Skip to content

Saturn v0.6.0

Pre-release
Pre-release

Choose a tag to compare

@github-actions github-actions released this 09 Nov 10:53
1b4919f

Saturn is a workbench for AI research on your own computer: the app, its research assistant and its local runtime in one download.

Preview build. The macOS app is not signed with an Apple Developer ID (Apple Silicon only), and the Windows installer is signed with Saturn's self-signed certificate, which Windows does not trust, so both systems warn the first time you open them.

Install

macOS (Apple Silicon)

  1. Download Saturn-0.6.0-mac-arm64-unsigned.dmg, open it, and drag Saturn into Applications (the .zip has the same app).
  2. Open Saturn. When macOS says it can't verify the developer, go to System Settings → Privacy & Security and click Open Anyway.

Windows (x64)

  1. Download Saturn-0.6.0-win-x64-setup.exe and run it.
  2. If SmartScreen stops it, click More info → Run anyway.
  3. Optional: confirm the installer is Saturn's. In PowerShell, (Get-AuthenticodeSignature .\Saturn-0.6.0-win-x64-setup.exe).SignerCertificate.Thumbprint should print D285DA425077C16A3D115D61F9748AFA574BC786.

Then open Settings → Models and connect a model, add a project folder, and write your first question (the first launch walks you through it). The interface is in English or Traditional Chinese, following the system or chosen in Settings → General.

Updating: from a version before the rewrite in Rust and Tauri (0.3.1 or earlier), install this one by hand; your data in ~/.saturn-app keeps working. Later versions update themselves: the update downloads in the background and installs when you quit Saturn, and it must carry a signature from Saturn's updater key.

Check a download (optional): each file has a .sha256 beside it, and a signed build attestation that proves it was built from this repository's source by its public release workflow:

shasum -a 256 -c Saturn-0.6.0-mac-arm64-unsigned.dmg.sha256
gh attestation verify Saturn-0.6.0-mac-arm64-unsigned.dmg -R Saturn-Technology-Project/Saturn-App

The signed attestations are attached too (*.sigstore.json); with one, gh attestation verify <file> --bundle Saturn-0.6.0-mac-arm64.sigstore.json -R Saturn-Technology-Project/Saturn-App checks a macOS download offline (-win-x64.sigstore.json for Windows).

Saturn-0.6.0-sbom.cdx.json lists the crates, npm packages and model this release is built from (CycloneDX 1.5, with their hashes), for your scanner or your security team.

saturn-worker-0.6.0-linux-x64 and -linux-arm64 are the worker Saturn installs on a Linux SSH host when you press Install worker in Environment; you do not need to download them. Saturn checks each against the hash built into this release before sending it, and the host needs libssl 3 (Ubuntu 22.04, Debian 12, RHEL 9 or later).

Changes

What changed in each version, and whether it affects you, is in CHANGELOG.md.

What changed

Security

  • Permission rules now govern Claude Code and Codex turns. The rules under Settings → Permissions and in a project's .saturn/settings.json or .claude/settings.json were consulted only by Saturn's own agent loop; a Claude Code turn asked by its own approval setting alone, and a Codex turn asked only in Review mode. A deny now refuses the call and tells the program why, an allow runs it without a card in any mode, and an ask asks even in Bypass, naming the rule on the card. Claude Code's Bash is judged as the code tool and its file tools as files, by action and path; a Saturn skill is judged by its own name. Codex runs its own shell commands and edits without asking Saturn, so a rule cannot reach those.

Fixed

  • Windows: Saturn no longer opens a terminal window beside the app. The app started its runtime as a console program, so Windows gave the runtime a console window of its own, which showed its log for as long as Saturn ran. The runtime is now started without one, and every program the runtime starts in turn (git, Python, ssh, MCP servers, the desktop helper, nvidia-smi and the rest) is started without a window too, since a program started by a process without a console would otherwise get a visible one. The runtime's log still reaches the app, which shows its last lines when the runtime fails to start. Windows has only been type-checked for this change, not run.
  • Windows: uv and Python on PATH are found, and the managed environment can be made. Environment looked for programs named uv and python3 as Unix spells them, so on Windows it found neither uv.exe nor python.exe and offered no way to make the environment; and its jobs ran through /bin/sh, which Windows does not have. Programs are now looked for as Windows spells them (the Store's empty python.exe alias does not count), the environment is made with the Python that was found, and the jobs run through cmd with its own quoting. Type-checked for Windows, not run there.
  • The training template's checkpoints now carry the GPU's random state, and a resume reports only what it put back. train_loop.py saved torch's CPU generator alone, so a run resumed on CUDA would have drawn its dropout and sampling noise from a fresh generator while its report said rng was restored; in fact a resume on a GPU stopped before that with "RNG state must be a torch.ByteTensor", because loading the checkpoint onto the device moved the generator's state there too. The checkpoint now holds each generator by name, Python's, NumPy's, the CPU's and every CUDA device's, and the resume report names each one restored (rng:cpu, rng:cuda) and lists under notRestored any the checkpoint held but this machine could not take, such as CUDA state from another number of devices. The skills on resuming say the same, and that a resumed GPU run agrees with an uninterrupted one only to within its kernels' nondeterminism. A program copied from the earlier template keeps its old behavior until the two functions are copied again.
  • A remote job's output comes back as fast as SSH carries it, and a long training log no longer stops it. Each exchange with the worker carried at most 128 KiB per output and the next waited a second, so a busy log fell minutes behind, and any output past 64 MiB stopped synchronization. Parts are now up to 1 MiB and follow one another while more is waiting, and logs, metrics and resource readings may reach 1 GiB each. Install the worker again on each SSH host after upgrading.
  • The research assistant is made only once. When two screens asked for the assistant at the same moment on a first start, each could make one, and chats were split between them. They now share the one being made.
  • The Workspace terminal keeps what a command set for the next one. Each command runs in a new shell, so source .venv/bin/activate, export or a shell function used to be gone by the next command, and Python ran outside the virtual environment. Each terminal panel now carries its environment, functions and shell variables from one command to the next, so an activated environment stays active and deactivate restores the old PATH. (An alias carries over only in bash; zsh does not apply aliases inside a script.) A new project's terminal starts fresh. The environment stays inside Saturn's runtime and is never sent to the app. Not on Windows yet.

Changed

  • Lists of choices show which option the pointer is on. In the dark theme the highlighted option of a drop-down list (the assistant's Ask, Approve, Auto and Agent modes, a permission rule's allow, ask and deny) had no background of its own, since the highlight color was the list's own color, so only its explanation turning white marked it. The highlighted option is now tinted in both themes and its explanation stays muted, a list whose options carry explanations is wide enough for each to sit on one line, and the check beside the chosen option sits at the option's middle rather than on its first line.
  • The Environment screen's managed-environment card lines up. Its first line (the .venv, whether it is in use, Use and Sync from lock file) was indented further than the Other packages field under it and carried a stray divider, and the card sat flush against the job-timeout field. The line now aligns with the field, "in use" is a badge rather than part of the sentence, the card stands off from the field, the timeout says what it applies to, and a project without a lock file is told so beside the Sync button instead of only in its tooltip. The job card under it is laid out the same way, with its log in its own box.
  • A remote run is followed over one SSH connection instead of one a second. Saturn used to open a new SSH connection every second to ask the worker for new output. It now keeps one connection open while the run goes on, and the worker sends what the run writes as it is written, so samples and logs appear within a fraction of a second and the host sees one connection per run rather than thousands. A stop goes down the same connection. If the connection drops, Saturn says so, catches up a second later from what it already has, and opens it again; the computation on the host is never affected. The worker on the host must be this version of Saturn, as before.
  • LM Studio is now Local server, and MLX is part of it. The provider under Settings → Models takes any model server on this computer with OpenAI-compatible endpoints: LM Studio, Ollama, vLLM, llama.cpp or mlx_lm.server, at the address you give. Saturn lists the models the server has loaded, leaving out embedding and reranking models, and sends no key. The separate MLX provider is gone, since it did nothing LM Studio's path did not; a model saved under it is a Local server model now, at the same address.
  • The semantic code index and its embedding models are gone. No screen had reached the index since the research interface, and the assistant's tools never used it; the project's files, the research records and memory are what the assistant reads. The bundled embedding model, its download during packaging, the Hugging Face connection it allowed and the Embedding model setting go with it, so the package is about 120 MB smaller and the runtime no longer carries SQLite. An index/ folder left in ~/.saturn-app can be deleted.
  • The MCP marketplace is gone, with its sign-in callback: nothing in the interface had listed or installed from it since the research interface. Servers installed from it earlier stay connected as ordinary MCP servers, and any MCP server is still added by its address or command under Settings → MCP & Skills.
  • Writing a finding in the Workspace is one card. What you understood takes the full width, with its kind and Save beside each other under it, ⌘↵ (Ctrl+Enter) saves, and an empty list says where findings are kept.
  • The Workspace terminal is denser. Lines sit at 1.45 times the text size instead of twice it, commands are closer together, and "Clear terminal" sits in the top corner instead of taking a row of its own, so more output fits in the panel.

Added

  • Home offers to start from two evaluations you already have. A project without results shows "Compare two evaluations" above the first-study checklist, and the results step offers it too: the assistant previews both files, imports them once you agree, pairs them by sample ID and lists the samples that regressed, so the question and hypothesis can follow once there is something to explain.
  • F12 follows a definition into an installed package. Going to the definition of torch.nn.Linear or a standard-library function opens that file read-only in its own tab, at the definition. Only the selected Python's packages and the Rust toolchain's and crates' sources are read this way; anything else outside the project still is not.
  • The research assistant can have several chats in a project. + starts a new chat, and the chat's name at the top of the panel opens the project's chats, most recently used first, where one is opened, renamed or deleted. A new chat is named after its first message, and each keeps its own messages and mode. A project opens on the chat it had open last, and a new chat is only kept once a message is sent in it. Your existing chat is kept and named after its first message.

What's Changed

  • Workspace: the terminal keeps its environment, a denser terminal, and one card for a finding by @rickwengdev in #80
  • Assistant chats, F12 into installed packages, faster remote output, and first-hour help for new researchers by @rickwengdev in #81
  • Permission rules reach Claude Code and Codex; the code index, embedding models and marketplace are removed by @rickwengdev in #82
  • Local server replaces LM Studio and MLX; the About usage summary is removed by @rickwengdev in #83
  • Remote runs: one SSH connection follows a run and streams its outputs by @rickwengdev in #84
  • Research: the training template keeps the GPU's random state, and a resume reports only what it put back by @rickwengdev in #85
  • Environment: the managed-environment card lines up with its fields and says what it needs by @rickwengdev in #86
  • Interface: the highlighted option of a list is tinted, and its explanation stays muted by @rickwengdev in #87
  • Windows: no console window, and uv and Python found on PATH by @rickwengdev in #88
  • Release: 0.6.0 by @rickwengdev in #89

Full Changelog: v0.5.1...v0.6.0