StoryboardHandler (#37): breaks the project's script into a numbered
scene-by-scene breakdown (description + image prompt per scene) via
OpenAI in JSON mode (added json_mode support to OpenAiDriver for
reliable structured output), stored as a JSON asset.
VideoGenerationHandler (#38): generates one still image per storyboard
scene via StabilityAI — the scene *visuals*, not motion/video clips
(no text-to-video model integrated; see the task comment for why).
VideoEditingHandler (#39): the step that actually produces a video
file — assembles the scene stills into an MP4 via ffmpeg's concat
demuxer, timed to the voiceover audio's duration (probed via ffprobe)
or a fixed per-scene duration if there's no audio, then muxes the
audio track in. Verified against the real ffmpeg/ffprobe binaries on
this box, not just Process::fake().
Asset previews (#45, #46): a reusable partial renders each asset
inline by type — image/audio/video get real playback elements, text/
JSON gets a truncated <pre> block, everything else a download link.
Wired into both the awaiting_approval step view (so approving/
rejecting is an actual content review, not just a word count) and a
general project assets grid.
Also fixes a latent bug in AssetStorage::url(), found while wiring up
previews: it called Storage::url() on the default 'local' disk, which
has no configured URL driver and throws in production — only worked
before because Storage::fake() is more permissive than the real local
driver. Now serves assets through a new authenticated streaming route
instead (VideoProjectAssetController).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
AiProviderDriver: one generic interface (generate(prompt, options) ->
AiGenerationResult) rather than per-capability interfaces, since text
(OpenAI chat completions), image (StabilityAI stable-image), and
speech (ElevenLabs TTS) generation all reduce to prompt-in/content-out.
Drivers throw AiGenerationException on failure; AiProviderManager
resolves the right driver from an AiProvider record's `provider`
column and guards on is_active. Real request/response shapes per each
provider's API, verified with Http::fake() — no live API keys
available yet to smoke test against the real services.
AssetStorage: stores generated content under
video-projects/{id}/{type}/v{n}.{ext}, versioned per (project, type[,
step]) slot so retries don't clobber earlier output, plus
contents/url/delete and pruneOldVersions for cleanup.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
