cadcli

name	cadcli
description	Inspect, search, summarize, preview, and safely edit DWG/DXF CAD drawings with the cadcli CLI. Use when the user asks about CAD files, floorplans, DWG/DXF contents, layers, blocks, entities, text, previews, or CAD edits.

cadcli

Use cadcli for progressive CAD file discovery. Start broad, then narrow, then render visual state only after you have vocabulary and evidence to guide the view:

cadcli overview <file> --json          # human clues: layers, labels, blocks, hints
cadcli info <file> --json              # size, format, counts, bounds
cadcli query <file> --schema --json    # learn query tables before writing SQL
cadcli search <file> "query" --json    # fuzzy discovery from overview terms
cadcli query <file> --sql "select text, x, y from texts" --json
cadcli entities <file> --limit 20 --json # raw drilldown only when needed
cadcli render <file> -o <name>-renders --diagnose # visual/state bundle
# write/update <file-with-.md> with evidence, inferences, and open questions

When working with a drawing, build a human-readable understanding of what is in it. Do not stop at listing CAD primitives such as layers, blocks, handles, entity types, or raw coordinates. Use those details as evidence, then explain the drawing in the terms a person would use: rooms, entrances, furniture, doors, fixtures, workstations, corridors, core/service areas, labels, structural elements, mechanical/electrical items, or other project-specific objects.

Do this as evidence-based discovery. Say what the file appears to contain and cite the clues that support it: layer names, block names, text samples, entity counts, and coordinates/handles where useful. Separate evidence tiers clearly. Treat text labels, block names, layer names, and counted inserts/entities as direct CAD evidence. Treat proximity, clusters, likely room grouping, gaps in walls, or repeated furniture arrangements as spatial inference. Treat project name, floor number, tenant, building name, or drawing purpose from the filename as filename context. Treat anything learned only from an SVG/PDF/rendered view as visual context. Do not present filename context, visual context, or spatial inference as if the DWG explicitly declared it.

Avoid imposing generic assumptions. Different drawings encode meaning differently: one file may identify rooms through text labels, another through room-tag blocks, another through closed polylines, and another not at all. Follow the file's own conventions and translate the underlying CAD structure into concepts users can understand. Phrases like "appears to", "likely", and "not directly labeled" are appropriate when the evidence is inferential rather than direct.

Persist observations alongside the drawing. For every DWG/DXF you analyze, maintain a Markdown sidecar in the same directory as the drawing unless the user asks otherwise. Use exactly one naming format: replace the CAD extension with .md. For example, building-floor16.dwg gets building-floor16.md; site-layout.dxf gets site-layout.md. Use it as working memory: record commands run, direct evidence, spatial inferences, visual/filename context, open questions, and corrected conclusions. Update it as you learn; do not duplicate huge JSON outputs. Keep the sidecar useful to the next agent: concise, cited, and organized by evidence tier.

Use each command for a different job. overview is for clues and vocabulary. search is for fuzzy discovery when you have a term from the overview or the user. query is for tabular/repeatable questions. entities is for raw drilldown after you know what type or layer matters.

Always run cadcli query <file> --schema before writing SQL for a file. The query engine exposes narrow in-memory tables: summary, metadata, layers, blocks, entities, texts, and inserts.

For a question like "how many rooms are there", use this pattern:

cadcli overview <file>
cadcli query <file> --schema
cadcli query <file> --sql "select id, text, layer, x, y from texts where text like 'Room%'" --json

Adapt the query terms to the drawing's language and labels. For Hebrew drawings, text may already be normalized by cadcli, so a room query may use text like 'חדר%'. For drawings where rooms are block tags rather than text labels, query inserts instead of texts.

If the drawing shows furniture or door layers, inspect their INSERT entities and block names. Use cadcli query for questions like "where are the inserted chair blocks", "which labels start with room", or "which layers have the most entities". If it shows cryptic layers like A2 or A3, infer cautiously from the entity mix and repeated block/text evidence rather than from the layer name alone.

When answering the user, communicate in ordinary language. Use terms like "rooms", "furniture", "doors", "labels", "support spaces", "architectural linework", and "evidence". Do not lead with implementation details like entity types, raw JSON fields, SQL tables, or file internals unless they are needed to justify the answer or the user asks for them.

For visual/layout questions, prefer cadcli render over raw SVG. Start with a diagnostic bundle when the user asks about entrances, adjacency, active areas, room grouping, or anything that requires seeing the drawing:

cadcli render <file> -o <name>-renders --diagnose

Inspect the generated overview.png, architecture.png, no-furniture.png, furniture.png, evidence-marked.png, and evidence.json. The most useful slices are usually: overview for general context, architecture for walls/core/openings, furniture for objects and labels, and evidence-marked for cited labels/blocks. Treat conclusions from these images as visual/spatial evidence, not direct CAD labels.

For targeted visual checks:

cadcli render <file> -o core.png --around-label "Server Room" --radius 3000
cadcli render <file> -o architecture.png --layers "wall,core,A-WALL"
cadcli render <file> -o marked.png --mark-label "Server Room,Electrical" --mark-block "desk,door"

For raw SVG checks, render an SVG preview:

cadcli view <file> -o preview.svg

For lightweight previews from the normalized model:

cadcli svg <file> -o preview.svg

For edits, find handles first, then write an edited copy. Never overwrite by default; use --overwrite only when the user explicitly asks for in-place edits:

cadcli entities <file> --type TEXT --json
cadcli search <file> "office" --json

cadcli edit <file> --set-text 'New label' --text-id <id> -o edited.dwg
cadcli edit <file> --set-layer A-TEXT --layer-id <id> -o edited.dwg
cadcli edit <file> --set-color '#ff0000' --color-id <id> -o edited.dwg
cadcli edit <file> --move <id> --dx 10 --dy 0 -o edited.dwg
cadcli edit <file> --rotate <id> --angle 90 --origin 0,0 -o edited.dwg
cadcli edit <file> --scale <id> --factor 2 -o edited.dwg
cadcli edit <file> --copy <id> --dx 10 --dy 0 -o edited.dwg
cadcli edit <file> --delete <id> -o edited.dwg

cadcli edit <file> --add-line 0,0:10,0 --new-layer A-WALL -o edited.dwg
cadcli edit <file> --add-circle 5,5:2 -o edited.dwg
cadcli edit <file> --add-text 'Label' --at 5,5 --height 2.5 -o edited.dwg
cadcli edit <file> --points '0,0;10,0;10,5' --closed -o edited.dwg

Edit formats: points are x,y or x,y,z; line is from:to; circle is center:radius; arc is center:radius:startAngle:endAngle; color is bylayer, byblock, an AutoCAD color index, #rrggbb, or r,g,b.