Open Science 0.8.1
A patch release that keeps you oriented during long-running work — category-level context breakdowns, durable unread task badges, and a searchable, expandable file library — while hardening artifact provenance, provenance recovery, and R package management.
Open Science is an open-source, model-agnostic AI workbench for scientific discovery — a self-hosted desktop app pairing a plan-and-execute agent with persistent compute, durable project/session storage, and inspectable research artifacts.
v0.8.1 is a patch release on top of v0.8.0. There are no breaking changes and no Roadmap phase status changes. It sharpens day-to-day orientation — showing where the context window is spent, surfacing finished tasks that need attention, and making the project file library faster to search and browse — and stabilizes the artifact provenance, upload migration, and managed-environment behavior introduced in v0.8.0. website Open Science v0.8.1 release notes Open Science Github
Highlights
See where your context window goes. The context-usage indicator now breaks the window down by category — system guidance, conversation, tool activity, connectors, and skills — using a model-aware estimator, so you can tell what is consuming the budget before a request. The agent-reported total stays authoritative; only the category breakdown is estimated locally. Never lose track of a finished task. A terminal result from an agent turn now marks its conversation unread unless the focused window already shows it, and a blocking approval request asks for native attention. Ordinary completions and failures keep using the system notification and unread badge paths. Find and browse files faster. The project file library gains source-scoped filename search, compact list and grid views, and a large expand modal that preserves scroll position, active filter, loaded pages, and thumbnails when moving between the inline panel and the full view. Hardened artifact and provenance stability. Generated files survive branch switches even when finalization lands in the same instant as the terminal message; prepared runs recover after a provider turn ends but before files attach; legacy uploads migrate cleanly into immutable versions without stranding live sessions; and finalization failures surface a stable recoverable code instead of matching human-readable error text. R stays pinned during package installs. Installing an R package through the managed environment can no longer replace the live notebook kernel, and provenance no longer implies that every observed package change came from the current operation. Claude Code keeps its built-in web tools. WebFetch and WebSearch are now explicitly part of every normal Claude Code session contract, rather than depending on an upstream adapter fallback that could change silently.
New Features
Category-level context-usage estimates — a model-aware estimator selects the right tokenizer for each model family and tracks cumulative system, message, tool, connector, and skill context, replacing snapshots instead of double-counting them. Durable unread task state and attention alerts — completed turns mark their conversation unread unless the focused window freshly proves it visible, blocking approvals request native attention, and platform-native badge behavior reaches macOS, Windows, and supported Linux shells. Artifact search and list view — search within all artifacts, your uploads, or a selected session, and switch between grid and compact list views without reloading metadata. Expandable project file library — a new expand affordance opens the file library as a large centered modal without remounting, so scroll position, filter, pagination, and thumbnails are preserved. Read-only package inspection — a dedicated inspection tool lets the agent check whether a Python or R package is installed and which version is present, without reaching for the mutation-oriented package tool. Built-in web tools for Claude Code — normal Claude Code sessions explicitly select the complete built-in tool preset, keeping web fetch and search available while the app permission broker stays in control.
Improvements
Public product documentation now leads with quantified skill and connector counts, an explicit Remote Compute (SSH) entry, and CLI FAQ answers, so the capability inventory and common questions are easier to discover. Installing or uninstalling a package through the managed environment now reports the verified before/after state of the explicitly requested packages, making repeated installation and version-change outcomes easier to explain.
Bug Fixes
Historical finalization ownership — every published runtime event now carries a UUID-backed lifecycle namespace, so a new artifact event can no longer collide with an older event after an app relaunch, and prompt ownership is required when an event id is reused. Generated files across branch switches — artifact finalization is now a true no-op only when the session metadata, flat message, and conversation graph all carry the finalized version ids, so switching branches no longer hides a generated file. Prepared artifact run recovery — an atomic prepared finalization intent is persisted before publishing the renderer event, so a crash after the provider turn ends can still prove ownership, seal write ingress, finish compatibility publication, and restore session metadata. Legacy upload migration — live migrations now publish through an atomic copy, persist a path-free session, and merge only proven identity deltas into renderer state, so a save, sibling upload, or later delete can no longer strand a session on stale paths. Stable finalization diagnostics — the finalize-run response now carries one stable ownership-persistence-race failure code for the single recoverable case, while prompt, role, frame, branch, runtime-segment, and compatibility failures stay terminal. Graph upload data paths — upload path decoding now transforms both the flat message projection and every conversation-graph message, so legacy sessions without a graph no longer fail immutable identity validation on startup. R package upgrades — the complete installed r-base version and build are pinned, a supervised dry run rejects any r-base link/unlink action, and the full identity is verified after every real install or uninstall regardless of exit code. DOI and trial-query handling — DOI suffix cleanup is now linear-time character trimming instead of a regex, and backslashes are escaped before quotes when constructing clinical-trial phrase filters, closing reported exponential and injection-shaped inputs. PowerShell resolution on Windows — Windows PowerShell is resolved from the system root with a fallback, so managed cache ACL preparation and notebook shell commands no longer fail when system directories are absent from PATH.