Reported issues for AnkusDrive

Pod holds 17 of 20 GitHub reports that passed its relevance review. This can include external user reports, maintainer-confirmed bugs, and concrete feature gaps. Treat them as evidence to inspect, not a count of distinct defects.

Back to AnkusDrive.

Most discussed

Agent loses airtight-shell invariant across iterative geometry edits

Context

While using the DriftPin MCP server to model a Makita router → Dewalt shop-vac adapter through an LLM agent (Claude Code, Opus 4.7), I observed a repeated failure mode that I don't think is solely the agent's fault — it's a class of problem where DriftPin's tool surface does not push the agent toward verifiable geometric invariants.

Full transcript: https://gist.github.com/mpetne/16dc4d2119ce41a24faf3d2c449ee79d

Failure pattern across three iterations

Each iteration "fixed the…

Read the thread · 2026-06-04 · closed · external user · 5 comments

macOS: no SU2 case builder, so "CFD degrades to SU2" is still aspirational (#237 item 3)

Found by the agent-eval run of 2026-07-31 (scratch/agent_evals/). Affects the "prove it with a CFD solve" workflow (#222) on macOS specifically.

Problem: on macOS, no built-in CFD case can actually solve — but every affordance says CFD is available

Verified live on Apple Silicon with the Multipass VM provisioned (OpenFOAM v2512 installed) and running:

solve_capabilities reports the cfd family any_available: true (SU2 resolves natively), and docs/MACOS.md says *"CFD alone…

Read the thread · 2026-08-01 · closed · 2 comments

Modeling: pocket/hole carry the #282 degenerate-cut warnings but have no `strict` escalation

Source: follow-up from the #282 implementation (the audit it asked for found this and deliberately left it).

Problem

#282 taught the subtractive path to name its two degenerate outcomes. _h_boolean_op got the full treatment — removed_volume, volume_ratio, a warnings list, and a strict: true param that upgrades annihilation/miss to an error for scripted recipes (driftpin/worker.py:2790, driftpin/worker.py:2834).

The audit found pocket/hole share the defect and they…

Read the thread · 2026-08-22 · closed · 1 comment

add_primitive silently ignores name= — every test that names a primitive is naming nothing

Found incidentally while building #234: a two-sided test asserting "a hand-modelled part called M6Screw must NOT receive a standard-part designation" failed, because the object was actually still labelled Cylinder.

The problem

@handler("add_primitive") in driftpin/worker.py accepts a name parameter by convention (callers pass it, and other add_* handlers honour it), but the handler contains no name or Label handling at all. The value is accepted and discarded, and the…

Read the thread · 2026-08-02 · closed · 1 comment

em_fullwave waveguide_sweep: returns ok:true with nulls when port placement exceeds guide length

Environment (reproduced): AnkusDrive 0.5.5 (main @ 17a90aa), macOS arm64, Python 3.14, openEMS source build 0.0.36.post1.dev206. Runner ankusdrive/em_fullwave_gpl_runner.py driven directly over the stdin/stdout sentinel protocol. openEMS itself is fine; these are problems in how the runner sets up the case and in what the gate measures.

Summary

Ports sit at 10*mesh_res–15*mesh_res from each end of the guide. When cells_per_wl is small, mesh_res grows until the port planes…

Read the thread · 2026-09-14 · closed · 0 comments

em_fullwave dipole_s11: mesh resolution is coupled to f_stop, so the resonance depends on the sweep window

Summary

In _dipole_s11, mesh_res = lam0/30 with lam0 = C0/f_stop. f_stop also sets the Gaussian excitation (f0=(f_start+f_stop)/2,…

Read the thread · 2026-09-14 · closed · 0 comments

em_fullwave dipole_s11: radius_mm and gap_mm are inert — both are below mesh resolution

Summary

_dipole_s11 accepts radius_mm and gap_mm, but the mesh never gets extra lines at either feature, so neither one changes the result.…

Read the thread · 2026-09-14 · closed · 0 comments

em_fullwave: fc_ratio compares the port's analytic cutoff to itself — it is not an FDTD gate

Summary

_waveguide_sweep reports fc_ratio = fc_crossing_ghz / fc_analytic_ghz. The runner docstring, the em_fullwave_submit handler and the…

Read the thread · 2026-09-14 · closed · 0 comments

Most recent

setup_status / doctor install hints are wrong inside the MCPB bundle

Found while verifying the MCPB bundle (#346) in Claude Desktop, 2026-09-12.

The defect

setup_status and ankusdrive doctor recommend commands like:

pip install 'ankusdrive[mbd]'       # mujoco
pip install 'ankusdrive[topology]'  # topopt/solidspy
pip install 'ankusdrive[optics]'    # rayoptics + optiland

When AnkusDrive runs from the .mcpb bundle, none of these can work:

The server's interpreter is a uv-managed .venv inside Claude's extension directory…

Read the thread · 2026-09-12 · closed · 0 comments

Tests: test_mcp_boot.py skips on the self-hosted Linux lane, so #279's boot check has never run there

Source: found while wiring test_mcp.py into CI for #288 — reading the job logs rather than trusting the green checkmarks.

Problem

tests/test_mcp_boot.py is the test #279 was written to produce: proof that an interpreter can serve MCP over stdio, with FreeCAD deliberately unresolved, so "MCP is broken" stays distinguishable from "FreeCAD isn't found". On the self-hosted Linux lane it has never run. From the Test suite (FreeCAD 1.1, self-hosted) job:

SKIP test_mcp_boot — the…

[Read the thread](https://github.com/gchen19/AnkusDrive/issues/317) · 2026-08-30 · open · 0 comments

### fem_run blocks the MCP channel — route it through the submit→poll job pattern

`fem_run` is the last long-running solve that still blocks the MCP channel. Its own docstring says so:

ankusdrive/mcp_server.py:3512 def fem_run(analysis, workdir="/tmp/ankusdrive_fem") -> dict: """Run the CalculiX solver on an analysis. Blocks until the solve finishes.


Every other heavy family has already moved to the submit→poll job pattern — `study_submit`, `optimize_submit`, `cfd_*_submit`, `molding_fill_submit`, `acoustic_fem_submit`, `em_fullwave_submit` — with `job_status`…

[Read the thread](https://github.com/gchen19/AnkusDrive/issues/308) · 2026-08-30 · closed · 0 comments

### Install: Windows first-run path failed for a real user (Python 3.14, manual README steps) — harden the quickstart

**Source:** external install report (2026-08-13, Windows 11 Home build 26200, Python 3.14).

## What happened

The reporter followed the README steps manually on Windows: installed FreeCAD fine, but "had some issues with the MCP step" and eventually **gave up on the local install entirely** — downloaded the repo zip and had Claude Cowork build a cloud Ubuntu 24.04 / Python 3.11 sandbox instead. The local Windows install never succeeded.

Contributing factors identified or suspected:

1. **mcp…

[Read the thread](https://github.com/gchen19/AnkusDrive/issues/279) · 2026-08-13 · closed · 0 comments

### Install: `driftpin doctor` should preflight the MCP server import, not just FreeCAD/solvers

**Source:** external install report (2026-08-13).

## Problem

In the field report, `driftpin ping` and `driftpin doctor` both **passed** while `driftpin mcp` was completely broken (mcp 2.0.0 ImportError — see companion issue "unpinned mcp dependency"). The CLI paths never import the `mcp` package, so `doctor` certified an install whose entire reason for existing — serving 240+ MCP tools to an LLM client — did not work.

This is the silent-default defect class again: an absence (working MCP…

[Read the thread](https://github.com/gchen19/AnkusDrive/issues/278) · 2026-08-13 · closed · 0 comments

### Bug: unpinned `mcp` dependency — mcp 2.0.0 breaks `driftpin mcp` with `No module named 'mcp.server.fastmcp'`

**Source:** external install report (2026-08-13, cloud sandbox: Ubuntu 24.04, Python 3.11, DriftPin 0.4.0, FreeCAD 1.1.3 AppImage).

## What happens

`pyproject.toml` declares `"mcp>=1.2"` with no upper bound (`pyproject.toml:54`). A fresh `pip install` today resolves **mcp 2.0.0**, whose server module was restructured — `mcp.server.fastmcp.FastMCP` no longer exists. Result:

- `driftpin mcp` fails immediately: `ImportError: No module named 'mcp.server.fastmcp'` (import at…

[Read the thread](https://github.com/gchen19/AnkusDrive/issues/277) · 2026-08-13 · closed · 0 comments

### slice_estimate MCP tool omits density_g_cc/filament_dia_mm — same unfollowable error as #238

Found by the wrapper-drift audit done while fixing #238 (PR #263). Same failure mode, same layer, different tool — and the only other hit in an AST sweep of all 281 MCP tools.

## Problem

`driftpin/analysis/slicing.py::slice_estimate` takes `filament_dia_mm: float = 1.75` and `density_g_cc: float | None = None`. The worker handler at `driftpin/worker.py:13614` forwards `**p`. The FastMCP wrapper at `driftpin/mcp_server.py:6713` exposes **neither** — it forwards a fixed eight kwargs.

So…

[Read the thread](https://github.com/gchen19/AnkusDrive/issues/264) · 2026-08-04 · closed · 0 comments

### add_primitive silently ignores name= (44 call sites naming nothing)

Found while building #234's two-sided designation test ("a hand-modelled part called `M6Screw` must NOT get a designation") — the test failed because the object wasn't actually called that.

## Problem

`@handler("add_primitive")` in `driftpin/worker.py` never reads `p.get("name")`, and the `add_primitive` MCP tool in `driftpin/mcp_server.py` has no `name` parameter at all. Callers passing `name=` get no error and no name — the object's `Label` stays `Box` / `Cylinder` / `Sphere`.

44 call…

[Read the thread](https://github.com/gchen19/AnkusDrive/issues/247) · 2026-08-02 · closed · 0 comments

### cost_estimate MCP tool omits price_usd_kg/density_kg_m3 — its own error message is unfollowable

Found by the agent-eval prep of 2026-07-31 (`scratch/agent_evals/`): a direct MCP call `cost_estimate(volume_mm3=30429, material="aluminum", ...)` fails with

ValueError: no price for 'aluminum'; pass price_usd_kg


## Problem

The error's instruction cannot be followed over MCP. `driftpin/analysis/cost.py::cost_estimate` accepts `price_usd_kg` (and `density_kg_m3`) overrides, and the worker handler forwards `**p` — but the FastMCP wrapper's signature…

[Read the thread](https://github.com/gchen19/AnkusDrive/issues/238) · 2026-08-01 · closed · 0 comments

The remaining reports are on [the project's issue tracker](https://github.com/gchen19/AnkusDrive/issues).

Reported issues for AnkusDrive

Most discussed

Agent loses airtight-shell invariant across iterative geometry edits

Context

Failure pattern across three iterations

macOS: no SU2 case builder, so "CFD degrades to SU2" is still aspirational (#237 item 3)

Problem: on macOS, no built-in CFD case can actually solve — but every affordance says CFD is available

Modeling: pocket/hole carry the #282 degenerate-cut warnings but have no strict escalation

Problem

add_primitive silently ignores name= — every test that names a primitive is naming nothing

The problem

em_fullwave waveguide_sweep: returns ok:true with nulls when port placement exceeds guide length

Summary

em_fullwave dipole_s11: mesh resolution is coupled to f_stop, so the resonance depends on the sweep window

Summary

em_fullwave dipole_s11: radius_mm and gap_mm are inert — both are below mesh resolution

Summary

em_fullwave: fc_ratio compares the port's analytic cutoff to itself — it is not an FDTD gate

Summary

Most recent

setup_status / doctor install hints are wrong inside the MCPB bundle

The defect

Tests: test_mcp_boot.py skips on the self-hosted Linux lane, so #279's boot check has never run there

Problem

Modeling: pocket/hole carry the #282 degenerate-cut warnings but have no `strict` escalation