vyges-loom-testbench

Watch an AI run open silicon sign-off — live.

· Live dashboard →

A capable model is shown the Vyges Loom sign-off engines through vyges mcp and — with nothing but each engine's self-description — picks the right tool, forms its arguments, and runs it. Every result is the engine's own real, content-addressed sign-off output. No mock-ups, no handwaving.

The testbench exercises 14 read-only engines through vyges mcp: cdc, char, drc, em-ir, extract, gds-view, glitch, lec, lvs, meas, opendb, power, sta-si, thermal.

How many of them the model drives correctly is whatever the badge above reports from the last run — not a number kept here.

The live demo (blinky)

A local browser dashboard where each engine lights up as the AI drives it — pending → AI is choosing the tool → running → PASS with the engine's real headline (timing met, IR-drop OK, LVS match, …):

# Install the Vyges CLI + engines once (https://vyges.com), then:
export PATH="$HOME/.vyges/bin:$PATH"

# Watch the AI drive it (uses your GitHub Models token):
GITHUB_TOKEN=$(gh auth token) python3 demo/live_demo.py --driver github --model openai/gpt-4.1

# …or the deterministic replay (no model, instant):
python3 demo/live_demo.py --driver echo

Opens http://localhost:8756. Stdlib only — no pip installs.

What each engine shows

Engine	Real result you'll see
sta-si	timing met · WNS · max frequency
extract	net count · total capacitance
lvs	layout = schematic
power	dynamic power
em-ir	IR-drop within limit
thermal	peak temperature vs limit
glitch	hazard count
lec	equivalent / not
gds-view	layout rendered

How it works

vyges mcp exposes each installed Loom engine as a tool with a typed, self-describing interface. A driver forms one call per engine; the harness runs it and validates the loom-result envelope (right tool, status: ok, content-addressed input_hash, plus the expected result). Three drivers: echo (deterministic replay), github (GitHub Models), anthropic. The agentic drivers see the whole surface and must choose correctly from the descriptors alone — a legibility test as much as a functional one.

CI + published dashboard

.github/workflows/test.yml, run manually (workflow_dispatch):

• deterministic — replays known-good calls; gates the run.
• agentic — one GitHub Models model drives the surface; advisory (never fails CI).
• report — publishes the matrix to the live dashboard.

Model access uses a models: read token via the MODELS_TOKEN secret (falls back to the Actions token where the org has GitHub Models enabled).

Coverage

The read-only engines listed above run against bundled Apache-2.0 fixtures under fixtures/. drc, cdc, and char are documented placeholders pending heavier inputs (a PDK DRC deck / a multi-clock netlist / ngspice + PDK models).

License

Apache-2.0. Bundled fixtures are copied from the corresponding Loom engine repositories.