Relay AI is generally available. Pipelines can now detect flaky tests, auto-retry transient failures, and post AI-generated summaries on pull requests.

• Flaky-test detection — statistical analysis across branches and matrix legs flags tests with intermittent failures
• Auto-retry — configure retry.when: relay.ai.flaky on any step to retry only likely-transient failures
• PR summaries — relay/ai-summary@v1 posts root-cause analysis and suggested fixes on failed builds
• Quarantine workflow — move persistently flaky tests to non-blocking jobs without stopping the pipeline
• Insights dashboard — new Flaky Tests and Timing Regressions views under Insights

Enable Relay AI in Settings → AI or add an ai: block to relay.yaml. See the Relay AI docs for configuration examples.

This release expands runner coverage and significantly improves remote cache performance for monorepos and matrix builds.

• Self-hosted ARM64 runners — register self-hosted-arm64 agents for Apple Silicon and Linux ARM builds in your VPC
• 2x faster cache restores — content-addressable cache layer reduces median restore time from 90s to under 45s on large node_modules trees
• Cross-branch cache sharing — cache hits now apply across feature branches when lockfile hashes match
• Matrix deduplication — identical matrix legs share a single cache upload
• Runner labels UI — filter and manage self-hosted pools from Settings → Runners

Also in this release:

• Deploy gates support required reviewer groups per environment
• OIDC token lifetime extended to 15 minutes for multi-step cloud deploys
• Fixed a bug where preview environments were not destroyed when a PR was closed via merge queue
• Fixed a bug where cache keys containing ${{ hashFiles() }} were truncated on Windows runners
• Fixed a bug where concurrent matrix jobs could exhaust small runner pools without queue visibility

Learn more in the Runners docs and API Reference.

Preview environments now support per-repo TTL policies and idle detection so stale PR previews do not linger after merge or abandon.

• TTL policies — set default lifetime in hours under Settings → Environments → Previews
• Idle detection — previews with no HTTP traffic for 72 hours are queued for teardown with a Slack or email warning
• Bulk cleanup — admins can purge all previews for closed PRs from the Environments dashboard
• Cost visibility — preview compute minutes appear as a separate line item in Usage

Also in this release:

• Preview URLs support custom subdomains on verified domains
• Fixed a bug where redeploying a preview could leave the previous deployment running
• Fixed a bug where monorepo path filters ignored nested preview: overrides

Configure TTL defaults in the API Reference.