The first live babysitter trial replayed PR #292 locally but published nothing: Bun was installed on the runner, outside the agent's sandbox mounts. Its infrastructure report then failed because the safe-output tool rejected the required empty `review` field. The workflow still appeared successful.

Expose verified Bun 1.4.2 through the existing read-only tool-cache mount and install pinned dependencies at the working head before verification. Allow the exact GitHub release-asset domain dprint uses. Make irrelevant report/proposal fields optional, normalize omissions in the trusted finisher, and retain exact diff, meaningful blocker reason, protected-path and branch-lease checks.

Validation: 158 maintenance tests pass; pinned gh-aw 0.86.2 strict compilation, formatting and repository CI checks pass. The staged binary reports Bun 1.4.2 locally. Independent review checked sandbox visibility, dependency-install timing and the output contract. A live sandbox trial is still required after this lands.