Runmoor macOS Host Runners
Availability: host execution is unreleased. Check your installed version and
release notes before using runmoor init --backend host.
The opt-in host backend runs separate, single-job GitHub Actions runner processes
on one macOS 14+ Apple Silicon computer. It requires no Docker, Tart or Guest
Agent. Runmoor's normal default on macOS remains Tart; select host explicitly.
Runmoor never falls back to host execution when another backend fails.
Verification limits: automated contract tests, mocked execution and builds cover this backend. Actual Mac host execution, unsigned Xcode builds and live GitHub jobs have not been validated. Existing Docker, Tart and package checks do not prove host execution compatibility.
Set up and route jobs
Use a dedicated macOS CI account. Install Xcode, required SDKs and language tools under that account, and complete their normal license/setup steps yourself. Runmoor does not create accounts, install tools or require a privileged helper. Ordinary build/test workflows and unsigned Xcode builds are intended workloads.
Create a new configuration with an existing GitHub credential reference:
Supply the credential separately as described in configuration. Repository and organization targets support the existing PAT and GitHub App options. GitHub repository access, runner groups and fork-execution policies remain authoritative. Never grant untrusted workflows access to this account.
A newly generated pool is named macos-host. Its scale set is
runmoor-macos-host, with labels runmoor-macos-host, macOS and ARM64:
Existing and manually authored labels are preserved. A minimal host pool can use:
The connection must be defined separately. Host pools reject image,
image_source, runner_path, DinD settings and image-only initialization.
Runmoor owns execution paths. image commands remain Tart operations.
Admission and runner updates
Omitted host allocations use 2 CPUs and 4096 MiB per runner. On smaller budgets they shrink to available capacity, with minimums of 1 CPU and 1024 MiB. Explicit values are never silently reduced. Omitted concurrency follows both CPU and memory budgets. Host execution does not inherit Tart's two-VM ceiling; more than two runners can be admitted when shared capacity permits.
CPU and memory reservations control admission; they do not enforce limits on host process usage. Host jobs share the global budget with Docker jobs, Tart VMs and runner preparation. Minimum idle defaults to zero. Free-disk reserve defaults to 10240 MiB, or 20480 MiB when Tart is also configured. Low disk blocks new work without evicting active jobs.
The host preparation timeout defaults to five minutes:
Idle runners do not consume their job timeout. The job timeout starts with assignment; if removal finds a busy job whose start is unknown, Runmoor uses the runner creation time as a conservative bound.
Runmoor selects official macOS arm64 runner releases, verifies their SHA-256 and
archive contents, and prepares an immutable distribution. Omitted or latest
runner versions follow managed updates; exact versions remain pins. Existing
jobs retain their original verified distribution while a replacement prepares.
Use runmoor runner update --pool macos-host for an immediate check and
runmoor config show --resolved for the committed version. Known support expiry
blocks new acquisition; it does not rerun failed jobs.
Disposable directories and trust
Every execution receives a fresh registration, runner installation, HOME, temporary directory and work directory. It handles at most one job. PATH and DEVELOPER_DIR are inherited when set. HOME, TMPDIR, TMP and TEMP point to the execution's disposable directories. Other manager environment variables and management credentials are excluded from job launch configuration.
Separate directories are not a security boundary. The manager and jobs run under the same macOS account. A job can access account files, personal data and Keychains available to that account, including another job's resources. Support is limited to trusted developers and small-team workflows.
Shared caches, signing Keychains, Simulator lifecycle, GUI automation and global tool settings are workflow/operator responsibilities. Runmoor does not isolate or manage those resources. Jobs that daemonize or detach outside the managed process group are unsupported; their lifecycle remains the operator's responsibility.
Cleanup and recovery
Use the normal control commands: pause, resume, reload, drain, stop and pool-scoped force-stop. The launchd user service uses the same lifecycle. A manager-only restart reconciles verified surviving work. Normal completion, cancellation and timeout terminate only the verified execution group and its children before removing execution directories.
Missing supervisors, changed process identity, replaced directories and uncertain
termination keep reservations and recovery diagnostics. stop --force does not
authorize deleting foreign resources or signalling unrelated same-account
processes. Preserve the installation's state/data and uncertain resources; inspect
the reported safe error code and restore a matching paired backup when available.
Do not repair ownership markers or delete a directory based only on its name.
Drain and stop before a binary upgrade or rollback. Keep paired state/data backups and wait for cleanup to finish before copying them. Restoring both together with a compatible binary preserves verified host distributions. Do not edit ownership markers or restore an older backup while later resources remain active. An older binary cannot safely open newer state. Runmoor preserves unresolved records until recovery and completed history for seven days. Diagnostics retain safe lifecycle information within seven days and 256 MiB; raw workflow output, runner logs and credentials are not copied into them.
Report reproducible problems through GitHub Issues with version, platform, stable error code and sanitized status. Community support has no response SLA, deadline, throughput guarantee or performance SLA.