Release

Two human steps. Everything else is automation.

No next branch. Features and fixes both land on main. Beta is a tag.

Cut a release

1. From core/, set [workspace.package] version in Cargo.toml to the version you will tag (0.2.0 or 0.3.0-beta.1). check-release-tag.sh rejects a tag if the Cargo version is missing or differs.
2. Merge that commit to main.
3. Tag and push:

Beta: v0.3.0-beta.1 with Cargo version 0.3.0-beta.1. -rc and every other suffix are rejected.

4. Wait for the Release workflow. The tag run validates the tag and commit, builds the six CLI and daemon archives using the shared kache/R2 cache, builds and pushes ghcr.io/getployz/ployz-cloud:<tag>, and opens a draft GitHub release (--prerelease on beta tags) with ployz-cloud-compose.yml (pinned to the tag) and ployz-cloud.env.example attached for self-hosting.
5. Fill ## Notes. Click Publish. That click is the review gate. Drafts are not public downloads.

The ghcr.io/getployz/ployz-cloud:<tag> image is pushed during the draft run and is pullable by exact tag before Publish; only Publish announces it and moves the channel pointers.

Automatic releases run the workflow version stored in the tagged commit. For recovery using the current workflow, dispatch release.yml from main with tag and its expected commit sha.

main pushes populate the shared R2 compiler cache through the checks selected for that change. Release archive checks run when packaging inputs, Cargo manifests, Cargo.lock, or core/.cargo change. Cloud reuses a native SDK and config WASM artifact only when their source-input hash matches exactly, and builds from source on a miss. These checks do not publish releases. Only main pushes receive R2 write credentials, and a main push without them fails rather than silently skipping cache writes. PR, tag, release, scheduled, and manual runs use separate R2 read-only credentials. GitHub Actions variables KACHE_S3_BUCKET, KACHE_S3_ENDPOINT, and KACHE_S3_REGION select the bucket; secrets KACHE_S3_ACCESS_KEY_ID and KACHE_S3_SECRET_ACCESS_KEY provide write access. Create a separate R2 token with Object Read only permission scoped to the cache bucket, and store its credentials as KACHE_S3_READ_ACCESS_KEY_ID and KACHE_S3_READ_SECRET_ACCESS_KEY. Same-repository PRs reuse the shared cache with these read-only keys. Both credential pairs must be configured. GitHub does not expose these secrets to fork PRs.

Confidence before Publish

ci.yml selects checks for PR and main changes and reports one CI ready result. Require that check in branch protection after the workflow lands. Rust lint and tests run independently; Cloud static checks, SDK preparation, frontend builds, and tests have separate results. The ignored cluster suite runs through cluster.yml nightly or by manual dispatch, with per-suite logs, five-minute suite limits, and no whole-batch retries. It is informing: a red nightly does not block a tag or Publish.

The live agent eval runs only on a developer's machine, never in GitHub Actions, and no release waits on it. From dashboard/, pnpm eval:live runs every pinned model on every task, 8 trials each, and writes evals/results/gate.json, which lists every blocker against the floors in evals/agent/metrics.ts. Point PLOYZ_EVAL_LAST at an earlier run's evals/results/summary.json to also flag pass^1 drops. It needs Docker and signed-in claude and codex CLIs.

No-model agent replays are also local only. Nothing under dashboard/evals runs in GitHub Actions, and rust-tests skips the ployz mcp transcript replays. Run both before a release. From dashboard/, pnpm sdk:eval, then pnpm eval:replay replays the golden trajectories through the sidebar. It needs Docker and cargo build -p ployz-testkit --bin ployz-fake-fleet. pnpm eval runs everything under dashboard/evals, with live trials still skipped. From core/, cargo nextest run -p ployz --run-ignored only -E 'test(/transcript_replays/)' replays every agent tool transcript through ployz mcp.

Before Publish, use the beta on real Machines through Cloud: enrol, deploy, attach a domain and certificate, and upgrade. There is no scripted qualification; when the informing cluster suite and those Machines disagree, the real Machines are the authority. Testkit bugs do not block a release.

What Publish does

scripts/promote-release.sh runs on release: published.

• Writes one-line pointer files (v0.2.0) on the channels branch: the tag's line pointer (v0/stable or v0/beta) and the unscoped one (stable or beta). A stable tag also writes both beta pointers.
• A pointer only moves to a higher semver tag. Publishing an older-line fix moves that line's pointers only; an older tag moves nothing.
• Only when the unscoped stable pointer names the published tag: regenerates Formula/ployz.rb from checksums.txt and pushes getployz/homebrew-ployz.

Promotion runs one at a time. GitHub keeps only one waiting run, so publishing three releases in quick succession cancels the middle one's promotion: re-run any cancelled Promote published release run. Re-running is safe; pointers only move forward.

Needs repo secret HOMEBREW_TAP_TOKEN (write access to the tap). Channel updates use GITHUB_TOKEN. Publish then dispatches ployz.sh, which deploys install.sh plus the channels branch files to Cloudflare Pages.

Install

stable and beta are the only channels.

The installer reads the unscoped https://ployz.sh/stable (one line, vX.Y.Z) or /beta (vX.Y.Z or vX.Y.Z-beta.N). Missing or invalid channel files fail the install, including a prerelease on stable. Pins skip the channel fetch.

Artifacts stay on GitHub Releases. ployz.sh is the pointer plus CLI installer.

Cloudflare

The repository's .github/workflows/ployz-sh.yml Direct-Uploads the staged site to the existing ployz-sh Pages project. Installer changes on main deploy immediately. Pointer updates deploy because Publish dispatches this workflow (channels has no workflow file, so a push there cannot). Production always uses --branch=main.

Needs repo secrets CLOUDFLARE_ACCOUNT_ID and CLOUDFLARE_API_TOKEN.

Apex and channel URLs must serve these bodies.

Homebrew

Goreleaser does not touch the tap (--skip=homebrew); scripts/promote-release.sh writes the formula. Bottles 404 if the formula is pushed while the GitHub release is still a draft. The tap updates only when a stable Publish leaves the unscoped stable pointer on that tag.

Machine daemon

ployzd install on Linux installs or replaces a Machine daemon. It accepts --version stable, --version beta, or an exact version. A channel reads the daemon's own line pointer, ployz.sh/v<major>/<channel>, and keeps an installed daemon that is newer than the pointer; only an exact version moves a Machine backwards or across lines. Use --software-only for ordinary replacement after the Machine has already been prepared. Setup downloads and verifies the CLI release's daemon as a temporary bootstrap, then that daemon installs the selected Machine release through this interface.

Management transport

Each of the six archives contains a single binary, covered by checksums.txt; Homebrew, the CLI installer, and Machine installation copy one file. The iroh management transport is in-process in the CLI, the daemon, and every native SDK binding, so no helper is packaged and no extra systemd unit is installed.

Clients reach Machines through the Ployz-hosted Ployz Relay at relay.ployz.dev (relay/). It is deployed separately from these releases; the hostname is a compiled constant in ployz-core.