PaperlayerView exact revision on arXiv

Source-linked AI summary

DeciCAD: Harness-Grounded Learning of Action Policies for CAD Agents

Yufeng Peng, Yixin Zhang, Ziyi Liu, Zhikun Xu, Jiarong Yang, Sijie Wang, Wenjie Feng, Qiyu Kang

arXiv:2610.21755v1cs.AIcs.CLcs.LGcs.RO

TL;DR

CAD agents make long-horizon decisions over persistent designs, and the harness determines which actions and state information they use. DeciCAD provides semantic CAD actions, persistent state, decision-oriented observations, and feedback, with a learning pipeline using SFT and geometry-grounded online RL. Across four CAD benchmarks, it improves geometric accuracy while reducing output tokens relative to a coding harness.

  • Problem

    CAD agents operate through state-dependent sequences, while harnesses can represent actions and observations differently; how to preserve needed decisions without irrelevant representation burden remains an open design question.

    Source: §1 · pp. 1–2

  • Method

    DeciCAD grounds semantic CAD actions in persistent state and structured feedback, then learns policies from causally conditioned trajectory supervision with SFT and geometry-grounded online RL.

    Source: §1 · p. 2

  • Results

    Across four CAD benchmarks, DeciCAD produces more accurate final geometry than competitive CAD agents and a general-purpose coding harness while reducing model-output tokens by 45–88% on text-to-CAD tasks.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The findings show that a harness shapes both the CAD operations a model can execute and the distinctions its policy must learn.

    Source: §5 · p. 8

Abstract

from arXiv · show

Large Language Model agents operate through closed-loop action policies whose decision space is defined by an agent harness. We investigate how this operating space shapes agent behavior and policy learning in computer-aided design (CAD), a professional domain characterized by persistent state, geometric constraints, and long-horizon procedures. We instantiate this perspective in DeciCAD, a domain-native harness that exposes semantic CAD actions, decision-oriented observations, environment feedback, and persistent state transitions. Building on this harness, we develop a learning pipeline that extracts causal decision supervision from interaction trajectories and improves the policy through supervised fine-tuning (SFT) and geometry-grounded online reinforcement learning (RL). Across four CAD benchmarks, DeciCAD achieves the best F1 and intersection over union (IoU) while reducing model-output tokens by 45--88% relative to a general-purpose coding harness. Under matched supervision, its SFT policy improves F1 from 0.6883 to 0.7437 and outperforms the coding policy; subsequent online RL further raises geometry availability from 66.5% to 73.5% and F1 to 0.7880. These results demonstrate that the decision space induced by a harness materially affects both agent execution and policy learning in CAD.

1 Introduction

CAD agents make sequential, state-dependent decisions, so their harness shapes what they can do, observe, and learn. DeciCAD uses semantic actions and decision-oriented feedback, and the reported results show stronger geometry performance with lower output-token use.

1 Introduction

  • The harness determines which CAD operations a model can express, which state it can inspect, and how execution outcomes shape later decisions.The same model, task, and backend can therefore be presented with substantially different decision spaces.Source: §1 · pp. 1–2
  • A useful CAD harness must reduce irrelevant implementation choices while preserving the actions and observations needed for correct decisions and error recovery.Over-compression can hide state differences or remove operations required for modeling and recovery.Source: §1 · p. 2
  • The Semantic Decision Bottleneck compresses behaviorally equivalent implementations and irrelevant state while preserving decision-relevant actions and observations.It is proposed as a design principle for the Harness-Induced CAD Decision Space.Source: §1 · p. 2
  • DeciCAD grounds semantic CAD actions in a persistent document and returns structured feedback with decision-oriented observations.Its learning pipeline reconstructs causal action supervision, uses SFT for initialization, and supports online RL with geometry-grounded rewards.Source: §1 · p. 2
  • Across four CAD benchmarks, DeciCAD produces more accurate final geometry than competitive CAD agents and a general-purpose coding harness, while reducing model-output tokens by 45–88% on text-to-CAD tasks.Matched training settings also yield stronger long-horizon CAD action policies in DeciCAD than in code-level interaction.Source: §1 · p. 2

2 Related Work

Related work studies agent interfaces, CAD generation, interactive modeling, and learning from interaction. This paper connects these directions by treating the harness as a joint action–state representation that affects both execution and policy learning.

2.1 Agent Interfaces and Decision Abstraction

  • Prior tool-learning and interactive-agent work shows interface relevance but generally treats interfaces as tool provision or runtime infrastructure, unlike this paper’s focus on the harness-defined policy space.The paper frames the harness as the joint action–state representation presented to the model.Source: §2.1 · p. 2

2.2 CAD Generation and Interactive Agents

  • Prior CAD work spans modeling histories, language-conditioned command sequences, executable code, and interactive systems with rendering, repair, or feedback.Code is expressive but makes program organization and API realization part of the model output.Source: §2.2 · p. 3
  • Interactive CAD systems place decisions in evolving design state; this paper complements them by comparing semantic-action and coding harnesses under fixed models, tasks, and terminal geometry.The cited systems include CAD-Assistant, ToolCAD, and CADMate.Source: §2.2 · p. 3

2.3 Policy Learning from Interaction

  • Because action validity in stateful environments depends on prior trajectories and feedback, the paper aligns causal supervision, student rollouts, and environment feedback within one decision space.This alignment supports a controlled study of how representation affects imitation and online learning.Source: §2.3 · p. 3

3 Method

The method formalizes how a CAD harness shapes the actions, observations, and feedback available to an agent, then applies this framework to DeciCAD and harness-grounded policy learning. Its design compresses implementation choices and visible context while retaining information needed for action selection and recovery.

3.1 From a CAD Backend to a Model-Facing Decision Process

  • A harness induces the model-facing decision space by mapping model actions to backend operations and returning observations and feedback; different interfaces can create different sequential prediction problems over the same CAD kernel.The model's visible history includes the task, previous model actions, and harness responses, but not the hidden backend state.Source: §3.1 · p. 4

3.2 Semantic Decision Bottleneck

  • The Semantic Decision Bottleneck favors compact actions and context that preserve necessary operations and state distinctions for correct action selection and error recovery.DeciCAD instantiates this principle with domain-specific action schemas, concise execution feedback, and targeted inspection.Source: §3.2 · pp. 4–5

3.3 DeciCAD: A Domain-Native CAD Harness

  • DeciCAD uses semantic CAD actions, persistent state-grounded execution, and decision-oriented responses to expose task-relevant design transitions and execution evidence.Targeted inspection actions provide additional state information when needed, rather than repeatedly supplying a complete state dump.Source: §3.3 · p. 5, Semantic CAD Actions · p. 5, State-Grounded Execution · p. 5

3.4 Harness-Grounded Policy Learning

  • The learning pipeline reconstructs causal history-action pairs from accepted teacher trajectories, retains recovered failures as corrective supervision, and stratifies decisions by task difficulty, design stage, preceding outcome, and action type.Online updates use current-policy rollouts with evaluable terminal geometry; geometry quality supplies the primary group-relative signal, while execution and protocol evidence alone cannot make an all-failure group trainable.Source: §3.4 · p. 5

4 Experiments

The experiments test whether DeciCAD preserves CAD capability while reducing representation burden, and whether policies in its decision space can be learned and improved. Results compare harnesses across benchmarks, examine action and observation design, and evaluate supervised and online policy learning.

4.1 Experimental Setup

  • The evaluation covers four CAD benchmarks of increasing calibrated difficulty, with seven relative difficulty quantiles per benchmark and geometric metrics computed on valid assets.The policy-learning comparison uses the Text2CAD/DeepCAD train–test split; geometry availability measures the percentage of tasks entering geometric evaluation.Source: §4.1 · p. 6

4.2 Decision Sufficiency and CAD Capability

  • DeciCAD matches or exceeds the strongest coding harness in geometric fidelity while producing evaluable assets across all four benchmarks.Table 1 compares the same frontier model across four benchmarks; token counts are descriptive and not ranked.Source: §4.2 · p. 6, Table 1 · p. 6
  • At the hardest quantile of every benchmark, DeciCAD exceeds Claude Code; it leads 14 of 28 dataset–quantile combinations overall.Its F1 decline is flatter on P3D-Text-to-3D, indicating a more stable advantage under complex textual constraints.Source: §4.2 · p. 6, Figure 4 · p. 7

4.3 Validating the Semantic Decision Bottleneck

  • Claude Code’s output grows with interaction length, whereas DeciCAD’s output remains comparatively stable, with the reduction attributed to semantic action representation rather than early termination.Claude Code’s log-linear slopes are approximately 0.022–0.069, compared with approximately −0.001 to 0.015 for DeciCAD.Source: §4.3 · p. 7, Figure 5 · p. 7
  • On P3D-Text-to-3D, more raw state degrades geometry, while removing active inspection or coarsening failure semantics hides distinctions needed by subsequent actions.The observation ablation compares structured decision-oriented observation with coarse-failure, raw-state, and execution-feedback-only variants.Source: §4.3 · p. 7, Table 2 · p. 7

4.4 Harness-Grounded CAD Policy Learning

  • Under matched supervision, the DeciCAD policy produces more evaluable assets and better geometry across all reported metrics than the coding policy.SFT also substantially improves the coding policy, which acquires program construction, execution, and repair behavior.Source: §4.4 · pp. 7–8, Table 3 · p. 8
  • Online feedback improves DeciCAD across availability and all three geometric metrics, while the coding policy does not achieve the same gains under the matched inference budget.The evaluation uses matched training and evaluation conditions.Source: §4.4 · p. 8, Table 4 · p. 8
  • DeciCAD SFT and GRPO reduce Invalid Reference and No Progress events, while coding GRPO increases No Progress from 18.9 to 30.4 events per 100 decisions.In DeciCAD, SFT reduces those categories from 18.0 and 33.4 to 4.6 and 21.8, and GRPO further lowers them to 2.1 and 8.8.Source: §4.4 · p. 8, Figure 6 · p. 8

5 Conclusion

The paper concludes that harness design shapes both the CAD operations agents can execute and the distinctions their policies must learn. It also identifies scope limits in the model, backend, dataset, and evaluation measures, and outlines the appendices’ supporting theory, implementation, and evidence.

5 Conclusion

  • DeciCAD instantiates the Semantic Decision Bottleneck through semantic actions, persistent state, and decision-oriented feedback; the study concludes that harness design is central to CAD agents.The comparison is limited to one CAD backend and one frontier model, while policy learning uses one student model on one dataset; mesh metrics omit editability, constraint satisfaction, and feature-tree validity.Source: §5 · p. 8

A Overview of the Appendix

  • The appendix connects the decision-space argument to implementation and evidence, covering theory, DeciCAD’s realization, evaluation details, and the Agentic CAD system.The overview includes the official FreeCAD integration and supported backends.Source: A Overview of the Appendix · p. 11

B Theoretical Foundations

  • The theoretical section characterizes decision sufficiency through expressible actions and retained state distinctions, and favors compression only within the decision-sufficient region.Its framing connects partial observability with value-preserving state abstraction.Source: B Theoretical Foundations · p. 11

B.1 Decision Sufficiency

Decision sufficiency characterizes when a harness can support an optimal policy using only model-visible history. The theorem requires both an executable optimal operation at each reachable state and observations that preserve a common optimal action across states consistent with a history.

B.1 Decision Sufficiency

  • Action sufficiency requires the model-facing action space to retain an executable realization of an optimal backend operation at every task-reachable state.It need not retain every backend operation or surface form.Source: §B.1 · p. 11
  • Observation sufficiency requires hidden states consistent with a visible history to share a common optimal model action.This lets the policy select an action that is optimal across the states it cannot distinguish.Source: §B.1 · p. 11
  • A harness satisfying action and observation sufficiency admits a policy based only on visible history that attains the optimal return at every reachable decision point.The harness therefore preserves the optimal return available through the CAD backend.Source: §B.1 · p. 11
  • If observation sufficiency fails, a policy using only visible history must be suboptimal for at least one aliased state, with expected loss depending on state frequency and value gap.If action sufficiency fails, no exposed action can attain the optimal return at that reachable state.Source: §B.1 · p. 12

B.2 Semantic Decision Bottleneck

The Semantic Decision Bottleneck compares harness representations that preserve decision sufficiency, favoring compression only when it retains the actions and observations needed for optimal CAD decisions. It defines a semantic compression frontier rather than a unique or shortest interface.

B.2 Semantic Decision Bottleneck

  • Decision-preserving compression merges actions only when their bound operations have the same next-state distribution and reward, and identifies histories only when the resulting states share an optimal action.The compressed harness must also retain an executable representation of every consequence needed for optimal decisions.Source: §B.2 · p. 12
  • The Semantic Decision Bottleneck is the frontier where any further semantic compression violates action or observation sufficiency.Different action vocabularies or observation organizations may be incomparable; the frontier is neither a unique interface nor necessarily the shortest one.Source: §B.2 · pp. 12–13
  • Semantic redundancy can increase token use, but semantic compression and token compression are not equivalent, so the evaluation considers performance, observation ablations, and token consumption.Matched SFT and online RL experiments test whether the induced representation improves policy learning in practice.Source: §B.2 · p. 13
  • Automated exploration of the harness–policy Pareto frontier is left for future work, with deployment settings potentially assigning different weights to input, output, and inspection costs.The objectives are evaluated over the task distribution and policy-induced trajectories.Source: §B.2 · p. 13

C Implementation Details

DeciCAD implements a closed loop in which semantic actions are grounded in persistent CAD state and paired with compact feedback and targeted inspection. Its design coordinates action and observation compression while retaining the decision paths needed for task completion and recovery.

C.1 Semantic CAD Action Space

  • DeciCAD's semantic action vocabulary covers design operations from sketches and solid features to assembly and inspection, while collapsing behaviorally equivalent implementations and retaining geometry-determining parameters.Its compositional actions support long-horizon modeling policies within one semantic action space for the task distribution considered.Source: §C.1 · p. 13

C.2 Grounding Semantic Actions in CAD State

  • A shared spatial contract, stable object references, and verified postconditions ground each action in the current persistent CAD document.Valid transitions commit; failed actions roll back or preserve the prior state, with execution outcomes reported to the model.Source: §C.2 · pp. 13–14

C.3 Decision-Oriented Environment Observations

C.4 Implementing the Semantic Decision Bottleneck

  • DeciCAD combines semantic actions, state grounding, passive feedback, and active inspection to compress model-facing representations subject to action and observation sufficiency.The model generates decisions rather than expanded implementations and observes evidence for the next decision rather than the full CAD representation.Source: §C.4 · p. 15
  • The implementation separates the harness runtime, interface contracts, and backend execution layer so CAD backends can change while preserving the model-facing decision space.The components allow interaction policy, interface, and CAD execution logic to evolve independently.Source: §C.4 · p. 15

D Supplementary Experiments

The supplementary section completes the evaluation protocol and adds analyses beyond those reported in the main text. It retains the frozen tasks, models, and evaluator used in the main experiments.

D Supplementary Experiments

  • The supplementary section provides protocols and evidence not shown in the main text, including difficulty calibration, per-quantile geometry, and the geometry–context trade-off.It also includes qualitative assets, supervision construction, online-RL design, and a per-task trajectory example.Source: D Supplementary Experiments · p. 15
  • Its additional analyses examine harness effects on CAD capability, interaction representation, and policy learning.The main text reports aggregate geometry results, difficulty trends, output scaling, observation ablations, policy performance, and learned behavior.Source: D Supplementary Experiments · p. 15
  • All supplementary analyses retain the frozen tasks, models, and evaluator from the main experiments.Source: D Supplementary Experiments · p. 15

D.1 Complete Experimental Setup

The experimental setup spans four CAD benchmarks, compares DeciCAD with coding and specialized-agent baselines, and specifies shared geometry evaluation and policy-learning protocols.

Benchmarks and baselines

  • Four CAD settings cover parametric part modeling, complex text-conditioned geometry, and multimodal assembly.The benchmarks are CADPrompt, challenging subsets of CADInstruct and P3D-Text-to-3D, and P3D-Assembly-3D.Source: §D.1 · p. 15
  • The comparison includes zero-shot CadQuery, CADSmith, CAD-Assistant, Claude Code, and DeciCAD, with harness-level methods using Claude Opus 4.7 and identical task inputs and geometry evaluation.DeciCAD exposes semantic CAD actions and object-level feedback, whereas Claude Code provides a general coding harness.Source: §D.1 · p. 15

Difficulty calibration

  • Tasks are scored by a calibrated difficulty combining semantic complexity, reference-mesh complexity, and description ambiguity, then divided into seven equal-frequency quantiles within each dataset.The interaction term raises difficulty when complex geometry is also described ambiguously; quantiles express relative, not cross-benchmark absolute, difficulty.Source: §D.1 · p. 16

Geometry and execution evaluation

  • Geometry evaluation aligns scale-normalized meshes with ICP and computes CD and surface F1 from 10,000 sampled points, using 0.01 for the F1 threshold and IoU voxel resolution.Availability is measured separately from F1, IoU, and CD, which are computed only over valid assets; stratified end-to-end F1 assigns zero to missing, unparsable, or invalid outputs.Source: §D.1 · p. 16

Policy-learning protocol

  • Policy learning uses the Text2CAD/DeepCAD 800/200 split, with the same student, teacher, tasks, environment, evaluator, and optimization settings across harness tracks.The final test set is excluded from trajectory selection, balancing, hyperparameter adjustment, reward design, and checkpoint selection; GRPO starts from each harness's selected SFT checkpoint.Source: §D.1 · p. 16, Table 6 · p. 16

D.2 Extended Harness Comparison

Across calibrated difficulty levels and benchmarks, DeciCAD retains strong geometry performance while using less model-visible context than Claude Code on text-to-CAD tasks.

Difficulty-stratified performance

  • DeciCAD leads 14 of 28 dataset–quantile cells and exceeds Claude Code at Q7 on every benchmark.It ranks first in five of seven P3D-Assembly-3D quantiles and leads from Q5 through Q7 on CADInstruct; on P3D-Text-to-3D, F1 is 42.7 versus 30.5 at Q4 and 42.4 versus 35.4 at Q7.Source: §D.2 · p. 17
  • The advantage is not confined to easy tasks: DeciCAD is flatter on P3D-Text-to-3D and exceeds Claude Code at the hardest quantile of every benchmark.CADInstruct is non-monotonic, so difficulty is a stratification variable rather than a deterministic predictor of failure.Source: §D.2 · p. 17

Geometry–context trade-off

  • DeciCAD attains higher F1 while consuming substantially less input context than Claude Code on the three text-to-CAD benchmarks.P3D-Assembly-3D enlarges context for both harnesses, yet DeciCAD retains a competitive performance–cost trade-off.Source: §D.2 · p. 17

Output scaling with interaction length

  • The output reduction does not arise from early termination, as interaction turns and cumulative output are compared at task level.The passage reports this analysis in Figure 5 in the main text.Source: §D.2 · p. 17

Decision allocation

  • DeciCAD allocates a larger share of observable decisions to Modeling, while Claude Code allocates more to Planning and Verification combined.The allocation pattern holds across datasets and difficulty levels.Source: §D.2 · p. 17

Qualitative comparison

  • Representative DeciCAD assets more consistently preserve requested profiles, openings, and repeated features than the compared agents.The qualitative examples are from P3D-Text-to-3D and CADPrompt and match aggregate F1, IoU, and Availability trends.Source: §D.2 · p. 17

D.3 Harness-Grounded SFT

The SFT pipeline trains each next action from its preceding interaction history, using audited trajectories from coding and DeciCAD harnesses. DeciCAD SFT improves geometry metrics and asset availability.

Causal decision supervision

  • Each teacher action is paired with the model-visible history immediately before it, including prior actions, tool results, and available tool schema.The target is the teacher's next coding response or semantic CAD action.Source: §D.3 · pp. 17–18
  • When a trajectory contains a local failure, the failure remains in the history and supervision targets the subsequent recovery decision.The construction follows the principle that interactive policies learn from histories induced by preceding actions, not future trajectory information.Source: §D.3 · p. 18

Data composition

  • The DeciCAD dataset contains 642 audited tasks, split into 542 training and 100 development tasks; the Coding dataset contains 499 geometry-valid tasks, with 471 for training and 28 for development.DeciCAD tasks retain up to 11 regular and three corrective decisions; the Coding set yields 3,651 training targets and 207 development targets.Source: §D.3 · p. 18, Table 7 · p. 18
  • Both tracks share source tasks, teacher and student models, causal truncation, context limit, and optimization protocol, while retaining the natural granularity of their decision spaces.DeciCAD examples are stratified by difficulty, CAD stage, document state, preceding action and outcome, history pattern, and target action class.Source: §D.3 · p. 18

Checkpoint selection

  • Checkpoints are saved every 25 optimizer steps, and the final checkpoint is selected on frozen development tasks using closed-loop Availability, F1, IoU, and CD.Training and token-level validation loss monitor optimization.Source: §D.3 · p. 18

SFT outcomes

  • DeciCAD SFT raises Availability from 61.5% to 66.5%, F1 from 0.6883 to 0.7437, and IoU from 0.6487 to 0.6631, while reducing CD from 0.01023 to 0.00782.Coding SFT raises Availability from 12.5% to 60.5% and F1 from 0.5902 to 0.6460, while CD changes from 0.00972 to 0.01130.Source: §D.3 · p. 18

D.4 Geometry-Grounded Online RL

This section presents geometry-grounded online RL for improving DeciCAD’s CAD metrics and valid-asset availability, alongside a persistent workflow for creating and managing editable CAD documents. It also covers the multi-agent control plane, backend contract, and FreeCAD integration, including worker-isolated FreeCAD processes and bridges; the SDK supports SolidWorks and AutoCAD, while experiments use FreeCAD.

D.4 Geometry-Grounded Online RL

  • Online RL uses a completion-and-recovery stage followed by a geometry-oriented stage, with geometry-first reward, gated GRPO updates, and a frozen-parent constraint.The second-stage reward weights geometry, process, and protocol terms at 0.94, 0.05, and 0.01; updates require nonzero geometry reward and reward standard deviation of at least 0.02. Small features in long-horizon tasks remain limited by terminal reward resolution, and similar scores can mask differences in feature count, topology, repeated structures, and constraints.Source: §D.4 · pp. 19–20
  • On a shared task, DeciCAD GRPO completes modeling, persistence, export, and inspection to produce geometry consistent with the reference, unlike the incomplete or degraded comparison trajectories.Coding SFT produces no completed trajectory; Coding GRPO produces a thin plate, while DeciCAD SFT leaves no evaluable asset after repeated failed Sketch actions. The example shows that recovery counts alone do not establish useful recovery.Source: Figure 13 · p. 20, §D.4 · p. 20
  • DeciCAD GRPO raises Availability from 66.5% to 73.5%, F1 from 0.7437 to 0.7880, and IoU from 0.6631 to 0.7036, while reducing CD from 0.00782 to 0.00570.Coding GRPO raises Availability from 60.5% to 65.5%, but F1 falls from 0.6460 to 0.6253, IoU from 0.6286 to 0.5987, and CD increases from 0.01130 to 0.02253. The results associate stable geometry improvement with how credit propagates through the harness-induced decision space.Source: §D.4 · p. 20

E DeciCAD: An Interactive Agentic CAD System

  • DeciCAD is a terminal-native application that brings natural-language interaction, project resources, CAD backends, and streamed execution into one persistent workspace.It supports long-horizon multi-agent projects, a backend SDK, and an official FreeCAD integration.Source: E DeciCAD: An Interactive Agentic CAD… · pp. 20–21

E.1 From Natural Language to Editable CAD Documents

  • Users can request modeling, revision, inspection, or export, attach project resources or CAD skills, and control runtime settings through slash commands.Actions and results appear in temporal order; successful calls update the persistent document, failures inform recovery, and subsequent actions continue from the current state. The native document remains editable for inspection and continued engineering work.Source: §E.1 · pp. 21–22

E.2 Control Plane for Multi-Agent CAD Work

  • An application-level control plane coordinates agents, tools, permissions, resources, and project state across compatible CAD systems.It authorizes writes only against observed current revisions, rejects stale writes, and serializes overlapping scopes while allowing disjoint work to run concurrently; its event stream supports replay and session restoration.Source: §E.2 · pp. 22–23

E.3 Adapting a CAD Backend with the DeciCAD SDK

  • The DeciCAD SDK defines canonical tools and typed contracts so backend integrations share schemas, validation, resource declarations, and output formats while mapping operations to native CAD APIs.The application handles orchestration and permissions, while backend developers implement native operations and report resource effects; contract and health checks verify compatibility before activation.Source: §E.3 · pp. 23–24

E.4 FreeCAD as the Reference Backend

  • The reference FreeCAD backend uses an isolated backend process and authenticated bridge to execute CAD operations in FreeCAD’s GUI thread while exposing the stable SDK contract.Hierarchical inspection retrieves selected document details, and modeling tools return compact results and resource effects while preserving FreeCAD’s editable feature history.Source: §E.4 · p. 24
  • Each worker uses a separate FreeCAD process, Addon bridge, and backend process, isolating mutable CAD state across tasks while enabling parallel services and evaluation.Distinct endpoint descriptor files provide independent loopback ports and authentication tokens; the isolation retains the interactive product’s backend contract.Source: §E.4 · p. 25

E.5 Additional Backends

  • SolidWorks and AutoCAD also support the shared SDK contract, with backend-specific tools loaded on connection and orchestration shared across systems; all experiments use FreeCAD.Table 9 summarizes capabilities across the three backends.Source: §E.5 · p. 25, Table 9 · p. 25

Related papers

Loading 2610.21755v1…