v0.15.0

Open Science v0.15.0

Chinese (Simplified and Traditional) interface translations, a configurable reviewer model, cross-drive folder browsing with an editable path bar, bulk skill management, save-as-skill from a conversation turn, direct user skill folder support, a host viewImage capability, CLI Codex device login, persisted branch source lineage, GLM-5.3, and reviewer correction attribution — alongside separated notebook approval and execution state, hardened skill package authoring, Crashpad crash diagnostics, and faster streamed-text rendering.

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.

v0.15.0 is a feature release. It introduces full interface translations for Chinese (Simplified and Traditional) covering settings, onboarding, workspace, and common components, with a runtime language switcher in Settings. The reviewer gains a configurable model policy: follow the active model or pin a dedicated provider, model, and reasoning effort with an isolated runtime for the complete review chain, plus durable assessment snapshots so correction attribution and review history survive restarts. Local folder access expands to cross-drive browsing with an editable path bar and a drive switcher. Skills gain bulk enable/disable management with filters, a save-as-skill action that turns a completed conversation turn into a reusable skill through the active specialist, and direct user skill folder support that validates out-of-band copied packages. A new host viewImage capability lets agent scripts attach cropped, orientation-normalized images to REPL results. The CLI adds Codex device login for headless servers. Session branch lineage is now persisted so a branched session remembers its source. GLM-5.3 joins the GLM Coding Plan catalog. On the reliability side, notebook tool presentation splits approval from execution state with stable presentation identity across replays, skill package authoring gains a deterministic inspector and structured validation, local Crashpad crash diagnostics are enabled across desktop platforms, and streamed-text rendering overhead is reduced.

✨ Highlights

• Chinese (Simplified and Traditional) translations. The full interface — settings panels, onboarding, workspace, composer, home, and common components — is now available in zh-Hans and zh-Hant, with a runtime language switcher in Settings and locale-aware date and time formatting. (#1196)
• Configurable reviewer model. A new Reviewer model policy under Settings lets the reviewer follow the active model or pin a dedicated provider, model, and reasoning effort. A pinned selection owns one fully resolved backend and an isolated ACP runtime for the complete initial-review, correction, and scoped re-review chain; the actual model used is persisted on each review. (#1199, #1225)
• Cross-drive folder browsing. Local folder grants now work across all mounted drives — drive letters on Windows, volumes on macOS, and mounts on Linux. The grant dialog gains an editable path bar (click to type a path directly) and a drive switcher; the Files tab gains a matching drives group. (#1168)
• Save as skill from a conversation turn. A completed conversation containing a reusable workflow can now be turned into a skill directly: Save as skill in the composer menu asks the active specialist to evaluate the conversation and author the skill, with the control prompt hidden from the transcript. (#1180)

🚀 New Features

• Chinese (Simplified and Traditional) translations — complete i18n infrastructure with natural-language keys, full zh-Hans and zh-Hant catalogs, a runtime language switcher in Settings, and locale-aware date/time formatting across settings, onboarding, workspace, and common components. (#1196)
• Configurable reviewer model routing — an application-wide Reviewer model policy under Settings that follows the active model or pins a provider, model, and reasoning effort; pinned execution owns one fully resolved backend and an isolated ACP runtime for the complete review chain; the actual runtime model is persisted on the review row. (#1199)
• Reviewer correction attribution and review history — reviewer-correction attribution is preserved across runtime, persistence, and reload; versioned reviewer assessment snapshots keep the checks and outcomes that produced each reviewer card; correction status renders compactly without duplicating feedback; correction_failed stays distinct from the fix-limit outcome. Includes immutable migration 0004_review_assessment_snapshots. (#1225)
• Persisted branch source lineage — branching a session into a new session now records the source session, agent frame, and message branch, so lineage survives restarts. Existing sessions without a source are treated as having none. (#1224)
• GLM-5.3 on the Coding Plan — GLM-5.3 is now the default model in the GLM Coding Plan catalog with its 1M context window and low/high/max thinking levels. (#1227)
• Host viewImage capability — agent scripts can attach images to REPL results with host.viewImage(source, options), supporting current-project artifacts/uploads and session workspace paths, explicit cropping, EXIF orientation normalization, and PNG/JPEG-only output with source and size budgets. (#1197)
• Bulk skill management — a dedicated Skills › Manage skills page lists imported and personal skills with source, status, and text filters, select-all, a persistent selected count, and a Selected view that gathers selections across filters; enable/disable selected applies atomically with full-batch validation. (#1202)
• Cross-drive folder browsing — the grant-folder dialog and Files-tab browser enumerate all mounted drives; the grant dialog's breadcrumb bar swaps to an editable path input on click; a leading drive crumb opens a drive/volume switcher; any absolute folder on any mounted drive can be granted. (#1168)
• CLI Codex device login — open-science codex login runs the native Codex CLI's device-code flow directly in the terminal with app-owned credential storage, so headless or remote servers can complete login from a terminal-only session; respects the configured proxy policy. (#1200)
• Direct user skill folders — directly copied skill packages in the personal or imported directories are validated and projected across all agent frameworks; canonical directory names are the invocation names; a fail-closed policy rejects reserved prefixes, bundled collisions, and unsafe metadata; a rebuildable compatibility index avoids rehashing unchanged packages. (#1193)
• Save-as-skill turn — Save as skill in the composer menu dispatches an app-owned continuation through the active specialist that evaluates the conversation and authors a skill; the hidden control node is excluded from transcript, history, export, and replay while the assistant response stays visible; works across Claude Code, OpenCode, Codex Responses, and Codex Bridge. (#1180)
• Hardened skill package authoring — the Host skills scripting surface gains a deterministic package inspector with recursive POSIX file inventory for exact SKILL.md reads, structured machine-readable validation reports, create-only and exactly-one-match edit semantics, and certified camelCase shapes across all four runtime routes. (#1194)
• Separated notebook approval and execution state — notebook tool presentation splits into prepared code, waiting for approval, executing, and terminal phases; a one-shot execution authorization joins the approved tool call to the run so replayed approvals no longer duplicate visible code; the implicit Python/R execution deadline is removed; agent cancellation propagates to admitted notebook executions. (#1156)

🔧 Improvements

• Streamed-text rendering overhead is reduced with lower per-frame render cost. (#1128)
• Capability projection is centralized so additive capability fields extend without breaking older projections. (#1226, #1221)
• The command module inventory is unified across application commands. (#1203)
• README Core Capabilities were backfilled with entries from past major releases. (#1205)
• CONTRIBUTING gains fork and upstream setup steps. (#1182)

🐛 Bug Fixes

• Prepared replay accepted after framework switch — a prepared skill replay left over from a different agent framework could be rejected after switching backends. Prepared replays are now accepted across framework switches. (#1223)
• viewImage requires workspace — the viewImage capability projection now requires a session workspace, preventing capability advertisement where image sources cannot resolve. (#1220)
• Local Crashpad enabled — local Crashpad crash diagnostics are now enabled across desktop platforms so crashes leave inspectable dumps. (#1219)
• Completion footer hidden during continuations — the completion footer could appear while a continuation was still streaming, suggesting the turn was complete early. The footer is now hidden during continuations. (#1218)
• Concurrent catalog additions tolerated — concurrent additions to the command catalog could race and fail; the catalog now tolerates concurrent additions. (#1208)
• Conversation aligned with composer — the conversation panel is aligned with the composer width for consistent visual rhythm. (#1209)
• Auditor prompts hidden from transcript — reviewer auditor prompts could appear in the local transcript; they are now hidden. (#1105)
• Decoded $DATA paths contained — decoded Windows $DATA alternate-stream paths could escape storage containment; they are now contained. (#1191)
• Cross-host request races prevented — file-browser requests across host switches could race and return stale results; races are now prevented. (#1192)
• Skill catalog refreshed after local changes — local skill changes outside the app (edits, copies) now refresh the catalog reliably, with observer bursts coalesced and polling reduced. (#1186)
• AppImage mounted before installed command — the installed CLI command now mounts the AppImage before running, so launcher paths resolve correctly. (#1198)

📦 Install

Requirements: macOS 12+ (Apple Silicon or Intel), Linux x64, 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 can update itself in place.

Download the appropriate package from the Assets section below:

Platform	Package
macOS (Apple Silicon)	DMG for ARM64
macOS (Intel)	DMG for x64
Linux	AppImage or Debian package for x64
Windows	Installer for x64

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 not notarized and may require approval through macOS Privacy & Security.

Windows — first launch (unsigned build). No Authenticode certificate yet, so SmartScreen shows a bypassable "unrecognized app" prompt (More info → Run anyway). 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: a local-first desktop, localhost-web, headless, CLI, and task-SDK surface over persistent projects and sessions with selectable message branches, branching into a new session with persisted source lineage, reversible archiving, project pinning, and persistent side conversations; selectable Claude Code, OpenCode, and Codex agent frameworks behind a shared provider turn-adapter interface; production subagent delegation with durable messaging, restart recovery, structured output, artifact and review evidence, and camelCase delegation APIs; review-gated session plans; a unified composer lane; hot-switching of compatible models and providers; multi-provider model configuration with DeepSeek V4 Pro Responses and GLM-5.3 support; a configurable reviewer model policy with an isolated review runtime and durable assessment snapshots; persistent Python/R/REPL kernels with separated approval and execution state and remote SSH execution; a host viewImage capability; immutable artifact versions with artifact lineage access; multi-format previews; file-based skills with conversational creation, save-as-skill turns, direct user folder support, bulk management, and imports; permissioned scientific connectors with immutable invocation identities; opt-in review and bounded correction; personal specialist agent profiles; scoped permission management; local folder access with cross-drive browsing; Chinese (Simplified and Traditional) interface translations; global proxy settings; prominent update reminders; conversation export as Markdown and PDF; a project-scoped command palette; CLI Codex device login; smooth live response rendering; collapsible side panels; split-view file preview; desktop notifications; a cross-surface notification message center; structured agent clarification cards; live session status; and session keyboard shortcuts.
• 🚧 Partial: R remains managed-only; provider choice remains constrained by the active framework's endpoint compatibility; remote compute is SSH-only; skills remain local; specialist roles are personal-only (no cross-machine sharing); and review is opt-in and record-scoped.
• 🗺️ Roadmap: a unified model gateway, a public skills commons and cross-machine specialist sharing, Slurm/cloud-GPU execution, stronger sandboxing and credential isolation, and collaborative research workflows.

🐢 Known Limitations

• R is managed-only. A bring-your-own R interpreter path is not built yet.
• Remote compute is SSH-only. Slurm and cloud GPU submission are not built yet.
• Provider choice is per framework, not one unified gateway. The available protocol depends on the selected agent backend.
• Hot-switching applies only to registered compatible targets. Framework, auth-lane, wire-route, or unsafe capability changes still require a reconnect.
• Code reconstruction is LLM-generated. It does not replace deterministic reproduction; portable environment locks and full-fidelity session replay remain open.
• Specialist roles are personal-only. There is no shared public commons or cross-machine specialist forking yet.
• The task SDK is a first-generation surface. Task creation, polling, artifact retrieval, run progress, and cancellation work; broader orchestration remains open.
• Switching agent backends cannot transfer in-flight tool state. Existing conversation history can replay, but a running action is not migrated.
• Skills are local only. There is no shared public commons, cross-machine forking, or user-facing version pinning yet.
• The reviewer is opt-in and record-scoped. It does not replace domain-specific validation of citations, units, statistics, or methods.
• Scoped permissions cover allow-grants only. Network sandboxing, directory-level file access control, and a credential vault are not built yet.
• Windows builds are unsigned. SmartScreen may warn on first launch; official macOS builds are notarized.
• No local GPU compute backend.
• No multi-user real-time collaboration.

🙏 Acknowledgements

Thanks to @ewen-poch, @wen2zhou, @roxi3906, @daanveer-tech, @LJLJ111, @LiuZhaoyan, @justemu, and everyone in Discord, X, and Discussions.