Skip to content

The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.

3.1 KB
Raw Blame

description: >- How OpenHuman updates itself: one signed manifest, a bounded download retry, and a single update that covers the app and its core. icon: arrows-rotate

Auto-update

OpenHuman updates the whole app in one step. The core is linked into the desktop shell, so there is no separate core update. Updating the shell upgrades both.

What you see

When a newer build is available, the app downloads it in the background. A prompt appears once the download is ready and moves through ready to install, installing and restarting. If you are current you see up to date. A failure says so instead of retrying forever.

Before installing, the app shuts down the in-process core under the same restart lock it uses elsewhere. That way replacing the bundle never races a core that still holds file handles, which keeps a .app replacement on macOS clean.

Where updates come from

There is one endpoint, a manifest published with every release:

https://github.com/tinyhumansai/openhuman/releases/latest/download/latest.json

Every artifact is signed, and the public key is compiled into the app. A build whose signature does not verify is refused. There is no update server to trust and no second channel.

Retries

A download gets three attempts: one try plus two retries, with a delay of two seconds times the attempt number.

Only transient network failures are retried. Signature failures, filesystem errors and "no artifact for this target" are not. Downloading again cannot fix a bad signature, and looping on a failed verification would waste bandwidth and could hide tampering. The policy lives in crates/openhuman-app/src/app_update.rs, where classify and is_transient_updater_err decide.

Periodic checks

The core runs the schedule, set under [update] in config.toml:

Key Default What it does
enabled true Turns periodic checks off when set to false.
interval_minutes 60 How often to check. The minimum is 10 minutes.
restart_strategy self_replace self_replace restarts in place after staging. supervisor stages the update and leaves the restart to whatever manages the process.
rpc_mutations_enabled true Whether a bearer-authenticated RPC client may call the mutating update methods.

Use supervisor for a headless or containerized core, where something else owns the process lifecycle and the update should not restart it.

Headless and self-hosted cores

A core you run yourself (openhuman-core serve) checks and stages updates through the update namespace instead of the desktop prompt. That is why rpc_mutations_enabled is a separate switch. The old desktop commands check_core_update and apply_core_update still exist for frontend compatibility but do nothing.

See also