Skip to content

AIPOCH Open-Science v1.3.2

Latest

Choose a tag to compare

@github-actions github-actions released this 20 Nov 10:42
· 45 commits to main since this release
10f9348

AIPOCH Open-Science v1.3.2

Lighter long conversations, Kilo Gateway as an official provider, image model catalogs you can check, reliable first messages in agent sessions, and PDF translation that keeps the native layout.

AIPOCH Open-Science is an open-source, local-first, model-agnostic AI research workbench for scientists and researchers. It enables reproducible, inspectable research across models with scientific AI agents, Python and R execution, scientific data connectors, and cross-platform support for macOS, Windows, and Linux.

v1.3.2 follows v1.3.1 with a polish-and-reliability release. Long conversations and sessions use less memory, so large research sessions stay responsive. Kilo Gateway joins as an official provider — add it with or without an API key and see the free models your connection is eligible for. Official image services for OpenAI, MiniMax, Bailian, and Volcengine Ark can refresh their bundled model list from the provider before you save. Fresh default Agent, Claude Code, and Codex conversations no longer fail before their first model request, and the built-in Agent starts scheduled sessions with their intended prompt. PDF translation recovers wrapped text in reading order and keeps formulas, links, and graphics in place, Notebook memory pressure announces itself without occupying the workspace, and the home screen names the specific reason a use-case download failed.

✨ Highlights

  • Long conversations use less memory. Streamed replies no longer accumulate hidden performance entries, and conversation bodies you are not viewing are reclaimed once they are safely saved — large sessions stay responsive without losing your place, pending edits, or active work. (#4500)
  • Kilo Gateway is an official provider. Add it with or without an API key: an anonymous catalog refresh shows the free models your connection is eligible for, and the onboarding provider step can be skipped to decide later. Existing credentials and main-model choices are preserved on upgrade. (#4465)
  • Image model catalogs you can check. Official image services for OpenAI, MiniMax, Bailian, and Volcengine Ark can refresh their bundled model list from the provider before you save, keeping your active model when it is still available. (#4474)
  • Reliable first messages in agent sessions. Fresh default Agent, Claude Code, and Codex conversations no longer fail before their first model request, and the built-in Agent starts scheduled sessions with their intended prompt. (#4505, #4490)
  • PDF translation that keeps the native layout. Wrapped titles, captions, definitions, and mathematical statements are recovered in reading order, and formulas, links, and graphics stay in their original places in the translated document. (#4492)

⚠️ Scope & Safety

  • Adding Kilo without a key contacts Kilo anonymously. The catalog refresh that shows eligible free models is a request to Kilo's servers that carries no account identity. Image and model requests after setup go to Kilo like any other provider connection. See data recipients. (#4465)
  • Checking an image model catalog contacts the provider. A refresh sends the configured connection's credentials to that provider to list its image models, and a failed check cannot retract what was already sent. Run it only for connections you configured. See data recipients. (#4474)

🚀 New Features

  • Add Kilo Gateway as an official provider with an optional API key; an anonymous catalog refresh shows the eligible free models, and saving without a key keeps the model choice open. (#4465)
  • The desktop onboarding provider step gains a Skip action that adds Kilo without a key and leaves the model decision to you; upgrades add any missing Kilo configuration without touching existing provider credentials or main-model choices. (#4465)
  • Official image services for OpenAI, MiniMax, Bailian, and Volcengine Ark can check the provider's live image-model catalog on demand: matching bundled models are shown before saving, the active model is preserved, and a model kept as default but missing from the refreshed catalog is marked. (#4474)
  • Bailian image-catalog requests are scoped to the workspace region, and China-linked credentials are limited to the China catalog. (#4474)

🔧 Improvements

  • Long conversations and sessions use less memory: streamed replies no longer accumulate hidden performance entries, inactive conversation bodies are reclaimed once their durable copy matches, and live session updates reuse serialized copies for delivery and replay. (#4500)
  • Notebook memory pressure is announced by a dismissible notice with a six-second expiry instead of occupying the workspace top row or opening a modal; refusal to start applies only to actual execution decisions, and recovery is announced once it completes. (#4499)
  • The home screen names the specific reason a use-case download failed — a missing manifest, unavailable package, timeout, network, server, verification, disk-full, or storage problem — instead of a generic failure. (#4498)
  • Replay keeps historical conversations clean: the ask and capture actions for completed steps appear on hover or keyboard focus, with full-size targets on touch or hybrid-pointer devices. (#4483)

🐛 Bug Fixes

  • Fresh default Agent, Claude Code, and Codex conversations no longer fail before their first model request when Notebook authority preparation resumes an already-attached or history-less native session. (#4505)
  • The built-in Agent opens scheduled sessions with their first prompt instead of interrupting an attached writer during Notebook credential refresh. (#4490)
  • PDF translation recovers text that native extraction split across rows or page boundaries — titles, captions, definitions, and mathematical statements — and keeps formulas, reference text, links, and embedded graphics in their original locations during writeback. (#4492)
  • Notebook startup no longer misreads a closed WSL command pipe as a failure. (#4500)
  • On Windows, the built-in Agent opens Reviewer sessions through the Reviewer service's named pipe using authenticated Streamable HTTP, and Reviewer cleanup closes the provider session while keeping review checkpoints in memory. (#4507)

📦 Install

Desktop client. Requirements: macOS 12+ (Apple Silicon or Intel), Linux x64 or ARM64, or Windows 10/11 x64. On first run, the onboarding wizard checks the environment and can install and configure an app-managed agent runtime. Once installed, the app updates itself in place.

Download the appropriate package from the Assets section below, or from the official download page. On macOS, the desktop app is also installable via Homebrew: brew install --cask open-science.

Platform Package
macOS (Apple Silicon) DMG for ARM64
macOS (Intel) DMG for x64
Linux AppImage or Debian package for x64 / ARM64
Windows Code-signed installer for x64

Command line via Homebrew (macOS and Linux). The standalone CLI and local Web UI live in the official tap — bundled Node runtime, no separate Node.js needed, and upgrades through the normal brew flow:

brew install aipoch/tap/open-science
open-science start --no-open
open-science url

Command line via npm. With Node.js 22.13+, install the CLI (and Task SDK) globally; updates then follow your normal npm upgrade flow:

npm install --global @aipoch/open-science
open-science --help

Command line standalone (all platforms, no toolchain). The Assets below also include standalone CLI archives — macOS and Linux arm64/x64 tarballs and a Windows x64 zip — with a bundled Node runtime, so they install and run without Homebrew, npm, or a local Node.js. Linux packages additionally arrive as standalone .deb and .rpm installers.

macOS — first launch. Official release builds are Developer ID signed and notarized by Apple, so they open like other trusted applications. A locally built copy is notarized differently and may require approval through macOS Privacy & Security.

Windows — first launch. Official release builds are code-signed; Windows recognizes the publisher and no SmartScreen warning appears. Verify that the package came from the official release page before continuing.

Build from source instead:

npm install
npm run build:mac   # or: build:linux / build:win

🧭 What's in this release (maturity)

  • ✅ Implemented: everything shipping in v1.3.1, plus Kilo Gateway as an official provider with optional API key and anonymous free-model catalog refresh, on-demand image-model catalog checks for official image services, lower memory use in long conversations and across sessions, Notebook memory-pressure notices that don't occupy the workspace, specific use-case download failure reasons, hover/focus actions in replay, and the fixes described above.
  • 🚧 Partial: remote compute covers direct SSH, Slurm, and user-connected Kubernetes, Hugging Face Jobs, Modal, and Ray targets, each requiring user-provisioned accounts, software, and permissions; provider choice remains constrained by the active framework's endpoint compatibility; WSL2 Bash remains an opt-in Windows x64 preview; the skills commons ships marketplace discovery but not cross-machine forking or user-facing version pinning; reproducibility ships replayable per-version verification, session replay, and RO-Crate packaging, while solver-exact equivalence remains open; review is opt-in and record-scoped.
  • 🗺️ Roadmap: a unified model gateway, cross-machine skill forking and version pinning, stronger sandboxing and credential isolation, and collaborative research workflows.

🐢 Known Limitations

  • Replay reproduces recorded steps rather than re-executing them, and code reconstruction is LLM-generated — solver-exact equivalence remains open, and replayable verification is the deterministic check (starting one from the web client currently requires the desktop app).
  • PDF annotations are bound to the file version they annotate and do not sync across machines.
  • AI-assisted outputs — study snapshots and the reviewer — reflect the configured main model and the record they cover; they are aids, not validation of citations, units, statistics, or methods.
  • Network sandboxing covers the notebook and compute runtimes, not the whole system — on Windows it applies only after the sandbox's one-time administrator setup, and Strict mode narrows Notebook egress to a manually managed whitelist whose built-in destinations deserve a review before you rely on them.

The full limitations and security register lives in docs/known-limitations.md.

🙏 Acknowledgements

Thanks to @ShuangChenyue, @liguohhh and everyone in Discord, X, and Discussions.


Full Changelog: https://github.com/aipoch/open-science/commits/v1.3.2