Skip to content

OpenAI4S v0.2.0

Latest
Latest

Choose a tag to compare

View all tags
@Nobody-Zhang Nobody-Zhang released this 25 Aug 08:48
v0.2.0
c5e80a3
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

OpenAI4S v0.2.0 is the first multi-platform release: one version, two desktop packages — macOS and Linux — both carrying the same embedded Python and offline science stack. (A Windows/WSL2 package is built and under stabilization; it ships in a future release.)

✨ Highlights

  • Two desktop packages — the Apple Silicon .dmg and a relocatable Linux x86_64 tarball with the same embedded Python and science stack.
  • Auto Mode (Stages 0–12) with a Guardian review boundary — autonomous runs route boundary actions to a read-only, digest-bound reviewer instead of failing closed.
  • Honest completion evidence — a crashed or unrun cell can no longer be rendered as a clean success; completion is reconciled against what actually executed.
  • Team Server M1–M4 — multi-tenant foundation, governance, cluster orchestration, and persistent Slurm sessions.
  • MCP Streamable HTTP transport, the Volcengine DataPro connector, Doubao web search, and Anthropic Messages SSE streaming.
  • 603 bundled Skills — the pinned 561-recipe bioSkills collection joins the 42 curated directories, installable anywhere via npx.
  • The interrupt-signal train — a running R cell is now reliably stoppable on every platform (tgkill the main thread, reset inherited SIG_IGN at spawn, never drop a delivered SIGINT).
  • Docker/Kubernetes deployment, the trajectory-ledger workbench view, read-only session sharing over an outbound relay, environments-as-a-transaction, and openai4s --version.

📦 Downloads

Asset Platform sha256
OpenAI4S-0.2.0-macos-arm64.dmg macOS 12+, Apple Silicon e3665e5e548da35540cb75f42fde99173593acc3986a799a28d6f6b28e3ce864
OpenAI4S-0.2.0-linux-x86_64.tar.gz Linux x86_64 2eb33f7bf1399752fcd7f6375c4158d2822259899123a46eb2321a35701b7caa

Verify a download against the attached SHA256SUMS: shasum -a 256 -c SHA256SUMS (macOS) or sha256sum -c SHA256SUMS (Linux).

  • macOS — drag the app from the DMG into Applications. The image is not notarized: on first launch use System Settings → Privacy & Security → "Open Anyway" (macOS 15+), or right-click → Open (macOS 12–14).
  • Linux — tar -xzf anywhere, then ./OpenAI4S to start (or ./install.sh for a per-user PATH entry and menu item; no root needed). Install bubblewrap so cells run sandboxed.
  • Windows — the WSL2-based package is not part of this release while it is stabilized; it ships in a future release. Meanwhile, install inside WSL2 with pip install openai4s, or unpack the Linux tarball in your WSL2 distribution.
  • PyPI — pip install openai4s works on every platform, including Intel Macs and arm64 Linux.
  • The R kernel is not bundled in any package (it needs a conda environment): install micromamba/mamba/conda, then run openai4s setup.

What's Changed

Platforms & packaging

  • docs: macOS DMG startup guide + README refresh (unblocks stale PyPI-publish governance tests) by @Nobody-Zhang in #35
  • feat(release): Linux and Windows desktop packages, and close an xattr read of denied files by @Nobody-Zhang in #58
  • fix(release): stop anchoring the Linux bundle to PyPI, which can never hold it by @Nobody-Zhang in #67
  • fix(windows): stop wsl.exe's stdout becoming the launcher's return value by @Nobody-Zhang in #68
  • feat: add supported Windows WSL2 launcher by @Nobody-Zhang in #81
  • feat(deploy): official Docker and Kubernetes deployment by @Nobody-Zhang in #89
  • release(v0.2.0): version bump, CLI --version, and the v0.2.0 README sweep by @Nobody-Zhang in #120

Agent & kernel runtime

  • fix: improve Conda environment discovery by @Linmj-Judy in #37
  • Promote v0.3 to main: dual-loop hardening, R as a first-class channel, and two review rounds by @Nobody-Zhang in #52
  • Add Anthropic Messages SSE streaming by @stau-7001 in #79
  • fix: land the completion-evidence reconciliation train — honest completion reporting, serve/stop lifecycle, suite hermeticity by @Nobody-Zhang in #82
  • fix(kernel): preserve R fallback capture and refresh contributors by @Nobody-Zhang in #90
  • feat(auto-mode): Stages 0–12, and the review pass that hardened them by @Nobody-Zhang in #103
  • fix(kernel): never drop a delivered SIGINT, and install the Skill library with npx by @Nobody-Zhang in #111
  • fix(kernel): make the interrupt SIGINT actually land — tgkill the main thread, reset inherited SIG_IGN at spawn by @Nobody-Zhang in #119

Web workbench & API

  • feat(webshare): share sessions read-only via an outbound relay tunnel by @Nobody-Zhang in #38
  • fix(webui): route web sharing through the versioned API prefix by @Nobody-Zhang in #42
  • refactor(server): introduce declarative route registry for kernel API by @wangyu-sd in #87
  • ui: complete the trajectory ledger view by @Nobody-Zhang in #91

Science: skills, connectors & compute

  • science: add BioProBench skill (with review fixes) — supersedes #36 by @Nobody-Zhang in #39
  • Feat/retrosynthesis workflow triage by @wangyu-sd in #70
  • feat: add the Volcengine DataPro connector and Doubao web search over a new MCP Streamable HTTP transport by @Nobody-Zhang in #88
  • fix(doubao,mcp): a fully-read response died on its own closed socket by @Nobody-Zhang in #96
  • Add model-backed retrosynthesis planning and deployment by @wangyu-sd in #98
  • feat(skills): integrate pinned bioSkills collection and fix CI test bugs by @Linmj-Judy in #108
  • feat(benchmark): add tool bring-up verifier and workflow cases by @Linmj-Judy in #109
  • fix: MCP tools discovery crash, false-ready env steps, uv fallback by @Nobody-Zhang in #115
  • feat(model-backend): formalize bring-up & admission; ship protein-design MCP; fix CI test bugs by @Linmj-Judy in #117

Team Server & orchestration

  • feat(team): Team Server M1–M4 — multi-tenant foundation, governance, cluster orchestration, persistent Slurm sessions by @Nobody-Zhang in #92
  • fix(team,orchestration,kernel): the defects a max-effort review found by @Nobody-Zhang in #104
  • release: next → main — Team Server M1–M4, auto mode Stages 0–12, cluster orchestration, and the fixes two review passes found by @Nobody-Zhang in #110

Security & trust

  • Trust Foundation (0.1.x Trust Patch): P0 correctness, secrets, durability, contract v1 by @Nobody-Zhang in #41
  • Tenth-round review fixes: terminal-evidence integrity, keychain boundary, resume cursors, and per-directory docs by @Nobody-Zhang in #44
  • fix(security): survive squash-rewritten fingerprints without widening a path rule by @Nobody-Zhang in #57
  • fix(compute): fail closed in the BYOC confinement probe, and describe the boundary that now exists by @Nobody-Zhang in #62
  • fix(security): an inline gitleaks:allow cannot cover the commit before it by @Nobody-Zhang in #63
  • fix(security): deny the whole read class over $HOME in the BYOC profile by @Nobody-Zhang in #65
  • v0.3 trust-boundary batch: P0-0 through P1-B, audited against production call chains by @Nobody-Zhang in #69

Docs, CI & tooling

  • docs(readme): credit joint research laboratory by @ClarenceYC in #53
  • fix(ci,contract): exempt dependabot by branch, and unpin a raced response shape by @Nobody-Zhang in #54
  • chore(lint): move ruff to astral-sh v0.16.0 and pin the rule set explicitly by @Nobody-Zhang in #55
  • test(byoc): probe the git directory a worktree actually has by @Nobody-Zhang in #56
  • docs(readme): name the originating lab in the citation block by @Nobody-Zhang in #60
  • chore: sync next with main after the v0.3 squash and the July 31 gate fixes by @Nobody-Zhang in #61
  • chore(ci): retire the history secret scan and the Linux sandbox job by @Nobody-Zhang in #64
  • fix(contract): stop publishing a thread race as the example-session shape by @Nobody-Zhang in #66
  • docs(readme): update joint laboratory attributionUPDATE: orgnization info by @ClarenceYC in #74
  • docs: refresh the root README and correct 28 directory READMEs against the code — plus the five source defects that audit found by @Nobody-Zhang in #77
  • chore: run the offline suite — and the response capture — in parallel by @Nobody-Zhang in #107

Dependencies

New Contributors

Contributors

  • @Nobody-Zhang — 43 PRs — primary maintainer
  • @Linmj-Judy — 4 PRs — bioSkills collection, tool bring-up verifier, model-backend bring-up, conda env discovery
  • @wangyu-sd — 3 PRs — retrosynthesis workflow & planning, declarative kernel-API route registry
  • @ClarenceYC — 2 PRs — joint research laboratory attribution
  • @stau-7001 — 1 PR — Anthropic Messages SSE streaming
  • @YuyangSunshine — BioProBench skill (authored PR #36, landed via #39)
  • @dependabot[bot] — 7 dependency PRs

Commit trailers also credit the Claude models used throughout development as AI pair-programming assistants.

Full Changelog: v0.1.0...v0.2.0

Contributors

dependabot, YuyangSunshine, and 5 other contributors
Assets 5
Loading