Skip to content

v0.1.0-alpha.3

Pre-release
Pre-release

Choose a tag to compare

View all tags
@armpro24-blip armpro24-blip released this 10 Aug 14:51
· 41 commits to main since this release
v0.1.0-alpha.3
4962c5b
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Experimental pre-release. This is the first published release of the
AIENG Workbench line. v0.1.0-alpha.1 / v0.1.0-alpha.2 were internal
drafts that were never tagged or published anywhere; the version history
starts here in public. The Python packages carry version 0.1.0a2.

What this is

An agent-driven CAD/CAE workbench: you connect an MCP-capable coding agent
(Claude Code, Codex, Cursor, …) and drive real 3D CAD modeling, static
structural FEA, and sizing optimization by prompt — the workbench enforces its
boundaries at the tool layer (modeling-plan confirmation, approval gates on
solver runs and destructive operations) instead of asking you to learn a
command vocabulary.

Two Python packages plus a Docker image:

Artifact Where What it contains
ghcr.io/armpro24-blip/aieng-workbench:latest GHCR (published) All-in-one image (backend + built web workbench + CalculiX); every published tag is smoke-validated on main
aieng-format 0.1.0a2 this repo (aieng/); PyPI publication planned The .aieng package format library: Shape IR, schemas, validation, evidence/credibility model, CLI
aieng-workbench-mcp 0.1.0a2 this repo (aieng-ui/backend/); PyPI publication planned The MCP server + FastAPI backend: CAD/CAE/optimization tools, approval gating, web viewer API

Capabilities (evidence-backed)

  • Real CAD, no API key — cad.execute_build123d runs your build123d code
    against the OpenCASCADE kernel and produces actual STEP/STL/GLB geometry,
    with named parts, colors, a 4-view contact-sheet thumbnail, incremental
    append/replace_part/remove_part editing, and a deterministic
    quantitative geometry report (proportions, symmetry, floating parts).
  • Fast parametric edits — cad.edit_parameter replaces a named constant
    and re-executes: sub-second to seconds, no LLM, with a regression_diff
    verdict (clean / collateral / topology-changed) on every edit.
  • Static structural FEA (CalculiX) — setup patch → preflight → deck
    generation → approval-gated solve → result extraction → field regions.
    Meshes default to quadratic tetrahedra (C3D10); every mesh carries a
    measured accuracy band, and a completed run on an unreliable mesh is
    downgraded, not presented as a result (verified against beam theory:
    linear-tet default was ~2× off on stress; quadratic lands within ~3%).
    Modal, buckling, and steady-state thermal analysis types exist with
    explicit linear-analysis honesty boundaries.
  • Mesh convergence — cae.mesh_convergence runs a GCI (Richardson) study
    with per-metric apparent order and verdicts.
  • Sizing optimization with the real solver in the loop —
    opt.sizing_sweep (one parameter) and opt.doe_sizing_study
    (multi-parameter, full-factorial/LHS) solve each variant with real static
    FEA and rank honestly; failed variants are reported, never recommended.
  • Topology optimization — built-in SIMP (2D default, experimental 3D)
    with honest coarse-limitations recording and Shape-IR writeback.
  • Binding durability — CAE loads/constraints bind to @face: pointers
    that now survive dimensional edits, hole cuts, and unrelated part
    replacements     (stable face identity + evidence-based re-verification);
    a genuinely ambiguous change (e.g. a face split in two) refuses honestly
    instead of guessing.
  • Credibility tiers on every result-bearing output —
    critique_finding < surrogate_prediction < proxy_assembly_result < executed_solver_result, with automatic downgrade when evidence does not
    support the claim.

Install

Docker all-in-one (recommended — published on GHCR, bundles the full CAD
stack and CalculiX):

docker pull ghcr.io/armpro24-blip/aieng-workbench:latest
docker run --rm -it -p 8000:8000 -p 8765:8765 -v aieng-data:/data \
  ghcr.io/armpro24-blip/aieng-workbench:latest

MCP server without cloning, straight from this repository (Python 3.11+):

uvx \
  --from "aieng-workbench-mcp[full] @ git+https://github.com/armpro24-blip/cad-cae-copilot.git@v0.1.0-alpha.3#subdirectory=aieng-ui/backend" \
  --with "aieng-format @ git+https://github.com/armpro24-blip/cad-cae-copilot.git@v0.1.0-alpha.3#subdirectory=aieng" \
  aieng-workbench-mcp \
  --approval-mode client \
  --data-dir ~/.aieng-workbench

PyPI publication of aieng-format / aieng-workbench-mcp is planned; the
pip install --pre path activates once the packages are on the index (the
release workflow that publishes them is already in the repository,
owner-gated).

Full wiring for Claude Code / VS Code / Codex: aieng-ui/backend/MCP_SETUP.md.
Prompt phrasing that works: docs/prompt-guide.md.

The optional real-CAD/FEA stack (build123d/OCP, gmsh, CalculiX) is heavy; the
Docker image bundles all of it, while the Python path runs with honest
degradation (stubbed CAD smoke, preflight reports the missing solver) until
you install the extras.

Honesty boundary (read before trusting any output)

.aieng records evidence and context; it does not certify engineering
correctness and does not advance engineering claims automatically. The
workbench is not production-certified CAD/CAE software; every
result-bearing output carries a credibility stamp and production_ready: false unless explicitly certified by a human. Solver claims are only true
when cae.run_solver actually executed and result artifacts exist. Assembly
connections are proxies (no contact physics, no bolt preload). No
stability/semver guarantee at alpha.

Verification at tag time

  • Backend suite: 1701 passed, 0 failed (Windows, full run; ubuntu CI green).
  • aieng core suite: ~3068 passed (full-suite CI).
  • Packaging smokes (installed wheel/sdist for both packages, clean venv): green
    locally on Windows and in CI on ubuntu.
  • Real-ccx verification gate (NAFEMS cases + CAD→mesh→deck→ccx→FRD
    integration): green in CI.
  • Canonical value demo (50 N cantilever: CAD → CAE → sizing): reproducible via
    aieng.value_demo_check, with mesh-convergence step (GCI 1.31%,
    extrapolated tip deflection at 99.2% of beam theory).

Known limitations

  • Real-geometry work needs the optional CAD stack (build123d/OCP); the pip
    package without extras runs the stubbed smoke path only.
  • Assembly CAE is a simplified proxy model (v0): no real contact, no
    preload, solver deck generation best-effort.
  • Mesh-to-CAD reconstruction (topology-optimization writeback) is lossy and
    explicitly not production CAD.
  • No parametric history / constraint solver — edits are source-level
    (named-constant substitution), which is a deliberate alpha scope choice.
Assets 2
Loading