Skip to content

About

Let AI agents take a design from brief to proof — does it behave, can we make it, what does it cost

Topics

Resources

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Repository files navigation

Casys Digital Thread

Casys Digital Thread is an experimental, local-first engineering control plane. It turns reviewed project intent into traceable technical artifacts and bounded evidence through registered operations: deterministic engineering tools compute, humans retain consequential decisions, and the Workbench projects persisted state read-only.

Status: active research / alpha. The repository contains implemented contracts, registered engineering operations, a native read-only Workbench, and dated local pilot evidence. Coverage remains partial. A successful tool run is not by itself a whole-product, manufacturing, conformity, release, or certification verdict.

The engineering loop

Human intent and consequential decisions
                  |
                  v
Agent proposes registered operations
                  |
                  v
Server selects parsing, lowering, provider, runtime and recovery
                  |
                  v
Exact artifacts -> persisted evidence -> bounded verdict
                  |
                  v
Read-only Workbench projection

The project supports idea-first, CAD-first, and reverse-engineering entry paths. Each path must recover explicit intent before it can make equivalent requirement claims. SysML, geometry, simulation, measurement, evaluation, and manufacturing evidence remain distinct records; one does not silently stand in for another. The current product boundary is described in Product direction.

“Local-first” means that project state, the control plane, the Workbench, and registered local execution paths are designed to live on the operator's machine. It does not mean that every installation is offline: agent providers, ERP systems, container registries, or other reviewed integrations may still use the network.

Authority model

Actor Owns Does not own
Human Intent and consequential decisions Provider envelopes or solver payloads
Agent Proposals, plans, queues, and execution of registered operations Provider, tool, arguments, lowering, runtime, or self-approval
Server Sequences, profiles, parsers, lowering, recovery, and canonical publication Human intent
Workbench Passive GET + SSE projection Commands, MCP authority, or provider credentials

Agents must start with AGENTS.md, then the agent workspace reference. The registered catalogue and backend code are authoritative; UI copy, examples, and internal planning records are not a second authority model.

Choose a path

Goal Start here
Understand the product and its three judgement branches Product direction
Walk through a dated engineering project Walk through an engineering project
Verify a new design from scratch Verify a new design
Walk a new Behave-only wall hook from this checkout Verify a new wall hook from source
Preview the read-only Workbench Preview the native Workbench
Understand the local runtime and ports Local runtime and ports
Find the implementation behind a capability Codebase map
Browse all documentation by purpose Documentation guide

A bounded local demonstration how-to walks a new Behave-only wall hook from this checkout. It is not a product, release, or certification tutorial, and it does not maintain historical wall-hook-wh01, dl04, or dl05 state. Other walkthroughs remain how-to guides because they solve concrete tasks and may depend on explicitly named local or dated evidence.

Validate a source checkout

The repository CI currently uses Deno 2.9.2 and Node.js 24. Docker is required for the provider topology, but not for the contributor gates below.

npm --prefix src/ui ci
deno task verify

This entry command runs fmt, lint, check, check:ui, test, then verify:docs. It does not run verify:evidence, verify:thread:presentation, or verify:task-catalog.

See Validate a source checkout for the full source-validation sequence.

Ordinary Digital Thread start is cold Deno (deno task start). Do not treat docker compose up -d as the atelier start or a clone-only quick start: CapabilityRuntime providers are H1 JIT launch groups, not the root Compose project. The complete provider-backed atelier currently has additional prerequisites: published images must be accessible, local microVM worker images must be prepared, and the optional ERPNext bridge expects its external repository, environment, and Docker network. The local runtime reference and the task-specific how-to guide state the exact boundary for each path.

Repository map

Path Purpose
server.ts, src/ Project control, domain contracts, orchestration, persistence, and projections
desktop/ Native desktop shell and packaged chat/runtime integration
src/ui/ Read-only React + Vite Workbench
docker-compose.yml, config/ Maintainer diagnostic Compose topology and reviewed runtime configuration; H1 JIT groups are not this file
scripts/ Gates, probes, runners, and local serving commands
state/fixtures/ Checked-in, explicitly labelled test and demonstration fixtures
state/local/ Ignored local project, Thread, CAS, and execution state; never commit it
docs/how-to/ Goal-oriented procedures
docs/reference/ Exact contracts, runtime facts, providers, and code locations
docs/explanations/ Product and architecture reasoning
docs/annex/project-dossiers/ Dated, non-authoritative observations from engineering projects

Engineering provider servers are maintained in separate repositories and run here from reviewed images. Related public building blocks include mcp-syson, mcp-build123d, mcp-calculix, mcp-modelica, mcp-spice, and constraint-solver.

Contributing and security

Read CONTRIBUTING.md before opening a change. Security issues belong in the private reporting path described in SECURITY.md, not in a public issue. Before changing repository visibility, complete the public release checklist.

License

Copyright © 2026 Casys AI.

This repository is licensed under the GNU Affero General Public License v3.0 only (AGPL-3.0-only). Provider images, external systems, bundled runtimes, and third-party dependencies retain their own licences and notice obligations; this repository's licence is not an aggregate licence claim for them.

About

Let AI agents take a design from brief to proof — does it behave, can we make it, what does it cost

Topics

Resources

Contributing

Security policy

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages