---
name: provider-model-refresh
description: Use the September 2026 image, video, speech and Avatar V adapters with explicit model/host contracts.
title: provider-model-refresh
canonical_url: https://skilld.dev/gh/openmontageinc/openmontage/provider-model-refresh
last_updated: 2026-11-12T11:03:20.000Z
---

> Index of every page on this site: [llms.txt](https://skilld.dev/llms.txt)

> **Skill from skilld.dev.** Follow the user's requested action.
> If the user asked to fork this Skill, follow the fork workflow below. Do not execute the Skill instructions while copying it.
> Otherwise, follow the instructions below for this session. You do not need to install anything.
>
> If the user asked to install this Skill, run `npx skilld install openmontageinc/openmontage/provider-model-refresh`. Install writes the Skill files into the project, so every session loads them.
>
> ## Fork workflow
>
> A fork creates an editable local Skill with its original author and licence. The request authorizes copying and local installation.
> 1. Check `./skills/provider-model-refresh`, the project lockfile, and selected Agent targets together. If the local directory or installed Skill exists, stop. Never overwrite an existing directory or Agent target.
> 2. Read [source metadata](https://skilld.dev/api/v1/skills/openmontageinc/openmontage/provider-model-refresh) once. Use sourceUrl, sourceCommit, skillPath, sourceGone, and license. If the source is gone or its path is missing, stop. If license is null, read licence files at the source commit.
> 3. Fetch only the source commit into a temporary Git repository. Do not clone full history. Derive repository_url from sourceUrl, including repository renames. If sourceCommit is absent, resolve the sourceUrl ref once. Set source_commit to that actual commit. Run these commands in one shell call:
>
> ```sh
> git init --quiet "$temporary_dir"
> git -C "$temporary_dir" fetch --quiet --depth=1 "$repository_url" "$source_commit"
> git -C "$temporary_dir" checkout --quiet --detach FETCH_HEAD
> ```
>
> Read applicable licence declarations and notices at that commit. If copying is not permitted, report the restriction and stop.
> 4. Inspect source entries together, then copy the directory containing skillPath into `./skills/provider-model-refresh`. Use the user's path if selected. Keep the original SKILL.md, relative links, scripts, binary assets, and executable modes. Exclude .git metadata. Reject symlinks and paths outside the Skill directory. After checking entries, use cp -a where available. A regular source directory needs no custom copy script. Do not save this page wrapper as SKILL.md.
> Preserve author credit, notices, and applicable licence files from repository or parent directories. Add PROVENANCE.md with the Skill page, source URL, actual commit, original path, and licence. Retain any existing PROVENANCE.md and record new provenance separately. Batch source inspection, copying, and provenance work where practical.
> 5. In the project root, run `skilld install ./skills/provider-model-refresh --mode copy --plain`. If skilld is unavailable, use `npx skilld install ./skills/provider-model-refresh --mode copy --plain`. This known command needs no help lookup. Install does not support --json. Use detected Agent targets, or add --agent for the targets the user selected. Install the local path, never the upstream selector. If installation fails, preserve the local copy and report the exact failure.
> 6. Confirm the local lockfile source and installed Agent copies once. Report the local path, actual commit, and Agent targets. After edits, reinstall the same local path. Upstream updates must not replace it. Do not publish or push unless the user asks.

# Provider model refresh

For media production, first follow AGENT_GUIDE.md and the selected pipeline.
These adapters are discoverable through the normal BaseTool registry. Inspect
their input schema and status before making a request. An API credential indicates
configuration, not verified account entitlement. No paid live test is implied.

## Routing

Use `preferred_tool` to require an exact adapter and `hosting_provider` to require
an API host. `preferred_provider` remains a legacy ranking preference. Select an
exact `model` for images/video or `model_id` for speech. An unavailable exact
model must fail rather than silently substitute a different model.

| Model | Direct tool | Aggregator tools |
| --- | --- | --- |
| GPT Image 2.5 Flare / Sunburst | openai_image | openai_fal_image, atlas_refresh_image |
| Gemini 3 Pro Image / Nano Banana Pro | google_imagen | gemini_fal_image, gemini_replicate_image, atlas_refresh_image |
| Ideogram 4.5 / Precise Edit | ideogram_image | ideogram_fal_image |
| Wan 3.0 | — | wan_fal_video, wan_atlas_video, wan_replicate_video |
| H3 Max | — | h3_max_video (fal) |
| Kling O3 video (Standard / Pro / 4K) | kling_official_video is separate | kling_o3_fal_video |
| Kling Image O3 | kling_official_image is separate | kling_o3_fal_image |
| LTX-2.5 Fast / Pro | ltx_api_video | — |
| Eleven v4 / v4 Turbo | elevenlabs_tts | fal_elevenlabs_tts (v4 only) |
| Gemini 3.8 Flash / Flash-Lite TTS | gemini_tts | — |
| Lyria 3.5 | google_music | — |
| Sonic 3.6 | cartesia_tts | — |
| Inworld TTS-2 / Flash | inworld_tts | — |
| Qwen Image 2.1 | qwen_image_21 (local) | — |
| HeyGen Avatar V | heygen_avatar | — |

## Image requests

GPT Image 2.5 uses separate `gpt-image-2.5-flare` and
`gpt-image-2.5-sunburst` IDs. Use Flare for speed and Sunburst for demanding
edits. Describe what changes and what must remain. Direct edits use local
`image_path`/`image_paths` and optional `mask_path`. Keep alpha output in PNG or
WebP. The old GPT Image 2 default remains available for compatibility.

Google Pro's exact ID is `gemini-3-pro-image`; it is distinct from
`gemini-3.1-flash-image` (Nano Banana 2). Use up to 14 reference images, and
1K/2K/4K on supported routes. Replicate fallback is disabled.

Ideogram direct `generation_mode=precise_edit` preserves unaffected pixels;
use `image_path` followed by up to four references in `image_paths`. With a
mask, only three references are allowed. Black edits, white preserves. Precise
Edit cannot specify output size. `dry_run=true` validates and prices the exact
request without generating. On fal, `precise_edit` sets `edit_precision=high`.

Qwen Image 2.1 requires the actual local model directory and compatible
Diffusers. It never downloads weights implicitly. Check the Qwen Research
License for the intended use. Transparent generation validates the returned
alpha channel rather than adding an opaque alpha channel.

## Video and jobs

`operation` selects text_to_video, image_to_video or reference_to_video.
Replicate Wan supports text and first-frame image only. Wan on Atlas uses
`refers` objects and `duration=-1` for automatic length; fal uses reference URL
arrays and nullable duration. Do not transfer native parameters between hosts.
Pass route-specific controls in `provider_params`; checked-in schemas under
tools/provider_contracts validate the complete request before submission.

LTX-2.5 supports text/image/audio generation, not retake or extend. Frame rates
are 24, 25, 48, 50; long durations require Fast at 720p/1080p and 24/25 FPS.
`duration=null` lets the model choose length, but cannot combine with a last
frame. H3 Max is separate from the older H3 endpoint.

Kling O3 on fal exposes five video modes on each of `kling-o3-standard`,
`kling-o3-pro`, and `kling-o3-4k`: `text_to_video`, `image_to_video`,
`reference_to_video`, `video_to_video_reference`, `video_to_video_edit`.
The separate `kling-o3-image` supports `text_to_image` and `image_to_image`.
Select `preferred_tool=kling_o3_fal_video` (or `kling_o3_fal_image`) and an
exact model/operation; use `FAL_KEY`, not the official Kling credentials.
For `reference_to_video`, provide `start_image_url` for the first frame and
`image_urls`/`image_paths` for `@Image1` identity references. `image_url`
belongs to `image_to_video` (and its local `image_path` alternative); it is
rejected on the reference route to avoid duplicate references. For V2V,
specify edit or reference explicitly and pass `video_url`; edit has no
`duration` or `aspect_ratio`. Never auto-select O3 for a legacy Kling request.

Set `job_path` to persist a checkpoint. On timeout, pass the returned
`resume_job` to the same tool; this polls the existing job without a second
paid submission. A submission timeout without an ID is ambiguous: check the
provider dashboard before submitting again. Preserve provenance and outputs.
Unknown pricing is reported explicitly, never treated as a free service.

## Speech and music

Direct Eleven IDs are `eleven_v4` / `eleven_v4_turbo`; fal uses `eleven-v4`.
fal v4 accepts at most 5,000 characters and has no speed/style control.
Direct v4 accepts up to 10,000. Voice catalogs and rights differ by host.
Gemini TTS uses Interactions with speech annotations; configure each dialogue
speaker's voice and turn text separately. Its current adapter outputs WAV.
Cartesia uses API version 2026-08-14 and WAV. Inworld uses a portal-issued
Base64 credential with Basic auth, MP3, and at most 2,000 UTF-16 code units.
Lyria duration is a prompt target, not guaranteed exact timing.

## Avatar V

Use `heygen_avatar` in the avatar-spokesperson pipeline. First `list_looks` or
`inspect_look`; generation checks `avatar_type=digital_twin` and
`supported_api_engines` before POST. Avatar V is explicitly requested with
`engine=avatar_v`; it never falls back to IV. Provide exactly one of
`script` (+ `voice_id`), `audio_url`, or `audio_asset_id`. A reference look must
be a digital twin in the same verified group. Use an authorized existing
presenter and preserve HeyGen's consent/account requirements.

## Evidence and limitations

The original refresh contracts were inspected 2026-10-03; each schema records
its official source. Kling O3 fal video/image schemas were refreshed 2026-10-25.
See docs/provider-update-plan-2026-10-03.md for release research. Public Kling 4
generation endpoints remain unverified, so do not invent a model ID. Live voice
agents require a separate streaming/session integration beyond batch narration.

## October 23 follow-up

For current Google retirement handling, Eleven v4 transports, and newly supported
providers, read [provider-refresh-oct23](../provider-refresh-oct23/SKILL.md).
The v4 narration HTTP payload is dialogue inputs/settings, not legacy TTS
text/voice_settings; v4 Turbo is WebSocket-only in the direct file adapter.
