reads that flag.

--interactive (user_pick_point) was not run in this session: the seat is shared with other agents and nobody was confirmed at the keyboard, and an unanswered Utility.GetPoint holds AutoCAD's command line — every COM client gets RPC_E_CALL_REJECTED — until someone presses Esc, past the 60 s timeout. The pick path was executed live earlier in the track (Task 23's measurement: GetPoint answers in WCS while GetEntity answers in the current UCS). Each plot also dropped AutoCAD's plot.log into the repo root; it is ignored now.

Live COM smoke — mechanical parts and the sheet (tracks B + G)

scripts/smoke_mech_com.py, AutoCAD 2026 (25.1s (LMS Tech)), Windows 11, 2026-09-23, in new documents only (an xref source saved and closed first, then the sheet) — apply_layer_set("mech") → apply_titleblock (A3, zones, marks) → draw_part → add_view(section) → dimension_part → insert_std_part → weld_symbol_prims → extract_records → draw_bom_table → add_balloon → xref_attach / xref_manage("list") / xref_manage("detach") → run_critique(focus=None) → drawing_export_dwg("R2018"), all through the COM engine, exit 0, every scratch document closed afterwards:

{
  "autocad": "25.1s (LMS Tech)",
  "part_id": "103",
  "views": 2,
  "section_omitted": [],
  "dimensions": 12,
  "bolt_handle": "2F7",
  "xdata_chunks": 1,
  "parts_list_rows": 1,
  "parts_list_representation": "native",
  "balloon": "31A",
  "xref_attached": "XREF_SOURCE",
  "xref_listed": 1,
  "xref_detached": true,
  "xrefs_after_detach": 0,
  "dwg": {
    "path": "%TEMP%\\mech_smoke_*\\mech_smoke.dwg",
    "exists": true,
    "bytes": 31687,
    "magic": "AC1032"
  },
  "critique_issues": 0,
  "critique": []
}

The first runs found the four live defects listed under Fixed (the MECH layer, BLOCKREFERENCE, the dimension text_position, the refused LINE read) and one harness fault: the first call after Documents.Add can be refused with RPC_E_CALL_REJECTED while the new document initialises, and a refused cleanup then left the scratch document open. The smoke retries refused calls for up to 90 s (a refused call was not executed) and tracks its documents by the document list rather than by an assumed name, because SaveAs renames the document it writes.

Live COM smoke — architecture (track F)

scripts/smoke_arch_com.py, AutoCAD 2026, Windows 11, in a new document only — draw_plan_from_spec(EXAMPLE_SPEC) (walls, openings, stair, two room labels, chains, three schedules in one transaction) → read_plan → rooms_detect → run_critique(focus=None), all through the COM engine, exit 0, the scratch document closed afterwards:

{
  "autocad": "25.1s (LMS Tech)",
  "walls": [
    "EXT",
    "INT"
  ],
  "openings": [
    "D1",
    "D2",
    "W1",
    "W2"
  ],
  "omitted": [],
  "stairs": 1,
  "schedules": [
    {
      "kind": "doors",
      "representation": "native",
      "rows": 2
    },
    {
      "kind": "windows",
      "representation": "native",
      "rows": 2
    },
    {
      "kind": "rooms",
      "representation": "native",
      "rows": 3
    }
  ],
  "rooms": {
    "01": {
      "label_mm2": 27743750.0,
      "detected_mm2": 27743750.0,
      "hand_mm2": 27743750.0,
      "agrees": true
    },
    "02": {
      "label_mm2": 21993750.0,
      "detected_mm2": 21993750.0,
      "hand_mm2": 21993750.0,
      "agrees": true
    }
  },
  "faces_detected": 2,
  "confidence_min": 1.0,
  "critique_issues": 0,
  "critique": []
}

The first run exited 1: the first room label was refused because the room reader found 0 wall segments on A-WALL-E-N — the entity_list narrowing listed under Fixed. The run above is the one after the fix.

Live COM smoke — understanding (track H)

scripts/smoke_understand_com.py, AutoCAD 2026, Windows 11, on the synthetic plant pair only — take_snapshot of pid.dxf opened live, the same P&ID saved as DWG and read back through the read-only DWG route, take_snapshot of layout.dxf, pipe_rows / cable_rows on the live snapshots against the pair's truth, then one tag moved 10 mm and diff_snapshots of the layout before and after, all through the COM engine, exit 0, every document closed:

{
  "autocad": "25.1s (LMS Tech)",
  "export_s": 0.104,
  "live_pid_source": "live:pid.dxf",
  "dwg_pid_source": "live:pid.dwg",
  "model_space": {
    "pid_file": {
      "ARC": 1,
      "INSERT": 15,
      "LINE": 8,
      "LWPOLYLINE": 8,
      "MTEXT": 8,
      "TEXT": 31
    },
    "pid_live": {
      "ARC": 1,
      "INSERT": 15,
      "LINE": 8,
      "LWPOLYLINE": 8,
      "MTEXT": 8,
      "TEXT": 31
    },
    "pid_dwg": {
      "ARC": 1,
      "INSERT": 15,
      "LINE": 8,
      "LWPOLYLINE": 8,
      "MTEXT": 8,
      "TEXT": 31
    },
    "layout_file": {
      "CIRCLE": 18,
      "LINE": 25,
      "LWPOLYLINE": 4,
      "MULTILEADER": 8,
      "TEXT": 26
    },
    "layout_live": {
      "CIRCLE": 18,
      "LINE": 25,
      "LWPOLYLINE": 4,
      "MULTILEADER": 8,
      "TEXT": 26
    }
  },
  "paper_space_layouts": [],
  "takeoffs": {
    "scale_verdict": "schematic",
    "within_10pct": 0.21818181818181817,
    "runs_checked": 9,
    "pipe_mismatch": [],
    "loads_checked": 4,
    "cable_mismatch": []
  },
  "diff": {
    "moved": "44",
    "added": 0,
    "removed": 0,
    "changed": 1,
    "moved_is_the_change": true
  }
}

The smoke works only in documents it brought in from its own temporary folder, known by full path; each live snapshot, the DWG save and the move first check, in the same COM-thread call, that the active document is the smoke's own, and on the way out it re-activates the document that was active when it started.

[1.5.1] — 2026-08-06

A patch for three defects found after 1.5.0 went to PyPI, two of them by the adversarial pass over the rewritten README. 1.5.0 is yanked: its diameter and radius callouts put wrong numbers on drawings at default settings, and that is not something to leave installable.

Fixed

• The HTTP bind guard failed open on the launch path it was written for. _guarded_run_async read the bind address from the host kwarg and treated an absent one as loopback. It is not: fastmcp run forwards that kwarg only when --host was passed, and otherwise fastmcp resolves the host from fastmcp.settings.host, which is environment-backed. So FASTMCP_HOST=0.0.0.0 fastmcp run server.py:mcp --transport http bound every interface — 149 tools including file open/save and AutoCAD command execution, with no token — while the guard inspected the literal string "127.0.0.1" and waved it through. v1.4.0's changelog claimed the guard "now fires on every launch path"; it fired, and then looked at the wrong host. It now resolves the effective host, and says so in a warning if it cannot.
• [com] and [full] could not be installed off Windows. Both extras listed pywin32>=306 with no environment marker, and pywin32 publishes Windows wheels and no sdist — so pip install autocad-mcp-pro[full] died at resolution on Linux and macOS, and [full] is the extra this project's own README recommended to contributors. CI never caught it because every Linux lane installs [pdf] or syncs from uv.lock, whose resolution markers hide the gap. Both entries now carry ; sys_platform == "win32", and a release-consistency test fails if a Windows-only distribution loses its marker again.
• drawing_settings({"dimscale": ...}) reported success and changed nothing headlessly. It writes the $DIMSCALE header variable, which is what AutoCAD consults when it draws a dimension — so the call worked on the live backend. ezdxf renders from the dimstyle table entry and never reads the header, so the same call produced two different drawings on the two engines while both said ok. The header value is now folded into the per-dimension override on all five dimension tools, including dimension_aligned and dimension_angular, which do not share the tolerance-override path and were missed by the obvious fix. DIMSCALE 1.0 writes no override.

Changed

• README: the "What you get" table qualified every engine-conditional family (boundary tracing, WIPEOUT/REVCLOUD/MTEXT masks, CHSPACE, typed hatch edge paths, headless undo) instead of listing them as if both engines had them, and the checkout install line no longer recommends an extra that cannot resolve on the reader's platform.

[1.5.0] — 2026-08-06

Added

• Selection filters — 3 tools. selection_window (AutoCAD's ssget window/crossing), selection_polygon and selection_filter (QSELECT). Window returns only entities wholly inside; crossing also returns the ones straddling the edge — the modes are reported back and window is the default, because the over-selecting mode must never be the implicit one. A zero-area box is refused rather than answered with an empty list, which would read as "nothing there". selection_filter takes named parameters rather than a query string deliberately: ezdxf's query language answers an unknown attribute with an empty result, so LINE[nosuchattr=='x'] is indistinguishable from "no matches". Here a typo is a TypeError at the boundary, and filtered_by reports which filters actually ran. min_area excludes shapes that have no area at all rather than treating them as zero, and a negative min_area is refused — no area is negative.

• Boundary tracing — 2 tools, and five deliberately not built. boundary_trace is AutoCAD's BOUNDARY/BPOLY: a seed point in, the nearest enclosing closed polyline out. It never calls edgeminer.find_all_loops — measured at 63.9 s for 40 edges and a TimeoutError at 60, because the docstring's O(n!) is real — and walks one loop per seed-adjacent edge instead, which is 0.0057 s at the same size. Straight edges are split where they cross, so a construction line drawn across a shape actually divides it; without that the line shares no endpoint with anything and is invisible to the edge graph. Curved edges are not split, and the capability says so. Two things the obvious implementation gets silently wrong, both now tested: edgesmith.lwpolyline_from_chain emits n+1 points with the last duplicating the first and closed left False — a polyline that draws like a square and behaves like an open path to hatch association, area queries and entity_offset; and edgesmith.loop_area works off edge endpoints, so two semicircular arcs closing into a circle measure 0.0 against a true 314.159. The area now comes from the flattened curve and each curved edge keeps its bulge. Not built: every REGION and 2D-boolean tool the plan listed. ezdxf cannot author a REGION — add_region() returns one with zero ACIS bytes — and the substitute, greiner_hormann, takes flat vertex lists. Measured on a square with one edge bulged into a semicircle, routing it through that boolean loses 28.2% of the area and 12.5% of the perimeter. That is the same magnitude as the analysis_measure_area bug this release fixed, and shipping it would have re-introduced the class.

• Hatch depth — 3 tools. hatch_set_gradient, hatch_edit and hatch_add_boundary. hatch_edit leaves omitted parameters alone — a partial edit that resets the rest is data loss — and changed reports the attributes that actually moved, so re-setting a value to what it already was comes back empty rather than as a false positive. hatch_add_boundary takes typed edges (line/arc/ellipse) because a boundary API that only accepts vertex lists silently straightens every curve it is given; every edge is validated before any is written, so a malformed list refuses instead of leaving a half-built path. On COM it refuses with capability: "hatch_edge_paths": ActiveX appends boundary loops as existing objects, not as typed edges, so the fidelity this tool exists for cannot be expressed there.

• analysis_list_properties — AutoCAD's LIST. The raw DXF attribute set for one handle, which entity_get deliberately does not carry. It cannot be handed over as-is for two reasons, both fixed here: dxfattribs() holds Vec3 values and is not JSON-serialisable, and its coordinates are in the entity's own frame, so for anything with an extrusion they would contradict every other coordinate this server reports.

• Annotation objects — 4 tools. entity_create_wipeout (mask what is behind a closed polygon), entity_create_revcloud, text_set_background (MTEXT background mask) and text_find_replace. Each carries a refusal that was measured rather than assumed. A wipeout of fewer than three points encloses no area, so it is refused instead of returning a handle to a mask that masks nothing. A revision cloud is its arcs — ezdxf.revcloud's own helper stamps the RevcloudProps xdata and takes the bulge sign from the winding direction, so the result satisfies is_revcloud() — and a segment_length longer than the shortest edge produces no arcs at all, which would be a plain rectangle reported as a cloud; that case is refused. text_set_background refuses TEXT (only MTEXT has a background-fill attribute, so setting one on TEXT would report success and change nothing) and refuses a scale below 1.0 (a box smaller than its text is a stripe through the text). text_find_replace searches TEXT, MTEXT, ATTRIB and ATTDEF, including inside block definitions — leaving the ATTDEF behind would fix the drawing and reintroduce the old text on the next insert. It returns searched_types, because "no matches" and "that type was never searched" are different answers and a caller cannot otherwise tell them apart. DIMENSION text is deliberately out of scope: its text field holds the <> override placeholder rather than the measurement, so editing it would either do nothing or detach the dimension from what it measures — the response says so. On COM, entity_create_revcloud refuses with capability: "revcloud": ActiveX has no revision-cloud member and driving the command blind against a live drawing is unverified.

• The rest of the paper-space lifecycle — 8 tools. v1.4 could make a sheet and never unmake it. Added layout_delete, layout_rename, layout_copy, viewport_list, viewport_set_scale, viewport_lock, viewport_delete and entity_change_space (CHSPACE). Verified headless (ezdxf); the COM paths for the first seven are implemented against the ActiveX reference but never executed against a live AutoCAD, and each is marked so in the source. Most of the work went into refusals, because ezdxf's own guards stop short of where a tool boundary has to: layouts.delete("") deletes the first sheet rather than refusing, layouts.get("") resolves to whichever is active, and layouts.rename() skips the name validation new() performs, so Bad/Name and even "" land in the table and the file still audits clean. Deleting or renaming the current tab used to leave three different answers to "where does my next line go?" — layout_list said one thing, $TILEMODE another, and _msp() quietly wrote to model space; all three now agree. A handle from a deleted sheet refuses instead of dying with AttributeError (entitydb.get is documented not to filter destroyed entities, and _get_entity believed it). layout_copy is a construction — ezdxf has no layout-copy API and xref.load_paperspace is cross-document only — and every way it can go wrong is invisible to the Auditor, which reports zero errors on a duplicated tab order, two main viewports, and a hatch whose boundary still points into the layout it was copied from and dangles the moment that layout is deleted. All three are prevented structurally; skipped and associativity_dropped report what could not be carried over rather than letting a bare ok: true imply a complete copy. viewport_list reports the layout's own main viewport with is_main: true rather than hiding it — it is what remains after every drafting viewport is deleted — and refuses to report a scale on R12, where view_height and flags are dropped on export and get_scale() then returns a fabricated number (150.0 measured for a viewport storing no scale at all). The two writers refuse on R12 outright rather than reporting a success the file cannot keep, and viewport_lock read-modify-writes the flags bitfield, which also carries the UCS-icon and grid bits a bare assignment would silently drop.

• entity_change_space (CHSPACE), headless only. AutoCAD moves objects through a viewport, rescaling them so they look identical on screen. ezdxf's move_to_layout applies no geometric change whatsoever, so the obvious implementation reports success, turns a 100 mm feature into 100 mm of paper inside a 1:2 viewport, and writes a perfectly clean file. The transform is therefore the tool, and it uses viewport.get_transformation_matrix() rather than a hand-rolled one: the hand-rolled version ignores view_target_point and misplaced geometry by 25 mm in testing while is_top_view was still true and the twist angle still zero — past every guard one would naturally write. A regression test pins the result against what ezdxf's own rendering pipeline projects. It refuses rather than guessing: non-plan and twisted viewports (ezdxf returns a meaningless matrix for the first and rotates about the paper origin for the second), dimensions unless freeze_dimensions=true bakes the measurement in first (transform halves get_measurement() while the baked block text keeps the old value, so one of the two is always lying), ACIS bodies (transform is accepted and moves nothing), tables and proxies (transform raises, and hasattr(e, "transform") is True for all of them), viewports themselves, and entities already in the target space. Geometry that lands outside the viewport or off the sheet is moved and flagged (inside_viewport, on_sheet), not refused — AutoCAD allows it too. On the COM backend this refuses with capability: "chspace": ActiveX has no change-space member, so the alternatives are CopyObjects + delete or driving the command blind, and neither could be verified against a live seat.

• CAD_PROGID — the COM ProgID is configurable. "AutoCAD.Application" was hardcoded at both connection sites, so AutoCAD-compatible applications (BricsCAD, ZWCAD, GstarCAD) were unreachable even though they expose the same ActiveX object model. The setting is honoured on both paths deliberately: Dispatch is not a passive probe — it COM-launches the application and sets Visible = True — so honouring the setting on GetActiveObject and then falling back would start the very AutoCAD the operator said they were not using. An unrecognised ProgID fails loudly for the same reason, and the error names both the ProgID it tried and the variable to change. system_status now reports cad_progid, without which a non-AutoCAD operator reads an autocad_version key naming a product they do not own. Caveat, stated everywhere it appears: CAD_PROGID only changes which COM application the backend attaches to — every tool beyond the connection itself is developed and tested against AutoCAD alone, and clone compatibility is unverified, so treat a non-default ProgID as experimental. The candidate ProgIDs in .env.example are commented out and marked UNVERIFIED: none has been tested against a running seat.

• analysis_measure_entity — measure the drawing, not your memory of it. Neither backend could answer "what is the area of the polyline I just drew": analysis_measure_area shoelaced points the caller typed in, and the live COM backend did the same despite having the document, a HandleToObject lookup it already used fourteen times, and an ActiveX Area property. The only available workflow was also silently wrong — entity_get read LWPOLYLINE vertices through the default format and kept x and y, dropping the bulge that makes an edge an arc. Measured on a 100×100 square with one edge bulged into a semicircle: 10000 reported against a true 13926.9908, 28.2% low, perimeter 12.5% low, nothing in the response saying so. And props["length"] = ent.length() was dead code — LWPolyline has no length(), so the key never appeared at all. Bulge geometry is always circular, so the new engineering/measure.py is exact rather than approximate: shoelace plus each arc's signed circular segment, and arc length for perimeter. entity_get now reports area and length from the same maths. The payload states its own accuracy — exact, flatten_tolerance, assumed_closed (an open boundary is closed the way AutoCAD's AREA does), and self_intersecting, because the shoelace cancels crossed lobes and a bowtie measures 0.0. REGION/3DSOLID refuse with capability: "measure_area_acis" naming the live backend; LINE/TEXT/INSERT raise a plain error, since no engine can give them an area and a capability tag would falsely suggest switching helps. AREA/MEASUREGEOM now route here rather than at the points-based tool — a drafter typing AREA picks objects, they do not retype coordinates.

• Handle grounding: view_screenshot(overlay_handles=true). A screenshot showed the model what the drawing looks like and nothing it could act on — every modify tool takes a handle, and nothing connected "the circle at the top-left" to a hex string. Each entity is now labelled with its handle at its own bounding-box centre (taken from the renderer's extents, so the label lands where the entity is drawn even in a rotated frame). Off by default, because labels are ink on the drawing; past 40 entities the render is capped and the image itself says "handles shown for N of M" rather than letting a labelled subset read as the whole sheet. The live backend captures the AutoCAD window rather than rendering, so there is nothing to label there — it refuses with capability: "handle_overlay" instead of quietly returning an unlabelled image, since producing the overlay would mean creating and deleting entities in the user's open drawing.

• block_create_from_entities works on the live backend. It refused outright and pointed at system_run_command with _BLOCK, pushing callers through the free-text escape hatch for something ActiveX does directly (Blocks.Add + Document.CopyObjects). The tool's own docstring said the opposite — "works by using AutoCAD's BLOCK command (COM backend only)" — so the one engine that could not do it was advertised as the one that could. Both engines now also report skipped handles instead of swallowing them, and a call where nothing resolves fails rather than leaving an empty definition behind. COM path unit-tested against a fake ActiveX object, not hardware-verified.

• Undo and redo on the headless backend (EZDXF_UNDO_DEPTH, default 0). The plan called this "add redo". The larger half was that undo never worked either: _undo_stack was declared, read and cleaned up, but nothing ever appended to it, so drawing_undo could only answer "Nothing to undo" — while the tool summary said "Step back one operation" and the alias corpus routed UNDO and "take that back" straight at it. The old test asserted isinstance(result, dict), which a permanent failure satisfies. It is opt-in because it is not free. ezdxf keeps no journal, so a history step is a whole DXF snapshot: measured 3.1 ms / 19 KB at 10 entities but 130 ms / 658 KB at 5000, and switching it on costs 37× on entity creation (0.18 → 6.65 ms per call) even on a small drawing. Off, it costs nothing; on, you get real multi-step undo/redo. Both tools refuse with capability: "undo_history" naming the setting rather than pretending. The live COM backend uses AutoCAD's own undo and ignores it. Drawing after an undo discards the redo branch, as in AutoCAD — otherwise redo would restore a state that never existed, with geometry you had removed reappearing beside geometry you drew afterwards.

• BREAKING (behavior): geometry now goes into the layout you selected. layout_set_current("A3-Sheet") answered {"ok": true, "current": "A3-Sheet"} and then every entity created afterwards landed in model space anyway, on both engines — _msp() returned model space unconditionally, so the tool reported a state change that changed nothing. Proven: a line drawn straight after selecting a layout was in modelspace and not in the layout. That was also why a title block could not go on a sheet, so titleblock_apply_iso_a3 gains a layout parameter — where a title block belongs, since the border frames the printed sheet rather than the model — and restores your current space afterwards rather than leaving you on the sheet. Existing model-space workflows are unaffected: a fresh drawing starts in model space, and a refused layout switch leaves the current space where it was.

• ISO 286 transition and interference holes (K, M, N, P). fit="P7" used to raise by name. They are now derived by the standard's delta rule, ES = -ei + Δ with Δ = IT(n) − IT(n−1) — a rule over the shaft and IT tables already in the module, not new table data, which is why it can be pinned against published values (K7/M7/N7/P7 and K6/N6/P6 cross-checked by hand) instead of merely trusted. Unlocks the hole-basis pairs K7/h6, M7/h6, N7/h6, P7/h6.

• ISO 286 interference shafts (r, s, t, u). fit="s6" and fit="u6" — the standard press and drive fits — used to raise. Unlike K/M/N/P these are new table data, sub-stepped more finely than the main size steps (ISO 286 splits 18–30 into 18–24/24–30, 30–50 into 30–40/40–50, and every step above 50 mm in two or three; flattening that is a silent error of tens of microns at exactly the sizes press fits are used at). Because it is data rather than a rule, its provenance is tested: every value is cross-checked against ISO 286-1's own derivation formulae, an independent source. r reproduces as the geometric mean of p and s on 24/24 steps — with p pinned in this module before r existed — and s/t/u track their IT7-based formulae within 1 µm on every step from 18 mm up. The five places where the standard's rounding puts a published value 2–3 µm off its own formula are listed by name in the tests, so a reader holding ISO 286 checks five numbers rather than ninety-five. Shaft t is undefined below 24 mm in the standard and refuses there rather than being extrapolated. Unlocks H7/r6, H7/s6, H7/t6 and H7/u6.

• Headless performance lane (benchmarks/perf_suite.py + render_perf_chart.py): fixed wall-time workloads (2k individual creates, 10k build/export/reopen roundtrip, 10k region query, full premium quality pass) measured through the same backend methods the MCP tools call; machine fingerprint recorded, self-measurement only.

• Task-matrix chart (benchmarks/render_matrix_chart.py): tasks × servers status heatmap rendered from the published live-run reports, so coverage gaps are visible per capability instead of one folded score.

• Tool discovery (DISCOVERY_MODE=search, default off): replaces the advertised catalog with search_tools + call_tool, so a client sees two tools instead of 127. Ranking layers an AutoCAD command / synonym corpus (discovery/aliases.py, 132 tools · 148 command tokens · 583 synonyms) over fastmcp's BM25 index, and hits are returned as one compact line per tool rather than full JSON schemas. Measured on a 66-query golden set with a 15-query holdout: top-1 62/66 vs 23/66 for stock BM25, top-3 64/66 vs 33/66 (holdout top-3 14/15 vs 9/15); stock returns nothing at all for 21 of the 66, and this transform for none of them — that gap is the corpus, not the ranking, and no BM25 parameter closes it. (Both sides moved slightly during the release as tool docstrings were rewritten: better prose feeds both indexes, which is why the headline is the top-1 gap rather than a ratio between two scores over a shared corpus.) A risk ceiling keeps destructive tools out of read-only answers and says when it hid one. Off by default — flipping it changes what every connected client sees on list_tools, which needs its own migration note.

• @cad_tool — a discovery-metadata channel next to the tool contract. A decorator written above an mcp.tool registration attaches a cad key to that tool's MCP meta: a one-line summary for a search hit's preview, a cost from a closed set (read, safe, mutate, destructive, escape) for clients that want to gate or colour-code the surface before calling, and the tool's AutoCAD command names and synonyms — imported from discovery/aliases.py, never restated, so the corpus stays the one edit site. tags= and annotations= deliberately stay on the mcp.tool call, so the channel cannot perturb _tool_groups() or the registration-counting release gate. Misuse fails at import rather than silently no-opping: an unknown cost raises, and a decorator written below the registration (or on a bare function) raises with the ordering rule in the message. Applied to the three transaction_* tools so far; every other tool still falls back to the first line of its docstring in search results, which is why the rollout can be section-by-section instead of a 131-tool patch.

• EZDXF_CALL_TIMEOUT (default 120s, 0 disables): per-call timeout for the headless backend, mirroring COM_CALL_TIMEOUT. Every ezdxf method runs under one document lock, so a single hung call previously deadlocked every later tool call.

• uv.lock — the dependency graph CI actually installs. 95 packages, resolved universally over requires-python >=3.11, so one file serves the Linux 3.11/3.12 matrix and the Windows leg (platform-only packages like pywin32 carry markers). Every lint and test lane now installs with uv sync --locked, which fails rather than re-resolving when the lock and pyproject.toml disagree — a fresh CI run can no longer pick up an upstream release nobody chose, and the locked set is the set the suite is green on. The package job stays unlocked on purpose: it is the only job that proves the published dependency ranges still resolve for a plain pip install autocad-mcp-pro. The uv.lock line in .gitignore (added when the committed lockfile was stale) is gone with it — a lockfile git ignores is not a lockfile.

• Scheduled floating-dependency job (.github/workflows/deps-floating.yml): weekly and on demand, it ignores the lock, resolves the declared ranges fresh on 3.11/3.12 and re-runs the suite. Locked lanes cannot be broken by an upstream release, which also means they cannot notice one; this is where a range ceiling that needs moving shows up. It never gates a pull request.

• Benchmark matrix v3 — five tasks that can fail. The v2 matrix scored this server 10/10, which was not evidence of anything: every task in it exercised something the server was built around, so full marks were the only available outcome. benchmarks/tasks_v3.py keeps the v2 ten byte-for-byte (their ids and weights key the published v1.4 reports) and adds tool_discovery, token_budget, hatch_islands, selection_filter and measure_from_handle. Each verifies against a value derived independently of the code under test — a closed-form area, entities placed on purpose, a token ceiling fixed before the measurement. Three of the five failed while being written, and the defects they found are in Fixed below. Two are renamed from the release plan, and the renames are the point rather than bookkeeping. tool_discovery_bilingual lost its "bilingual": the Turkish query normalizer was written and then removed, because this repository is English throughout and the Turkish was inferred from the author's chat language, not required by the product — a task named "bilingual" would benchmark a capability that does not exist. measurement_massprops became measure_from_handle: no centroid or moment-of-inertia tool ships, since F2 was cut to entity_measure. The competitor reports are not re-scored. Those servers were pinned and run against v2 at v1.4 and have never been asked these five questions; the matrix chart renders the five rows blank with a not run legend entry rather than zero, because an invented zero is indistinguishable from a measured one. run_competitors grows --matrix v2|v3 (v2 stays addressable so a v1.4 report can be reproduced rather than only described), records which matrix ran in the report, and grows --publish, which writes the artifact-path-free file that results/published/ holds — that sanitisation used to be a manual edit.

• A capability refusal in the benchmark runner is now recorded unsupported instead of fail. The result enum has carried unsupported since v2 for exactly this, but a refusal is still an exception and was being caught by the generic handler — "your engine cannot reach this" and "your server got this wrong" are different findings about a competitor.

• Five checks added to correctness_suite.py (26, was 21), and the two kinds read differently in the A/B table. Three measurement checks are new capability — v1.4.0 misses them by not having the method (miss → pass) — and are there because all three broke during this release. Two dimension checks are repaired defects: v1.4.0 has the methods and gets them wrong (fail → pass). Result against v1.4.0: 21/26 → 26/26, five fixed, zero regressed.

Changed

• The backend contract is 19 modules instead of one 1854-line file, and optional methods are declarative. backends/base.py held a single ABC with 112 abstract methods; adding one made both backends fail at instantiation until both were written, and two test modules construct ComBackend at import time, so a half-finished contract edit took the entire suite red. The interface now lives in backends/contracts/, one module per domain, split along the section markers that were already in the file — the seams were not re-decided. AutoCADBackend is the composition of all 19 and stays the import surface, so from backends.base import AutoCADBackend and every isinstance check are unchanged, and the abstract-method count is identical before and after. The split surfaced something the file's name hid: roughly 740 of those lines are shared concrete implementation (premium meta-tools, GD&T, the settings facade), not interface. Those slices are mixins and are labelled as such rather than presented as pure contracts. New backends/capability.py carries @capability(key, reason=...), which marks a contract method optional and supplies a default. The default raises UnsupportedCapabilityError rather than returning a quiet {"ok": False} — a method nobody implemented has to be indistinguishable from a deliberate refusal, not from a soft failure, which is the distinction this release spent a milestone restoring. A key used by the decorator must be declared in both capability maps or tests/test_capability_contract.py fails; without that gate the decorator would just be a way to ship refusals system_capabilities never mentions, which is worse than the build break it replaces. Its first user is COM's entity_change_space, which now inherits the refusal instead of hand-rolling it. The refusal vocabulary moved into this module as well (the contracts need it and base imports the contracts); backends.base re-exports it, so there is still exactly one exception class object in play.

• BREAKING (env): the core tool profile was removed. TOOL_PROFILE is now lean or full; the discovery layer above solves the crowded-surface problem core existed for. TOOL_PROFILE=core still boots — it falls back to full with a warning naming the removal, and system_about reports the requested profile alongside the applied one.

• BREAKING (install): ezdxf>=1.4 is now the floor (was >=1.3). The 1.4 line is the only one this release is developed and tested against; 1.3 is no longer exercised by CI and is unsupported. Upgrade with pip install -U "ezdxf>=1.4".

• Runtime pins narrowed. A published library cannot pin exactly and stay co-installable, so what a user's pip install gets is still a range — and a range is the only guard that install has. They are therefore deliberately narrow. (uv.lock, above, is the exact guard on our side; the two are not substitutes for each other.)

  • fastmcp>=3.0 → fastmcp>=3.4.5,<3.5. fastmcp is now a thin meta-package (Requires-Dist: fastmcp-slim[client,server]) and 3.x moves fast enough that a new minor can reshape the API the server is built on. Tested against 3.4.5; the ceiling moves only after the suite is re-run.
  • ezdxf>=1.3 → ezdxf>=1.4,<2 (see the BREAKING note above).
  • pydantic>=2.0 → pydantic>=2.0,<3, so a future pydantic 3 cannot be pulled into an install of this release.

• Rendering requires an extra — documented, not changed. ezdxf.addons.drawing.__init__ imports .frontend, which does an unconditional top-level import PIL.Image. Pillow is therefore a hard requirement of every ezdxf render path (PNG screenshot and PDF export alike), not just of the COM window-capture path it is listed under. A bare pip install autocad-mcp-pro resolves 77 packages with neither Pillow nor matplotlib, so it cannot render at all; [pdf] (or [full]) is required, and satisfies Pillow transitively via matplotlib. Note that [com] alone installs Pillow but not matplotlib, so the ezdxf backend still cannot render under it. Pillow deliberately stays out of the core dependencies for now.

• CI lint pinned to ruff==0.16.1 (was an unpinned pip install ruff / ruff>=0.8). The version is now written in exactly one place — [dependency-groups] dev — and frozen there by uv.lock; the workflows install ruff from the lock instead of restating the pin, so the three copies that had to be bumped together no longer exist.

• Dev toolchain moved to a PEP 735 dependency group. [tool.uv] dev-dependencies, which uv now warns is deprecated, became [dependency-groups] dev — same four tools, byte-identical resolution. The standard spelling is also understood by pip >=25.1, so pip install --group dev gives contributors who do not use uv the same list without a second copy of it.

• CI and the release gate install from the lock. lint (dev group only — ruff without building the project), test-linux, test-windows and the release.yml test job now run uv sync --locked + uv run --no-sync, with uv itself pinned in the workflow env for the same reason ruff is. package, docker and registry-validate are unchanged. benchmark.yml is locked too: that lane publishes evidence, and an ezdxf release moving our own numbers between runs would make the diff unattributable.

• README rebuilt: PyPI-first install, a four-lane benchmark section with Python-generated visuals (scores, task matrix, performance, source review), v1.4 feature table, updated client/config/roadmap sections.

• render_live_chart now skips non-task-matrix reports so multiple lanes can share benchmarks/results/published/.

Fixed

• The COM paths were run against a live AutoCAD 2026, and four of them were broken. Every COM implementation added in this release was written against the ActiveX reference and shipped labelled unverified. Executing them against a real seat — a fresh instance, a throwaway drawing, nothing saved — found four defects that no amount of reading would have caught, one of them pre-existing:

  • AcadPViewport has no ViewCenter member. viewport_list died on it outright. Worse, viewport_create has carried the same line since v1.4 wrapped in a bare except: pass, so it raised on every single call and said nothing — every COM viewport has been looking at the model origin instead of the view_center_x/y the caller asked for. The member is Target, and the failure is no longer swallowed.
  • ModelSpace.AddWipeout does not exist. ActiveX has no wipeout constructor at all; WIPEOUT is a command. Now a typed refusal (capability: "wipeout") instead of an AttributeError.
  • AcDbMText.BackgroundFillColor does not exist — it raises on read as well as on write, with an int and with an AcCmColor object alike. BackgroundFill (the on/off flag) does work, so text_set_background enables and disables the mask on COM and refuses a colour request with capability: "mtext_background_color", rather than switching the fill on in the wrong colour and reporting success.
  • analysis_list_properties called a helper that does not exist on the class. After the fixes, all 25 verification calls behave correctly: 18 succeed, 1 refuses per-argument (deleting model space), and 6 refuse with the capability keys they declare. The remaining COM refusals (chspace, revcloud, boundary_trace, hatch_edge_paths) were confirmed to hold rather than leak a result. The source comments now say VERIFIED with the date instead of UNVERIFIED.

• The headless renderer does project model content through viewports. v1.4 said it did not. viewport_render was declared unsupported on ezdxf, every drawing_export_pdf(layout=...) response carried the note "projecting model content through viewports is COM-only", and the README table agreed — so a caller with no AutoCAD was told to go get one for something that already worked. Measured on ezdxf 1.4.4: a sheet holding nothing but a viewport renders the model geometry that viewport looks at, two viewports at 1:1 and 1:2 render the same circle at two sizes in the right places, and geometry outside a viewport's window is correctly left out. Understating a capability is the same class of untruth as overstating one, and it is the more expensive one for the user. What is genuinely missing is the viewport border, which the headless renderer does not draw — the capability now says exactly that (reason: viewport_borders_not_drawn), and the old note's claim that "viewport frames are rendered" was wrong in the other direction. Two tests that pinned the false claim were corrected rather than deleted, and the replacement proves the behaviour by subtraction: the sheet renders less ink once the model geometry is gone.

• Layout names are matched case-insensitively, the way the engine underneath already matched them. ezdxf resolves layout names case-insensitively in get, new, rename, delete and set_active_layout — but names() returns the stored spelling, so every guard written as name in doc.layouts.names() disagreed with the code it was guarding, in both directions. layout_create("SHEET") next to an existing Sheet walked past its own duplicate check and died on a raw DXFValueError from inside ezdxf; layout_set_current("SHEET") reported "Layout not found" for a tab that was right there. Both shipped in v1.4. One resolver now backs create/set_current/delete/rename/copy and viewport_list on both engines, and the current space is stored in the canonical spelling. Re-spelling a layout's own name (Sheet → SHEET) is a rename onto itself rather than a collision; ezdxf's rename refuses that outright, so it goes through a staging name.

• entity_change_space refuses an entity that lives on another sheet instead of half-applying. move_to_layout raises DXFValueError from deep inside ezdxf when the entity is not in the source layout — and it raises after the transform has already been applied, so the failed call left the line rescaled and translated by a viewport it never moved through, still sitting on the wrong sheet. Membership is now checked before anything is written, and the refusal names the space the entity is actually in. One bad handle in a batch no longer aborts the good ones.

• BREAKING (behavior): coordinates crossing the boundary are now WCS. They were not. DXF stores CIRCLE, ARC, LWPOLYLINE, TEXT, INSERT and a dimension's text position in a per-entity frame derived from the entity's extrusion vector — and the string "extrusion" appeared nowhere in this repository, so every such read returned a raw attribute rather than a world coordinate. Mirroring flips extrusion to (0,0,-1) and leaves the stored x unnegated, so after the server's own entity_mirror across the Y axis a circle drawn at (30, 20) reported center [30.0, 20.0] while the same response's bounding_box said min [-35, 15] / max [-25, 25]: the payload contradicted itself by 60 mm, silently, with no error anywhere. point_from_snap returned points on the wrong side of the axis, entity_select_smart found the entity where it wasn't, and entity_offset placed the new circle 60 mm away and grew it when asked to shrink. Reads are now normalised at the single _entity_info_dxf funnel, so all 38 call sites and every consumer of EntityInfo.properties are fixed with them. The write paths (entity_edit_geometry, entity_offset) got the matching inverse — mandatory, not cosmetic: read and write were equally wrong, which made a read-then-write round trip an accidental no-op, and fixing reads alone would have turned that no-op into a real 60 mm move.

  • Not touched, deliberately: LINE, MTEXT, ELLIPSE, SPLINE and POINT carry an extrusion attribute too and were always correct (ezdxf makes their ocs() a pass-through); entity_rotate was never affected — it preserves extrusion (0,0,1). The trigger was entity_mirror and foreign DXF files. A dimension's defpoint is DXF group code 10 and WCS already; only text_midpoint (code 11) is in the entity frame.
  • Tilted planes. An entity whose plane is not parallel to WCS XY cannot be described by an xy pair — a tilted circle of radius 5.0 has a 5.0 × 3.99 footprint. entity_get reports the WCS centre plus plane_normal and omits radius / start_angle / end_angle rather than reporting a number that is not true in the frame the rest of the dict uses, and entity_edit_geometry refuses with capability: "ocs_tilted_plane", pointing at entity_move (which transforms the whole entity and preserves its plane). New capability keys ocs_normalized and ocs_tilted_plane on both backends make the boundary pre-checkable via system_capabilities.
  • Named residuals (in the ocs_normalized reason rather than left silent): the OCS z component is not reported, HATCH boundaries are not reported at all (silence, not a wrong number), and TEXT rotation stays in the entity frame because a mirrored TEXT is mirror-imaged and no single scalar angle expresses that — measured 29.999 reported against a true WCS 150.0.
  • COM: ActiveX returns WCS natively for every point property except AcDbLWPolyline.Coordinates, which is OCS; that one read is now translated. Verified only against a fake COM object — this machine has a live AutoCAD and connecting to it during development is forbidden, so the path ships unit-tested, not hardware-verified. AcDb2dPolyline is deliberately out of scope: its Coordinates may be 3D triples, and changing the stride on a documentation reading alone would risk a path that works today.

• validate_path handed out working handles to C:\Windows. Windows names the same file through several anchors that compare unequal. The backslash spellings were rejected, but pathlib normalises the forward-slash forms into them only after .resolve() — so //?/C:/Windows/win.ini slipped past the pattern check and then failed to match C:/Windows in the system-directory loop. Proven: it returned a handle that os.path.samefile'd the real file. Non-local anchors are now rejected in every spelling. This removes no capability anyone had: the equivalent backslash paths were already refused, so it only makes one policy consistent across spellings.

• The AutoLISP guard read drawing text as code. The symbol denylist scanned inside double-quoted literals, so an engineering server refused its own vocabulary: (setq desc "Max load 500 kg") matched the load pattern and (setq layer "C:CENTER") matched the custom-command pattern. Literals are now blanked before scanning. Every known-dangerous vector still blocks, including (load "malicious.lsp") and (vl-catch-all-apply (quote command) ...), whose dangerous token is outside the quotes.

• The command denylist matched characters, not commands. A word-boundary regex over the whole string refused MTEXT 0,0 10,10 Do not erase this part — annotation text — while missing spellings AutoCAD accepts. SendCommand takes a macro: segments split on newline or ;, and only a segment's first token is a verb. Matching is now verb-scoped with decorators (_, ., -, ') stripped, the set gained the load/run channels it was missing (ARX, VBASTMT, VBAIDE, CUILOAD, SCR, SAVE, EXPORT, ETRANSMIT, …), and the refusal now names the offending verb — previously a user whose annotation was refused had no way to tell why. Single-letter aliases stay out on purpose: E is the ERASE alias but also ZOOM's Extents option, so blocking it would refuse _ZOOM\nE\n, the canonical zoom-extents macro.

• README no longer claims the denylist is a security boundary. It said "blocked by default, with regression tests for known bypass patterns"; the bypasses above had no tests, and no denylist over a command language that loadable modules extend can be complete. It now says what the guard is — a guardrail against an agent accidentally issuing ERASE ALL — and names the actual boundaries: ALLOWED_PATHS, not exposing the server over the network, and the typed tools. The same sentence is in the two tool docstrings, so it reaches the model that calls them. Not changed: blocking INSERT/SAVEAS as command strings while block_insert/drawing_save_as exist was flagged as incoherent and is not — the policy is scoped to the free-text channel, and the typed tools take validated arguments where the raw commands take file paths. The reasoning is recorded in security.py so it is not re-opened.

• config.settings.backend was a control that controlled nothing. _make_backend read os.environ["AUTOCAD_MCP_BACKEND"] directly, so the setting had no production reader at all and monkeypatch.setattr(config.settings, "backend", "ezdxf") — which two tests used — read like a safety guard while doing nothing. On a developer machine with AutoCAD open that meant auto, i.e. attaching to the operator's live unsaved drawing and creating entities in it. Settings.backend is now a read-only property that reads the environment live, _make_backend goes through it, and the test suite pins the headless backend for every test via an autouse fixture (opt out with @pytest.mark.live_backend). Assigning to config.settings.backend now raises AttributeError instead of silently protecting nothing. monkeypatch.setenv — what the correct call sites already used — is unaffected.

• A capability refusal now survives the trip to the client. The backends always knew exactly why they could not do a thing — they raise UnsupportedCapabilityError(capability, message), where capability is a key of the capability map, so system_capabilities answers the same question up-front. None of that reached anyone: FastMCP.call_tool rewraps a non-FastMCPError as ToolError(f"Error calling tool {name!r}: {e}") from e, and while the from e keeps the cause alive in-process (which is why cad_batch could classify refusals), __cause__ does not serialise. A remote client got an English sentence, and telling "this backend cannot" from "your arguments were wrong" meant substring-matching prose. A refusal now returns is_error: true with {"ok": false, "kind": "unsupported", "capability", "error", "tool", "backend"} in structured_content, produced by one new CapabilityRefusalMiddleware — no tool signature changed. Not done by rebasing the exception on FastMCPError: that was measured and changes nothing a client can observe, since neither the payload nor the error flag travels with the exception class. Non-capability errors are byte-for-byte unchanged. UnsupportedCapabilityError moved to backends/base.py (still importable from backends.ezdxf_backend, and the same class object) so both backends raise one type, and the five headless solid_* methods now raise it instead of returning {"ok": false} — same payload, but is_error stops being quietly false on a call that did nothing.

• BREAKING (behavior): cad_batch counted a step that declined by value as a success. A tool that returns {"ok": false, ...} rather than raising was recorded status: "ok" and folded into succeeded, so a batch could report succeeded: 12 over work that never happened — the same class of lie as the DWG bytes and the unverified rollback. Such a step is now status: "error" with kind unsupported (when it names a capability) or refused, counts toward failed, and honours on_error like any other failure. Batches that relied on, say, a no-op drawing_redo passing silently will now stop or roll back; that is the point. Read-only tools are excluded on purpose: validation_check / drawing_critique / drawing_preflight answer {"ok": len(issues) == 0}, so their ok: false means I found something, not I could not run — reading the two the same way would let a read-only check trigger a rollback that destroys the batch's own geometry. The discriminator is the readOnlyHint annotation each tool already publishes.

• BREAKING (behavior): drawing_audit said "0 errors" about documents it had just silently rewritten — and then threw the repairs away. Drawing.audit() is not a read: it fixes every fixable problem in place, so auditor.fixes is a log of mutations already applied. The tool discarded that list entirely, never marked the document dirty — so drawing_close(save=True) saw a "clean" document, skipped the save, and the repairs it had just made were lost — and rendered the errors it did report through str() on an ErrorEntry that defines no __str__, i.e. as <ezdxf.audit.ErrorEntry object at 0x...>. The tool now returns repaired / fixes / fix_count alongside errors / error_count, both as {code, name, message, handle} dicts, and marks the document dirty when it repaired something. The behavioral break: an audit that repairs now leaves unsaved changes, so a following drawing_close(save=True) rewrites the file where it previously did not. That is the repair finally persisting; the tool was already annotated readOnlyHint: false / cost: mutate. Audit can delete entities — an entity whose owner handle points nowhere is removed, not patched — and with the dirty flag fixed, that deletion now reaches disk on the next save where before it was accidentally discarded. ezdxf reports those particular fixes without the entity attached, so handle is null for them; the message names the type and handle in prose. Read fixes before saving. On the COM backend _AUDIT Y likewise repairs, but AutoCAD returns nothing machine-readable over SendCommand, so the counts come back as null with detail: "unavailable" and capability: "audit_detail" — never 0, which would claim nothing was wrong — plus the path of the .adt log. That backend also no longer leaves the user's AUDITCTL permanently set to 1.

• BREAKING (behavior): the ezdxf backend no longer writes DXF bytes into a .dwg name. drawing_save("part.dwg") previously called doc.saveas() and produced a file whose contents did not match its extension — AutoCAD cannot open it, and nothing in the response said so. Both save paths now refuse, driven by the extension, so drawing_save_as("x.dwg", fmt="dxf") is refused too. Refusals carry a machine-readable {"ok": false, "capability": "dwg"}, and dwg is now a declared capability in both backends (ezdxf: unsupported; COM: native). Save as .dxf, or use the COM backend for DWG.

• drawing_open no longer pretends it can read DWG. ezdxf detects format by content, so a DXF renamed .dwg (exactly what the bug above produced) opened "successfully" as a document that does not exist in that format, while a genuine DWG died with a raw OSError blaming the file rather than the backend. The tool now refuses .dwg unless the active backend declares the capability, and its parameter text no longer advertises DWG unconditionally. Workflows holding files produced by the old drawing_save bug must rename them to .dxf; the error message says so.

• TOOL_PROFILE was a silent no-op whenever a tool transform was attached. _registered_tools() read the post-transform tool list, so with a transform in place it saw 2 tools instead of 131 — the lean profile computed 0 disabled tools and system_about reported a collapsed inventory. It now reads the underlying registry.

• ruff format --check . failed on unmodified main. Ruff 0.16 began formatting Python code blocks inside Markdown, so the committed docs/superpowers/plans/2026-07-17-repository-cleanup-readme-benchmark.md started failing the format gate with no source change. Reformatted the file and pinned the CI ruff version so a future release cannot re-break the gate the same way.

• release.yml now runs ruff format --check . alongside ruff check ., matching the CI lint job — the release gate previously skipped formatting.

• Every headless diameter and radius dimension reported the wrong number, at default settings. add_diameter_dim / add_radius_dim measure to the point they are given; leader_length is a text placement. Passing centre + (radius + leader) as that point made the leader part of the measurement, so a true ⌀40 bore was dimensioned 60 with the default leader_length=10, and 108 at 34. Radius callouts were leader_length too large the same way. This is the worst class of defect for this project — a wrong number, silently, on the one part of a drawing that exists to carry numbers — and it shipped in v1.4.0 and earlier. Now the size comes from radius and location only places the text. The live COM backend was never affected: ActiveX's AddDimDiametric takes the two chord points and the leader length as separate arguments, so the two engines had been disagreeing by 2 × leader_length on the same call. Found while building the sheet in the README's hero image, whose own bore callout was wrong. Two checks added to correctness_suite.py (26, was 24); both report fail → pass against v1.4.0 rather than miss → pass, which is what distinguishes a repaired defect from new capability in that lane.

• A HATCH could not be measured, and the refusal listed HATCH as measurable. entity_measure had no HATCH branch, so a hatch fell through to the generic error — whose own text named HATCH among the measurable types. Whichever half was right, the pair could not both be. It now walks the boundary paths and reports the area the hatch actually fills: outer loops minus the islands inside them, 300 rather than 400 on a 20×20 square with a 10×10 island. Nesting depth is computed by containment rather than read off the boundary flags, because DXF marks a path external or outermost and never "two levels down" — and depth is exactly where AutoCAD's three island styles stop agreeing. hatch_style comes back with the number so it can be read against the drawing's own setting. On a curved edge flatten_tolerance is not the accuracy knob it looks like: ezdxf hands boundaries over as cubic Beziers, so tightening it converges on the Bezier's area, ~0.028% above the circle it stands for, which is why exact goes False and stays False.

• boundary_trace and entity_measure disagreed about the same shape. The boundary tools reported an area taken from a copy of the loop flattened at a fixed 0.01 sag while entity_measure read the stored bulges analytically: 139.2177 against 139.2699 for one figure, with nothing in either payload marking one as the approximation. The polyline the boundary tools store carries exact bulges, so the exact answer was always available; both now read it and round to the same six decimals. The flattened figure survives only where it is harmless — ranking candidate loops, where a 0.01% error cannot change which is smallest.

• A circle stored as two bulged vertices measured 0.0. Two semicircular arcs joined into a circle store as a two-vertex closed LWPOLYLINE with both bulges 1.0 — what AutoCAD writes for a JOINed pair of arcs, and what boundary_from_entities builds from any two-arc loop. polygon_area_perimeter treated fewer than three vertices as degenerate and returned zero, when the shoelace over two points is genuinely 0 and the two circular segments are the entire shape. Only zero or one vertex bounds nothing now.

• The v1.4 → v1.5 performance comparison was measured, and it is not flattering. The published perf report was recorded on CPython 3.14 and the new one on 3.11.15, which alone accounts for nearly all of an apparent 2–6× "speedup". Re-running v1.4.0 in a worktree on the same interpreter with the same suite file gives the real picture: entity creation is 28% slower (245 → 315 ms for 2,000 lines) and the 10,000-line roundtrip 22% slower. Fully attributed — EZDXF_CALL_TIMEOUT=0 returns v1.5.0 to 244.8 ms and 1,679 ms, v1.4.0's numbers exactly. The whole cost is the per-call asyncio.wait_for this release wrapped around every headless call so that one hung ezdxf call can no longer wedge a server whose document lock is a single asyncio.Lock. Deliberate trade, documented knob, and on the 1.6 roadmap to win back.

The five items carried into this release from 1.4 are all done: screenshot overlay + handle grounding, ezdxf undo/redo, COM block_create_from_entities, ISO 286 transition/interference holes, and titleblock on paper-space layouts. ISO 286 shafts r/s/t/u are deliberately not done — they need table data nothing here can derive, and the refusal says so rather than guessing.

[1.4.0] — 2026-07-23

Release infrastructure + the roadmap features shipped together. 474 tests, Ruff lint- and format-clean. Tool count: 122 → 131.

Added

• CI restored and extended (.github/workflows/ci.yml): ruff lint/format, Linux test matrix (3.11/3.12) with coverage, a Windows leg running the mocked-COM suite on the platform the COM backend targets, a package job (build + twine check + clean-venv wheel smoke incl. autocad-mcp --help), a Docker build/smoke job, and MCP server.json schema validation.
• Release pipeline (.github/workflows/release.yml): tag v* → test gate → tag/version match check → build → PyPI trusted publishing (OIDC) → GitHub Release with artifacts.
• MCP registry manifest server.json (schema 2025-12-11, name io.github.u-c4n/autocad-mcp, PyPI package) plus the mcp-name ownership marker in the README and an autocad-mcp-pro console alias so uvx autocad-mcp-pro works directly.
• Live competitor benchmark lane: a generic black-box MCP-stdio driver (benchmarks/adapters/mcp_stdio.py), reproducible pinned checkouts (benchmarks/competitors_env.py), and adapters for puran-water/autocad-mcp and beiming183-cloud/AutoCAD-MCP that run the same 10-task matrix over stdio with harness-side DXF verification. Published results (benchmarks/results/published/): autocad-mcp-pro 100.0, beiming183 50.0, puran-water 45.0; rendered to docs/assets/autocad-mcp-livebench.svg. Weekly/dispatch CI workflow.
• Tool profiles (TOOL_PROFILE=lean|core|full): capability-aware discovery applied in the server lifespan — lean is a curated ~46-tool drafting core, core hides raw escape hatches and long-tail tools, full (default) exposes everything. Reported by system_about.
• Paper space / layouts: layout_list, layout_create, layout_set_current, viewport_create (scaled model viewports) on both backends, and drawing_export_pdf(layout=...) for plotting a layout. Headless limitation is explicit: viewport model-content projection is COM-only (viewport_render capability).
• ISO 286 limits and fits (engineering/fits.py): authored table data (IT4–IT11, sizes 1–500 mm; shafts d/e/f/g/h/js/k/m/n/p, holes D/E/F/G/H/JS) with fit_lookup("H7", 20.0); dimension_linear/radius/diameter gained a fit parameter that resolves deviations from the measured nominal and appends the fit code to the dimension text. Out-of-scope letters raise a clear error naming the supported set.
• Opt-in native 3D solids (ENABLE_3D=true, COM only): solid_box, solid_cylinder, solid_extrude, solid_revolve, solid_boolean. Hidden from discovery and rejected while disabled; ezdxf reports the honest solid_3d capability boundary (ACIS cannot be generated headlessly).
• Release-consistency test suite (tests/test_release_consistency.py): pyproject ↔ version.py ↔ CHANGELOG ↔ README snapshot ↔ server.json versions, README tool/resource/prompt counts vs live registrations, per-section header counts vs decorators, Dockerfile COPY vs wheel only-include, and the README mcp-name marker.

Changed

• pyproject metadata completed for PyPI: SPDX license + license file, authors, keywords, classifiers, project URLs.
• build_dim_override: deviation/limit modes now pass DIMTM signed (-lower) instead of abs() — double-positive ISO 286 fits (e.g. p6 +0.035/+0.022) render correctly; the legacy "positive magnitude = minus deviation" contract is unchanged.
• Capability maps: paper_space is now native on both backends; new viewport_render feature (COM native / ezdxf unsupported); solid_3d reflects the ENABLE_3D gate on COM.
• README rebuilt around the live-run benchmark lane; CI/PyPI badges added; ruff format applied repo-wide and enforced in CI.

Fixed

• Dockerfile shipped a broken image: engineering/ and version.py were never copied, so engineering tools failed in containers. Both are now copied (plus README/LICENSE for metadata) and the healthcheck imports the engineering package.
• server.py section-header tool counts and the CLAUDE.md inventory were stale (six sections drifted); both now match the live surface and are locked by tests. Removed the reference to the non-existent .claude/skills/ directory.

[1.3.0] — 2026-07-17

Closed-loop quality, production annotation, and auditable delivery. 415 tests, Ruff lint-clean. Tool count: 116 → 122.

Added

• drawing_preflight normalizes production requirements, reports missing or conflicting facts before geometry starts, and emits a deterministic SHA-256 spec hash that drawing_plan can enforce.
• drawing_refine runs a bounded critique → repair → re-critique loop. Each repair round is isolated in its own transaction and rolls back if score regresses or hard-error count increases. Construction, duplicate, layer color/lineweight, untrimmed endpoint, and dimension-overlap repairs are supported; undefined GD&T datums remain explicitly manual.
• TABLE and MLEADER semantics via entity_create_table and leader_create_mleader: native ActiveX entities on COM and deterministic, portable LINE/LWPOLYLINE/MTEXT composites on ezdxf. Composite creation returns child handles and a logical ID; after DXF reopen, inspection uses the persisted standard child entities rather than a native TABLE/MLEADER object.
• system_capabilities and shared typed backend capability maps distinguish native, rendered, composite, snapshot, shared, and unsupported features.
• drawing_deliver creates DXF/PDF/PNG bundles, runs validator + critique, re-opens the canonical DXF for entity/type/layer/bounds parity, hashes every artifact, and writes manifest.json plus validation.json. Failed gates keep artifacts for diagnosis but never report delivery success.
• Benchmark v2: ten fixed vendor-neutral tasks, an adapter interface, an ezdxf/COM reference adapter, cooperative per-task timeout reporting, fixed-matrix coverage-aware scoring, machine-readable statuses, runtime/environment/capability metadata, and artifact SHA-256.

Changed

• ezdxf transaction snapshots are isolated from user-facing undo history, so a refiner rollback cannot consume an unrelated undo entry.
• Package metadata is the canonical version source used by system_about and delivery manifests.
• benchmarks/correctness_suite.py now shows usage for argless/--help calls instead of raising a traceback.

Fixed

• Install from source — pip install -e ".[full]" failed at "Preparing editable metadata" because the flat layout has no directory matching the project name, so hatchling could not infer which files to ship. pyproject.toml now declares the wheel file selection explicitly (server.py, config.py, security.py, backends/, engineering/). (#2)
• dimension_linear on the COM backend — always crashed with <unknown>.AddDimLinear: the AutoCAD ActiveX API has no AddDimLinear method. Linear dimensions are now created via AddDimRotated (identical argument order), with a mocked regression test. (#3)

[1.2.0] — 2026-07-07

Production-ISO parity + a measurable quality moat. 360 tests, ruff-clean. Tool count: 111 → 116.

Added

• drawing_settings — read or change common AutoCAD drawing settings by friendly name (units mm/cm/m/inch/feet, linear/angular precision, LTSCALE, DIMSCALE, text size, point mode/size, OSMODE, fillet radius) without memorising system-variable names. Call with no argument for a full snapshot. A convenience facade over system_get_variable / system_set_variable; cross-backend.
• In-place editing — entity_edit_text re-labels or resizes an existing TEXT/MTEXT entity, and entity_edit_geometry re-drives a CIRCLE/LINE/ARC (center, radius, endpoints, arc angles) — both preserve the entity handle, so the user can adjust a drawing without delete-and-recreate. Cross-backend.
• 2D GD&T (ISO 1101 / ASME Y14.5) — gd_frame draws a feature control frame (all 14 geometric characteristics, ⌀ zone prefix, Ⓜ/Ⓛ/Ⓢ material modifiers, multi-datum references) and datum_feature places a datum symbol. Frames are composed from LINE + TEXT so they render identically on both the COM and ezdxf backends (ezdxf's native TOLERANCE entity renders blank via matplotlib). No competing CAD MCP or surveyed text-to-CAD product ships 2D GD&T authoring.
• GD&T datum-consistency gate — a new gdt critique focus fails drawing_finalize when a feature control frame references a datum with no matching datum feature (a meaningless FCF per ISO 1101).
• ISO 129 dimension tolerances — dimension_linear / dimension_radius / dimension_diameter gain tol_upper / tol_lower / tol_mode (symmetric ±, deviation +a/-b, limit stacked, basic boxed) and a text_override (e.g. ⌀20 H7). Toleranced production dimensions are now possible — previously the single biggest functional gap.
• Scalar drawing-score + invalidity ratio (engineering/scoring.py) — the drawing_finalize payload now carries a 0-100 score, an invalidity_ratio, and an A-F grade over the union of the structural validator and the premium critique, so drawing quality is regression-trackable (MUSE / CadBench grade an Invalidity Ratio, not shape).

Fixed

• Honesty: the COM backend advertised a false all_entity_types capability (it authors 2D entities only, no 3D solids) — corrected to entities_2d.

[1.1.1] — 2026-06-20

Added

• selection_get (Entity Query, COM backend): reads AutoCAD's implied ("pickfirst") viewport selection — the entities the user highlighted with grips before invoking the AI — and returns their handles + EntityInfo. This lets the AI scope work to exactly what the user picked (dimension_auto(selection_get()["handles"])) instead of acting on the whole drawing. Resolves #1 — the layer-juggling workaround is no longer needed. Surfaces the PICKFIRST sysvar state so an empty selection is self-explanatory. The ezdxf headless backend has no viewport, so it returns ok=False with an empty handle list (same shape, never raises). Tool count: 110 → 111.

[1.1.0] — 2026-06-19

Correctness, cross-backend parity, and an enforced quality gate, landed across four audited sprints (see docs/analysis/). 318 tests, ruff-clean.

Added

• Premium drafting workflow (shared across both backends): drawing_plan, drawing_critique (ISO-128 focuses), point_from_snap, drawing_apply_iso_layers, dimension_auto, entity_select_smart, construction_xline / construction_clear.
• Deterministic geometry for exact OSNAP coordinates: point_intersection (line/line, line/circle, circle/circle) and point_tangent (external point → circle).
• Engineering / deterministic CAD layer: involute gear front view + section A-A, DIN 6885 keyed bore, ISO A3 title block, and the 8-step drawing_finalize validator.
• HTTP bearer-token auth: StaticTokenVerifier wired into FastMCP when MCP_AUTH_TOKEN is set.
• Mocked-COM test harness (tests/test_com_backend.py + Sprint-3/4 suites) — COM logic is now regression-tested on Linux CI without a live AutoCAD.
• entity_create_mtext rotation parameter; dimension_auto layer override; bounding_box / ARC length / TEXT-MTEXT rotation+char_height on both backends; BlockInfo.description.
• Security module (security.py), centralized config (config.py), .env.example, path validation, command/LISP sanitization, ruff config, pre-commit hooks, pytest-cov.

Changed

• drawing_finalize now enforces the premium critique in addition to the structural validator: leftover construction geometry, non-ISO-128 lineweights, untrimmed corners, duplicate entities, and dimension overlap block the gate (was advisory-only). strict_critique=True fails on warnings too.
• drawing_save_as derives the on-disk format from the file extension — part.dxf writes DXF, not DWG. ezdxf refuses to mislabel a .dwg file.
• Dimension and construction layers resolve from the active layer set (iso13567 → M-DIMEN-T-N / M-CONST-E-N, not the hardcoded DIM / CONSTRUCTION).
• Premium meta-tools lifted into the shared AutoCADBackend base class (single source of truth).
• Dead code removed (section.py, generate_tooth_profile, CommandResult, set_layer_active, …).

Fixed

• ezdxf dimension_aligned / dimension_angular raised TypeError (wrong ezdxf 1.4 args) — fixed.
• Full-circle polar array placed a duplicate copy over the original — fixed (divisor = count).
• entity_offset ignored side_x/side_y on both backends — now honored (and COM no longer leaks extras).
• COM entity_create_hatch built an associative hatch then deleted its boundary — now non-associative.
• Lineweight mm-vs-hundredths truncation wiped ISO-128 weights — fixed via normalize_lineweight.
• view_screenshot / drawing_export_pdf could SIGSEGV (matplotlib GUI backend in a worker thread) — now render headless via Agg (Figure + FigureCanvasAgg).
• entity_select_smart length_range silently rejected all ARCs (no length) — fixed.
• Gear tooth profile self-overlapped for high tooth counts (z ≥ ~42); section view drew duplicate bore lines; validator keyway heuristic was a permanent false-negative — all fixed.
• COM system_run_lisp always reported "nil"; system_set_variable didn't coerce numeric sysvars — fixed. system_about tool groups / _registered_tool_count no longer drift or surface -1.
• drawing_new bootstrap failures now surface as degraded instead of reporting success.

Security

• HTTP remote-bind guard now fires on every launch path (including fastmcp run server.py:mcp), not only the __main__ block — closing an anonymous-remote-bind gap.
• AutoLISP allowlist bypass-vector regression tests (newline injection, symbol aliasing, vla*/acet-*, c: custom commands).
• COM apartment leak bounded (CoUninitialize on teardown); transaction commit/rollback and system_run_lisp now respect the CMDACTIVE guard.

[1.0.0] — 2026-03-01

Added

• Initial release with 67 tools, 6 resources, 5 prompt templates
• Dual-engine architecture: COM backend (live AutoCAD) + ezdxf backend (headless)
• FastMCP 3.0 server with middleware stack (error handling, audit, timing, logging)
• Drawing management: new, open, save, save-as, export DXF/PDF, purge, audit, undo/redo
• Entity creation: line, circle, arc, polyline, rectangle, text, mtext, hatch, spline, ellipse, point, block reference
• Dimensions: linear, aligned, angular, radius, diameter
• Entity modification: move, copy, rotate, scale, mirror, offset, delete, rectangular/polar array
• Layer management: create, delete, set current, modify, freeze/thaw, lock/unlock, hide/show, isolate
• Block operations: list, insert, explode, attributes, create from entities, find references
• Analysis: entity stats, region search, distance/area measurement, bounding box, select by layer/type
• View control: zoom extents/window, screenshot
• Transaction support: begin, commit, rollback
• System tools: status, variables, command execution, AutoLISP evaluation
• 5 prompt templates: floor plan, P&ID, electrical schematic, mechanical drawing, quick drawing