OpenClaw Release Validation

Help a human test an immutable latest-main build against a selected real Gateway's state. Automate preparation, triage, and reporting; the human drives interactive surfaces and judges quality. This skill collects feedback, not a release go/no-go decision.

Choose the mode

• Campaign artifact: only when RELEASE_VALIDATION_ARTIFACT_PATH is present. Read artifact generation, write that artifact, and stop. GitHub is read-only; no skill-update request or interactive setup.
• Update campaign: for an explicit refresh request, read installed skill check, then campaign lookup/dispatch. Dispatch the isolated runner for the selected explicit tag, verify its issue, and stop. Never analyze or edit the canonical issue yourself.
• Validate release: default human-testing workflow below. It joins the campaign; it never creates or rewrites the issue body directly.

Human validation

1. Read installed skill check. Introduce the run briefly: choose a Gateway, protect it with a disposable OCM copy or approve an in-place update, prepare the pinned main build, then test and review feedback.
2. Read campaign lookup/dispatch. Require a current campaign and record its stable train and current beta. Gateway preparation pins the main test target separately.
3. Read Gateway preparation. Discover without mutation, let the tester choose source and mode, then prepare the selected target. Never use the caller's active checkout.
4. Only after readiness succeeds, offer local diagnostics for an isolated target, then read human testing to create the worksheet and guide one chosen surface at a time.
5. On finish validation, read cleanup and report approval. It handles blocked upgrades too. Review the complete batch before any post and preserve the exact structured-report contract.

Before readiness is terminal, show only campaign/beta identity, source/mode choice, and upgrade progress or errors. Do not expose testing instructions or create a worksheet for a Gateway that never became ready.

Boundaries

An isolated OCM copy must prove all candidate-writable paths remain contained. Do not bypass a failed containment check with manual copying. Copied channel credentials require explicit authorization to stop their current owner and a restoration plan. In-place mode changes real state: present verified backup/snapshot and dry-run evidence before obtaining update approval. Rollback requires separate approval because newer migrations can make code-only reversion unsafe.

Keep one worksheet for a ready candidate; only tester-authored results count as tested. A blocked candidate uses its final upgrade-report draft instead. For setup, OCM, backup, build or cleanup failures, use the private tooling-feedback packet; do not turn these into candidate findings. With no candidate evaluation, clean up and report that fact without producing a candidate vote or posting batch.

Diagnostics are opt-in, disposable-target-only, loopback-only and content-off. Public feedback excludes raw logs, telemetry, paths, Gateway names, identifiers, credentials and setup/cleanup details. Apply the report's narrow environment allowlist and sanitization before showing the complete draft. finish validation does not authorize posts; existing explicit batch approval does. The Discord summary is a copy-ready handoff, never an automatic send.