Skip to content

Self-hosted, Self-Improving, Federated Everything App — dev management, agent orchestration, persistent mind, digital twin, second brain, and creative pipeline from a single dashboard over Tailscale to port the user into the OS

5.8 KB
Raw Blame

Runway model options

PortOS uses Runway's public OpenAPI document (API version 2024-11-06) and model guide. The documented API has no model-list/discovery endpoint. In particular, /v1/models is not used. The public schema describes advertised capabilities, not an account's entitlements or remaining credits. Catalog refresh makes no generation request and needs no API key.

The bundled capability snapshot was verified on 2026-10-22. It offers 19 image models and 24 image-to-video models; 23 of the video models also accept text alone. Video edit adds Aleph 2 and Character performance adds Act-Two, for 26 distinct video models across these workflows.

Selectors filter by the active operation. Gen-4 Image Turbo requires a reference; Gen-4 Turbo video requires a start frame. The builder validates each model's prompt, reference and seed constraints and emits supported parameters only. Models accepting native negative prompts use that field; other video models receive a visible Avoid addition. Native provider audio is disabled where the API exposes that control.

Image settings retain existing Gen-4 720p/1080p choices. Other models offer automatic canvas selection that preserves requested aspect ratio, plus their exact documented canvas dimensions. Video sizes and lengths depend on both model and operation: WAN keyframe requests use auto ratios, text requests use explicit canvases, and models such as MiniMax accept a separate resolution field. Optional advanced API controls keep provider defaults. Unknown prices are displayed as unknown, not zero; consult Runway pricing before rendering.

Veo inclusion decision (2026-10-22)

The earlier Runway integration (#12344) deliberately excluded Veo as a project choice. This expansion reverses that choice in response to the newer request to offer all models on Runway's API within the supported workflows. veo3.1 and veo3.1_fast are therefore intentionally included; Gen-4.5 remains the default. The earlier catalog and PR record no safety or contractual basis for the exclusion. Inclusion does not assert account entitlement, waive provider terms, or enable generation without the normal submission flow.

Cache and refresh

All form instances share a five-minute browser cache, including concurrent loads. The server shares one 24-hour catalog cache across images and videos and persists it atomically in .cache/runway-model-catalog.json. This is disposable public metadata, not an application record; it contains no credentials or personal data, is not federated, and needs no database migration. A hash of the shipped capability snapshot invalidates older disk caches on upgrade. Changing credentials does not invalidate public model metadata, and no key is sent to the documentation source.

The Refresh models button bypasses freshness on both layers and updates open subscribers. Concurrent source refreshes coalesce. A source failure preserves the last good catalog, or uses the bundled snapshot on a cold start, marks it stale, and backs off automatic source retries for five minutes. A valid empty model list stays empty. A malformed document or unsupported API version cannot overwrite the last good catalog. Corrupt or incompatible disk caches are discarded.

Refresh can remove retired options. Known models with changed consumed request constraints and newly advertised model IDs are listed as requiring a PortOS capability update; refresh does not guess new payload semantics. Update server/lib/runwayCatalogSnapshot.js, its shared image/video projections and tests after checking the official schema. A refreshed removal also refuses submission before a paid POST. Generation does not perform a separate discovery lookup.

Workflow limits

Video Gen's Runway lane offers Video edit (aleph2) and Character performance (act_two) alongside Text and Image. Aleph takes a 2–30 second source video at up to 30 FPS, an optional prompt (1000 characters including the negative prompt), an optional guidance image at the start, and optional expand aspect ratio. It uses the default MP4 delivery and preserves source resolution up to 1080p. Act-Two takes a 3–30 second performance clip plus one character image or video, body-control and expression-intensity controls, and an optional supported canvas. It follows the performance rather than a text prompt. Video durations and streams are probed before upload; invalid or unreadable inputs fail before generation.

Uploads are staged durably using the normal queue ownership/cleanup rules. The worker creates ephemeral Runway uploads at dispatch time, propagates cancellation through registration and storage POST, and submits only after every upload succeeds. Video files must be 512 bytes–200 MB. Only local input paths are queued; 24-hour upload URIs are never persisted as retry inputs, and replay creates fresh uploads. Storage POSTs receive signed upload fields, not the Runway API key. Local owned inputs are cleaned on success/failure/cancel; Runway's remote uploads expire automatically. Reloaded forms require video files to be selected again before a new submission.

Specialized models stay out of text/image-only Music Video and Creative Director selectors. Other models' advanced video-reference controls, multiple timed Aleph keyframes, non-MP4 delivery, upscaling, HDR and avatars remain outside these forms. This integration does not claim to expose every Runway API operation.

Tests use public schema excerpts and mocked transport, including refresh, cache restart/expiry, source failures and capability-specific request bodies. No paid Runway generation is needed for validation.