Skip to content

Releases

Tensorseed 0.3.2

Choose a tag to compare

@flyingmongoose flyingmongoose released this 15 Nov 17:12

0.3.2 makes a fresh install work without hand editing, renames the
Flatpak app id for Flathub, and builds every release asset in CI,
including Windows ARM64 installers for Desktop.

Fixed

  • A fresh install follows the official Tensorseed catalog. With no config
    file, or a config that names no catalog of its own (for example one that
    only has the [torrent] settings Desktop writes), the CLI and Desktop use
    the official catalog: name tensorseed, publisher key
    RWS4/z1UPQ4ezmjrcpKqDYpqyn5h1WUS2oDs1B73Dh91kVzzNIZd03xd (key id
    CE1E0E3D543DFFB8), https://catalog.tensorseed.dev/, and the mainnet
    Chia store with strict chain checking. Desktop's Catalogs view no longer
    stops at "catalog.publisher_pubkey must be set". A [catalog] table,
    [[catalogs]], [[follow_lists]], file.base_url, or
    chia.catalog_store_id in your file (or a TENSORSEED_CATALOG__*,
    TENSORSEED_FILE__BASE_URL, or TENSORSEED_CHIA__CATALOG_STORE_ID
    variable) replaces the default, exactly as before.
  • catalog add and catalog follow-list follow on such a config first
    write the official catalog into the file, so adding a catalog does not
    stop following the official one.
  • tensorseed doctor reports a missing config file that follows the
    built-in official catalog as ok rather than a warning.

Changed

  • The Flatpak app id is now dev.tensorseed.Tensorseed (it was
    dev.tensorseed.desktop), because Flathub does not accept ids that end
    in .desktop. The metainfo, desktop file, and icons are renamed to
    match, the metainfo gains screenshots, branding colors, and release
    links, and a scalable icon is installed. The Flatpak build sets the
    Tauri identifier to the app id, so the single instance name stays under
    it; other builds keep dev.tensorseed.desktop.
  • Upgrading from a 0.3.x Flatpak: the new id installs side by side with
    the old one. Config, keys, catalogs, and pulled models are shared,
    because both use the same tensorseed folders, but Desktop's own
    preferences (desktop.json) start fresh. Remove the old one with
    flatpak uninstall dev.tensorseed.desktop once the new one runs.
  • tensorseed gui finds the Flatpak under either id.

Added

  • Windows ARM64 Desktop installers (NSIS and MSI) and a Windows ARM64 CI
    job for the CLI and Desktop. The in-app updater does not offer ARM64
    updates yet; install them by hand.
  • Release builds in CI: a v* tag builds every unsigned release asset
    (CLI archives and packages, Desktop AppImage, .deb, .rpm, Flatpak,
    and the Windows installers), writes SHA256SUMS, and attests build
    provenance for each file. Signing and publishing stay on your signing
    machine (docs/RELEASING.md).
  • A CI status badge in the README.

Docs

Tensorseed 0.3.1

Choose a tag to compare

@flyingmongoose flyingmongoose released this 15 Nov 05:03

Tensorseed 0.3.1 is a verified serving and catalog release. It adds
signed follow lists, claim expiry and revocation, a check-claim command,
a verify-only crate with a WebAssembly build and a JavaScript package, and
an endpoint check that no longer fails honest sellers. Tensorseed Desktop
is 0.3.1 too: it gets the library changes and is the first version offered
to 0.3.0 installs through the in app updater. Formats and commands may
still change before 1.0.

Highlights

  • check-endpoint no longer fails honest sellers. On one CPU host,
    0.3.0 passed an honest llama.cpp seller 1 run in 11 against its own
    reference set, because llama.cpp logprobs drift from run to run and one
    failing challenge failed the run. References now hold per-token medians
    over several passes (reference-logprobs --samples N, default 5), the
    run is decided over the whole set (ts.set-decision/1: any bad
    signature, more than --max-fail-fraction failing challenges, default
    0.25, or a median deviation above --max-median, default 0.35 nats),
    and tensorseed calibrate-reference writes your own server's spread into
    the reference file. 55 of 55 honest runs passed, and other
    quantizations, a smaller model under the claimed name, and tampered
    responses failed every run. --max-fail-fraction 0 keeps the 0.3.0 rule.
  • tensorseed check-claim <url-or-file> checks a seller's
    ts.served-weights/1 claim end to end: its hash and wallet signature,
    that it is served at its own public_base_url, that its catalog is one
    you trust with the same publisher key, manifest, and files, no takedown,
    catalog freshness, and the commercial serving mark recomputed from the
    catalog's license. Exit 0 pass, 1 fail, 2 inconclusive; --json writes
    ts.claim-check/1.
  • Claim expiry and revocation. An optional expires_at (set with
    --expires-in on claim-serving and serve-inference --sign) and
    signed revocation lists (tensorseed revoke-claim, served at
    /.well-known/tensorseed/revocations.json). check-claim,
    check-endpoint, and verify-response honour both; an unreadable
    seller list makes a passing seller check inconclusive.
  • Signed follow lists (ts.follow-list/1). A publisher signed list of
    catalogs that a marketplace, a community, or one user can follow
    instead of adding each catalog by hand: catalog follow-list create,
    sign, verify, follow, and unfollow. Followers pin the list key
    and the accepted rev, and a list only touches the entries it added.
  • Verify anywhere. tensorseed-verify is now the verify-only crate
    for claims, revocations, signed responses, catalogs, and follow lists,
    with no networking or serving dependencies, and it builds for
    wasm32-unknown-unknown. The WebAssembly build exports the claim and
    response checks, and the @tensorseed/verify JavaScript package (built
    and tested in this release, not published to npm yet) wraps it for
    Node and browsers.
  • Smaller fixes. verify-response checks URLs like check-claim
    (url_mismatch), the response skew window is configurable
    ([verify] max_skew_secs, --max-skew, 1 second to 7 days) and
    verify-response --at checks stored responses, catalog export --catalog exports catalogs added with catalog add, catalog info
    shows an empty publisher contact as "none published", and the release
    archives ship every guide.

The full list of changes is in
CHANGELOG.md.

Compatibility with 0.3.0

Read this before mixing 0.3.0 and 0.3.1 on one machine or between a
seller and a buyer.

  • Reference files: 0.3.0 reference files, commitments, and reports
    load unchanged in 0.3.1. Reference files written by 0.3.1 with
    --samples above 1 (the new default is 5) or calibrated with
    calibrate-reference are rejected by 0.3.0 binaries.
  • Claims with expires_at: claims without an expiry hash and verify
    exactly as in 0.3.0, in both directions. A claim with expires_at is
    rejected by 0.3.0 binaries (unknown field). There is no default
    expiry, so leave out --expires-in while your buyers still run 0.3.0.
  • Configs with follow lists: the new config keys [[follow_lists]]
    and follow_list in [[catalogs]] make 0.3.0 refuse the config.
    Remove them (or run catalog follow-list unfollow) before going back
    to 0.3.0. Desktop and the CLI share one config, so this applies to
    both.
  • Hand-written claims: no claim hash changes, and every claim
    Tensorseed has written is unaffected. A hand-written claim with an
    explicit null for an optional field, without
    catalog.publisher_key_id or license.id, or with unknown nested
    fields loaded in 0.3.0 and is rejected in 0.3.1.
  • verify-response --claim <URL> now fails with url_mismatch when
    the claim is not served at its own public_base_url (for example, a
    claim fetched through a local proxy). Pass the claim file instead.
  • --max-skew of 0 or above 604800 (7 days) is now refused. Use
    verify-response --at for stored responses.
  • Reports: ts.claim-check/1, ts.check-endpoint/1, and
    verify-response --json gain optional fields and the error labels
    claim_expired, claim_revoked, and url_mismatch; 0.3.0 reports
    still load.

Downloads

Tensorseed Desktop

File For
Tensorseed.Desktop_0.3.1_x64-setup.exe Windows x86_64, NSIS installer. Installs for you, no administrator rights.
Tensorseed.Desktop_0.3.1_x64_en-US.msi Windows x86_64, MSI. Installs for all users, needs administrator rights.
Tensorseed.Desktop_0.3.1_amd64.AppImage Linux x86_64, any distribution. Make it executable and run it.
Tensorseed.Desktop_0.3.1_amd64.deb Debian and Ubuntu, x86_64.
Tensorseed.Desktop-0.3.1-1.x86_64.rpm Fedora and RHEL, x86_64.
Tensorseed.Desktop_0.3.1_x86_64.flatpak Flatpak bundle, x86_64. Not on Flathub yet.

Install the Flatpak bundle with
flatpak install --user Tensorseed.Desktop_0.3.1_x86_64.flatpak. It uses
the GNOME 51 runtime, which Flatpak fetches from Flathub. It includes the
tensorseed CLI (flatpak run --command=tensorseed dev.tensorseed.desktop).

Command line

The CLI archives and packages as in 0.3.0: static Linux builds for x86_64
and ARM64, a glibc 2.17 Linux build, Windows x86_64 and ARM64, .deb and
.rpm packages for x86_64 and ARM64, and signed source archives
(tensorseed-0.3.1-src.tar.gz and .zip).

There is no macOS build.

Updates

  • AppImage, MSI, and NSIS from 0.3.0: Desktop offers 0.3.1 when it
    starts (unless you turned the check off in Settings) or when you choose
    Check for updates. It shows the version and these notes and
    installs only after you confirm, and only if the installer is signed
    with the Tensorseed Desktop update key (id B81047C9B22BB400) for
    0.3.1. You can also install 0.3.1 by hand from the files above.
  • .deb and .rpm: these never update themselves. Install the 0.3.1
    package with your package manager.
  • Flatpak: the bundle never updates itself. Install the 0.3.1 bundle
    the same way as 0.3.0.
  • CLI: download the new archive or package.

Windows SmartScreen

The Windows installers are not code signed yet, so Microsoft Defender
SmartScreen may show "Windows protected your PC" when you start one, and
your browser may call the file uncommon. Check the download against
SHA256SUMS first (see "Verify your download" below), then choose
More info and Run anyway.

Serving inference

Tensorseed does not bundle an inference engine. To serve a model, install
llama.cpp (llama-server, for GGUF models) or vLLM (vllm, for
safetensors checkpoints; Linux with a supported NVIDIA or AMD GPU, or WSL 2
on Windows) yourself, then use tensorseed serve-inference. Desktop does
not serve inference yet; use the CLI. See the
seller guide
and, for marketplaces, the
integration guide.

Verify your download

Every download above is listed in SHA256SUMS, which is signed with the
Tensorseed release key (minisign, key id F5D2D3C4C27D459D). Download the
file, SHA256SUMS, and SHA256SUMS.minisig, then:

minisign -Vm SHA256SUMS -P RWSdRX3CxNPS9Vl1JNjDnZ04OmM/3xrfp9vqG5FO8TNpK8Sbg6ENAI4P
sha256sum -c --ignore-missing SHA256SUMS

The trusted comment must read tensorseed 0.3.1 release SHA256SUMS. On
Windows, see
RELEASE-KEY.md.
The .sig files next to the Desktop installers are the update signatures
that Desktop checks; you do not need them to verify a download.

Tensorseed 0.3.0

Choose a tag to compare

@flyingmongoose flyingmongoose released this 14 Nov 21:29

Tensorseed 0.3.0 adds Tensorseed Desktop, an app for Linux and Windows
that browses the catalogs you follow, checks every entry signature and the
Chia anchor natively, pulls models and datasets with visible progress, and
seeds them back. It shares one config, one job database, one library, and
one seeder with the tensorseed CLI on the same machine, and it sends no
telemetry. The CLI is 0.3.0 too. Formats and commands may still change
before 1.0.

Highlights

  • Tensorseed Desktop. Catalogs and entry pages with signature and
    anchor status, license and commercial serving marks, and file lists.
    Pull with a progress bar, Pause, Resume, and Cancel. A Downloads view
    shared with the CLI (jobs started in either one can be resumed in the
    other), a Library with Verify, Open folder, and Delete, a Seeding view,
    a tray icon with live counts and Pause all and Resume all, settings for
    every [torrent] option, and Hugging Face sign in.
  • One seeder per machine. serve, tensorseed seed, and Desktop
    share a lock, so two seeders never fight over one machine. Desktop
    attaches to a running serve and can pause and resume its revisions.
  • Signed in app updates for the AppImage, MSI, and NSIS builds (see
    "Updates" below).
  • New default download folder. download_dir now defaults to a
    tensorseed folder in your Downloads folder (~/Downloads/tensorseed
    on Linux, %USERPROFILE%\Downloads\tensorseed on Windows) instead of
    the current directory. A download_dir set in your config is used as
    written, and copies pulled under the old default stay where they are.
  • Windows Firewall rule, opt in. Both Windows installers offer a
    checkbox, off by default, that adds an inbound rule for Tensorseed
    Desktop so peers can connect while you seed. Without it, Windows asks
    the first time Desktop seeds.

The full list of changes is in
CHANGELOG.md.

Downloads

Tensorseed Desktop

File For
Tensorseed.Desktop_0.3.0_x64-setup.exe Windows x86_64, NSIS installer. Installs for you, no administrator rights.
Tensorseed.Desktop_0.3.0_x64_en-US.msi Windows x86_64, MSI. Installs for all users, needs administrator rights.
Tensorseed.Desktop_0.3.0_amd64.AppImage Linux x86_64, any distribution. Make it executable and run it.
Tensorseed.Desktop_0.3.0_amd64.deb Debian and Ubuntu, x86_64.
Tensorseed.Desktop-0.3.0-1.x86_64.rpm Fedora and RHEL, x86_64.
Tensorseed.Desktop_0.3.0_x86_64.flatpak Flatpak bundle, x86_64. Not on Flathub yet.

Install the Flatpak bundle with
flatpak install --user Tensorseed.Desktop_0.3.0_x86_64.flatpak. It uses
the GNOME 51 runtime, which Flatpak fetches from Flathub. It includes the
tensorseed CLI (flatpak run --command=tensorseed dev.tensorseed.desktop).

Command line

The CLI archives and packages as in 0.2.0: static Linux builds for x86_64
and ARM64, a glibc 2.17 Linux build, Windows x86_64 and ARM64, .deb and
.rpm packages for x86_64 and ARM64, and signed source archives
(tensorseed-0.3.0-src.tar.gz and .zip).

There is no macOS build.

Windows SmartScreen

The Windows installers are not code signed yet, so Microsoft Defender
SmartScreen may show "Windows protected your PC" when you start one, and
your browser may call the file uncommon. Check the download against
SHA256SUMS first (see "Verify your download" below), then choose
More info and Run anyway.

Updates

0.3.0 is the first release with Desktop, so install it by hand from the
files above. In app updates start with 0.3.1.

  • AppImage, MSI, and NSIS: Desktop checks for a new release when it
    starts (you can turn this off in Settings) and when you choose
    Check for updates. It shows the new version and its notes and
    installs only after you confirm, and only if the installer is signed
    with the Tensorseed Desktop update key (id B81047C9B22BB400) for that
    exact version. The check downloads one small file from the GitHub
    release and sends nothing about your machine.
  • .deb and .rpm: these never update themselves. Install the new
    package from the next release with your package manager.
  • Flatpak: the bundle never updates itself. Install the next
    release's bundle the same way; once Tensorseed Desktop is on Flathub,
    flatpak update will handle it.
  • CLI: download the new archive or package from the next release.

Serving inference

Tensorseed does not bundle an inference engine. To serve a model, install
llama.cpp (llama-server, for GGUF models) or vLLM (vllm, for
safetensors checkpoints; Linux with a supported NVIDIA or AMD GPU, or WSL 2
on Windows) yourself, then use tensorseed serve-inference.
Tensorseed finds the engine on PATH or through inference.llama_server
and inference.vllm in your config. Desktop does not serve inference yet;
use the CLI. See the
install guide
for each OS.

Verify your download

Every download above is listed in SHA256SUMS, which is signed with the
Tensorseed release key (minisign, key id F5D2D3C4C27D459D). Download the
file, SHA256SUMS, and SHA256SUMS.minisig, then:

minisign -Vm SHA256SUMS -P RWSdRX3CxNPS9Vl1JNjDnZ04OmM/3xrfp9vqG5FO8TNpK8Sbg6ENAI4P
sha256sum -c --ignore-missing SHA256SUMS

The trusted comment must read tensorseed 0.3.0 release SHA256SUMS. On
Windows, see
RELEASE-KEY.md.
The .sig files next to the Desktop installers are the update signatures
that Desktop checks; you do not need them to verify a download.

Tensorseed 0.2.0

Choose a tag to compare

@flyingmongoose flyingmongoose released this 13 Nov 17:55

Tensorseed 0.2.0

Tensorseed mirrors open Hugging Face models and datasets for preservation.
0.2.0 lets one client follow several signed catalogs at once and adds a
browser explorer at explore.tensorseed.dev
that verifies the catalog signature and its Chia anchor without installing
anything. Formats and commands may still change before 1.0.

Highlights

  • Multi-catalog: follow several signed, chain-anchored catalogs with
    [[catalogs]] or tensorseed catalog add. Revisions that agree across
    catalogs share one swarm, and list shows how many catalogs agree.
  • Catalog explorer at explore.tensorseed.dev: browse and search the
    official catalog, open an entry, and check the minisign signature and
    the Chia DataLayer anchor in the browser (WebAssembly built from the
    same Rust verification code as the CLI). Add another catalog by URL,
    key, and store id to compare.
  • Default download order is the BitTorrent swarm, then Hugging Face, then
    HTTP mirrors, configurable with [fetch] source_order.
  • tensorseed hf login, logout, and status, and automatic reuse of an
    existing Hugging Face CLI token.
  • pull --include and --exclude globs for partial pulls.
  • .deb and .rpm packages, shell completions, and man pages.

Download targets

The same five archives as 0.1.0 (Linux x86_64 and ARM64 static musl,
Linux x86_64 glibc 2.17, Windows x86_64, Windows ARM64), signed source
archives (tensorseed-0.2.0-src.tar.gz and .zip), and new .deb and
.rpm packages for x86_64 and aarch64. All are listed in SHA256SUMS,
signed with the release key (minisign, key id F5D2D3C4C27D459D).

Added

  • Read-only catalog explorer (explorer/, Svelte 5, TailwindCSS,
    DaisyUI, built with pnpm, no third party scripts) deployed at
    explore.tensorseed.dev. It verifies the catalog with a WebAssembly
    build of tensorseed-verify (tensorseed-wasm), shows chain status,
    per-entry files, hashes, licenses, and torrent details, and has a
    "use another catalog" form with agreement across catalogs.
    Catalog signature and chain checks moved into the tensorseed-verify
    crate so the CLI and the explorer share one implementation.
  • tensorseed pull --include GLOB and --exclude GLOB (repeatable) select
    files by path. Globs use the same rules as serve (* matches any
    run of characters, including /). --exclude wins when both match.
    The swarm uses librqbit only_files; HTTP mirrors and the Hub use the
    same filter. pull.json records the full file list plus skipped and
    unselected. verify on a partial pull passes when selected files
    match; missing unselected files fail only with --strict. serve and
    --seed-after seed the selected subset with only_files, and do not
    announce missing files as a complete torrent.
  • Default download order is the BitTorrent swarm, then Hugging Face,
    then HTTP mirrors. [fetch] source_order is a list of "swarm",
    "hub", and "mirrors" (validated, no duplicates) and applies to
    pull and serve auto-fetch. --no-swarm, --swarm-only, and
    dead-mirror memory are unchanged. A 401, 403, or 404 from the Hub
    falls through to the mirrors without marking the host dead.
  • Hugging Face tokens are also loaded from tensorseed hf login and from
    the Hugging Face CLI (HF_TOKEN_PATH, $HF_HOME/token, or
    ~/.cache/huggingface/token). Resolution order: token_env, then
    token_file, then the tensorseed login file, then the CLI token.
    tensorseed hf login reads a token from a hidden prompt or stdin and
    writes ~/.config/tensorseed/hf_token (0600 on Unix, a user-only ACL
    on Windows). hf logout removes that file. hf status calls
    whoami-v2 and prints the username and token source, never the token.
  • just dist-packages builds .deb and .rpm packages (x86_64 and
    aarch64) from the static musl binaries with nfpm. Each package
    installs /usr/bin/tensorseed, a disabled tensorseed-serve systemd
    unit running as the tensorseed system user, and
    /etc/tensorseed/tensorseed.toml.example (copied to tensorseed.toml
    on first install). just dist includes the packages in SHA256SUMS.
  • tensorseed completions <shell> prints a bash, zsh, fish, or PowerShell
    completion script. just generate-docs also writes man pages
    (tensorseed.1 and one page per subcommand). Linux packages install
    bash, zsh, and fish completions (fish to
    /usr/share/fish/vendor_completions.d) and the man pages to
    /usr/share/man/man1. Release tarballs and zip archives include
    completions/ and man/.
  • Follow several catalogs at once with [[catalogs]] (name, label,
    publisher key, store id, base URL, chain check, optional priority).
    A config with no [[catalogs]] still uses the top-level [catalog],
    [file], and [chia] block as the sole subscription. Revisions are
    keyed by Hugging Face repo and commit; agreeing hashes share one swarm
    and can use every agreeing catalog's HTTP sources. list prints
    "N of M catalogs agree". The same repo and commit with different
    hashes is a hard error unless --catalog is given. Tombstones are per
    catalog: a revision remains pullable and seedable if another
    subscription still lists it. serve seeds a revision that is on disk
    and still valid in at least one non-tombstoned subscription.
    tensorseed catalog list, catalog add, and catalog remove manage
    subscriptions; catalog info and catalog chain-status accept
    --catalog.

Fixed

  • The BitTorrent DHT binds IPv4-only when the configured listen address is
    IPv4, and stores its routing table under Tensorseed's data directory
    (%LOCALAPPDATA%\tensorseed\dht.json on Windows). A previous default
    used librqbit's dual-stack [::] bind and its own cache path, which
    failed to start the engine on Windows and aborted pull before HTTP
    sources could run. If the engine still fails to start, pull now
    falls back to HTTP unless --swarm-only is set.

Verify a release

Release archives ship with SHA256SUMS and SHA256SUMS.minisig, signed
by the release key (minisign, key id F5D2D3C4C27D459D). This is not the
catalog publisher key. Do not verify binaries with the catalog key.

RWSdRX3CxNPS9Vl1JNjDnZ04OmM/3xrfp9vqG5FO8TNpK8Sbg6ENAI4P
minisign -Vm SHA256SUMS -P RWSdRX3CxNPS9Vl1JNjDnZ04OmM/3xrfp9vqG5FO8TNpK8Sbg6ENAI4P
sha256sum -c --ignore-missing SHA256SUMS

The public key file is also published as tensorseed-release.pub (and in
keys/tensorseed-release.pub in the repository).

Known limitations

  • The Windows ARM64 build has not been run on ARM Windows hardware.
  • There is no macOS build; build from source with Rust 1.87 or newer.
  • Catalog and on-disk formats may change before 1.0.
  • takedown and doctor still print "not implemented yet".

Tensorseed 0.1.0

Choose a tag to compare

@flyingmongoose flyingmongoose released this 13 Nov 13:05

Changelog

0.1.0

Tensorseed mirrors open Hugging Face models and datasets for preservation.
A signed catalog lists each mirrored revision (files, sizes, SHA-256
hashes, license, and the Hugging Face commit). The catalog is anchored on
Chia DataLayer, and clients check the signed file catalog against the
chain. Downloads use the BitTorrent swarm first; HTTP mirrors and Hugging
Face are failover.

This is an early release. The software works today for pulling, verifying,
and seeding, and we are looking for volunteer seeders. Formats and
commands may still change before 1.0.

Highlights

  • Strict chain checking is the default once a store id is set:
    tensorseed catalog chain-status reports whether the file catalog
    matches the on-chain head.
  • tensorseed pull uses the BitTorrent swarm first. HTTP mirrors and
    Hugging Face fill only what the swarm cannot deliver. If no peer
    connects within about 30 seconds, or the download stalls for a minute,
    the client switches to HTTP. Pulls are outbound only and do not keep
    seeding unless you pass --seed-after.
  • Dead HTTP mirrors are remembered. A connect failure or repeated
    transient failures marks an origin dead for later files and later runs,
    with a one hour cooldown by default ([fetch] dead_mirror_cooldown_secs).
    After the cooldown it is retried automatically. --retry-dead-mirrors
    retries immediately. A 404 for one file does not mark the host dead.
    If every HTTP source for a file is marked dead, the client tries them
    anyway.
  • Seeding is opt in per item. Pull the repos you care about, then run
    tensorseed serve to share them. The daemon does not download the rest
    of the catalog unless you turn on auto-follow.
  • Windows uses standard directories (%APPDATA%\tensorseed for config,
    %LOCALAPPDATA%\tensorseed for data).
  • Anyone can run an independent catalog (own publisher key, own Chia
    store, own host). See docs/PUBLISH-A-CATALOG.md.
    Nobody but the official publisher can write to the official
    tensorseed.dev catalog. Following several catalogs in one process is
    planned for 0.2.0. To request a repo in the official catalog, open a
    "Request a catalog mirror" issue (gated and restricted-license repos
    are refused).

Download targets

Prebuilt archives:

  • Linux x86_64, static (x86_64-unknown-linux-musl)
  • Linux ARM64, static (aarch64-unknown-linux-musl)
  • Linux x86_64, glibc 2.17 (x86_64-unknown-linux-gnu)
  • Windows x86_64 (x86_64-pc-windows-gnu)
  • Windows ARM64 (aarch64-pc-windows-gnullvm); the file format is
    confirmed, but this build has not been run on ARM Windows hardware

There is no macOS build. macOS users can build from source with
cargo install or cargo build --release (Rust 1.87 or newer).

Verify a release

Release archives ship with SHA256SUMS and SHA256SUMS.minisig, signed
by the release key (minisign, key id F5D2D3C4C27D459D). This is not the
catalog publisher key. Do not verify binaries with the catalog key.

RWSdRX3CxNPS9Vl1JNjDnZ04OmM/3xrfp9vqG5FO8TNpK8Sbg6ENAI4P
minisign -Vm SHA256SUMS -P RWSdRX3CxNPS9Vl1JNjDnZ04OmM/3xrfp9vqG5FO8TNpK8Sbg6ENAI4P
sha256sum -c --ignore-missing SHA256SUMS

The public key file is also published as tensorseed-release.pub (and in
keys/tensorseed-release.pub in the repository).

Known limitations

  • pull downloads whole revisions. A file the swarm left partial is
    currently fetched again in full over HTTP.
  • A client follows one catalog per config file. To follow a third party
    catalog, use a separate config and --config.
  • Catalog and on-disk formats may change before 1.0.
  • takedown and doctor still print "not implemented yet".