Enterprise Testing

Prove the changed contract with the smallest meaningful check, complete required checks, then stop. Broaden or repeat for new changes, failures, or unresolved risks. Use test-audit when authoring or reviewing tests; do not create tests that merely mirror a reversible documentation or style edit.

Read the touched scope's AGENTS.md and the selected testing guide. Run commands from the repository root. Use Node.js 24 or newer and repository-pinned pnpm with matching installed dependencies. Report missing prerequisites; dependency-independent checks prove only their own scope. Never run npm run precommit.

Select proof

Run pnpm check:workspace and applicable pnpm format:check with the installed graph. The root formatter does not include .agents/skills/**/*.md; inspect adapted skill Markdown separately. Keep autoreview implementation and test changes in openclaw/agent-skills. Use the local checks guide for formatting and generated-artifact procedures. Resolve <name> to existing owner/sibling tests; do not run placeholder commands. When all conformance and integration contracts need proof, pnpm test selects both; browser and credentialed coverage still depend on their documented selection and prerequisites.

Real-runtime boundaries

Use only owned disposable resources and authorized credentials. Scope selection variables to the chosen process; preserve shared dependencies, services, databases, kubeconfig, and unrelated clusters. Do not run untrusted contributor tooling against a credentialed local environment; use an approved isolated CI environment. Setup and credentialed execution need their existing authorization.

Follow repository integration requirements and each suite's guide for images, model selection, database roles, credential placement, networking, and cleanup. Do not bypass missing credentials, disable isolation, or substitute a fake for requested runtime proof. Selected required cluster/runtime cases must pass without skips. Fixture or in-memory success does not prove genuine gateway, Codex WebSocket, model execution, or production deployment. A green PR aggregate does not prove manual credentialed lanes ran.

Diagnose CI

gh pr view <pr> --json headRefOid,statusCheckRollup
gh run view <run-id> --json status,conclusion,headSha,url,jobs
gh run view <run-id> --job <job-id> --log

Bind evidence to the PR head, actual run SHA, run ID, and job ID. A pull-request run may test a merge commit: record that relation rather than assuming its SHA equals the branch head. Inspect the latest attempt for that source; confirm whether a newer run superseded a cancellation. Fetch relevant leaf logs and result artifacts, not just the rollup status. Use CI result accounting and coverage to interpret selected lanes and skips.

• Product: supported behavior fails with valid setup; reproduce narrowly, repair the owning implementation, and rerun that proof.
• Harness: selection, setup, assertions, or result accounting are wrong; repair the harness without weakening required outcomes.
• Infrastructure: provisioning, image availability, networking, or capacity fails; retain evidence and retry the affected job only when justified.
• Credentials: approved credentials, access, or protected-environment approval is missing; report the gap without exposing values or substituting mock proof.

Do not infer a category solely from a check name. State uncertainty when logs are unavailable. Route unrelated failures with evidence instead of expanding the task; never rerun a broad workflow merely to erase a failure. Report commands, source/run/job identity, pass/fail/skip counts, proof limits, and remaining blockers.

Provenance

Adapted from OpenClaw; see source and intentional adaptations.