Open Science Desktop v0.3.3

Open Science Desktop v0.3.3 brings the bundled agent runtime up to date, makes clusters that ask for a password or a one-time code usable, and adds a reviewer agent — plus fixes for two field-reported bugs.

Features

Bundled runtime: OpenCode 1.17.13 → 1.18.12

• The model catalog and reasoning-effort levels now match a current OpenCode instead of a nine-release-old one, so efforts such as max appear where the provider offers them. (#74)
• Settings → Runtime now shows which OpenCode version the app is running.
• A per-agent effort pinned under the old runtime is reconciled away once the catalog is known: from 1.18 a model's efforts come from the catalog, and some models lost levels the previous runtime synthesized — the runtime accepts such a value and then silently applies nothing.

Interactive SSH sign-in for 2FA clusters

• Institutional clusters that ask for a password or a one-time code on every connection were unusable, not merely slower: non-interactive ssh can never answer that prompt. The app now opens one authenticated connection per host over a real terminal, relays the server's own prompts to a dialog, and routes every later ssh, scp and sbatch — the app's own probes and the agent's commands alike — through it. One sign-in covers the working session. (#73)
• Prompts are relayed verbatim rather than pattern-matched, so Duo, PAM and campus OTP flows nobody anticipated still work. Secrets go into the terminal, never a process argument, and nothing is persisted.
• The agent can reach the same dialog mid-run through a new ssh_connect tool, so a run continues instead of dying on "Permission denied".
• Connection sharing is a unix feature; on Windows the Compute card says so instead of offering a sign-in that would need repeating per command.

Reviewer agent and opt-in auto-review

• A bundled read-only reviewer agent plus a /science-review command, orchestrating the existing traceability, stats-integrity and domain-check skills; its findings render as a structured card in the thread. (#72)
• Auto-review is off by default. When enabled, a turn that actually wrote or edited workspace files earns one read-only reviewer turn — never a review's own turn, never an interrupted or errored one, one at a time with a queue.

Per-agent reasoning effort

• Each agent can be pinned to its own reasoning effort alongside its model, so a reviewer can think hard while a titler stays cheap. The composer's slider only reaches the turn you send, so subagents need this. (#71)

Fixes

• Windows: projects showed "no sessions" even with sessions inside them. A project's path came from Rust's canonicalize() (the verbatim \\?\D:\… form) while the sidecar reported D:/…, so every match between the two failed — which also broke project-scoped memory, the active-project highlight, and move-to-project. Paths are now handed out in native form and compared by a normalized key; installs that already persisted a verbatim path repair themselves without re-picking a folder. (#76)
• Zoom no longer restructures the desktop into phone layout. Page zoom shrinks the CSS viewport, so at 175% a 1200 px window reported ~686 px and the whole shell flipped to the phone layout — sidebar as an overlay drawer — because the user asked for bigger text. (#63)
• A finished answer no longer keeps a spinner turning under it. OpenCode re-emits a turn's user message just after the turn ends, which re-locked the session the instant it completed; nothing cleared it until a server poll ~15 s later, which also rebuilt the whole thread.
• A provider-blocked prompt now explains itself. OpenAI's content filter answers Request blocked. when it rejects the prompt — and on the Responses API a prompt is the whole conversation, so every retry fails identically. The app now says so and names the ways out. Error text also reaches debug.log, redacted so a credential echoed back by a provider cannot land in a file users attach to bug reports.
• Auto-review queue: a pane finishing several file-changing turns while another session was being reviewed no longer earns duplicate paid reviews, and an interrupted or errored turn no longer silently cancels a review earned by an earlier one.
• Interactive sign-in: the managed ssh configuration is no longer handed to platforms that cannot share a connection, and one ssh_connect call no longer starts two competing sign-ins.
• Security: both open Dependabot alerts closed (postcss GHSA-r28c-9q8g-f849, a second brace-expansion advisory GHSA-mh99-v99m-4gvg), with every affected major line pinned above both advisories' ranges.

Downloads

• macOS Apple Silicon: Open.Science_0.3.3_aarch64.dmg
• macOS Intel: Open.Science_0.3.3_x64.dmg
• Windows: Open.Science_0.3.3_x64-setup.exe or Open.Science_0.3.3_x64_en-US.msi
• Debian/Ubuntu: Open.Science_0.3.3_amd64.deb
• Fedora/RHEL: Open.Science-0.3.3-1.x86_64.rpm

The macOS packages are Developer ID signed, notarized, and stapled. They should open normally without using xattr.

Contributors

Thanks to @ShellLayer, @mdanh-bio, @aayushkrm, @SIITW, @Alfie2028 and @ChengjieShan for the reports and requests, and to @almosme for #70, whose diagnosis pointed at the Windows path mismatch.

Full Changelog: v0.3.2...v0.3.3