infrastructure

name	infrastructure
description	EDA tool detection, wrapper deployment, and MCP configuration for analog/mixed-signal + RF design environments. Use when setting up a workstation, verifying tool availability before a domain flow, or generating per-tool install scripts with TCL modulefiles.
version	1.0.0
author	chuanseng-ng
license	MIT
allowed-tools	Read, Write, Bash

Skill: Infrastructure Setup

Invocation

• If invoked by a user presenting a setup task: immediately spawn the analog-chip-design-agents:infrastructure-orchestrator agent and pass the full user request and any available context. Do not execute stages directly.
• If invoked by the infrastructure-orchestrator mid-flow: do not spawn a new agent. Treat this file as read-only — return the requested stage rules or sign-off criteria.

Purpose

Detect open-source and proprietary analog/RF EDA tools, generate installation scripts for missing open-source tools, deploy output-filtering shell wrappers that emit compact JSON instead of raw SPICE .lis/.measure and DRC/LVS logs, configure MCP server templates, and validate the environment before any domain orchestrator begins work.

---

Supported EDA Tools

Open-Source

• ngspice (ngspice) — SPICE simulator; loads Verilog-A via OSDI
• Xyce (Xyce) — parallel SPICE; harmonic balance; large MC
• gnucap (gnucap) — general circuit analysis
• Qucs-S (qucs-s) — schematic + simulation front-end (RF capable)
• xschem (xschem) — analog schematic capture and netlisting
• Magic (magic) — layout editor, DRC, and extraction
• Netgen (netgen) — LVS
• KLayout (klayout) — layout viewer, DRC/LVS, GDS/OASIS
• OpenVAF (openvaf) — Verilog-A → OSDI compiler
• ADMS (admsXml) — legacy Verilog-A → C model compiler
• openEMS (openEMS) — FDTD electromagnetic solver
• FastHenry (fasthenry) / FastCap (fastcap) — inductance/capacitance field solvers
• gmsh (gmsh) — mesh generation for EM
• scikit-rf (Python package skrf) — S-parameter / network analysis
• PySpice (Python package PySpice) — scripted SPICE corner/Monte-Carlo control
• cocotb (Python package cocotb) — analog↔digital co-simulation control
• Hdl21 (Python package hdl21) — programmatic analog hardware description
• gdstk (Python package gdstk) / gdsfactory — layout generation
• Verilator (verilator) / Icarus Verilog (iverilog) — digital + RNM co-sim
• GTKWave (gtkwave) — waveform viewer
• open_pdks — sky130 / gf180mcu / ihp-sg13g2 PDK installer
• freepdk45 / asap7 — predictive academic PDKs (NCSU 45 nm / ASU-ARM 7 nm FinFET); installed from their own distributions, not via open_pdks; detect-only, non-manufacturable
• uv (uv) — fast Python package/project manager (required for PySpice/cocotb installs)

Proprietary (detect only — never install)

• Cadence Virtuoso (virtuoso) — schematic/layout/ADE
• Cadence Spectre / Spectre X / APS (spectre) — analog/RF simulation
• Cadence Xcelium AMS / AMS Designer (xrun) — Verilog-AMS / mixed-signal
• Cadence Quantus QRC (quantus) — parasitic extraction
• Cadence Pegasus (pegasus) — DRC/LVS
• Cadence Liberate (liberate) — characterization
• Synopsys HSPICE (hspice) / PrimeSim (primesim) / FineSim (finesim) — SPICE
• Synopsys Custom Compiler (custom_compiler) — schematic/layout
• Synopsys StarRC (StarXtract) — extraction
• Synopsys IC Validator (icv) — DRC/LVS
• Siemens Calibre (calibre) — DRC/LVS/PEX/PERC
• Siemens Symphony / AFS (afs) — AMS / Analog FastSPICE
• Siemens Solido (solido) — variation / ML characterization
• Keysight ADS (ads) — RF circuit + EM
• Ansys HFSS (hfss) — 3D EM
• Sonnet (sonnet) — planar EM

---

MCP Architecture — Two Tiers

Tier 1: Batch MCP servers (short, self-contained runs)

MCP config	Tool	Typical duration
mcp-ngspice.json	ngspice single analysis	seconds–minutes
mcp-xyce.json	Xyce single analysis / HB	minutes
mcp-magic.json	Magic DRC / extract	minutes
mcp-klayout.json	KLayout DRC/LVS	minutes
mcp-netgen.json	Netgen LVS	seconds–minutes
mcp-openvaf.json	OpenVAF Verilog-A compile	seconds
mcp-openems.json	openEMS short solve	minutes (set TOOL_TIMEOUT_S)

The adapter is plugins/infrastructure/tools/mcp-adapter.py.

Tier 2: Interactive session MCP servers (stateful, query-based)

MCP config	Tool	Exposed tools
mcp-ngspice-session.json	ngspice interactive control	load_netlist, run_analysis, measure, alter_param, run_corner, close

The adapter is plugins/infrastructure/tools/mcp-session-adapter.py.

Full-flow / long EM solves — do NOT use MCP

Large Monte-Carlo sweeps and 3D EM solves run for tens of minutes to hours and write structured output files. Launch via Bash and read the output files (.mt*, .csv, Touchstone .sNp) directly.

Execution Hierarchy (per domain agent)

1. Tier 2 session MCP — when iterating over an already-loaded netlist (corner/MC/ECO loops)
2. Tier 1 batch MCP — one-shot single-analysis invocations
3. Wrapper script — if MCP not configured; wrapper emits compact JSON
4. Direct execution — last resort; raw logs consume significant context

---

Stage: tool_discovery

Domain Rules

1. Run which <command> and <command> --version (or -v) for every open-source tool above.
2. Python interpreter detection — run once at the start, before checking any Python packages.
   • Step A — module system probe: if $MODULESHOME is set or modulecmd exists, run module avail 2>&1, match python/python3, load the latest, resolve PYTHON_EXEC with which python3, set python_env.type = "module" and record python_env.module_name. Keep the module loaded for the whole run.
   • Step B — PATH fallback: which python3 → PYTHON_EXEC; FAIL if empty (required for all wrappers). Classify system (/usr/bin/python3) vs custom.
   • Step C — finalize: set PYTHON_BIN_DIR = dirname(PYTHON_EXEC); record python_env { exec, type, bin_dir, module_name }; add a python3 entry to the tools array.
3. Python packages (PySpice, cocotb, skrf, hdl21, gdstk, uv): for custom/module Python use "$PYTHON_EXEC" -m pip show <pkg>; for system use the PATH tool. Report MISSING if absent.
4. For proprietary tools: check PATH with which <primary-executable> only (see executable names above); never install; record PROPRIETARY_ONLY if found, MISSING otherwise.
5. Record each tool as FOUND, MISSING, or PROPRIETARY_ONLY; capture the exact version for each FOUND.
6. Write results to tool-status.json before advancing.

QoR Metrics to Evaluate

• tools_detected: count of FOUND tools (target ≥ 8 for a functional open-source analog flow)
• tools_missing: count of MISSING open-source tools
• proprietary_found: count of PROPRIETARY_ONLY tools detected

Output Required

• tool-status.json with python_env object and a tools array of { tool, command, status, version, path }

---

Stage: module_discovery

Domain Rules

1. Detect classic TCL Environment Modules ($MODULESHOME set or modulecmd in PATH); else set module_system: "none", WARN, write empty module-status.json, advance.
2. Run module avail 2>&1; match entries against the tool→module map (case-insensitive substring), e.g. spectre→cadence/spectre, hspice→synopsys/hspice, calibre→mentor/calibre, virtuoso→cadence/virtuoso, ngspice→ngspice, magic→magic, xschem→xschem, klayout→klayout, xyce→xyce.
3. For each FOUND tool with a module available: upgrade status to FOUND_PREFER_MODULE; for MISSING with a module: MISSING_LOAD_MODULE. Add module_names/versions_available.
4. Generate load-modules.sh (default to latest version; comment alternatives). Never auto-run it; print: "Review and source load-modules.sh, then re-run the flow".

QoR Metrics to Evaluate

• module_system_detected: bool
• tools_found_via_modules: count of FOUND_PREFER_MODULE + MISSING_LOAD_MODULE

Output Required

• module-status.json, updated tool-status.json, load-modules.sh (when modules found)

---

Stage: tool_installation

Domain Rules

1. Never auto-run installs — only generate per-tool install-<toolname>.sh in install-missing-tools/.
2. If python3 is missing: FAIL immediately and escalate (required for all wrappers).
3. Generate a script for every MISSING tool; skip FOUND/FOUND_PREFER_MODULE/MISSING_LOAD_MODULE/PROPRIETARY_ONLY.
4. Python-package scripts (PySpice, cocotb, skrf, hdl21, gdstk, uv) read python_env.exec and use "$PYTHON_EXEC" -m pip install <pkg>; never bare pip for custom/module Python.
5. Always generate a TCL classic modulefile per tool (no extension) under $EDA_MODULEFILES_ROOT/<tool>/<version> with at least PATH and LD_LIBRARY_PATH.

Package Name Mapping (open-source)

Tool	apt	brew	source / notes
ngspice	ngspice	ngspice	—
Xyce	build-from-source	—	https://github.com/Xyce/Xyce
xschem	xschem	build-from-source	https://github.com/StefanSchippers/xschem
magic	magic	—	http://opencircuitdesign.com/magic
netgen	netgen-lvs	—	http://opencircuitdesign.com/netgen
klayout	klayout	klayout	—
openvaf	build-from-source	—	https://github.com/pascalkuthe/OpenVAF
openEMS	openems	—	https://github.com/thliebig/openEMS-Project
PySpice/cocotb/skrf/hdl21/gdstk	<PYTHON_EXEC> -m pip install <pkg>	same	Python packages
uv	curl -LsSf https://astral.sh/uv/install.sh | sh	uv	prefer standalone installer

Output Required

• One install-<toolname>.sh per MISSING tool in install-missing-tools/

---

Stage: wrapper_deployment

Domain Rules

1. Deploy the wrapper scripts to plugins/infrastructure/tools/ and chmod +x each; on permission denied, FAIL and escalate with sudo chmod +x.
2. Every wrapper must emit JSON conforming to the schema below regardless of exit code, and must handle a missing tool gracefully (status: "FAIL").
3. Never suppress the tool's original exit code.

Wrapper JSON Output Schema

{
  "tool": "<tool-name>",
  "exit_code": 0,
  "status": "PASS|FAIL|WARN",
  "summary": {},
  "errors": [],
  "warnings": [],
  "raw_log": "/tmp/<tool>-XXXXXX.log"
}

summary carries tool-specific metrics (e.g. ngspice: measured specs + convergence; magic/klayout: DRC/LVS counts; openvaf: compile status; openEMS: S-param/passivity).

Output Required

• Executable wrapper scripts in plugins/infrastructure/tools/

---

Stage: mcp_configuration

Domain Rules

1. Emit MCP config snippets for the batch + session configs listed in the MCP Architecture section.
2. Batch configs use "command": "python3" with mcp-adapter.py (never point directly at the wrapper); session configs use mcp-session-adapter.py --tool ngspice.
3. Resolve absolute adapter/wrapper paths at runtime (realpath/pwd); never leave /absolute/path/to/.
4. Write snippet files to plugins/infrastructure/mcp/; do not modify .claude/settings.json automatically — instruct the user to paste the mcpServers block.

Output Required

• Batch MCP configs in plugins/infrastructure/mcp/ (mcp-ngspice.json, mcp-xyce.json, mcp-magic.json, mcp-klayout.json, mcp-netgen.json, mcp-openvaf.json, mcp-openems.json)
• Session MCP config (mcp-ngspice-session.json)
• Adapter scripts present in plugins/infrastructure/tools/ (mcp-adapter.py, mcp-session-adapter.py)
• Printed snippets with resolved absolute paths

---

Stage: environment_validation

Domain Rules

1. Python environment check first: read python_env; if type == module, verify the module is still loaded; if custom/system, verify the path matches python_env.exec (WARN on mismatch).
2. Re-run tool checks using the same Python-aware detection as tool_discovery.
3. Verify wrapper scripts exist and are executable; verify MCP snippets + adapters are present.
4. FAIL if any critical-path tool (ngspice or Xyce, plus magic/klayout for layout flows) is still MISSING.
5. For MISSING_LOAD_MODULE critical tools: WARN and instruct to source load-modules.sh.
6. Print the final sign-off summary: tools detected, tools via modules, wrappers deployed, MCP servers configured.

Sign-off Checklist

• tool-status.json written (includes python_env)
• module-status.json written (even if "none")
• install-<toolname>.sh generated for all MISSING tools
• load-modules.sh generated if module-available tools found
• Wrappers deployed and executable
• mcp-adapter.py and mcp-session-adapter.py present
• MCP config snippets written with resolved absolute paths and printed
• No critical-path tools with status MISSING

Output Required

• Printed environment validation report
• tool-manifest.json with the final confirmed state

---

Memory

Infrastructure memory is opt-in (default off) and environment-keyed — see the infrastructure orchestrator's Infrastructure Memory section and memory/README.md. When enabled (design_state.pipeline_config.track_infrastructure == true or --track-memory), upsert a record into memory/infrastructure/experiences.jsonl after environment_validation with a key_metrics.tool_versions map. Distillation into memory/infrastructure/knowledge.md is handled by the memory-keeper skill.