SiliconWeaver

An agentic, evidence-driven path from hardware intent to verified RTL and open-PDK GDS.

English · 简体中文

SiliconWeaver is an open-source chip agent that connects a user-confirmed hardware specification to verified RTL generation, bounded repair, and managed LibreLane physical implementation. It is designed around explicit contracts, durable run identities, immutable evidence, and fail-closed release gates.

natural-language intent -> confirmed specification -> verified RTL
                        -> managed physical flow -> GDS + QoR + attestations

SiliconWeaver is a research and engineering preview. Its open-source checks are useful evidence, but they are not commercial foundry signoff and do not by themselves establish tapeout readiness.

Many AI hardware demos stop when a model emits syntactically plausible RTL. SiliconWeaver treats generation as the beginning of the workflow:

• Intent is explicit. Interfaces, timing, reset behavior, architectural constraints, requirements, assumptions, and verification policy are bound to a user-confirmed specification.
• Expected behavior is protected. The generated RTL cannot rewrite the frozen oracle, expected outputs, or existing test epochs to make itself pass.
• Repair is evidence-driven. Simulation and formal counterexamples are normalized into bounded feedback, with regression and rollback on every candidate.
• Physical runs are managed. Long LibreLane jobs execute in the background with durable identities, resumable state, QoR extraction, artifacts, and integrity attestations.
• Physical planning is explicit and evidence-bound. Optional tool-driven planning probes a sealed geometry candidate set before the final Lane run; the public default remains off and the bundled bootstrap policy is not publishable calibration evidence.
• Read and write operations are separated. Status and report commands are pure reads; mutations such as start, stop, retry, or reconcile are explicit.

flowchart LR
    U["User intent"] --> P["Pi-Lane interactive agent"]
    P --> S["Confirmed JSON specification"]
    S --> O["Durable workflow orchestrator"]
    O --> R["RTL-Closure"]
    R --> V["Oracle quorum, simulation, formal, lint, synthesis and repair gates"]
    V --> B["Read-only RTL publication bundle"]
    B --> F["Optional fail-closed physical planning"]
    F --> L["lane-agent"]
    L --> C["LibreLane 3.0.5 Classic Flow"]
    C --> E["GDS, QoR, reports, hashes and attestations"]
    O -. "bounded status and durable events" .-> P
    R -. "explicit run identity" .-> O
    L -. "explicit run identity" .-> O

The repository contains two independently versioned Python distributions:

They share a runtime environment but keep their commands, state roots, run identities, releases, and artifacts distinct.

SiliconWeaver deliberately avoids treating every successful command as a successful chip:

See the detailed English Pi-Lane architecture and Chinese end-to-end pipeline reference.

• Linux on x86_64 or aarch64 with Docker;
• Python 3.12 and uv (recommended);
• Node.js 24 and npm for the pinned Pi agent and monitor;
• an OpenAI-compatible model endpoint for RTL generation;
• storage and compute appropriate for LibreLane and the selected open PDK.

The tested compatibility set is:

Clone and install both components into one Python environment:

git clone https://github.com/vpromise/SiliconWeaver.git
cd SiliconWeaver/pi-lane

uv venv .venv --python 3.12
uv pip install --python .venv/bin/python --editable "../rtl-closure[test]"
uv pip install --python .venv/bin/python --editable ".[test]"
npm ci --ignore-scripts --no-audit --no-fund

source .venv/bin/activate
export PI_LANE_PI_EXECUTABLE="$PWD/node_modules/.bin/pi"

Run non-mutating checks before configuring a model or downloading EDA assets:

pi-lane --version
rtl-closure --version
rtl-closure validate --spec "$PWD/specs/simple_counter.json"
lane-agent doctor --pdk sky130A --json

Configure the model backend through the following environment variables. Keep the API key in your shell's secret manager or protected environment; never add it to a specification, fixture, log, or Git commit.

RTL_CLOSURE_PROVIDER
RTL_CLOSURE_MODEL
RTL_CLOSURE_BASE_URL
RTL_CLOSURE_API_KEY

Then launch the interactive agent:

pi-lane

The following setup commands are intentionally separate because they may download substantial images or open-PDK assets:

lane-agent image-install
lane-agent pdk-install --pdk sky130A

Long physical flows must be started through the managed background controller; do not wrap them in an additional polling loop. Use the workflow and child run IDs returned by the commands for all later status, event, log, and artifact queries.

SiliconWeaver/
├── pi-lane/                 # interactive agent and physical-flow control plane
│   ├── src/pi_lane/         # Python package and public CLIs
│   ├── extensions/          # read-only Pi monitor
│   ├── prompts/             # constrained physical-design agent prompt
│   ├── specs/               # canonical smoke specification
│   ├── examples/            # broader synthetic examples
│   └── docs/                # architecture, development, release and Chinese docs
├── rtl-closure/             # verified RTL generation and repair engine
│   ├── src/rtl_closure/     # specification, oracle, pipeline and verification code
│   ├── examples/            # standalone public specifications
│   ├── docker/              # pinned verification-image source
│   └── docs/                # detailed Chinese pipeline reference
└── .github/                 # monorepo CI, dependency updates and issue templates

Fast checks do not start model requests or physical-design runs:

cd pi-lane
ruff check src tests
PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 python -m pytest
npm run check

cd ../rtl-closure
ruff check src tests
PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 python -m pytest

Read CONTRIBUTING.md before opening a pull request. Changes must use only synthetic or redistributable fixtures, preserve required gates, and never include credentials, proprietary RTL, commercial PDK material, or NDA logs.

This repository intentionally excludes benchmark datasets, hidden evaluation, comparison baselines, score aggregation, evaluator credentials, and commercial PDK material. It does not import or invoke ORFS integrations. See pi-lane/OSS_SCOPE.md and rtl-closure/OSS_SCOPE.md.

Current non-goals include:

• commercial foundry signoff or a tapeout guarantee;
• automatic use of proprietary PDKs or licensed EDA tools;
• model-quality rankings or benchmark leaderboards;
• treating bounded formal traces as universal behavioral proof;
• silently migrating incompatible persisted state or protocol schemas.

Please do not report secrets, private endpoints, proprietary designs, or vulnerabilities in a public issue. Follow SECURITY.md and use GitHub private vulnerability reporting.

SiliconWeaver is licensed under the Apache License 2.0. Third-party dependencies and container inputs retain their own licenses and notice requirements; see THIRD_PARTY.md and the component notices.