Assemble video-link import: async pipeline, premium gate, cost controls

[windoze95]

What this is

PR2c — the final assembly of the video-link import feature (Phase 2.5). It wires the three merged foundation PRs into a working end-to-end pipeline behind a paywall:

• Video-import foundation: cache, async-job model, premium quota + kill switch #86 cache/quota/config + async-job model
• Video import PR2a: ScrapeCreators client (platform detect + TikTok) #87 ScrapeCreators client (platform detect + TikTok)
• Video import PR2b: ffmpeg frame sampling + ffmpeg in Docker image #88 ffmpeg frame sampler (+ ffmpeg in the Docker image)

A user pastes a TikTok/Instagram/YouTube/Facebook/Pinterest link; we pull the caption + transcript + a downloadable copy of the video, sample representative frames, and run a single multimodal Claude extraction over frames + transcript/caption together to reconstruct the recipe.

Pipeline (async)

POST /v1/recipes/import/video → premium gate → create job → return 202 {job}.
Goroutine: FetchVideo → per-video cache check → daily-budget kill switch → SSRF-guard the scraper media URL → Sample frames → ExtractRecipesFromMedia(frames, transcript+caption) → save recipe → cache the extraction.
GET /v1/recipes/import/video/:id → owner-only status poll (returns recipe_id when done).

Cost controls ("don't let users bleed me")

• Process-once cache (VideoExtractionCache, keyed platform:video_id): a viral video is extracted once; every later importer is served from cache at $0.
• Per-user monthly quota: free 2 / premium 20 (CanUseVideoImport).
• Global daily-budget kill switch: fresh extractions stop once SUM(cost_usd) since UTC-midnight reaches VIDEO_IMPORT_DAILY_BUDGET_USD (default $25).
• Feature flag: the endpoint returns 503 until SCRAPECREATORS_API_KEY is configured — ships dark.
• Frame count is the master cost dial (~$0.0048/frame on Sonnet); cost is over-estimated so the kill switch trips early.

Notable decisions

• Interface ripple: VisionProvider.ExtractRecipesFromMedia gained a contextText arg (transcript/caption); ImportFromFiles passes "".
• SSRF: the scraper-supplied media URL is validated with the existing ValidateExternalURL before the sampler fetches it. Follow-up: harden the sampler's HTTP client against redirect-hop rebinding (the validator is TOCTOU-imperfect on its own).
• Quota on accept: usage is incremented when a job is accepted (not on completion) and cache hits still count — a deliberate product/cost choice; documented for review.
• Bug fix: monthly usage reset now also zeroes VideoImportsUsed in-memory.

Tests

All offline and race-clean (go test ./... -race):

• Service: fresh extraction, cache hit (no sample/extract), budget-exceeded (fails before sampling), fetch-failure, not-configured, unsupported platform.
• Handler: gate 403, unconfigured 503, bad platform 400, accept 202 + poll → done, non-owner 404.

Deploy note (not in this PR)

At deploy, add SCRAPECREATORS_API_KEY to the prod ECS task definition to switch the feature on, then run a real TikTok end-to-end check. Only TikTok is implemented in the client so far; IG/YouTube/Facebook/Pinterest detect but return "not yet supported".

🤖 Generated with Claude Code

https://claude.ai/code/session_01BU4UWZutHd1AnK3XAf7H19

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.