rfauto v0.13.0 — 电路图深化 · ADS 调参回路闭环 · LNN 代理 · CI 全绿

更新日志

v0.13.0（2026-11-02）

新增：

• 引脚级原理图：rfauto diagram schematic 从 ADS 网表生成带引脚、连线和网名的原理图
• 图转 SVG：rfauto diagram to-svg 把 .drawio 转成 SVG，不装 draw.io 也能在服务器或报告里看图
• 叠层 3D 预览：rfauto diagram stackup --preview 在横截面图之外另出 3D 预览 PNG
• 评审标注：多人标注可合并成一份图，按评审人配色、署名和计数
• 测试连接图：新增 CISPR 25 传导发射场景，可选专业电气元件形状
• ADS 调参：有源链与 HB 负载牵引模板支持出图调参；反标注可嵌入 S 参数曲线图；优化器可直接调用 ADS 电路仿真作为目标函数
• 设计报告：rfauto design-report 新增 Pareto 前沿、两次运行的参数与指标对比两节
• 校准设计：rfauto cal-wizard trl 按频带给出 TRL 校准件尺寸、带宽覆盖核算和测试连接图
• 运维：rfauto remote board 查看服务器与本地求解队列；性能画像一键生成；源码变更后自动预跑受影响测试（devtest_speculative --watch）
• 代理模型：新增液态神经网络（CfC）、复数 stamp 族（ngo_stamp_rf）、多端口参数化 VF、SVGP、按参数导出 SPICE 网表、ONNX 导出
• 优化算法：TuRBO 多重启进入竞赛引擎榜；SAAS 后验可选 NUTS 采样；多目标优化支持约束条件（c-EHVI）
• 架构借鉴：吸收 DeepSeek Super-agent（MIT）的事件驱动执行、探活与信道分域思路，落地为投机测试预跑、rfauto watchdog 和 CLI/JSON 输出的统一规范；另有编排、信道两篇设计笔记（docs/explanation/agent-orchestration.md、docs/explanation/cli-json-channel.md）

修复与改进：

• 串行终门固定 BLAS 单线程，消除多线程环境下的偶发测试超时
• 对拍门未通过时不再误报为“ADS 段失败”，两种失败分开提示
• CLI 命令 297 条，MCP 工具 166 个，单元测试 30738 条（README 徽章同源）

当前版本 v0.13.0，上一版本 v0.12.1。

---

Changelog

v0.13.0 (2026-11-02)

Added:

• Pin-level schematics: rfauto diagram schematic renders ADS netlists
  as schematics with pins, wires and net names
• SVG export: rfauto diagram to-svg converts .drawio to SVG so diagrams
  can be viewed on servers and in reports without draw.io
• 3D stackup preview: rfauto diagram stackup --preview renders a 3D
  preview PNG alongside the cross-section view
• Review annotations: annotations from multiple reviewers merge into one
  file, with per-reviewer colors, attribution and counts
• Test-setup diagrams: new CISPR 25 conducted-emission scene, with
  optional professional electrical component shapes
• ADS tuning: active-chain and HB load-pull templates now support
  diagram-driven tuning; back-annotation can embed S-parameter curve
  plots; the optimizer can call ADS circuit simulation as its objective
• Design reports: rfauto design-report gains Pareto-front and
  two-run parameter/metric comparison sections
• Calibration design: rfauto cal-wizard trl proposes Thru/Reflect/Line
  dimensions, bandwidth coverage checks and a test-setup diagram for a
  given frequency band
• Operations: rfauto remote board shows the server and local solve
  queues; one-command performance profiling; affected tests pre-run
  automatically when source files change (devtest_speculative --watch)
• Surrogates: Liquid Neural Network (CfC), complex stamp family
  (ngo_stamp_rf), multi-port parametric VF, SVGP, per-parameter SPICE
  netlist export, ONNX export
• Optimization: TuRBO multi-restart joins the competition leaderboard;
  SAAS posterior supports NUTS sampling; multi-objective optimization
  supports constraints (c-EHVI)
• Architecture credit: event-driven execution, liveness-probe and
  channel-separation ideas absorbed from DeepSeek Super-agent (MIT) land
  as speculative test pre-runs, rfauto watchdog and unified CLI/JSON
  output rules; two design notes in
  docs/explanation/agent-orchestration.md and
  docs/explanation/cli-json-channel.md

Fixed and improved:

• The serial test gate pins BLAS to a single thread, removing flaky
  timeouts in multi-thread environments
• A failed comparison gate no longer reports "ADS stage failed"; the two
  failure kinds now have distinct messages
• 297 CLI commands, 166 MCP tools, 30738 unit tests (same source as the
  README badge)

Current version: v0.13.0 (previous: v0.12.1).

---

安装：pip install rfauto-0.13.0-py3-none-any.whl（或解压源码包后 pip install .）。

附件：sdist 源码包（rfauto-0.13.0.tar.gz）+ wheel 二进制包（rfauto-0.13.0-py3-none-any.whl），均已过 twine check 与发布包成员卫生门。

rfauto v0.12.1 — guided onboarding, capability map, docs-only CI fast path

rfauto v0.12.1 — documentation release: a guided onboarding path, a nine-domain capability map, and bilingual figures.

Scale: 16 EM/EDA engines · 75 parameterized device templates · 341 closed-form calculators · 87 calibration anchors · 280 CLI commands · 157 MCP tools (+10 resources) · 28000+ unit tests — no commercial license required.

Start here: download Source code (zip) below, then follow the eight-step Quick start in the README (about ten minutes, no license needed).

---

v0.12.1 — documentation release: guided onboarding, capability map, bilingual figures

This is a documentation and workbench-polish release on top of v0.12.0. No API or
solver behaviour changed.

Documentation

• Both READMEs rebuilt around a nine-domain capability map and an eight-step,
  license-free quick start (download → install → doctor → first run → read
  artifacts → workbench → tune → real engine), with measured outputs quoted
  verbatim from a real run.
• New: docs/tutorials/getting-started.md walkthrough is now the canonical
  first-hour path, linked from both READMEs together with the other 17 tutorials
  (real simulation, workbench tour, optimization loop, campaign, datasets and
  health gates, measurement alignment, uncertainty, verdict semantics, remote
  solving, DAG, compose/league, recast, MCP in practice, extending).
• Four new reference figures in Chinese and English: capability map, engine
  ladder and arbitration, the six quality gates, and the calculator distribution
  by design domain (counts computed from the registry, not hand-written).
• Refresh of stale counts across reference pages (152 → 157 MCP tools) and
  the calculator count is now stated precisely: 341 = 340 regular + 1
  experimental.
• Bilingual workbench screenshots (dashboard, template gallery, tuning, run
  detail, microwave tools) taken from a live local run.

Workbench

• Relative timestamps ("5 min ago") and the brand tagline are now localized —
  English mode no longer shows Chinese fragments.
• Tuning, run detail and gallery screenshots in the docs match the shipped UI.

Recipe

• recipes/wilkinson_pd_v1.yaml now carries tuned nominal dimensions (λ/4 arm
  18.00 mm, 70.7 Ω arm width 0.448 mm, 50 Ω feed 1.151 mm). The documented first
  run reproduces −12.43 dB in-band S11, −3.59 dB insertion loss and 27.4 dB
  isolation with the zero-license fake channel.

Scale

16 engines · 75 parameterized device templates · 341 closed-form calculators ·
87 calibration anchors · 280 CLI commands · 157 MCP tools (+10 resources) ·
28000+ unit tests (no commercial license required). CI: 7/7 jobs green on both
Windows and Linux.

CI

• Docs-only fast path: a changes job classifies each push. Documentation-only
  pushes run the quality gates, the docs-site regeneration check and the docs
  content gates (README numbers, citations, capability cards, gallery, pending
  verifications, template reference) — the heavy matrix (two pytest shards,
  preflight, supply chain, docker build) is skipped. Measured on the first
  docs-only run (40286868258): heavy jobs skipped, both fast-path jobs green;
  kernel_cards is skipped too because its inputs are code-face. Any change under
  src/tests/recipes/scripts/configs — or under docs/templates/**, whose
  meta.yaml files are test inputs — still runs the full matrix. Escape hatches: [full-ci] in the commit message, or full_ci on
  a manual dispatch; an unresolvable baseline falls back to the full matrix, and a
  classifier failure explicitly defaults to full (fail-safe).

Verified gates on this commit

• Windows full suite: 28,000+ unit tests green (local gate), plus the four
  documentation gates run in CI on the fast path.
• Docs-only CI path measured end to end: 8.2 minutes from push to green
  (quality + docs-site regeneration + 119 docs content tests), with the two
  pytest shards, merge preflight, supply chain and Docker build skipped.
• Full matrix green on the code commits in this release (UI i18n fix; CI change):
  quality, merge preflight, license gate, supply chain, Docker smoke and both
  pytest shards all pass — the changes classifier itself ran in production on
  every run.

rfauto v0.12.0 — EM semantics layer · IPD template families · 340-calculator kernel

rfauto v0.12.0 — the RF/microwave design-automation framework with an LLM orchestration layer that is never allowed to invent physics.

Scale this release: 16 EM/EDA engines · 75 parameterized template families · 340 deterministic calculators · 87 calibration anchors · 280 CLI commands · 157 MCP tools (+10 resources) · 28000+ unit tests (no commercial license required).

Source code: download Source code (zip) / Source code (tar.gz) below, then see the Quick start in the README (6 steps, zero license).

---

[0.12.0] - 2026-10-26

Added

EM semantics layer (feedback-driven)

• em doctor: a 16-item pre-solve checklist (workdir hygiene, ASCII path
  guards, unit sanity, mesh/window traps, janitor integration) that catches
  the failure modes that silently waste solver hours; delivered as service,
  CLI (rfauto em doctor) and MCP tool.
• Run-artifact freshness verification: every solve point can now carry a
  closing manifest (file hashes + sidecars); rfauto runs verify reports
  four states (fresh/stale/legacy/error) instead of trusting directory
  mtime, and stale artifacts are flagged — not silently consumed — by
  metrics, diff, sweep and comparison paths.
• rfauto em diff: four-section comparison of two runs (parameters,
  manifests, S-parameters on a common frequency grid, freshness) rendered
  in plain language.
• DOE sweep planner (rfauto em sweep) with case-level freshness summary,
  lock-owner awareness and resume support; resource janitor for orphan
  solver processes and engineering locks.

New device families and calculators

• Three GaAs IPD lumped-element template families: octagonal spiral
  inductors, MIM capacitors and interdigital capacitors, each with
  synthesis, nominal-parameter anchors and offline geometry audits.
• Six new closed-form CALC families (registry now 340 calculators):
  terminated-line metrics, filter prototypes, network conversions,
  Leeson-phase-noisy oscillators, branchline couplers, de-embed metrics —
  each cross-checked against independent closed-form references.
• Metric binding groups: series/shunt family detection with automatic
  same-family Q reporting and convergence gates.

Port semantics guard

• oe_port_semantics: five geometric criteria for MSLPort launchers
  (box spans the strip, strip fully inside the excitation cross-section,
  measurement plane clear of the feed near-field, mesh landing, ladder
  traps) enforced at model-build time — rendered scripts fail loudly
  instead of producing quietly wrong S-parameters. Wired into all TL,
  SIW, hairpin, C4, antenna2 and varactor renderers.

EDB freshness and docs

• EDB fast-lane artifacts now carry freshness sidecars end-to-end
  (build → solve → extract → dataset), with an eighth gate criterion that
  skips — not fails — unverifiable legacy runs.
• Docs overhaul: 75 template reference cards, a per-design-flow calculator
  usage guide (48 keys), refreshed architecture/workflow/trust-layer
  figures, and an anchor-guide rebuild (87 anchors).

Changed

• MCP surface now 157 tools (+10 resources); CLI 280 leaf commands.
• Test suite grew to 28,352 unit tests (no commercial license required).
• Anchors registry: 87 anchors across 69 template families.

Fixed

• Wilkinson test corpus had series/shunt width semantics swapped; the new
  port-semantics guard caught it at model-build time.
• Slot template measurement plane moved out of the feed near-field band
  (first real catch of criterion #347).
• A leaked test fixture could make the runs-root convergence tests flip;
  fixture removed and the convergence semantics documented.