r2g-skills: Instructions file for Claude Code

Agent-with-OpenROAD — Project Guide

AI-driven open-source EDA flow: natural-language spec → GDSII via OpenROAD-flow-scripts (ORFS), with full signoff (DRC, LVS, RCX), then a training-ready graph dataset for GNN predictors. Implemented as the r2g-skills Claude Code skill collection — four sub-skills (signoff-loop + def-graph from the 2026-07-07 split; eda-install added 2026-07-08; rtl-acquire ingested 2026-07-09. The three design plans were retired once completed — see git log for them, not a path in the tree):

• eda-install — detects the machine and installs + verifies the toolchain the others run (ORFS + openroad/yosys, iverilog, klayout, magic/netgen, sky130A PDK, torch venv). One command, bootstrap.sh (detect → plan → install → pin env.local.sh → verify); no-sudo conda path by default.
• rtl-acquire — the RTL corpus supplier, UPSTREAM of the others: discovers/screens/acquires RTL at corpus scale (local trees, repo manifests, keyword search) and expands it synth-only into pre-layout netlist_graph.pt graphs with dedup, quality scoring, and publish gating. Owns acquire + corpus publish + the one-click promote of a synth-proven candidate into a signoff-loop full-flow project (scripts/promote/promote_candidates.py, 2026-07-10); BORROWS env (_env.sh), synth (run_orfs.sh, ORFS_STAGES=synth), the graph format (def-graph netlist_graph.py), and failure learning (knowledge.sqlite, runs stamped flow_scope='synth_only'; frontend classes land as synth-frontend-* events).
• signoff-loop — drives the flow RTL→GDS with full signoff and the self-improvement loop (the two memory DBs + engineer_loop) that eliminates DRC/LVS violations and closes timing at Fmax.
• def-graph — converts the clean, signed-off physical design (the ORFS 6_final.odb/.def/ .spef + platform liberty/LEF) into PyTorch-Geometric graph datasets: five graph views (b–f), emitted as HeteroData by default (2026-07-16; R2G_GRAPH_KIND=homo for the legacy flat tensors), the shared tech-lib/LEF/DEF parser, and feature (X) / label (Y) extraction — labels are congestion, wirelength, per-path timing slack, IR drop, and SPEF-derived RC parasitics (the last a y[N,6] node label + a separate rc_edge_* parasitic edge set, merged 2026-07-07).

Each skill has ONE heart; everything else is plumbing — read the two ⭐ sections below:

1. signoff-loop → The Closed Learning Loop — the two memory DBs (knowledge.sqlite = what resulted, journal.sqlite = what was done) + engineer_loop, the autonomous driver that closes the wheel unattended (flow → fix → learn → A/B-promote) and learns repair recipes that transfer across designs/platforms.
2. def-graph → The Dataset-Construction Pipeline — three composable stages (labels → features → graphs) keyed to the same DEF so X and Y join, whose failure mode is a plausible CSV with silently wrong values.

This file is orientation; the skills document how to run/debug/tune. Don't duplicate skill content or per-run results here — when you fix a bug, update the relevant sub-skill under r2g-skills/ (signoff-loop/ for flow/signoff/learning, def-graph/ for dataset construction), not this file. Prefer editing existing scripts/ over adding new ones; use the documented steps, not ad-hoc shell, in production.