Noyce IDE：面向 DO-178C 认证固件的 AI 原生工作台

The AI-native workbench for firmware that has to certify.

Requirements, source, DO-178C evidence, hardware tooling and fifteen specialist agents —
in one Code-OSS window, against the code that is actually open.

_{macOS (Apple silicon) · Windows x64  |  DO-178C · ISO 26262 · MISRA C:2025  |  Local or cloud models}

_{Recorded in the running IDE on a real STM32G431 motor controller. Real clicks, real data.}

See it work  ·  DAL A run  ·  What's new  ·  Every surface  ·  Install  ·  Architecture

Why Noyce

Safety-critical firmware is usually spread across a requirements manager, two static analysers, a traceability spreadsheet, a coverage tool, a CI dashboard, a chatbot and a vendor IDE. None of them knows what the others measured, so the certification evidence gets stitched together by hand at the end.

Noyce keeps it in one place.

See it work

Short clips recorded in Noyce IDE 2.0.6 on the H-Bridge project. Click any frame to play it.

▶ Watch the full 95-second demo

More clips: coupling, document reader, firmware memory, HardFault

A real DAL A run

The STM32G431 H-Bridge motor controller was taken from prototype to a DO-178C DAL A document set inside the IDE. Every fix that run demanded shipped in 2.0.6. The result is reported the way an auditor would read it — including what is still open.

%%{init: {'theme':'dark','themeVariables':{'pie1':'#4ade80','pie2':'#facc15','pie3':'#6b7280','pie4':'#374151','pieStrokeColor':'#111','pieOuterStrokeWidth':'0px','pieSectionTextColor':'#111'}}}%%
pie showData
    title Annex A · 71 Level A objectives
    "Satisfied" : 5
    "Partial" : 36
    "Open" : 27
    "N/A" : 3

_{Still open on this project: low-level requirements · system requirements · target and timing tests · independent human review and SQA · coupling analysis · object-code analysis · qualified build · authority engagement.}

The rules the IDE holds itself to

A tool that did not run claims nothing. Coverage that was analysed but never measured, CBMC with zero properties and sample fallbacks all count as did-not-run.

Zero findings counts only after a scan. A clean result needs a scan that actually ran.

AI earns no independence credit. Reviews record author and reviewer from the project roster; AI-assisted reviews are labelled and never satisfy independence.

What's new in 2.0.7

What's new in 2.0.6

Full release notes →

Every surface

Every screenshot is captured from the running Code-OSS Electron build with the H-Bridge project open. Nothing here is a mockup. Expand a section to browse it.

🏁  Getting started  _{1 surface}

Recent projects, project creation and import (STM32CubeIDE, TI Tiva/CCS, MPLAB X, generic folder), and one-click access to every surface.

🤖  AI  _{3 surfaces}

AI Orchestrator: a kanban board of specialist agents

Objectives become cards routed to specialist agents along a handoff chain. The header reports the live model, its context window, concurrency, sprint progress and items awaiting review. Cards show what actually ran — tokens, rate, evidence — and never claim work that did not happen.

🔗  Requirements and traceability  _{6 surfaces}

Requirements & Evidence

The requirement register read from the life-cycle documents and @req annotations, with a per-requirement evidence ledger linking design, source, tests and reviews.

🛡️  DO-178C compliance  _{8 surfaces}

Compliance Dashboard

DO-178C objectives for the selected level, graded from mapped evidence: requirements linked, design allocation, verification, analysis findings and structural coverage, with the Annex A objective table below. Objectives route to their specialist agent, and one action assembles the signed Certification Evidence Package.

Compliance Run: one reproducible record

A run freezes the configuration (commit, tree state, level), executes every evidence producer and writes a hashed, append-only record. Runs compare at the level of objectives, so a regression reads as "A-7 objective 5 stopped being satisfied".

Certification Readiness: readiness score, blockers and a ranked closure queue	Evidence Records: verification cases, problem reports, CCB, change sets, MISRA deviations, reviews, roster, tool qualification
DO-178C Documents: the 17-document set, authored, completed and baselined	Document Reader: formal .docx output with cover, TOC and numbered sections
Immutable Audit Trail: hash-chained ledger of every review, decision and export	Review Workflow: reviews with author / reviewer independence

🔬  Verification and analysis  _{11 surfaces}

MC/DC structural coverage, measured

Statement, decision and MC/DC obligations come from the project's own source. Coverage is measured with LLVM -fcoverage-mcdc on the host and attributed to the module under test, and the test environment is stated alongside the numbers.

🧭  Code and architecture  _{6 surfaces}

Project Graph

Functions, files and macros as an entry-flow graph, with code ownership (project / vendor / generated), libraries, callers, callees and macro use for the selected symbol.

🔌  Hardware  _{10 surfaces}
📈  Signals and lab  _{8 surfaces}

_{Surfaces that need a connected target (probe, serial, logic analyser) say so and show labelled simulated data until one is attached.}

The full feature list

Install

Download the installer for macOS (Apple silicon) or Windows (x64) from the latest release, or:

# Homebrew (macOS)
brew tap Hitheshkaranth/noyce https://github.com/Hitheshkaranth/noyce_ide
brew install --cask noyce-ide

# GitHub Packages
oras pull ghcr.io/hitheshkaranth/noyce-ide:2.0.7      # or :latest

Then open the vendor project you already have — STM32CubeIDE, TI Code Composer / TivaWare, MPLAB X, or any folder — and point AI Models at a local server (LM Studio, vLLM, Ollama) or a cloud provider.

Build from source

The source lives in Hitheshkaranth/noyce_ide. Prerequisites: Node 22, Python 3, Git, and on Windows the Visual Studio Build Tools with the VC v142 Spectre libraries.

npm ci
npm run build                       # webview bundle (tsc + vite)

# macOS / Linux, with an already-prepared Code-OSS checkout
npm run workbench:launch -- /path/to/firmware

# Windows, first time
npm run codeoss:bootstrap:windows   # prepare third_party/code-oss
npm run workbench:launch

npm run workbench:doctor (or workbench:doctor:windows) diagnoses missing prerequisites.

npm run check          # build + cargo check on the Rust sidecar
npm run test:all       # unit, integration and quality-invariant suites
npm run smoke:ui       # Playwright smoke (CI gate)

Architecture

flowchart LR
    FW["Firmware project<br/>CubeIDE · CCS · MPLAB"] --> IDE
    subgraph IDE["Noyce IDE · Code-OSS + Electron"]
        direction TB
        UI["React 19 surfaces<br/>shadcn/ui · Tailwind 4"]
        AG["15 specialist agents<br/>Orchestrator · RAG"]
        RS["Rust sidecar<br/>import · analysis · probe-rs · serial"]
        UI <--> AG
        UI <--> RS
    end
    IDE <--> M["Your models<br/>LM Studio · vLLM · Ollama · cloud"]
    IDE --> T["Real tools<br/>CBMC · CodeQL · LLVM · cppcheck"]
    IDE --> N[".noyce/<br/>documents · evidence · audit ledger"]

Loading

The project stays clean. Everything the IDE generates — evidence, documents, the audit ledger, analysis scratch — goes under <project>/.noyce/, which carries its own .gitignore. The IDE never edits your files unless you approve an agent change set.

Distribution

This repository ships the installers and the noyce-ide.tech website. The source is in Hitheshkaranth/noyce_ide.

License

Proprietary. © 2026 Noyce IDE. All rights reserved.

_{Built for engineers who ship firmware that has to be right.}