Runway

Build me a thin video-generation workbench over hosted models, in place of
Runway. Requirements:

- A local web page on localhost:3300: Node + Express, a prompt box, a model
  picker, and a gallery of past generations.
- Generate via the Replicate API (token in .env) against current open video
  models (Wan/LTX/Hunyuan class). Pick one provider and stick to it; fal.ai
  only if it has a model Replicate lacks.
- Poll the job, download the mp4 to ~/generations/YYYY-MM-DD/, and record
  prompt, model, params, cost, and file path in SQLite (better-sqlite3).
- Image-to-video: upload a still and pass it as the conditioning frame.
- Show a month-to-date spend total computed from per-job cost so usage stays
  honest.
- No accounts, no telemetry; prompts and outputs stay local except the API
  calls themselves.
- Out of scope: editing timelines, inpainting, and anything realtime. Do not
  build model hosting or a job queue beyond simple polling.
- README: where to get the Replicate token, rough dollars per second of video
  on the default model, and a note that this rents models rather than replacing
  Runway, whose frontier video models and compute cannot be rebuilt solo.

# AGENTS.md — Build guide for Runway

## Project scope
Wrap open-source/video APIs to generate or edit clips, manage credits, and export assets.

Catalogue verdict: no. A UI around video models is buildable; the video-generation model quality, compute, safety, editing stack, and continuous research are not a solo afternoon project.
Use the implementation prompt below to define the deliverable. Complete each phase's acceptance checks before extending the scope.

## Working agreement
- Inspect the repository and its existing instructions before choosing paths, dependencies or commands. Keep one coherent stack and explain changes to the proposed architecture.
- Plan a vertical slice that accepts a real input and produces the useful output described below. Persist only the state the prompt calls for; respect memory-only and upstream-managed workflows. Use fixtures only when they are clearly labelled.
- After scaffolding, document the actual install, development, check and build commands in README and keep them synchronized with the package or project manifest. Do not report commands as successful unless they ran.
- Work in small steps. At handoff, list implemented flows, checks actually performed, remaining blockers, and any credentials or provider setup the owner must supply.
- Do not publish, spend money, contact customers, delete source data or run irreversible migrations without the project owner's authorization.

## Prerequisites
- GPU/hosted video API
- storage
- video rendering
- safety filters
- prompt/workflow UI

## Stack and architecture
- A local web page on localhost:3300: Node + Express, a prompt box, a model picker, and a gallery of past generations.
- Model prompts, source clips, model settings, jobs, and exported videos; keep stable source IDs and timestamps.
- Interface for this generative video workflow: a local web page with input, progress, review, and export views.

## Security and data integrity
- Keep the original media file untouched when decoding or rendering fails; show the failing job and allow retry.
- Validate imported records and file paths; keep user data within the storage location named in the prompt.
- State the practical limit: frontier video models.

## Agent implementation rules
- Project rule, data: Model prompts, source clips, model settings, jobs, and exported videos; keep stable source IDs and timestamps.
- Project rule, behavior: Generate via the Replicate API (token in .env) against current open video models (Wan/LTX/Hunyuan class). Pick one provider and stick to it; fal.ai only if it has a model Replicate lacks.
- Project rule, recovery: Keep the original media file untouched when decoding or rendering fails; show the failing job and allow retry.

## Optional agent skills and references
No installable skill is listed for this guide; use the project rules and the stack documentation.

Read the linked SKILL.md and its dependencies before adding a skill. Select only the skills matching this project's runtime and task; their documentation does not supply API access, credentials or approval to perform external actions. Pin the reviewed revision where the tool supports it. Follow the chosen agent's documented project-level installation mechanism.

## Distribution ideas
These are optional planning notes. Obtain the owner's approval before publishing or contacting anyone.
- Demonstrate this generative video operation with sample or user-owned data: Generate via the Replicate API (token in .env) against current open video models (Wan/LTX/Hunyuan class). Pick one provider and stick to it; fal.ai only if it has a model Replicate lacks.
- Show a concrete result or export: Poll the job, download the mp4 to ~/generations/YYYY-MM-DD/, and record prompt, model, params, cost, and file path in SQLite (better-sqlite3).
- State the practical limit: frontier video models.

## Engineering roadmap
1. Phase 1, architecture and data: A local web page on localhost:3300: Node + Express, a prompt box, a model picker, and a gallery of past generations. Model prompts, source clips, model settings, jobs, and exported videos; keep stable source IDs and timestamps.
2. Phase 2, implement: Generate via the Replicate API (token in .env) against current open video models (Wan/LTX/Hunyuan class). Pick one provider and stick to it; fal.ai only if it has a model Replicate lacks.
3. Phase 3, implement: Image-to-video: upload a still and pass it as the conditioning frame.
4. Phase 4, review and output: Show a month-to-date spend total computed from per-job cost so usage stays honest. Poll the job, download the mp4 to ~/generations/YYYY-MM-DD/, and record prompt, model, params, cost, and file path in SQLite (better-sqlite3).
5. Phase 5, recovery and acceptance: Keep the original media file untouched when decoding or rendering fails; show the failing job and allow retry. Verify this invariant with a saved fixture: A failed or cancelled job cannot be presented as finished; exported video duration and frame size must match the saved job settings. State the practical limit: frontier video models.

## Paid-product capabilities outside this build
- frontier video models
- compute
- asset history
- model updates
- commercial rights workflows
- editing polish

## Implementation prompt
Build me a thin video-generation workbench over hosted models, in place of
Runway. Requirements:

- A local web page on localhost:3300: Node + Express, a prompt box, a model
  picker, and a gallery of past generations.
- Generate via the Replicate API (token in .env) against current open video
  models (Wan/LTX/Hunyuan class). Pick one provider and stick to it; fal.ai
  only if it has a model Replicate lacks.
- Poll the job, download the mp4 to ~/generations/YYYY-MM-DD/, and record
  prompt, model, params, cost, and file path in SQLite (better-sqlite3).
- Image-to-video: upload a still and pass it as the conditioning frame.
- Show a month-to-date spend total computed from per-job cost so usage stays
  honest.
- No accounts, no telemetry; prompts and outputs stay local except the API
  calls themselves.
- Out of scope: editing timelines, inpainting, and anything realtime. Do not
  build model hosting or a job queue beyond simple polling.
- README: where to get the Replicate token, rough dollars per second of video
  on the default model, and a note that this rents models rather than replacing
  Runway, whose frontier video models and compute cannot be rebuilt solo.

## Completion evidence
Demonstrate the prompt's acceptance scenarios against the scoped workflow. Include setup from a clean checkout and failure recovery. Check persistence across restart and export/restore only for the state the prompt says to store; for memory-only tools, confirm that temporary content is discarded as specified. Record actual results and remaining limitations. A detailed plan alone does not establish a working replacement.