Skill: Publish Evidence

The orchestrator owns this verdict and human-verification step. It derives the verdict from completed evidence; publishing makes that verdict inspectable and never reruns a test.

Declare the verdict

• Passed: every claim has an observable assertion in the completed test run.
• Failed: an assertion disproves at least one expected outcome.
• Incomplete: requirements, tooling, or evidence are missing. A skip is always Incomplete, never Passed.
• Prose, screenshots, and recordings do not decide the verdict.

Make every claim auditable

• Show the test name and verdict, each claim's assertion evidence, the relevant test artifacts, the source test run, and the reproduction command.
• Publish one compact sticky comment linking to a report with a section for every claimed test. The report must show assertion evidence and original records. When the review app is not configured, keep those sections in the comment. If a claim has no inspectable evidence, report the PR Incomplete.
• Write the <!-- test-evidence --> marker. The publisher recognizes old sticky markers only to update comments created before the migration.

Publish the PR head

Run checks on the final PR head after any rebase or cherry-pick. Test runs are bound to a commit SHA, so history rewrites require rerunning and republishing. After a multi-test run, publish the complete selection together:

pnpm evals:e2e --publish --pr <n> --all
# Or select runs explicitly; repeat --test-run for each claimed run.
pnpm evals:e2e --publish --pr <n> --test-run <dir|name> --test-run <dir|name>

--all selects stored runs matching the current PR head. Include a DocShot receipt with --docshot <image.png.review.json>. It must match that commit too. Names and captions provide the default report; --title and repeatable --gap are optional context. No separate report-writing step is required.