On this page
On this page
Release process
Evidence to keep
Evidence to keep
Keep the Full Release Validation summary: release index, child IDs, slowest-job
tables. Classify failures as product, harness/tooling/provenance,
infrastructure/credential, or wrapper; only confirmed product failures change
Code SHA. Diagnose once, fix if needed, retry narrowly, then reassess—not automatic
all. Decide blocker/flake for each failed test. Every selected failure blocks
unless granted below; narrow evidence never authorizes publication. FRV must succeed.
Linux/Windows/macOS Gateway install and upgrade lanes are required for beta, stable, and full validation. Record actual conclusions and retain diagnostics; failures block. Never label failed, omitted, or deferred lanes passed.
| Release | Record |
|---|---|
| Regular | Code SHA and Release SHA, even if identical; retain successful full parent and exact prepared artifacts for both roles. Later changelog-only reuse: also policy, complete changed paths, green Code parent, and Release parent. |
| Extended-stable | Branch, exact release SHA, accepted producer, parent ID/attempt, workflow ref, every child, frozen-target compatibility repairs/intentional omissions. |
Retain:
- Attempt-one
full-release-execution-plan-<parent-run-id>and currentfull-release-validation-<parent-run-id>-<attempt>: source facts, registry observations/exact selection, immutable upload identity, admission-time freshness. Reuse preserves distinct original/current records without restamping. Authenticate original successful admission and guarded upload—not overall sealer success; interrupted sealers may upload complete plans. Later attempts never replace originals. release-package-under-testfromOpenClaw Release Checks.- Docker release-path
.artifacts/docker-tests/. - Package Acceptance
package-under-testand Docker acceptance artifacts. - Cross-OS artifacts per OS/suite.
- QA/runtime parity and selected Matrix, Buzz, Telegram, Discord, WhatsApp, Slack artifacts.
Readers authenticate retained evidence without refreshing observations or expiring it against today's clock. Admission neither authorizes publication nor promises unchanged registry state; final writers retain trust, approval, content, and readback checks. Pre-fanout budgeting cannot prevent later job/attempt growth exceeding final artifact limits: fail sealing, never truncate evidence.
Bounded publication recovery
A nonempty plugin subset may consume authenticated all-publishable admission, retaining original release/channel/route, Windows operands, and sealed artifact identities. No replacement admission, qualification change, or broader authorization results. Each batch needs new parent authorization and final-effect current authority checks.
| Path | Required evidence and completion |
|---|---|
| Prepared npm | Authenticate complete original manifest before selecting entries. Exclude published packages only after exact registry-byte verification. Before core publication, after plugin child completion, verify complete original npm roster; missing/mismatched packages block. |
| ClawHub | Use original authenticated recovery seal/transaction roster, not replacement prepared manifest. Parent names original run/attempt and dispatches existing recovery workflow. Sealed payload binds original child, selected plugins, current parent, FRV run. Keep environment/credential; verify current parent authorization before every request and exact public bytes. Subsets claim no excluded completion; full-scope parent waits for entire original roster before core. Manual recovery without parent inputs retains full-roster behavior. |
Recovery records validated successor attempt in its existing result artifact before waiting; parent-requested recovery also updates postpublish attempt roster. Timeout or failed observation retains last status as incomplete—not vanished attempt or permission for replacement publication. Original sealed manifest stays unchanged.
Original publisher never ran: authenticated job evidence permits the existing prepared-artifact consumer to restore original bytes. Successful pack jobs authenticate source despite enclosing child failure. Authenticate complete admission/seals before selection; keep original source separate from new child/current parent authorization. Do not invent successful preparation, prior publication, or historical authorization. Fresh authoritative registry checks must find every selected package absent; pending, failed, unknown, or ambiguous states stop for original-attempt reconciliation. Never repack. Parent keeps child active through authorization/exact public readback even with optional ClawHub waiting/core publication disabled. Core waits for verified batch. Ordinary releases retain authorization after core startup; both use the same transaction authorization owner.
Prepared npm quota: preview checks current workflow token's REST core allowance before emitting matrix/evidence handoff:
- Budget 512 requests/package: 256
consume+ 256authorize; leave 1,024 headroom for final authority/surrounding work. Preview itself: 1,024 requests maximum. - Commands enforce ceilings across
gh/artifact HTTP, pagination, retries, archive requests; exhaustion fails closed, never skipping authority/bytes. Accounting counts actual attempted requests; oldfetch-metadata-onlytelemetry is partial. - Oversized selection reports supported smaller selection/reset time before outputs. Example: 98 packages need 51,200 allowance; 5,000 remaining admits at most seven. This does not reduce per-package authentication or promise 98 completions/hour. Resume smaller batches after observed reset with fresh parent authorization; bytes and per-package receipts stay unchanged.
- Allowance is observation, not global reservation. Other workflows, parent/ClawHub, action-internal requests, and manual retries share quota; coordinate separately. Headroom does not enforce their budgets. Later exhaustion requires original-attempt reconciliation, never blind republication.
Release-tooling waivers
Trusted P owns scripts/lib/release-validation-waivers.mjs, not C or Q
(default Q=C). Operator-approved entries require exact version, child role, job name,
and short reason; optional targetSha and evidence: { runId, runAttempt } restrict
candidate/execution. Checked-in list is authoritative; retain release-specific
authorization/diagnostics in PR or handoff.
Land policy, pin P, run/revalidate FRV with unchanged C/Q. Candidate workflows must route state operations to admitted P; changing P cannot retrofit old workflows. State/manifest/summary retain raw failures/reasons. Grants neither skip tests nor mark passes; failed aggregates must contain only granted causes. Missing, cancelled, timed-out, incomplete, mismatched required evidence and ungranted failures block. Package, dispatch, and seal integrity remain mandatory.
Kova grants cover approved thresholds only—not crashes, correctness, isolation, data
integrity, install/update failures. Separate qualificationExemption means
waived-not-run, not failed-job grant or reused measurement. Require:
- P authenticates historical receipt and seals exemption into coverage.
- Q supports
FULL_RELEASE_QUALIFICATION_EXEMPTION_CONTRACTversion1, consuming verified admission before dispatch. Older frozen Q needs explicit repair: frozen C plus reviewed main-ancestor Q in a fresh request, or backport/requalify new C=Q. Changing P alone cannot suppress old Q dispatch. - Plan omits only exempted lane; manifest records not-run, empty run ID, supporting receipt.
Strict consumers verify the record; historical receipts never qualify new bytes.
Retain historical empty telegram_waiver support without rewriting evidence.
FRV must conclude success under pinned policy before publication; publishers cannot
waive failed FRV or change P only at publication. Beta confidence deferrals remain
separate not run records.
Receipt-backed beta aggregate recovery
If a full beta's original plan artifact is unavailable,
.github/workflows/full-release-receipt-recovery.yml is an explicit trusted-main
operation—not restoration/reinterpretation of historical plan, observations, or failure.
- Closed
openclaw.full-release-receipt-recovery/v1input: new request UUID, repository, original qualification-admission descriptor, parent run/attempt, complete sealed candidate binding, all five original beta child selections, original publication descriptors. Preserve the originally admitted C/Q tuple; original publication selection, complete coverage, and frozen upgrade baselines remain authoritative. No new test selection/current-registry baseline resolution. - Authenticate original admission/dispatches, current policy-passing child attempts and complete composite receipts, sealed candidate/package bytes, all npm/SDK/plugin/Docker producer receipts/payload digests. Only admission-authenticated version selects P-owned waivers. Recheck exact aggregate causes, retain raw conclusions, seal matched waivers; strict readers recompute and reject added/omitted grants. Reused children keep dispatch origin; verify original parent's exact reuse-digest witness with ordinary aggregation validator. Ordinary green-only child reuse is unchanged.
- Missing payloads, incomplete history, changed attempts, or lost authority fail closed. Per-child complete rerun-history budget: ten minutes; ordinary reuse: two. Job-inventory timeouts/retry limits unchanged. No test/preparation dispatch; Actions/content read-only. Install only trusted P dependencies with lifecycle scripts disabled; never build/execute C.
- Existing publication-admission owner uses current trusted P to read committed candidate metadata and collect newly timestamped registry observations. Successful aggregate records separate source/admission and original receipt request. Strict readers verify workflow, upload order, observation bytes, current operator authority, original evidence, trusted-main lineage. Publishers use unchanged Q-owned artifacts; P owns aggregation and current admission. Never relabel original parent successful; historical validation and plan restoration are unchanged.
- Producer hashes/inspects original payloads once and seals exact inventory. Readers authenticate successful recovery job, ordered uploads, observations; reconstruct full inventory from original descriptors, small manifests, live artifact metadata and require exact match. No repeated bulk npm/SDK/Docker/candidate downloads per plugin publisher. Fresh bounded plugin listing replaces per-package metadata requests only within that authentication, not immutable replay. Selected consumers still download/verify original bytes; current producer attempts, expiry, authority remain mandatory. Immutable replay contains raw small evidence only, with unchanged limits.
- Parsed payloads retain archive size/integrity limits. Opaque Docker and candidate image/registry archives stream SHA-256 over every byte, without retaining/expanding; authenticated metadata size is exact transfer bound. Recovery Docker transfers share four minutes across at most three attempts, rather than sixty seconds each. Metadata and other archive callers retain sixty-second caps. Binding, expiry, exact size, and full digest remain mandatory; missing/mismatched bytes never permit omission/rebuilding.
Workflow files
.github/workflows/full-release-validation.yml.github/workflows/full-release-receipt-recovery.yml.github/workflows/full-release-candidate.yml.github/workflows/openclaw-release-checks.yml.github/workflows/openclaw-live-and-e2e-checks-reusable.yml.github/workflows/plugin-prerelease.yml.github/workflows/install-smoke.yml.github/workflows/install-smoke-reusable.yml.github/workflows/openclaw-cross-os-release-checks-reusable.yml.github/workflows/package-acceptance.yml.github/workflows/openclaw-performance.yml.github/workflows/npm-telegram-beta-e2e.yml