The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.
Releasing the CLI
Running daemons self-update from a public GCS bucket (src/lib/selfUpdate.ts). Releasing is pushing a
tag: .github/workflows/release.yml bundles the CLI, publishes it to that bucket, and cuts the GitHub
Release. The tag IS the version — CI never bumps.
Tags are vX.Y.Z_cli, not plain vX.Y.Z: the backend now lives in this same repo under backend/
and releases as vX.Y.Z_backend (see ../backend/README.md), so the suffix is what tells each tag-push
trigger which workflow to run. It's stripped before anything treats it as a version — the published
version everywhere below (manifest, release title, ADAPTER_VERSION) is always a bare X.Y.Z.
make release-cli # bump the patch, tag, push — CI does the rest
make release-cli ARGS="--dry-run" # print the version it would cut and the notes, do nothing
make release-cli ARGS="0.2.0" # release an explicit version
make release-cli ARGS="--notes-file notes.md" # hand-written release notes
Nothing is built locally and no GCS credentials are needed: the only things cli/scripts/release-cli.sh
touches are git and a public HTTPS read of the manifest. See ../docs/cicd.md for
what the release workflow itself does, step by step.
Why the version comes from two places. cli/scripts/release-cli.sh takes the highest of the last
git tag and the version published under the cli key of harness/cli/metadata.json. The two can
drift — a scratch publish that slipped, or a manifest edited by hand, leaves the manifest ahead of the
last tag. Bumping from tags alone there would produce a version LOWER than what's already served —
every daemon refuses it (shouldUpdate / semverGt in src/lib/selfUpdate.ts) while the release
still reports success.
A release contains the one before it. The script refuses a checkout that lacks the last release's
commit (the highest v*_cli tag, and the tag of the version the manifest serves): it is still on
origin/main, so nothing else stopped it, and it would publish older code under a newer version to
every daemon. v0.3.82_cli was tagged that way on October 12, 2026, from a checkout three commits
behind v0.3.81_cli; its run was cancelled before it published. Update the checkout
(git fetch origin && git checkout --detach origin/main). A rollback, which means to do exactly this,
is make release-cli ARGS="--allow-older". The desktop, backend and web release scripts check the same.
Unlike autonomous-harness-desktop's make release, there is no --minor flag: shouldUpdate is a
plain "strictly newer" check with no major/minor-triggers-a-forced-update concept, so every release
just needs to outrank what's live.
package.json's version field is never touched by this script; the published version is baked into
the bundle via ADAPTER_VERSION at build time (see "The publishing step" below).
Core app DSHs ship inside that same release: Harness Monitor, Devices and Model Manager are
materialized at daemon startup before viewer restoration. Official older Store installations
migrate automatically, retaining their prior files, workspaces and conversations. Companions
refresh their generated package when the existing collection identity returns. No separate
harness dsh update is required for these core tools. Other Store apps retain user-driven updates.
Before uploading, the release script runs scripts/test-core-harness-upgrade.mjs against the
compiled bundle to verify existing-install migration, package assets and the Monitor viewer.
The publishing step
cli/scripts/upload-cli.sh is the publishing step, and CI is its only caller: release.yml runs it
with the version taken from the tag. There is deliberately no make target for it — publishing from a
laptop created no git tag and bumped from the remote manifest, so the repo stopped reflecting what was
served, and a stale local tag then broke the next make release-cli at git fetch --tags. If CI is
down, fix CI; a version with no tag is worse than a late one.
What the script does, for reading the workflow:
- Reads the current version from the live manifest (falling back to
package.jsonif the manifest is unreadable), then bumps or takes the version it was given. - Bundles with
ADAPTER_VERSION="$VERSION" npm run bundleand assertsnode dist/cli.js versionreads back that exact string before uploading anything. - Uploads
cli.jsandnotify.mjstoharness/cli/<version>/, then download-merge-reuploadsmetadata.jsonin a single write, touching only theclikey.
A failed build stops the release; nothing is uploaded and no version is consumed.
The curl | bash installer
cli/scripts/install.sh is what curl -fsSL https://cdn.autonomous.ai/harness/cli/install.sh | bash
runs: it installs the managed Node runtime, CLI bundle and hn from their published manifests.
After verifying a production CLI release, release.yml publishes this installer from the same
commit and checks that the CDN serves the exact source bytes. Releases to a scratch manifest leave
the production installer alone. The installer is a static, unversioned file; a maintainer can also
publish an installer-only correction:
make upload-cli-install-sh # cli/scripts/install.sh -> harness/cli/install.sh in the bucket
curl -fsSL https://cdn.autonomous.ai/harness/cli/install.sh | head -5 # confirm the CDN edge moved
Its command contract (flags, exit codes, what it refuses) is pinned by cli/src/scripts/install.spec.ts,
which runs with the rest of make cli-test. Not to be confused with cli/scripts/install-cli.sh, the
local dev loop below.
The npm wrapper of harness mcp
cli/packages/harness-mcp is @autonomous-ai/harness-mcp, for MCP clients and directories that install servers
with npx. It is not the CLI: it runs harness mcp from the installed CLI, and where there is none it offers one
tool that says how to install it. Its mcpName ties it to the Official MCP Registry entry. It changes rarely, and
apart from the CLI's releases:
npm login # an account in the npm organization autonomous-ai
cd cli/packages/harness-mcp && npm publish --access public
Its tests are cli/src/agentClient/mcpPackage.spec.ts, in make cli-test.
harness-mcp.mcpb, the same server as an MCP Bundle (Claude Desktop, the Official MCP Registry, Smithery), goes on
a CLI release as an asset, by hand, after the release workflow has cut it:
make mcpb ARGS="X.Y.Z" # writes cli/packages/harness-mcp/mcpb/dist/ and prints the sha256
gh release upload vX.Y.Z_cli cli/packages/harness-mcp/mcpb/dist/harness-mcp.mcpb --repo autonomous-ai/openharness
Publishing it, and every other listing, is in docs/research/2026-10-11-listing-submissions.md.
Local install (no upload)
bash cli/scripts/install-cli.sh bundles this working tree into ~/.harness/cli — the same
layout the public installer produces — and restarts the daemon on it, so harness on this machine
means your code without publishing anything:
make install-cli # bundle + install + restart the daemon
make install-cli ARGS="--no-restart" # skip restarting the daemon
make install-cli ARGS="--no-updates" # pin self-update off (see below)
Self-update stays on by default: the local build is labelled with the currently published version,
which the release you're level with cannot outrank, so your bytes survive until a genuinely newer
version ships — then the daemon quietly swaps back to it. --no-updates pins your build instead.
GCS layout and manifest shape
gs://s3-autonomous-upgrade-3/harness/cli/metadata.json
gs://s3-autonomous-upgrade-3/harness/cli/<version>/cli.js
gs://s3-autonomous-upgrade-3/harness/cli/<version>/notify.mjs
The manifest is always read straight off the GCS origin (CDN_ASSET_BASE_URL in upload-cli.sh does
not apply to it) — it's polled every ~60s by every running daemon, and this product's CDN caps any
cacheable response at ~31 days regardless of origin headers, so it must never be CDN-fronted. The bundle
files it points at are the opposite: immutable once published, so their url fields point at
cdn.autonomous.ai instead, and are uploaded with a long Cache-Control on purpose.
{
"cli": {
"version": "0.1.72",
"url": "https://cdn.autonomous.ai/harness/cli/0.1.72/cli.js",
"sha256": "<64 hex>",
"size": 2103552
}
}
The bucket must be public-read — that is bucket policy, not something the scripts set.
How a running daemon self-updates
- Polls
ADAPTER_UPDATE_URLeveryADAPTER_UPDATE_CHECK_MS(default 60 s) — seesrc/config/env.ts. - Compares the manifest's
cli.versionagainst its own — strictly newer only (semverGt), so republishing an old build can never downgrade a fleet. - Downloads
cli.js+notify.mjsand verifies each sha256/size against the manifest before anything is swapped in. A mismatch is discarded. - Swaps the bundle in when the daemon is idle and restarts, with rollback on a bad build.
There is no undo. A bad release is rolled back by publishing a higher version containing the
older code — never by deleting the new one, which just leaves every daemon pointed at a 404. Disable
self-update on one machine with ADAPTER_UPDATE_DISABLE=true.
hn updates
The installed daemon checks the separate harness/tui/metadata.json manifest on startup and
on the same update schedule. An installed hn moves only to a newer published version, after
checksum, size and executable-version checks. Replacing the binary is atomic; open clients
keep running and the next launch uses the update. Failures retry without blocking CLI updates.
harness update also checks hn even if the CLI is already current.
An hn release still uses release-tui.yml; publishing a CLI does not rebuild or republish hn.
The CLI updater discovers whichever hn release is latest, so an hn-only release reaches
existing installations too. ADAPTER_UPDATE_DISABLE=true covers both automatic updaters.
Checkout/local CLI builds, HARNESS_TUI_BIN overrides, symlinked managed binaries and hn development
version labels are left alone. A missing managed binary is repaired on the next update check when
the managed hn launcher is present; otherwise first use or the installer downloads it.
Older manual/development hn commands can bypass this managed binary. harness update diagnoses
these installations, including a different hn earlier on PATH. harness tui --install or
harness update --force verifies the latest download, backs up a recognized old Harness launcher
under ~/.local/bin/.hn-backup-*/hn, and replaces it with a launcher that follows harness tui.
An unrelated command is preserved; use harness tui or correct PATH in that case. Explicit
HARNESS_TUI_BIN overrides remain in effect until unset. The installer never writes through
an existing hn symlink.
Rolling out safely
Publish to a scratch manifest before touching the real one, then point a local build at it:
gh workflow run release.yml -f version=9.9.9 -f metadata_path=harness/cli/metadata-test.json
ADAPTER_UPDATE_URL=https://storage.googleapis.com/s3-autonomous-upgrade-3/harness/cli/metadata-test.json harness start
No daemon polls metadata-test.json, so nothing self-updates from it. The artifacts still go to
their real home (harness/cli/<version>/ is hardcoded in upload-cli.sh), so use an unmistakable
version and delete gs://s3-autonomous-upgrade-3/harness/cli/9.9.9/ afterwards.