| name | publish-evidence |
| description | Publish evidence, publish all tapes, update PR verification, audit a red tape. Use for the human-verification layer after @openwork/testkit runs. |
Skill: Publish Evidence
The orchestrator owns this human-verification step. Publishing makes the
agent-first verdict inspectable; it never decides pass/fail and never reruns a
test.
Make every claim auditable
- Show the spec name and verdict, each claim's assertion or fact, the relevant
frames, the source tape, and the reproduction command.
- Require one sticky-comment section per claimed spec. If a claim has no visible
tape section, report the PR
Incomplete.
- Keep both
<!-- photo-roll --> and <!-- fraimz --> markers.
Publish the PR head
After a multi-spec run, publish each tape whose gitSha matches the PR head,
once per roll:
pnpm evals --publish --pr <n> --roll <dir|name>
evals --publish judges pending vision claims on the selected tape, then
publishes it. It publishes existing @openwork/testkit tapes, not legacy
flows, and never reruns tests.
- Omitting
--roll selects the most recent roll; pass --roll explicitly
when several rolls exist so each spec's tape is published deliberately.
- Publishing replaces the sticky comment with the selected roll. Confirm the
final comment shows the spec and verdict you intend reviewers to see.
- Exit codes:
0 published, 1 failed claims published (or publish failed),
2 pending claims still need judging (set a vision key and rerun).
Refuse misleading evidence
- Never use
--force to hide a SHA mismatch. Re-run the spec on the PR head.
- Use
--force only to deliberately publish a historical or red tape. The
output is annotated; call the exception out explicitly. Red tapes are valid
human-verification artifacts and should be published when they explain a
Failed or Incomplete verdict.
- Read
BLOB_READ_WRITE_TOKEN from the environment or the Infisical fallback.
Without it, still post verdicts with a no-screenshots note.
