IntentForge

IntentForge is a deterministic CAD intent pipeline for turning simple engineering language into editable, explainable, validated parametric CAD models.

Phase 33 expands the declarative topology registry with spur_gear and standard_bolt alongside the Phase 32 industrial_flange. Family manifests define bounded parameters, supported features, closed adapter identifiers, evidence bindings, rule bindings, and safe arithmetic metric mappings. Geometry remains deterministic compiled Python selected through a closed adapter registry; manifests never execute code.

intentforge assurance build --profile static
intentforge assurance build --profile standard --dry-run
intentforge assurance validate output/assurance/assurance_case.json
intentforge assurance package output/assurance/assurance_case.json
intentforge review policies
intentforge review evaluate output/assurance/assurance_case.json --policy intentforge_standard_design_review_v1
intentforge review provenance output/assurance/review_decision.json --verify
intentforge review diff baseline-decision.json candidate-decision.json
intentforge review verify-offline output/assurance/audit_package
intentforge review cas-check output/assurance/audit_package
intentforge review cas-store output/assurance/audit_package --store output/review-cas
intentforge review chain-verify <head-package> --store output/review-cas
intentforge topology list
intentforge topology validate
intentforge topology schema industrial_flange
intentforge topology schema spur_gear
intentforge topology schema standard_bolt
intentforge topology build-json flange-intent.json

See Engineering assurance cases, Audit packages, Offline verification, Audit portability, Content-addressed audit, Audit chains, Review policies, Acceptance decisions, Decision provenance, Multi-variant differential audit, and Policy checks.

It is not a general text-to-CAD generator. The goal is not to produce geometry that merely looks right once. The goal is to preserve the design intent behind the model so later edits can update named parameters and active features without losing the original assumptions, constraints, and feature history.

The current implementation remains intentionally bounded:

wall_mounted_bracket / mounting plate
l_bracket / right angle bracket
industrial_flange / flat ring pipe flange foundation
spur_gear / involute-approximated external transmission gear
standard_bolt / metric fastener macro-geometry

IntentForge currently uses Python, Pydantic schemas, CadQuery, pytest, deterministic regex parsing, optional MCP wrappers, and an optional LLM intent translator. The deterministic CAD core does not depend on an LLM.

Pipeline

The current pipeline is:

natural language
-> structured intent
-> named parameter table
-> constraint graph
-> feature plan
-> CadQuery model
-> STEP/STL exports
-> validation report
-> edit intent
-> regenerated CAD

Each stage writes structured artifacts so a reviewer or coding agent can inspect what the system inferred, which dimensions are named, which assumptions were made, which optional features are active, and why validation passed or failed.

Related MCP server: Fusion360 MCP Server

Why This Is Different

Normal text-to-CAD demos often generate a one-off model from a prompt. That can work for visual sketches, but it is brittle when the design changes.

IntentForge keeps the editable model state explicit:

• important dimensions are named parameters

• optional features are represented by feature flags

• assumptions and unknowns are recorded

• feature steps include reasons

• validation checks produce structured reports

• edits modify the existing intent and parameter table instead of treating every change as a new prompt

Supported Scope

IntentForge currently registers five deterministic model families:

• wall_mounted_bracket / mounting plate

• l_bracket / right angle bracket

• industrial_flange / flat ring flange with a central bore and polar bolt-hole pattern

• spur_gear / external spur gear with a deterministic involute approximation

• standard_bolt / metric bolt macro-geometry with hexagonal or socket-cap head

Supported features:

• registry-derived JSON schemas with typed parameters and conservative safe bounds

• industrial flange outside diameter, bolt circle, bolt-hole diameter/count, thickness, and bore

• deterministic flange STEP/STL generation, validation, knowledge evaluation, and audit packaging

• spur-gear module, tooth count, pressure angle, face width, bore, derived pitch/root circles, and deterministic STEP/STL generation

• bolt nominal diameter, pitch, shank/thread lengths, closed head selection, stress-area metadata, and deterministic STEP/STL generation

• L-bracket base leg and vertical leg parameters

• L-bracket no holes or two holes per leg

• optional L-bracket triangular gusset

• no mounting holes

• two horizontal symmetric mounting holes

• four rectangular/corner mounting holes

• optional center cutout

• optional rounded corners

• optional edge fillets

• deterministic natural-language edits

• structured edit JSON

• benchmark suite

• optional MCP wrapper

• optional LLM intent translator with schema guardrails

• deterministic engineering knowledge rules and design rationale

• modular engineering knowledge rule packs

• engineering capability coverage and capability matrix

• engineering evidence bundles and deterministic trust reports

• deterministic run-level assurance cases and portable audit packages

• declarative engineering review policies, findings, conditions, and acceptance decisions

• frozen review-decision provenance with deterministic replay verification

• structural pairwise and multi-variant review differential audits

• standard-library-only offline audit-package verification

• canonical platform-neutral package JSON, paths, runtime identifiers, and hashes

• full SHA-256 package/object content addresses and immutable-by-address local storage

• optional predecessor binding across assurance claims, decisions, provenance, and CAS envelopes

• deterministic multi-package chain verification with changed/deleted/switched predecessor detection

• deterministic engineering reasoning over knowledge findings

• golden-case reasoning verification and recommendation consistency checks

• topology-informed feature recognition for supported generated models

• design review reports that summarize intent, parameters, validation, topology, recognized features, warnings, and artifacts

The flange profile is ASME B16.5-oriented only. It does not certify pressure class, material, facing, hub, gasket, loads, manufacturing, or service fitness. The spur-gear tooth profile is an involute approximation and does not certify ISO accuracy, contact, strength, wear, or noise. The bolt thread is represented by a cylindrical major-diameter envelope; no helical flank or strength-grade certification is claimed. External engineering review remains required.

Unsupported by design in this phase:

• arbitrary CAD objects

• unregistered or undeclared model families

• four-hole L-bracket patterns

• freeform L-bracket hole placement

• curved or adjustable L-brackets

• sheet-metal unfold patterns

• robust geometric inside-corner filleting for L-brackets

• required LLM parsing in the deterministic core

• LLM-generated CAD code or direct LLM geometry generation

• GUI

• SolidWorks, Fusion, or FreeCAD desktop control

• freeform hole placement

• circular or diagonal hole patterns

• full industrial CAD feature recognition from arbitrary solids

Installation

Create an environment and install the released package:

python -m venv .venv
. .venv/bin/activate
python -m pip install intentforge

Or install the source tree in editable mode for development:

python -m pip install -e .

IntentForge also supports uv for local development:

uv sync
uv sync --all-extras
uv run pytest
uv run intentforge doctor
uv run intentforge benchmark

Pip remains the standard install path for PyPI users. See docs/development.md for the full development workflow.

CadQuery is optional for non-CAD parser tests, but required for real STEP/STL generation:

python -m pip install "intentforge[cad]"

Optional MCP support is installed separately:

python -m pip install "intentforge[mcp]"

Optional LLM translation can be configured from the interactive client. You do not need to edit files by hand:

intentforge interactive
# then follow the first-run setup prompt, or run:
intentforge config setup

IntentForge also accepts OPENAI_API_KEY and the INTENTFORGE_LLM_* environment variables for non-interactive use. Do not commit real keys.

For the interactive terminal client (Claude Code-like experience):

python -m pip install "intentforge[client]"
intentforge interactive

Run the test suite:

python -m pytest

Check the local development environment:

python -m intentforge.cli doctor

Usage

Parse a prompt into structured intent artifacts:

python -m intentforge.cli parse "Make a wall-mounted bracket 120 mm wide and 60 mm tall with two screw holes."

Parse, build, export STEP/STL, and validate:

python -m intentforge.cli parse-build "Make a wall-mounted bracket 120 mm wide, 60 mm tall, 8 mm thick, with two screw holes."

Build an L-bracket:

python -m intentforge.cli parse-build "Make an L-bracket 100 mm base leg, 80 mm vertical leg, 40 mm wide, and 6 mm thick."

Build an L-bracket with holes on both legs:

python -m intentforge.cli parse-build "Make an L-bracket with two holes on the base and two holes on the vertical face."

Build and validate the bundled bracket example:

python -m intentforge.cli build-example bracket
python -m intentforge.cli validate-example bracket

Parse a natural-language edit:

python -m intentforge.cli edit-parse "Make it 150 mm wide but keep the same thickness."

Parse and apply an edit to the bundled bracket example:

python -m intentforge.cli edit-parse-apply bracket "Change it to four mounting holes."

Parse and apply an edit to the bundled L-bracket example:

python -m intentforge.cli edit-parse-apply l_bracket "Make the base leg 120 mm long."

Rejected edits are reported without exporting new edited CAD:

python -m intentforge.cli edit-parse-apply bracket "Change it to three mounting holes."

Run parse-build without exporting STEP/STL files:

python -m intentforge.cli parse-build "Make a wall-mounted bracket 120 mm wide, 60 mm tall, with two screw holes." --dry-run

Run an edit without exporting edited STEP/STL files:

python -m intentforge.cli edit-parse-apply bracket "Make it 150 mm wide but keep the same thickness." --dry-run

Use the optional LLM translator with the deterministic mock provider:

python -m intentforge.cli llm-parse "Make a wall-mounted bracket 120 mm wide with two screw holes." --mock-provider
python -m intentforge.cli llm-parse-build "Make a wall-mounted bracket 120 mm wide, 60 mm tall, with two screw holes." --dry-run --mock-provider
python -m intentforge.cli llm-edit-parse l_bracket "Add a triangular gusset." --mock-provider
python -m intentforge.cli llm-edit-apply l_bracket "Add a triangular gusset." --dry-run --mock-provider

Without a configured provider, LLM commands return a structured LLMProviderUnavailableError.

Run the parametric sweep harness:

python -m intentforge.cli sweep --max-cases-per-family 30

Run the edit preservation harness:

python -m intentforge.cli edit-harness

Run the adversarial rejection harness:

python -m intentforge.cli adversarial-harness

Run the unified technical harness and quality gates:

python -m intentforge.cli technical-harness --quick
python -m intentforge.cli technical-harness --include-demo

Recognize expected features from generated CadQuery topology:

python -m intentforge.cli recognize-features wall_mounted_bracket
python -m intentforge.cli recognize-features l_bracket

Generate a user-facing design review report:

python -m intentforge.cli design-review wall_mounted_bracket
python -m intentforge.cli design-review l_bracket

Demo

Run the release demo:

python -m intentforge.cli demo

The demo runs parse-build examples, accepted edits, an intentionally rejected vague edit, and the benchmark suite. It writes a traceable demo run under:

output/demo_runs/<run_id>/

The run directory contains demo_report.json, demo_summary.txt, generated CAD outputs, validation reports, edit reports, and benchmark reports.

Benchmark

Run the deterministic regression benchmark:

python -m intentforge.cli benchmark

The benchmark covers parsing, CAD generation, optional features, hole patterns, rejection behavior, natural-language edits, validation, output traceability, and family-level results for wall_mounted_bracket and l_bracket.

Reports are written to:

output/benchmark/benchmark_report.json
output/benchmark/benchmark_summary.txt
output/benchmark/runs/<run_id>/

API Contract

Workflow and MCP outputs use a standard response envelope with ok, request_id, run_id, object_type, operation, artifact references, validation summaries, quality gate summaries, warnings, metadata, and structured errors on failure.

Rejected prompts and edits are recoverable tool responses with ok: false, cad_exported: false, and no standard artifact refs. Compatibility keys such as latest_outputs, persistent_outputs, validation_valid, and message are still preserved.

See docs/api_contract.md.

LLM Translator

The LLM translator is optional. It may translate prompts or edits into structured IntentForge JSON, but it never writes CadQuery code and never directly generates CAD. Schema guardrails reject unsupported families, unsupported geometry, unsupported hole counts, arbitrary coordinates, and vague optimization requests before deterministic workflows can build or edit anything.

See docs/llm_translator.md.

Edit Preservation Harness

The edit preservation harness stress-tests the core differentiator in IntentForge: changing an existing design without losing intent.

It runs multi-step edit chains for both supported families and checks:

• changed parameters are the ones the edit requested

• preserved parameters remain unchanged

• optional feature state stays consistent

• regenerated CAD still validates

• topology inspection and volume delta checks remain coherent when applicable

• rejected edits do not export CAD

Run it with:

python -m intentforge.cli edit-harness

Latest reports are written to:

output/harness/edit_preservation_report.json
output/harness/edit_preservation_summary.txt

Persistent run artifacts are written to:

output/harness/edit_preservation_runs/<run_id>/

Adversarial Rejection Harness

The adversarial rejection harness verifies that unsupported objects, unsupported geometry, vague optimization requests, invalid dimensions, unsupported hole counts, and unsafe fallback prompts are rejected clearly.

It checks that rejected cases include an error message and do not export STEP/STL files.

Run it with:

python -m intentforge.cli adversarial-harness

Reports are written to:

output/harness/adversarial_report.json
output/harness/adversarial_summary.txt
output/harness/adversarial_runs/<run_id>/

Feature Recognition And Design Review

Feature recognition inspects generated CadQuery/OpenCascade topology for supported engineering features: through holes, center cutout, L-bracket solid connectivity, and triangular gusset where practical. It is topology-informed and conservative. Low-confidence findings are reported as warnings instead of pretending to be industrial-grade feature recognition.

Design review reports combine structured intent, parameters, feature plan, validation checks, topology metrics, volume delta checks, feature recognition, warnings, artifacts, and known limitations.

See docs/feature_recognition.md and docs/design_review.md.

Design reviews can also include deterministic engineering knowledge findings and Markdown rationale:

python -m intentforge.cli knowledge list
python -m intentforge.cli knowledge packs
python -m intentforge.cli knowledge packs-validate
python -m intentforge.cli knowledge coverage
python -m intentforge.cli knowledge coverage --json
python -m intentforge.cli knowledge coverage-validate
python -m intentforge.cli knowledge capability-matrix
python -m intentforge.cli knowledge evidence-list
python -m intentforge.cli knowledge evidence-validate
python -m intentforge.cli knowledge evidence-bundles
python -m intentforge.cli knowledge trust-report
python -m intentforge.cli knowledge trust-validate
python -m intentforge.cli knowledge reasoning-info
python -m intentforge.cli knowledge reasoning-verify
python -m intentforge.cli knowledge reasoning-benchmark
python -m intentforge.cli design-review wall_mounted_bracket --knowledge
python -m intentforge.cli design-review wall_mounted_bracket --knowledge --reasoning

The knowledge and reasoning layers are advisory. They encode explainable engineering heuristics, rule interactions, trade-offs, conflicts, priorities, and recommendations. The current rule database is organized into modular rule packs for traceability and packaging safety.

The capability coverage layer makes support claims machine-readable. It records which capabilities are supported, partially supported, or explicitly unsupported; which stages and rule packs contribute; and what implementation and verification evidence backs each claim.

The evidence layer resolves those evidence references into auditable evidence bundles and a deterministic trust report. It measures evidence completeness, unresolved references, boundary rejection evidence, limitation evidence, and deterministic report IDs. It does not add CAD support, run FEA, or produce a vague AI trust score. See docs/design_knowledge.md, docs/knowledge_rule_packs.md, docs/knowledge_coverage.md, docs/capability_matrix.md, docs/evidence_traceability.md, docs/evidence_bundles.md, docs/trust_report.md, and docs/engineering_reasoning.md.

Technical Harness

The technical harness orchestrator runs the benchmark, parametric sweep, edit preservation harness, adversarial rejection harness, volume delta checks, shape inspection checks, feature recognition checks, and engineering reasoning checks as one quality gate suite.

Run a faster local check:

python -m intentforge.cli technical-harness --quick

Run the full harness with the demo workflow included:

python -m intentforge.cli technical-harness --include-demo

Default quality gates require:

• benchmark pass rate >= 0.95

• sweep pass rate >= 0.95

• edit preservation rate >= 0.95

• adversarial rejection success rate == 1.0

• reasoning generation pass rate == 1.0

• unknown reasoning rule references == 0

• recommendation contradictions == 0

• recommendation applicability errors == 0

• reasoning report ID mismatches == 0

• rule pack load pass rate == 1.0

• active rule pack count == 4

• capability manifest validation == pass

• unknown capability references == 0

• evidence manifest validation == pass

• evidence bundle/trust report deterministic mismatches == 0

• supported capability evidence bundles complete

• orphan active engineering rules == 0

• supported capabilities have implementation and verification evidence

• active engineering rule count == 10

• duplicate pack and rule IDs == 0

• unsafe acceptances == 0

• unexpected failures and exceptions == 0

• review policy manifest validation == pass

• deterministic review fixture decisions match expected categories

• review decision and decision-bearing audit-package validation == pass

• review policy reference and deterministic ID mismatches == 0

• review provenance snapshot, execution-node, and replay mismatches == 0

• every review provenance record maps all 65 evidence definitions

• semantic pairwise and multi-variant diff identities remain deterministic

• five reviewed fixture packages pass isolated static-chain verification

• frozen offline package matrices contain 10 rules, 28 capabilities, 65 evidence records, and 54 policy checks

• the five fixture assurance records total 49 claims with no static replay mismatch

• portability violations and Linux/macOS/Windows canonical identity mismatches == 0

• three-block CAS chain validates with zero pointer or object-hash mismatches

• changed, missing, and switched predecessor simulations are all rejected

Reports are written to:

output/harness/technical_harness_report.json
output/harness/technical_harness_summary.txt
output/harness/technical_harness_runs/<run_id>/

HTTP API Server

IntentForge ships an optional local HTTP API server (Phase 15). It is a thin FastAPI layer over the same deterministic workflows, returning contract-compatible ToolResponse envelopes on every endpoint.

Install API dependencies:

python -m pip install "intentforge[api]"

Start the server:

intentforge serve
# or:  python -m intentforge.api.server
# with auth:  intentforge serve --token my-secret-token
# custom host/port:  intentforge serve --host 0.0.0.0 --port 9000

# Environment variables for python -m entry point:
# INTENTFORGE_API_HOST  (default: 127.0.0.1)
# INTENTFORGE_API_PORT  (default: 8765)
# INTENTFORGE_API_TOKEN (optional Bearer auth token)

API docs are at http://127.0.0.1:8765/docs.

Key endpoints:

• GET /health — health check

• POST /v1/parse — deterministic prompt parsing

• POST /v1/parse-build — parse, build STEP/STL, validate (supports dry_run)

• POST /v1/edit-apply — parse and apply edit (supports dry_run)

• POST /v1/llm/parse / /llm/parse-build / /llm/edit-parse / /llm/edit-apply — optional LLM translation

• GET /v1/runs/recent / /runs/{kind}/{run_id} — run metadata

• GET /v1/artifacts/{path} — safe artifact file serving

See docs/api_contract.md and docs/product_demo.md.

Interactive Terminal Client

IntentForge ships an optional interactive terminal client (like Claude Code's experience) with rich colored output, spinners, session tracking, and auto-completion.

Install client dependencies:

python -m pip install "intentforge[client]"

Start the interactive session:

intentforge interactive

On first launch, IntentForge offers to configure optional LLM translation. The setup wizard uses up/down arrow selection in an interactive terminal, and falls back to typed choices in scripts. It supports OpenAI, OpenAI-compatible endpoints, the deterministic mock provider, or skipping LLM setup entirely. Saved settings are written to ~/.intentforge/config.json with owner-only permissions where supported, and API keys are masked when displayed.

Available commands inside the session:

The client works without rich and prompt_toolkit — it gracefully falls back to plain text output and basic input() prompts.

Product Demo Workflow

Two demo scripts show how a real user or external agent would call IntentForge end-to-end via the HTTP API:

# Start the server first:
intentforge serve

# Minimal 5-step API client demo:
python examples/api_client_demo.py

# Full 7-step product workflow demo (parse → dry-run → build → edit → validate → rejection → artifact list):
python examples/product_workflow_demo.py

Both scripts fail clearly if the API server is not running. With auth:

python examples/api_client_demo.py --token my-secret-token

See docs/product_demo.md for the full walkthrough.

Phase 12 adds this harness command only; it does not create a new release tag.

MCP Usage

IntentForge can be exposed as an optional MCP tool server for coding agents:

python -m mcp_server.server

The MCP server is a thin wrapper around existing workflows. It does not duplicate parser, generator, validator, or editor logic. Optional LLM MCP tools only translate structured intent and still pass through schema guardrails.

Available tool functions include:

• parse_cad_prompt

• parse_build_cad_prompt

• parse_edit_prompt

• parse_apply_edit_prompt

• build_example_bracket

• validate_example_bracket

• list_recent_runs

• get_run_metadata

License

License: Apache-2.0. See LICENSE.

Project Structure

demo/           Release demo script and notes
docs/           Architecture, design intent, validation, benchmark, LLM, MCP, and roadmap docs
examples/       Bundled wall-bracket and L-bracket prompt, intent, parameters, constraints, feature plan, and edit examples
output/         Generated artifacts
src/benchmark/  Deterministic benchmark package and bundled prompt data
src/harness/    Topology, volume delta, sweep, edit-preservation, adversarial, and orchestrator harnesses
src/intentforge/ Core schemas, parser, planner, generator, validator, editor, knowledge, assurance, review, workflows, API, reports, and CLI
src/mcp_server/ Optional MCP wrapper around core workflows
tests/          Pytest coverage

Roadmap

Near-term roadmap:

• Phase 10: harden the new L-bracket family while preserving the same intent-first architecture

• add an electronics enclosure family

• add topological feature detection for generated solids

• add an LLM-assisted parser that emits the same structured schemas

• consider GUI or CAD-plugin integration after the core pipeline is more mature

See docs/roadmap.md for more detail.