Skip to content

OpenAI4S v1.0.0

Choose a tag to compare

@Nobody-Zhang Nobody-Zhang released this 09 Nov 06:53
4de8520

OpenAI4S v1.0.0 is a major update. It brings the OpenMHS simulated lab to the workbench: a Lab tab beside the Notebook, 746 installable simulated devices, and a one-command setup through a pinned OpenMHS submodule. It also adds Ollama models and web tools, an optional You.com search, a molecular-dynamics Skill, read-only notebook and science-search previews, and a round of provenance, workbench and model-wire fixes.

The database moves from schema 35 to 36. Read “Upgrading from 0.3.1” before starting this version on an existing data directory.

Highlights

  • The OpenMHS lab and device market. OpenMHS, the open-source Model Hardware Standard, now ships as the third_party/OpenMHS git submodule and is found automatically: no PYTHONPATH, no settings file. The Lab tab shows every lab device grouped by how the agent reaches it, MHS (the native mhs_* tools) or MCP (a connector per device), with live state, e-stop latches, the hash-chained receipts of every endpoint, and charts derived from them. The Device market installs any of 746 devices in one click: 299 hardware drivers served on fake backends and 447 simulations, each with an introduction. The same works from the shell with openai4s lab add <id>, plus search, show, list, status, remove, up, down, reset and setup. Three built-in simulated instruments (an enzyme-inhibition assay workstation, a flow reactor with inline HPLC, and an xArm 7 on its fake backend) demonstrate the three access paths. The lab keeps an OpenMHS Lab project whose fixed agent context directs discovery through the governed host tools. Live device descriptions remain untrusted tool data; an operator lab card lists their names, arguments and limits. Installing, start/stop, e-stop and operator reset are human actions; no tool reaches them (#250, #253).
  • Governed OpenMHS calls. MCP calls to an OpenMHS connector run through the reference host's durable command journal, host ledger and endpoint evidence. A device-bound software e-stop skips the ordinary approval prompt, and persisted MCP denies are preserved (#250).
  • Ollama models and web tools. Ollama cloud and local model profiles, plus Ollama web search and fetch over the stdlib transport. A web-tool key can be set independently of the main model (#233). An optional You.com search integration adds a keyless web_search backend (#232).
  • 612 bundled Skills. A new bilingual molecular-dynamics Skill covers OpenMM preparation, minimization and staged NVT/NPT runs, and MDAnalysis RMSD, RMSF and contact analysis (#231).
  • Bounded, honest previews. .ipynb artifacts open in a read-only renderer that fetches and executes nothing. Saved science-search results get a records view, and remote-task listings distinguish unknown, unavailable and empty states. Previews have hard read budgets and report complete, partial or unavailable. Update checks are explicit and bounded (#242).
  • More exact provenance. Frozen snapshot reads and pandas grouped outputs keep their exact input versions, and shared constants and string literals are never tagged as inputs to unrelated writes (#214, #235, #237).

Downloads

Download For
OpenAI4S-1.0.0-macos-arm64.dmg (SHA-256) macOS Apple Silicon (arm64); embedded Python and science stack; ad-hoc signed
OpenAI4S-1.0.0-linux-x86_64.tar.gz Linux x86_64; embedded Python and the default science stack
OpenAI4S-1.0.0-windows-x86_64.zip Windows x86_64 with WSL2; includes the same Linux payload
openai4s-1.0.0-evidence.zip Original publishing-workflow evidence: build manifest, component inventory, SBOM, provenance and validation references
OpenAI4S-1.0.0-macos-arm64-validation.json Supplemental local macOS build and validation record
SHA256SUMS SHA-256 digests for all six downloadable package/evidence/checksum files

Choose the package for your platform. Python packages are distributed through PyPI; container images through GHCR. The GitHub downloads focus on the desktop packages and their verification files.

Build sources and validation scope are recorded separately from source CI. The original evidence archive covers the publishing workflow, which omitted macOS. The macOS DMG was subsequently built from the same frozen source, validated locally and published with its supplemental JSON record above.

Verify downloads with sha256sum --ignore-missing -c SHA256SUMS on Linux, or shasum -a 256 --ignore-missing -c SHA256SUMS on macOS. To check the evidence archive's internal integrity, run openai4s verify-package openai4s-1.0.0-evidence.zip.

  • Linux: unpack the tarball and run ./OpenAI4S, or use ./install.sh for a per-user PATH entry and menu item. Install bubblewrap for sandboxed Cells. See the startup guide.
  • Windows: follow the WSL2 steps below. The release ZIP is complete; no split-file restoration is needed.
  • macOS Apple Silicon: download OpenAI4S-1.0.0-macos-arm64.dmg and its SHA-256 file, then follow the macOS installation steps below. This ad-hoc signed DMG embeds Python 3.13.15 and the science stack. No Developer ID signature or Apple notarization is included. Intel Macs should use the Python/source path; this DMG is arm64 only.
  • Python: install with python -m pip install --upgrade 'openai4s==1.0.0', then run openai4s serve. Python 3.10+ is required; the core remains standard-library-only. For the lab, run openai4s lab setup once: it downloads the pinned OpenMHS release into ~/.openai4s/openmhs.
  • Source with the lab: git clone --recursive --branch v1.0.0 https://github.com/PKU-YuanGroup/OpenAI4S.git && cd OpenAI4S && ./setup.sh. --recursive fetches OpenMHS; for an existing checkout, run git submodule update --init third_party/OpenMHS.
  • Containers: run docker pull --platform linux/amd64 ghcr.io/pku-yuangroup/openai4s:1.0.0. Use the 1.0.0 image tag for this release; older examples in the Docker guide still show 0.3.1. See the Docker guide for authentication, persistent data, and sandbox limits.
  • R: the R kernel is not bundled. Install micromamba, mamba, or conda, then run openai4s setup --profile standard or openai4s setup --only r.

Try the OpenMHS lab

git clone --recursive --branch v1.0.0 https://github.com/PKU-YuanGroup/OpenAI4S.git
cd OpenAI4S && ./setup.sh && source .venv/bin/activate
openai4s lab add labs/tiagocoutinho__julabo   # any id from https://mhs.openai4s.org/market/
openai4s serve                               # open the project "OpenMHS Lab", start a conversation, open the Lab tab

Every device runs on a simulator or a fake backend. Choose the agent surface per device with --via mcp (default) or --via mhs, or with the market's selector. The OpenMHS guide (中文) explains both surfaces, the lab files and the operator actions.

OpenMHS 实验室: git clone --recursive 后运行 ./setup.sh,再用 openai4s lab add <设备 id> 安装 OpenMHS 市场 中的任意设备。在工作台里也可以操作:打开“OpenMHS Lab”项目,进入“实验室”标签页,点“设备市场”,再点“安装”。所有设备都在模拟器或 fake 后端上运行,不连接真实硬件。

Install on macOS (Apple Silicon)

  1. Download OpenAI4S-1.0.0-macos-arm64.dmg and OpenAI4S-1.0.0-macos-arm64.dmg.sha256 into the same folder. In Terminal, change to that folder and run:

    shasum -a 256 -c OpenAI4S-1.0.0-macos-arm64.dmg.sha256

    The result must be OpenAI4S-1.0.0-macos-arm64.dmg: OK.

  2. Open the DMG, drag OpenAI4S.app into Applications, then eject the disk image. Launch OpenAI4S from Applications; it starts the local workbench and opens your browser.

  3. This build uses ad-hoc signing, with no Developer ID or notarization. If macOS blocks the first launch, use System Settings → Privacy & Security → Open Anyway for this verified app and confirm the prompt. See Apple's instructions.

  4. Configure your provider under Customize → Models. No model key is included. For an existing data directory, follow Upgrading from 0.3.1 below before starting the new version.

macOS 安装: 适用于 Apple Silicon(M 系列芯片)。下载 DMG 和同名 .sha256 文件,在两者所在目录运行上面的校验命令,确认 OK。打开 DMG,将 OpenAI4S.app 拖入“应用程序”,推出磁盘后启动应用;如果首次启动被系统拦截,在“系统设置 → 隐私与安全”中选择“仍要打开”。本包采用 ad-hoc 签名,不含 Developer ID 签名或 Apple 公证。首次启动后在“自定义 → 模型”中配置服务商。

The DMG passed all ten installed-runtime checks, normal LaunchServices startup, browser artifact viewing, enforced Seatbelt isolation and before/after signature verification on macOS 27.0.1 arm64. Its SHA-256 is 988ec93f1bd8e701911af04b3dfbdf670f8c9eaedc6930061531cd7c1040dc62. These results describe the local supplemental build; they do not claim a notarized Gatekeeper first-open flow.

Install on Windows (WSL2)

  1. Prepare WSL2 once. On x86_64 Windows 10 2004+ or Windows 11, run wsl --install -d Ubuntu-24.04 in administrator PowerShell. Restart if requested and create the Ubuntu user. WSL1 is not supported; the launcher requires working bubblewrap 0.8.0+ inside WSL.
  2. Download and verify. Download OpenAI4S-1.0.0-windows-x86_64.zip and SHA256SUMS. Run (Get-FileHash .\OpenAI4S-1.0.0-windows-x86_64.zip -Algorithm SHA256).Hash; it must match the ZIP's entry, ignoring case.
  3. Extract and start. Use Extract All to a local folder, then double-click OpenAI4S.cmd. Run it from the extracted folder, not the ZIP preview or a network share. The launcher verifies and installs the bundled payload, starts the daemon, and opens the browser.
  4. Configure a model. Open Customize → Models and enter your provider settings. No API key is bundled. For later starts, double-click OpenAI4S.cmd again. Commands such as OpenAI4S.cmd status, url, and stop are also available.

Windows 安装: 先安装 WSL2 和 Ubuntu 24.04,下载并校验完整 ZIP,用“全部解压缩”解压到本地文件夹,然后双击 OpenAI4S.cmd。首次启动后,在“自定义 → 模型”中配置服务商。更新时解压到新文件夹并运行新启动器;WSL 中的 ~/.openai4s 数据目录继续保留。

Upgrading from 0.3.1

  1. Stop and back up first. Run openai4s stop, confirm with openai4s status, and copy the whole data directory while the daemon is stopped, for example cp -a ~/.openai4s ~/.openai4s-before-1.0.0. Include any database WAL/journal files. For containers, stop the service and back up its /data volume.
  2. The first database-opening command migrates schema 35 → 36. Schema 36 adds the dataset_verification_receipts table. Dataset verification is now shown only with a Host-owned receipt bound to the exact version, source and bytes. Historical artifact sources are kept as they were, and no receipt is inferred for them, so their verification status reads as unknown. doctor and diagnostics inspect the schema without upgrading it.
  3. Keep your own backup. Downgrade is unsupported: 0.3.1 refuses the newer schema, and restoring your pre-upgrade directory is the way back.
  4. The lab is opt-in. Nothing changes for an installation that never runs openai4s lab. Set OPENAI4S_OPENMHS=off to turn OpenMHS discovery off entirely. The lab keeps its endpoint state under <data dir>/mhs-lab/ and stores connector/project configuration in the normal OpenAI4S data store. It creates no global command allow rules; normal host approvals apply. Child processes receive a filtered environment, and operator mutations are serialized across threads and processes.

The full upgrade guide is available in English and 中文.

What's Changed

OpenMHS lab and integration

  • Add the OpenMHS submodule, automatic discovery, the openai4s lab CLI, installable market devices on the MHS or MCP surface, the Lab panel's device market, setup and start/stop controls, and the auto-maintained OpenMHS Lab project (#253).
  • Route OpenMHS MCP calls through the governed host route, allow trusted device-bound e-stops, and preserve persisted MCP denies (#250).

Models, search and Skills

  • Send only the request parameters each model and wire accepts (OpenAI reasoning models, Anthropic, Gemini tool-history replay), and rejoin interleaved indexless tool-call fragments by call ID (#251, #243).
  • Ollama models and web tools; optional You.com search; the molecular-dynamics Skill (#233, #232, #231).

Workbench and runtime

  • Keep execution history, run location and modal state correct across delayed, stale and failed responses, and ignore replayed socket events before they apply side effects (#216, #238, #239, #248).
  • Clean up recovery candidates before journaling failures (#244).
  • Promote the next line: bounded artifact previews, read-only notebook and science-search views, truthful remote-task and capability states, explicit update checks, and installed-runtime release acceptance (#242).
  • Update dependency advisory floors and keep Python 3.10 CI (#252).

Known limitations

  • Validation scope: source CI and the full tagged publishing workflow passed at 4de8520c27f64fff467f5dbef735ad16e4f936ec. The publishing run passed 8 quality gates and attested 14 required source checks. Serial pytest passed 12,475 tests (83 skipped, 7 deselected); response-schema recapture also passed 12,475 tests, and all 237 response contracts matched. The Linux bundle passed all 10 installed-runtime checks, including the real browser and restart read-back. Windows archive/launcher checks, clean wheel runtime, evidence sealing and public download digests passed. A separate clean public PyPI installation verified 13 modules, 612 Skills, pinned OpenMHS setup, all 746 market entries, MCP tool discovery and MHS HTTP device discovery on fake backends. The published wheel/sdist and original Linux/Windows downloads match the sealed build receipts. The later macOS DMG is covered by its separate supplemental validation record and the updated SHA256SUMS. The GHCR publishing workflow passed its container runtime checks. An anonymous post-publication pull and startup also passed, including exact version, health and authentication. 1.0.0 and latest both resolve to sha256:95817316f1db79e88b8a70889e23dc650ae16f182ec9e29cfa058133c7bb9bf0 (linux/amd64).
  • The lab is simulation only. Market devices are contract-checked on their fake backends; that is not a hardware qualification. Driving hardware remains an explicit operator decision outside the lab, and a software stop never replaces a hardware emergency stop.
  • Deprecated bare @name references: the unpinned @name artifact spelling was announced for removal in 0.4.0. Because the workbench autocomplete can still insert a bare name for an artifact row without a version, the removal moves to 1.1.0; use the pinned @name#v-<version_id> form.
  • Lab packaging: the Python and desktop packages do not bundle OpenMHS; run openai4s lab setup with Git installed and network access to GitHub. The container image bundles neither OpenMHS nor Git: mount a public OpenMHS checkout and set OPENAI4S_OPENMHS to its path inside the container. The post-publication Lab check used the public PyPI installation on the host; container Lab integration was not exercised.
  • Native MHS tools in a running daemon: a device added on the MHS surface is reachable at once through the Lab routes; if the daemon was started before the lab was ever set up, its native mhs_* tools register on the next Lab action or restart.
  • Windows/WSL2: the public x86_64 ZIP passed post-publication runtime validation on Windows Server 2025 with Ubuntu 24.04 WSL2: the shipped launcher, authenticated Windows Chromium dashboard, nine science/runtime checks (including plots, immutable artifacts and restart read-back), status/stop, and enforced bubblewrap isolation. The test uses no model credentials; doctor correctly reports that a model key and optional R environment still need user configuration. Its launcher remains unsigned. Windows ARM and the full Windows 10/11 host matrix were not exercised.
  • Containers and network isolation: unprivileged containers may lack working bubblewrap namespaces. In that state, allowlist mode refuses new Cells. See the security and Docker guides.

Contributors

Thank you to everyone who contributed since v0.3.1:

  • @Nobody-Zhang — OpenMHS integration and lab, Ollama support, provenance, workbench reliability, the next promotion, dependencies and release engineering.
  • @dvd233 — exact input lineage through frozen reads and grouped outputs, and execution-history response ownership.
  • @JunkaiWang-TheoPhy — interleaved tool-call fragments, recovery-candidate cleanup, and replayed socket events.
  • @Gary-2048M — per-model and per-wire request parameters.
  • @Linmj-Judy — the molecular-dynamics Skill.
  • @brainsparker — the optional You.com search integration.

Full changelog

v0.3.1...v1.0.0