GcadClaw

Generate 2D GstarCAD drawings with feedback.

Note on naming. gcadclaw is the external, marketplace-style skill bundle (Windows setup-wizard payload, agents/openai.yaml persona) and is configured to install pygcadwin directly from PyPI. It is distinct from the in-app cadclaw agent in data/graphs/agents_config.json, which is an llm_tool agent driven by the gcad-coder-general and gcad-design2d-general skills. Both can coexist.

Purpose

Create or modify 2D GstarCAD drawings from natural-language requirements using the pygcadwin PyPI package, then prove the result with structured feedback artifacts.

Defaults

Use these defaults unless the user specifies otherwise. Do not ask for clarification just because the user omitted units, "2D", output format, base plane, origin, or common drafting conventions.

• Task type: a 2D GstarCAD DWG drawing generated through pygcadwin.
• Units: millimeters.
• CAD target: live GstarCAD through pygcadwin.Gcad.
• Output: .dwg plus feedback artifacts.
• Base drawing plane: XY model space unless the user requests paper-space layout.
• Up/elevation axis: positive Z for side, front, section, and elevation references.
• Origin: center of the main part or assembly unless a mating interface, pivot, fixed root component, or datum feature suggests a better origin.
• View selection: for ordinary part designs, default to a standard three-view orthographic drawing: top/plan, front/elevation, and right-side view. Replace one side view with a section/detail view only when it communicates the design better. For assemblies, include the top/plan arrangement plus front/side/section views needed to show stack-up, pivots, travel, or mating relationships.
• Geometry representation: closed 2D outlines for visible profiles; centerlines for axes, bolt circles, symmetry, and rotation; hidden/dashed lines only when useful; hatches for cut material in sections.
• Layers: create named layers for outline/body, cut/hole, centerline, hidden, dimensions, text/notes, construction, and optional part-specific layers.
• Dimensions and notes: include critical overall sizes, hole diameters, spacing, bolt circles, thickness/height notes, fillet/chamfer notes, view labels, section/detail labels, part labels, and assumptions.
• Assembly structure: fixed root part or datum, named part outlines/layers, explicit generated placements, pivot axes, mating holes, travel arcs, and labels for separate moving parts.
• Small plastic enclosure wall: 2.0-3.0 mm when unspecified.
• Cosmetic fillet: 1.0-3.0 mm when safe for local geometry.
• M3/M4/M5 normal clearance holes: 3.4/4.5/5.5 mm unless another standard is requested.
• Feedback: mandatory before/after entity evidence plus at least one PNG screenshot.

Ask one focused clarification only when the drawing is impossible, fit-critical, safety-critical, compliance-bound, or lacks any usable dimensions. Otherwise proceed with explicit assumptions and record them in the brief.

Required Workflow

1. Write a brief. Record intent, dimensions, units, assumptions, layers, expected entities, annotations/dimensions, output paths, and validation targets in brief.md.
2. Capture before state. Use scripts/capture_feedback.py or equivalent pygcadwin enumeration to write before_entities.json.
3. Plan actions. Prefer a Python script or pygcadwin.run_actions() sequence that uses typed APIs: ensure_layer, create_segment, create_circle, create_arc, create_polyline, create_rect, create_text, create_dimension, create_hatch, create_table, zoom_extents, save_as, and snapshot.
4. Execute drawing edits. Use pygcadwin.Gcad(create_if_missing=True, visible=True) and cad.context; keep calculations explicit and named before creating geometry.
5. Capture after state. Write after_entities.json and compare entity counts, layers, object names, handles, colors, linetypes, and lineweights against the brief.
6. Capture screenshot. Call ctx.zoom_extents() or a scoped view operation, then ctx.view.snapshot().save(<png>) or the snapshot tool. Verify that the PNG is nonblank. Do not mark a drawing-modifying task complete without a nonblank PNG screenshot.
7. Repair if needed. If validation or screenshot capture fails, classify the issue, make the smallest responsible change, rerun the failed check, and update feedback.
8. Report final feedback. Write feedback.md or feedback.json with files, checks run, screenshot path, repair attempts, assumptions, and remaining caveats.

Non-Negotiables

• Do not claim success from code generation alone.
• Do not skip screenshots for drawing-modifying tasks.
• Do not accept all-black/all-white screenshots as successful evidence.
• Do not use raw CAD commands when a typed pygcadwin API exists.
• Do not invent unavailable pygcadwin API names; inspect the installed package or PyPI project docs when uncertain.
• Do not report validation that was not actually run.

Feedback Artifacts

Every completed drawing workflow must persist:

• brief.md
• actions.jsonl
• before_entities.json
• after_entities.json
• review.png or a more specific PNG screenshot name
• feedback.md or feedback.json
• final .dwg path when save succeeds

Read references/feedback-loop.md before implementing or repairing a feedback-producing workflow. Read references/2d-pygcadwin-workflow.md when translating prose into pygcadwin drawing code.

External Prerequisites

Live drawing, DWG save, and screenshot workflows require Windows with GstarCAD installed and registered as a COM server. Install pygcadwin from PyPI with python scripts/setup_python_env.py; the PyPI package declares its Windows automation dependencies.

Repair Loop

When a check fails:

1. Read the failing output or screenshot capture error.
2. Classify it as source error, COM connection error, missing geometry, wrong layer/style, wrong scale, missing dimension/annotation, screenshot failure, or save failure.
3. Change the smallest responsible source/action.
4. Rerun the failed command and dependent validation.
5. Update feedback.md with the failed attempt and final state.

If screenshot capture fails after one repair attempt, report partial completion and keep entity evidence, action logs, and DWG output available. Do not call the workflow fully successful.

Python package setup

This Skill installs pygcadwin directly from PyPI:

text

https://pypi.org/project/pygcadwin/

When this Skill is used in a fresh Python environment:

1. Validate the import:

   bash

   python scripts/validate_env.py
   

2. If validation fails, install the PyPI package:

   bash

   python scripts/setup_python_env.py
   

3. Validate again:

   bash

   python scripts/validate_env.py
   

The expected import is pygcadwin. The skill does not bundle a local wheel for this package.