[executorrr86]

Summary

video_compose render runs a slideshow-risk gate before rendering, and it reads an optional scene_plan input. If an agent passes the scene_plan artifact as it is stored ({"version": ..., "scenes": [...]}), the gate loops over the dict's keys and scoring fails with 'str' object has no attribute 'get'. That error is only logged, so the check is skipped without notice. If the agent passes the same plan as its scenes list, the gate scores it and can block. This PR unwraps the artifact to its scenes, so both forms get the same verdict.

Related issue

None. Refs #518, which also normalizes this dict as part of a much larger change. This PR is the small piece that can be reviewed on its own.

Changes

• video_compose.py, _pre_compose_validation: if scene_plan is a dict, use its scenes (2 lines). If the dict has no scenes or the list is empty, the gate falls back to the cuts as before.
• video_compose.py, input_schema: declare scene_plan (array or object). execute() already read it and idempotency_key_fields already listed it, but the schema did not, so agents had no way to know the input existed.
• test_video_compose_scene_plan_artifact.py: 7 tests that use the real scorer, with no stubs.

Testing

Gate verdicts with renderer_family: cinematic-trailer (scores from score_slideshow_risk):

Case	No plan (cuts only)	Plan as list	Plan as artifact, main	Plan as artifact, this PR
6 video cuts (3.33 revise) + slideshow plan (4.25 fail)	pass	block	pass (scoring crashed)	block
5 text cards + 1 video cut (4.00 fail) + good plan (0.63 strong)	block	pass	pass (scoring crashed)	pass

• The new tests: 6 of 7 fail on main. The 7th passes on main only because the crash skips the check; the other tests catch that case.
• Full pytest tests with this change: 2924 passed, 17 skipped.
• Behaviour change to be aware of: a slideshow-like plan passed as the artifact now blocks, as it already did when passed as a list. A plan passed as a list, or no plan at all, gives the same result as before.

Checklist

• The change is focused on a single logical concern.
• I ran the relevant tests locally (make test-contracts / make test) where applicable.
• I updated docs/README if behavior or usage changed. (The new input_schema description documents the input.)
• No unrelated files (build artifacts, local config) are included in the diff.

🤖 Generated with Claude Code