| name | render-feed |
| description | Generate and verify paid fal.ai video renders for the DoomScrum swipe feed. Use when: "regenerate content", "render the feed", "new renders", "bake-off a model", "switch video model", "re-roll a clip", or any task that spends FAL money. Trigger: /render-feed. |
render-feed
Paid renders for the swipe feed. Every dollar is tracked; every clip is
transcript-verified before it counts.
Money rules
- Unit cost is quoted by
fal::unit_cost(cfg) (snapped clip duration × model
price). Spend cap lives in doomscrum.toml (max_total_spend_usd); raise
it only with the owner's say-so. Provenance = render JSONs in
.doomscrum/renders/<spec-sha>/<render-id>.json; the sum of their
cost_estimate_usd (all statuses, including retired) is total spend.
- To remove a render from the feed without losing spend provenance, set its
status to "retired" — never delete render dirs.
- FAL key: env
FAL_API_KEY/FAL_KEY or ~/.secrets. Never print it.
- Failed seedance moderation (422 after render) is unbilled; silent duds are
billed. Budget ~1.3x for re-rolls.
Model facts (verified on fal 2026-06-10 — re-verify before adopting)
Authoritative table: src/providers/fal.rs (model_price_per_second,
clip_duration, per-family request schemas) + comments in doomscrum.toml.
| Model | $/s w/ audio | Cap | Note |
|---|
bytedance/seedance-2.0/fast/text-to-video | 0.2419 @720p | 15s | hero: native word-synced on-screen captions |
fal-ai/sora-2/text-to-video | 0.10 | 12s | won 8s script-fit bake-off |
fal-ai/kling-video/o3/standard/text-to-video | 0.112 | 15s | 1080p, structured multi_prompt |
fal-ai/veo3.1/lite | 0.05 @720p, 0.08 @1080p | 8s ("4s"/"6s"/"8s") | cheapest dialogue; 9:16 native |
fal-ai/pixverse/v6/text-to-video | 0.060 @720p (0.025 @360p silent) | 15s | draft tier for prompt iteration |
Schema check for a new model:
curl -s "https://fal.ai/api/openapi/queue/openapi.json?endpoint_id=<id>".
Iron rule: no unverified paid render
After every render:
python3 scripts/check_script_fit.py <mp4> "<planned script>" [--words-json out.json]
--words-json saves word-level timings (caption overlays for the demo).
COMPLETE (exit 0) or it gets re-rolled/retired. The planned script comes from
the storyboard in .doomscrum/storyboards/. Word budget is
(duration - 2) * 2.0 words (src/distill.rs); do not pad scripts past it.
Commands
cargo build --release
./target/release/doomscrum generate --root . [--model <id>] [--spec <filter>]
Run long renders in the background with the absolute binary path and
--root pinned — cwd drift breaks relative paths. rm is hook-blocked;
use /usr/bin/trash.
Render gotchas (earned, not theoretical)
-
Exhausted fal balance masquerades as everything else (2026-06-11): a
locked account produced storage 403s, LTX 422s at result-fetch, and
600-poll queue timeouts across models before any request said "Exhausted
balance". When two unrelated failure shapes appear in one batch, probe
with a cheap submit and read the 403 body before debugging models.
Our spend cap tracks OUR ledger, not the prepaid fal balance.
-
ltx-2.3 is banned from the content mix (verified 2026-06-11): 1 of 4
clips had catastrophic diction ("no vibes merges, capiche" -> "No merges.
Copy, sheep."), 1 cut its last word. veo3.1/lite carries the cheap weight.
-
Transcription fallback: fal storage upload may 403 independently of
the queue; check_script_fit.py falls back to Deepgram through the exact
Mint route derived from MINT_BASE_URL, using
__mint.deepgram.default__ — direct binary upload, no storage hop and no
raw Deepgram key in the agent process.
-
Scripts must spell numbers as words and skip exotic interjections
("capisce") — voice models garble them and the transcript gate eats the
miss. The scriptwriter prompt enforces this; don't hand-write scripts
that violate it.
-
Seedance output moderation rejects "unhinged" / "deadly serious" phrasing.
-
Accented narrators garble words unless the prompt demands "crisp, clearly
intelligible English".
-
fal rejects data: URIs — upload via
https://rest.alpha.fal.ai/storage/upload/initiate + PUT.
-
Stale storyboards for archived specs masquerade as budget bugs — trash
.doomscrum/storyboards/ before a full regen.
