Skip to content

v1.3.2

Choose a tag to compare

View all tags
@github-actions github-actions released this 10 Oct 16:14
· 1081 commits to main since this release
Immutable release. Only release title and notes can be modified.
v1.3.2
e7d27a4

OMK v1.3.2

Install in one line, check your machine, and keep every notification

OMK 1.3.2 adds a checksum-verified installer for the standalone binary, an
omk doctor that tells you what blocks a first session, and a first run that
opens sign-in instead of failing your first prompt. It also fixes messages that
were lost while a session was compacting. All seven public workspace packages
move together to 1.3.2.

This is a patch release under OMK's version policy: fixes and additions, with no
breaking change to the SDK or CLI. A few defaults behave differently; read the
compatibility notes below.

What's new

  • A verified one-line install. Once this release's assets are published:

    curl -fsSL https://github.com/dmae97/omk/releases/latest/download/install.sh | sh

    The installer checks the archive against the release's SHA256SUMS, which it
    always takes from GitHub over HTTPS, even when --base-url names a download
    mirror (--trust-mirror opts into the mirror's file, with a warning). It
    installs into ~/.omk/versions/<version>, switches ~/.omk/bin/omk with an
    atomic rename, and never replaces an installed version directory. It needs
    curl. Release assets now include SHA256SUMS and a signed build-provenance
    attestation for the binaries and the installer:
    gh attestation verify omk-linux-x64.tar.gz -R dmae97/omk. See the
    quickstart.

  • omk doctor. It checks the runtime, the agent directory, credentials,
    the model a first session would pick, the bash sandbox, fd and rg, and the
    network, and prints the fix for your machine. It is read-only, connects to
    nothing without --online, and exits 0, 1 (a check failed) or 2 (usage).
    Use --json for scripts.

  • A first run that starts with sign-in. With no model configured, OMK opens
    the /login selector before your first prompt. If a Claude Code or Codex CLI
    login already exists, it names the omk provider adopt command that reuses it.

  • No more lost notifications during compaction. A background-task
    notification or other extension message that arrived while the session was
    compacting, summarizing a branch, switching models for a retry or preparing a
    prompt failed with PromptExecutionBusyError, so the agent never woke up for
    it. After a manual /compact it could also be answered from the
    pre-compaction context. Such messages now reach the agent once, after that
    work, on the current context. Text you type in the TUI during compaction is
    sent when compaction ends instead of returning to the editor queue.

  • Faster start. omk --version no longer loads the runtime (635 ms to
    62 ms for the Node CLI on the measurement VM). fd and rg download after
    the first frame, so a stalled network no longer delays the first screen
    (about 11 s to 1.2 s).

  • A clearer sandbox check. The bash sandbox counts as available only when
    bwrap actually runs a command. Hosts that block unprivileged user namespaces
    get setup steps for that host instead of a raw bwrap error. See
    Bash sandbox setup.

  • omk run explain. Explains a verified run from one ledger snapshot: DAG
    dependencies, unfinished ancestors, candidate and receipt bindings, and claim
    closure. It changes nothing. The SDK equivalent is RunCoordinator.explain.

  • Opt-in RTK output filter. OMK_RTK_OUTPUT=1 summarizes long successful
    vitest run and tsc --noEmit output from the bash tool through RTK, which
    you install separately; the full output stays in a file. Off by default. See
    RTK output filter.

  • Model catalog. Mistral Large 4 on Mistral and OpenCode Zen, a catalog
    refresh to 1,932 coding routes across 40 providers, and 61 image models. See
    the ai changelog.

Compatibility

  • With no default model configured, credentials are now ranked stored login,
    then a provider API key, then ambient cloud credentials. With AWS credentials
    in the environment and ANTHROPIC_API_KEY set, a first session uses Anthropic
    instead of Amazon Bedrock. A configured default model is unchanged.
  • On hosts where bwrap cannot create user namespaces, omk doctor and the
    bash tool report the sandbox as unavailable with setup steps, instead of
    failing at the first command with a raw bwrap error. Bash still refuses to
    run without a working sandbox.
  • Follow-up input that arrives while another prompt is still being prepared
    waits for that prompt's run instead of failing. Registered extension commands
    run at once in that window; other slash text is still refused while the
    session is busy.
  • The installer is Linux and macOS only; on Windows use WSL2. It does not verify
    the provenance attestation itself; run gh attestation verify for that.
  • The 1.3.1 changelog entries now record what the 1.3.1 packages contain:
    changes that were still listed under Unreleased when v1.3.1 was tagged are
    listed under 1.3.1.

Upgrade

curl -fsSL https://github.com/dmae97/omk/releases/latest/download/install.sh | sh
# or, with Node.js 22.19 or newer
npm install -g open-multi-agent-kit@1.3.2 --ignore-scripts
omk --version
omk doctor

Restart any running OMK process after upgrading. A release is complete only when
the v1.3.2 tag is reachable from main, its GitHub Release exists, and npm
latest points to 1.3.2 for all seven public packages.

How this release was checked

These are scoped checks, not a certification:

  • The local AdaptOrch command verifier ran 18 gates on the release candidate:
    build, the whole-repository npm run check, the onboarding, session-routing,
    compaction, model and verified-run test files, the installer end-to-end
    tests, built-CLI and terminal smoke tests, and the private-home boundary, with
    a negative control that must fail.
  • A real claude-haiku-4-5 session in the interactive TUI received a
    background-task notification that landed during automatic and during manual
    compaction. The build before these fixes, with 1.3.1's session code, lost the
    first and answered the second from the pre-compaction context.
  • Mutation tests: each of 26 single-line reversions of the session-routing and
    TUI changes made at least one test fail.
  • Installer: 11 kit checks and 12 hardening tests pass; the earlier installer
    draft fails the hardening tests.
  • A reviewer from a different model family approved the installer, the session
    input changes and the onboarding follow-ups, after up to three rounds.

Not tested: macOS, native Windows, NixOS and setuid-bwrap hosts, and the
release workflow itself, which runs only on a tag.

Assets 11
Loading