OpenClaw Release CI

Use this with $release-openclaw-maintainer and $openclaw-testing when a release candidate needs full validation, install/update proof, live provider checks, or CI recovery.

Guardrails

• No version bump, tag, npm publish, GitHub release, or release promotion without explicit operator approval.
• After compaction, resume, or new steering, rewrite the effective goal and current phase from the latest explicit operator instruction. Do not merge old scope back into the active release.
• Hold the release scope once a release branch or Code SHA exists. Validate and ship that exact release; do not turn moving main into a second work queue.
• Record every active validation run as the immutable tuple Validation SHA + Tooling SHA + rerun group. Validation SHA maps to the Code SHA for product validation or the Release SHA for changelog-only validation; it is not a third release identity. A branch or temporary ref is context and transport.
• Freeze the candidate SHA/ref and Tooling SHA/ref once. Main lineage authorizes the initial Tooling SHA selection; it does not authorize replacing that tooling after main advances.
• Apply a release firebreak after the Code SHA is frozen. Admit only confirmed product defects, package/provenance defects in the bytes to publish, security defects, or failures that make publication impossible. Queue other findings for postpublish confidence or the next beta.
• Use trusted main workflow revisions as immutable dispatch sources. Do not adopt newer main code, repair unrelated main CI, wait for broad main health, or expand a release fix because the workflow source lives on main.
• Once publication binds the Tooling SHA to an exact protected lightweight release-publish/<12sha>-<provenance-run> tag, that live tag-to-SHA mapping remains authoritative when main advances. The suffix records tag-creation provenance; it is not the current parent run id.
• Touch main only for an operator-requested change or the smallest critical main-owned blocker that prevents this release and cannot be handled from the release branch. If the required main landing policy is blocked by unrelated main failures, report that blocker and keep independent release work moving instead of healing broader main.
• Validate provider secrets before dispatching expensive full release matrices.
• Do not set GitHub secrets from unvalidated 1Password candidates. If a candidate returns 401/403, leave the existing secret alone and report the exact missing provider.
• Use $one-password for secret reads/writes: one persistent tmux session, targeted items only, no secret output.
• Watch one parent run plus compact child summaries. Avoid broad gh run view polling loops; REST quota is easy to burn.
• Fetch logs only for failed or currently-blocking jobs. If quota is low, stop polling and wait for reset.
• Treat live-provider flakes separately from code failures: prove key validity, provider HTTP status, retry evidence, and exact failing lane before editing code.
• A model-list response proves authentication, not billing or inference entitlement. Mandatory live providers must pass a real completion probe before release dispatch. Fix the credential first; do not add an alternate auth path merely to bypass a failed release credential.
• Full Release Validation separates exact-child dispatch, Release Decision, and Diagnostic Drain. With fail_fast=false, it makes zero child cancellation calls; Diagnostic Drain follows every selected child to terminal unless the collector itself is cancelled or loses GitHub API access. With fail_fast=true, Release Decision may cancel only the exact still-active child that owns a blocking failure.
• Same-parent continuation requires the original root to have been dispatched with fail_fast=false. The controller verifies that exact logged input before any rerun mutation.
• A parent that produced its own sealed candidate artifacts cannot be continued: GitHub reruns make those prior-attempt artifacts unavailable. Keep the candidate and Tooling SHAs frozen, supersede that parent, and start a fresh all-group Full Release Validation.
• After dispatch, one immutable execution-plan artifact records the original parent attempt, exact child tuples and titles, selected coverage, gates, and reuse identity. The same bytes are saved under an exact run-ID cache key. Decision, Drain, manifest writing, evidence validation, and final verification consume the artifact for their current attempt. A collector retry restores the cached plan, validates it, re-uploads its artifact, and adopts the same children; missing plan state is an orchestration failure, not permission to reconstruct the plan or redispatch.
• Reused evidence is not trusted merely because plan sealing found it. Release Decision repeats the sealed target SHA, evidence SHA, policy, changed paths, selected run, root run, source manifest, trusted tooling identity, and exact-child checks before returning passed.
• Parent retries select the newest Decision and Drain artifacts independently; both must bind the same immutable plan even when their source attempts differ.
• Child retries are part of the same immutable plan only when the child run ID, workflow path, ref, Tooling SHA, dispatch title, event, target, candidate, and validation inputs remain exact. Newer child attempts replace matching jobs; jobs absent from a newer attempt carry forward. A duplicate job identity, missing attempt, regressed attempt, or changed tuple fails closed.
• Use pnpm frv status|continue --failed|verify for attempt-aware recovery. The controller is stateless: the immutable execution plan, exact GitHub run attempts, Diagnostic Drain, and final manifest are the only authorities. It never writes a tag, package, registry entry, release candidate, or publication.
• Post-merge controller proof must use the reviewed landed SHA on protected main through the non-release FRV Proof Broker and FRV Proof Fixture. Dispatch the broker with the merged pull request number and exact landed commit. The broker must require that pull request's merge commit to match the landed commit, prove the landed commit is identical to or an ancestor of its trusted workflow SHA, and repeat authority, merge, and ancestry checks immediately before rerunning the fixture. Require the exact fixed no-op fixture run to advance from its intentional attempt-one failure to an attempt-two pass. The broker must emit its receipt without creating a release candidate, release artifact, publication, repository ref, replacement parent, or other workflow mutation. This is the hosted GitHub failed-job rerun proof; focused controller tests own immutable plan eligibility, green-attempt preservation, same-parent collection, and strict-verifier invocation. Never use a real Full Release Validation run for this proof. See Full Release Validation.
• Use one release operator, one transition-only watcher, and at most one investigator for the current failed surface. Do not build audit-review-plan trees around a single workflow transition.
• For regular beta/stable releases, treat the product-complete pre-changelog commit as the Code SHA. Full product validation and performance evidence bind to that SHA. The later Release SHA may reuse those results only when it is a descendant whose complete changed path set is exactly CHANGELOG.md.
• Extended-stable validates one exact branch tip; it does not reuse the regular Code-SHA/Release-SHA evidence model.
• In a sparse worktree or Testbox source sync, first confirm package.json, pnpm-lock.yaml, and every source path the selected check reads. If any are absent, that checkout cannot validate a release dependency or Docker lane: stop and use the repo remote changed gate or a full task worktree. When the inputs are present and a release fix changes package.json or pnpm-lock.yaml, rebuild only the task-owned disposable box with CI=true pnpm install --frozen-lockfile, then run an explicit require.resolve() probe before Docker or focused tests. The CI flag permits pnpm to recreate a prewarmed modules directory without an interactive confirmation. Do not weaken the lockfile or label sparse-checkout failures as product/Docker failures.
• If the candidate is rebased or its base SHA changes after warmup, stop the task-owned box and warm a fresh one before testing. Testbox source sync is relative to the warmed source tree; continuing can mix an old base file with a new candidate diff and produce false lockfile or Docker failures.
• Reused Testboxes are provenance-gated after their first successful run. Source-only edits may reuse the lease; base, dependency, wrapper, or Testbox workflow drift requires a fresh lease. Do not set OPENCLAW_TESTBOX_ALLOW_STALE=1 for release evidence.
• For a committed release candidate, warm the box with blacksmith testbox warmup ... --ref <candidate-branch-or-sha>. Do not rely on source sync to overlay committed branch changes onto the workflow's default ref.

Run identity and retry budget

Record Validation SHA, Tooling SHA/ref, target context ref, parent run id, attempt, and phase before watching or recovering Full Release Validation. Keep Code SHA and Release SHA separately in the lifecycle ledger. Record the immutable Release Publish parent receipt separately from tag provenance.

For the core and plugin npm mutations enforced by this foundation, re-read the exact protected lightweight tag and revalidate the exact parent run tuple immediately before each publish or dist-tag mutation. Reject a missing, moved, annotated, or wrong-SHA tag; a repository, workflow, run id, attempt, tooling identity, or parent-state mismatch; and any same-name branch. Never refresh either identity from current main. Treat other privileged writers as blocked until their dependent enforcement changes land.

• Conceptual phases map to current inputs as follows:

  • beta-publish: release_profile=beta, run_release_soak=false
  • postpublish-confidence: published package inputs with run_release_soak=true or explicit focused groups
  • stable-publish: release_profile=stable

• An all run without soak for an actual beta package on its matching canonical release branch or beta tag records coveragePolicy=npm-beta-v1. It keeps Linux/macOS/Windows Node, Control UI, plugin, package, install/update, cross-OS, QA parity, runtime-pair/restart, and tool-coverage gates. Native app CI, performance, and published-package Telegram are deferred to confidence. Beta all without soak also defers Package Acceptance Telegram, including beta-profile checks of main or alpha. Record deferred checks as not run, never passed. Stable/full, soak, and focused groups retain their coverage; selected children still require terminal evidence. An absent coverage policy retains historical full behavior.

• Keep at most one active parent for the same Validation SHA + Tooling SHA + rerun group + release profile + effective soak coverage. Stable/full always include soak. Distinct coverage profiles can run independently; concurrency does not cancel an older exact child automatically.

• Parent cancellation or timeout leaves adopted identity-checked children running. The operator must cancel an exact child explicitly when it is no longer useful. Do not infer a child identity from branch, title prefix, or latest-run order.

• Recover one failed surface with one diagnosis, one fix when needed, and one narrow retry. Then reassess the release decision. Do not automatically dispatch rerun_group=all.

• For a supported parent, pnpm frv continue --failed --run <parent-run-id> adopts any active newer child attempt, reruns failed child jobs in parallel, leaves green children untouched, then reruns the parent once to restore the immutable plan and seal a trusted all-group manifest. It does not start a second child retry while an attempt is active. Each child or parent rerun mutation is sent exactly once; ambiguous transport failures trigger only bounded read reconciliation. The controller never repeats the mutation, and provenance drift fails closed.

• Inspect without mutation:

  pnpm frv status --run <parent-run-id>
  pnpm frv verify --run <successful-parent-run-id>

• Parents whose immutable plan predates attempt-aware evidence cannot be continued. Start a fresh all-group Full Release Validation; never reconstruct old state or dispatch a replacement parent.

• Controller retries are ci, plugin-prerelease, install-smoke, cross-os, live-e2e, package, qa-parity, qa-live, npm-telegram, or performance. Never use the removed release-checks handle. qa is only a direct-child manual aggregate, not a controller retry API.

• Filtered retries fail closed unless the filter belongs to the selected group. Never turn an empty derived filter into an unfiltered broad run.

• A new all-group parent is justified only when shared orchestration changed, earlier evidence is invalid for the selected tuple, or the operator explicitly requests it. Record the invalidating event.

• Narrow child or group evidence does not by itself become publish authorization. Keep it in the evidence ledger for the release owner to judge against the current publish gate.

Preflight

Before full matrix dispatch, run both pnpm ui:i18n:check and pnpm native:i18n:check against the frozen trusted target in approved isolation. Bind both results to that exact SHA; either generated-locale drift blocks dispatch. Keep target execution outside the trusted dispatch helper—do not execute an arbitrary target checkout as helper code.

Before expensive full validation, also run pnpm ui:build on the same frozen trusted target with its frozen dependencies in approved isolation, outside the trusted dispatch helper. Record the target SHA with the successful production build, precompressed-asset verification, and startup/largest-asset budget results; any failure blocks fanout. Do not substitute a dev server or raise budgets to admit the target.

Before full release validation:

node .agents/skills/release-openclaw-ci/scripts/verify-provider-secrets.mjs --required openai,anthropic,fireworks
gh api rate_limit --jq '.resources.core'
git status --short --branch
git rev-parse HEAD

1Password service-account values are the first source for release provider preflight. Inject those exact targeted keys first, then run the verifier; use ambient env only when it was already intentionally injected for this release. The script prints only provider status and HTTP class, never tokens. The Anthropic check performs a tiny message completion so exhausted or non-billable credentials fail before the expensive release matrix.

Dispatch

An early standalone product-performance run is optional beta confidence. If useful, start it against the frozen Code SHA in parallel with release work:

# Optional early beta confidence; stable/full use the required parent child.
fail_on_regression=false
gh workflow run openclaw-performance.yml \
  --repo openclaw/openclaw \
  --ref main \
  -f target_ref=<code-sha> \
  -f profile=release \
  -f repeat=3 \
  -f deep_profile=false \
  -f live_openai_candidate=false \
  -f fail_on_regression="$fail_on_regression"

• Do not add a separate mandatory prepublish wait for this optional beta signal.
• Compare available Kova, gateway startup, and CLI startup metrics with earlier release evidence or clawgrit reports before publish/closeout.
• Call out any regression in the release proof. Treat a major regression as a release blocker until it is fixed, waived by the operator, or proven to be infrastructure noise.
• Full Release Validation requires blocking performance evidence for stable and full profiles. npm-beta-v1 defers the child; explicit performance and soak-enabled beta runs retain advisory performance coverage. Every selected performance child must finish and prove artifact-only publication.

Prefer an immutable trusted-main workflow revision, target the exact Code SHA:

• Keep trusted-workflow checks compatible with frozen release targets. If main adds a target-owned guard script or package command after the release branch cut, make the trusted workflow skip only when that target surface is absent. Repair the smallest trusted-workflow compatibility issue only when it blocks the release, then rerun validation. Do not port an unrelated runtime refactor, heal other main failures, or mutate the release candidate just to satisfy a newer main-only check.

TOOLING_SHA="<exact-main-ancestor-sha>"
node scripts/full-release-validation-at-sha.mjs \
  --sha <code-sha> \
  --target-ref release/YYYY.M.PATCH \
  --workflow-sha "$TOOLING_SHA"

For regular release/* validation, never raw-dispatch the workflow without target_context_ref (the helper's --target-ref records it). Canonical release/* and extended-stable/* workflow refs remain supported routes, but their Telegram child must retain the exact parent workflow ref and SHA through OIDC and attestation. Trusted-workflow release-branch CI passes target_ref + release_candidate_ref; never release_gate there — it requires workflow head == target. (The PR-head ci.yml fallback below is a different dispatch and does use release_gate=true.)

The release branch may advance after the Code SHA is frozen. The helper accepts that frozen SHA only while it remains an ancestor of the canonical release branch and its package version is either the branch's final version or a matching beta prerelease. Alpha remains on the Tideclaw path with a matching alpha branch and exact alpha tag. Extended-stable branches and all tags require an exact package-version match. Always pass the previously recorded full Tooling SHA for release-branch runs. Never replace it with a fresh main lookup. The Tooling SHA must declare the current release-isolation contract; older workflow revisions fail closed.

For immutable workflow proof on a moving main, use pnpm ci:full-release --sha <code-sha> --target-ref release/YYYY.M.PATCH --workflow-sha <tooling-sha>. Its canonical release-ci/* ref keeps evidence reuse enabled after proving the workflow commit is still on trusted main lineage. Pass -f reuse_evidence=false only when the operator intentionally needs a fresh full run.

After the Code SHA is green, commit only CHANGELOG.md and run the same helper against the Release SHA. The parent must report policy=changelog-only-release-v1, evidenceSha=<code-sha>, and changedPaths=["CHANGELOG.md"]; it should reuse the product matrix instead of dispatching child lanes. Npm preflight and package/install acceptance still run against the exact Release SHA and its new tarball bytes.

Current all-group FRV also owns read-only npm source/build/qualification and Docker preparation. Use its successful run as preflight_run_id; the candidate helper defaults to that run. Do not dispatch a second npm preflight unless recovering historical separate evidence. Regular final qualification records SDK reports for both beta and latest; review the acknowledgement for the actual publication channel. Prepared descriptors live in publicationArtifacts in the exact final manifest. Product evidence reuse never substitutes Code-SHA package or image bytes for the final Release SHA. A parent that produced these artifacts needs a fresh all-group FRV instead of same-parent continuation.

The SHA-pinned helper infers beta for matching beta release candidates and exact alpha tags, and stable for stable/correction versions, then passes the Validation SHA + Tooling SHA run identity. Canonical beta all without soak uses npm-beta-v1; main, alpha, and non-beta targets do not qualify for that policy. Run deferred native, performance, Telegram, broad live QA, and E2E as postpublish confidence with the exact published package and run_release_soak=true or explicit groups. Stable and full profiles force the release soak. Native artifact publication still requires its own build, signing, notarization, and promotion gates. Use a narrow rerun_group after focused fixes; never widen automatically. Publish with openclaw-release-publish.yml using release_profile=from-validation unless a maintainer intentionally wants to cross-check a specific profile; the publish workflow reads the effective profile from the full-validation manifest.

Extended-stable validation

For .33+, dispatch from and target the canonical branch. This direct route is intentional: downstream extended-stable evidence requires the canonical branch identity, while Telegram still authenticates the exact branch SHA:

RELEASE_SHA="$(git rev-parse HEAD)"
gh workflow run full-release-validation.yml \
  --ref extended-stable/YYYY.M.33 \
  -f ref=extended-stable/YYYY.M.33 \
  -f expected_sha="$RELEASE_SHA" \
  -f release_profile=stable

Accept only a complete rerun_group=all run whose branch, head/target SHAs, manifest workflowRef, and package versions identify the same commit. Save its successful run_attempt and require the final tag to resolve there. Reject release-ci/*, current-main, narrow, and earlier-attempt evidence.

Product failures need an approved backport. Frozen-target tooling failures need the smallest behavior-preserving repair. Provider, approval, runner, or log races keep the candidate unchanged. Record repairs and superseded runs; any branch change requires a new complete parent. Omit only an explicitly unsupported frozen-target scenario, never a required behavior or package.

Watch

Use the transition-only summary watcher instead of repeated raw polling:

node scripts/release-ci-summary.mjs <full-release-run-id> --watch

Do not start this watcher when the SHA-pinned helper is still the foreground owner. The helper reads the exact Release Decision artifact itself. On blocked_diagnostics_running, it exits nonzero immediately, keeps the temporary refs, and leaves Diagnostic Drain collecting the remaining terminal evidence. The watcher behaves the same way for separately dispatched parents: it reports the Release Decision blocker once and exits while the drain continues.

For a one-shot snapshot:

node scripts/release-ci-summary.mjs <full-release-run-id>

release-ci-summary accepts Full Release Validation parent runs only. Diverged release-branch logs: --first-parent plus a bounded count. Stop watchers before ending the turn or switching strategy.

Interpret state precisely:

• qualifying: no decisive blocker yet; selected children are still active.
• blocked_diagnostics_running: publication is blocked; Diagnostic Drain is still collecting independent failures. Diagnose now, but do not retry until the drain is terminal.
• passed: all required policy and exact-child evidence passed.
• blocked_complete: publication is blocked and all selected diagnostics are terminal.
• orchestration_error: GitHub API or collector failure prevented a verdict. This is not a provenance mismatch. Recover the collector against the same exact children; never redispatch tests to repair collection.
• cancelled_with_children: the collector was cancelled while exact children remained active.

The full-release-diagnostics-<run-id>-<attempt> artifact is the terminal failure and timing manifest. Use it after an early blocker instead of restarting all merely to discover what the still-running children found. The stable full-release-execution-plan-<run-id> artifact is the identity source within each collector attempt; retry attempts restore its immutable run-ID-cached bytes first.

Failure Triage

1. Confirm parent SHA and child run IDs.
2. List failed jobs only:

   gh run view <child-run-id> --repo openclaw/openclaw --json jobs \
     --jq '.jobs[] | select(.conclusion=="failure" or .conclusion=="timed_out" or .conclusion=="cancelled") | [.databaseId,.name,.conclusion,.url] | @tsv'

3. Fetch one failed job log. If rate-limited, note reset time and avoid more REST calls.
4. For secret-looking failures, validate a real completion from the same secret source before editing code. A successful model-list request is insufficient. Claude CLI subscription credentials are a separate native auth path; prove them in a clean-home CLI probe, never as a substitute for a required Anthropic API-key lane.
5. For live-cache failures, inspect whether it is missing/invalid key, empty text, provider refusal, timeout, or baseline miss. Do not weaken release gates without clear provider evidence.
6. Classify before editing:
   • confirmed product/code failure: fix the release branch, freeze a new Code SHA, and invalidate product evidence
   • harness/tooling/provenance failure: keep the Code SHA, fix the smallest owning surface, and retry only the failed surface with the required Tooling SHA
   • infrastructure/credential failure: keep both SHAs, repair the external prerequisite, and retry only the failed surface
   • wrapper/monitor failure: keep the child and candidate identities; record the wrapper result separately from the child result
   • changelog/release-note failure: change only CHANGELOG.md, keep Code SHA evidence, and repeat Release SHA proof
   • publish child/registry selector failure: keep Release SHA and resume the failed child; never rebuild an immutable version that already published Only the first class changes the Code SHA. After one diagnosis/fix/narrow retry, reassess instead of starting another all-group cycle.
7. If a required PR CI run is capacity-stalled with queued jobs and no active jobs, do not cancel unrelated work or accept a generic manual dispatch. First verify the PR head carries the current fallback schema: gh api 'repos/openclaw/openclaw/contents/.github/workflows/ci.yml?ref=<pr-head-branch>' --jq .content | base64 --decode | rg -q 'pull_request_number:'. If absent, refresh the PR head from main and use the new head SHA; let normal CI run before considering another fallback. From the PR head branch, dispatch the explicit exact-SHA fallback: gh workflow run ci.yml --repo openclaw/openclaw --ref <pr-head-branch> -f target_ref=<full-pr-sha> -f pull_request_number=<pr-number> -f include_android=true -f release_gate=true. It runs on GitHub-hosted runners and is accepted only when its run title is CI release gate <full-pr-sha>. Record the stalled Blacksmith run and the fallback run in release evidence. If Blacksmith Build Artifacts Testbox is the only remaining required gate and remains queued without a runner, that completed exact fallback may cover it because CI's build-artifacts job already builds, packages, and smoke tests the artifacts. Do not use this coverage after the artifact workflow starts or completes non-successfully.

Evidence

Record:

• release lifecycle ledger: Code SHA, Release SHA, and Tooling SHA for regular releases; canonical branch, exact SHA, and immutable tag for extended-stable
• evidence-reuse policy, coverage policy, and complete changed-path set
• active full parent run URL, attempt, workflow SHA, and any superseded parent with the exact replacement reason
• selected child run IDs and conclusions: CI, Release Checks, Plugin Prerelease, NPM Telegram, Product Performance; record deferred confidence as not run
• performance comparison result versus earlier releases when available
• targeted local proof commands
• provider-secret preflight result
• frozen-target compatibility repairs or omitted inapplicable scenarios, with their source PRs and invariant
• known gaps or unrelated failures

For lessons and recovery patterns, read references/release-ci-notes.md.