docs(skill): input-limit 400s and a Seedance 2.5 edit/extend recipe

SKILL.md

@@ -61,6 +61,7 @@ Each page is short and stands alone. Open one when its row applies.
 | 小红书选题与趋势 (keyword expansion → note search → comments → summary) | <https://beatapi.io/skill-refs/recipes/xiaohongshu-topic-research.md> |
 | Competitor accounts on 抖音 / TikTok / 小红书 | <https://beatapi.io/skill-refs/recipes/competitor-accounts.md> |
 | N 选 1 decisions with JEV | <https://beatapi.io/skill-refs/recipes/decide-with-jev.md> |
+| Edit or extend a video with Seedance 2.5 (`task_type`) | <https://beatapi.io/skill-refs/recipes/seedance-video-edit-and-drafts.md> |
 
 Direct APIs for developers (`/v1/responses`, `/v1/systemone`, OpenAPI):
 <https://docs.beatapi.io/>. Source: <https://github.com/BeatAPI/beatapi-skill>.

references/errors.md

@@ -13,7 +13,8 @@ user; when `true`, the same call may succeed later. A failed task carries the sa
 | 401 `missing_api_key` | The request had no key: add the `Authorization` header. |
 | 401 `invalid_api_key` | The key was rejected: ask the user to check it in their secure settings. Do not retry. |
 | 402 / `insufficient_credits` | Balance too low: send the user to <https://beatapi.io/dashboard/billing>. |
-| 400 `bad_request` | Inspect again and fix the named field. Not retryable. |
+| 400 `bad_request` | Fix the named field. If the message lists the values the model takes (e.g. `available: 16:9, 9:16`), use one of them. Not retryable: never resend unchanged. |
+| 400 naming an input file (e.g. `reference_videos[1] runs at 15 frames per second; this model takes 24 to 60.`) | The file is outside the model's limits. Send a file within the stated limit. Not retryable. |
 | 404 `not_found` | Use `suggestions`, or `GET /v1/models` for text models. Not retryable; do not loop. |
 | 429 | Wait `error.retry_after_seconds` (or `Retry-After`; MCP sees only the body). Free keys are rate limited before first top-up; batch calls. |
 | 5xx on a sync call (`retryable:true`) | It failed and was not charged. Retry once, then tell the user or try another capability. |

references/recipes/seedance-video-edit-and-drafts.md

@@ -0,0 +1,43 @@
+# Recipe: edit or extend a video with Seedance 2.5
+
+Seedance 2.5 makes a new video from references, edits one of them, or extends one. Inspect `model:seedance-2.5` before a paid run: its input schema and price are live. Full reference: <https://docs.beatapi.io/video-api/seedance-2-5>.
+
+Poll any task with the status call from the Run page (<https://beatapi.io/skill-refs/run.md>):
+
+```json
+{"reference":"model:seedance-2.5","operation":"status","task_id":"<id>"}
+```
+
+Repeat it every 8 seconds until `succeeded` or `failed`. The video is `data.output.media[0].url`.
+
+## Edit or extend a reference video
+
+Set `task_type` to say what `reference_videos` are for:
+
+- `reference`: makes a new video from them.
+- `edit`: edits one of them and keeps its length and aspect ratio. Leave `duration` and `aspect_ratio` out. The source must be 4 to 30 s.
+- `extend`: continues one. `duration` is the total length of the finished video, source included: a 5 s source extended by 5 s takes `duration: 10`.
+
+Without `task_type` the model decides from the prompt, and a prompt it reads as an edit fails if `duration` or `aspect_ratio` was set. Name the type.
+
+Refer to inputs by position: `@image1`, `@video1`, `@audio1`. For a new video from references, avoid "replace", "remove" or "extend" in the prompt, or set `task_type` to `reference`. For an edit, say exactly what changes and what stays.
+
+Edit:
+
+```json
+{"reference":"model:seedance-2.5","input":{"model":"seedance-2.5","task_type":"edit","prompt":"Make the jacket in @video1 red. Keep the person, the background and the camera move unchanged.","reference_videos":["https://example.com/clip.mp4"]}}
+```
+
+Extend a 5 s source to 10 s:
+
+```json
+{"reference":"model:seedance-2.5","input":{"model":"seedance-2.5","task_type":"extend","prompt":"@video1 continues: the car turns onto the coast road at sunset.","reference_videos":["https://example.com/clip-5s.mp4"],"duration":10}}
+```
+
+## Output format
+
+`output_format` is `mp4` (default) or `mov`, at the same price. Draft mode (`draft`) is not offered yet.
+
+## Inputs are checked first
+
+Before a task is created: images 300 to 6000 px per side, aspect 0.4 to 2.5, at most 30 MB; reference videos 2 to 30 s each, at most 30 s in total, 24 to 60 fps, at most 200 MB; audio 2 to 30 s each, at most 30 s in total, at most 15 MB. A file outside these limits gets a 400 that names the field, the index and the limit. Fix the file; never resend it unchanged.

references/run.md

@@ -19,6 +19,7 @@ Over MCP, pass the same JSON to `capabilities_run`.
   OpenAI's API does. Send a unique `idempotency_key` per task as a top-level field of
   the Run body, next to `reference` and `input` (or as an `Idempotency-Key`
   header), and reuse it only to retry the same task.
+- Reference media (images, videos, audio) is checked before a video task is created. A file outside the model's limits, or a reference image over 50 MB, gets a 400 that names the field, the index and the limit, and no task is created.
 - **Sync** capabilities (social data, web search/read/map, text models, JEV)
   return the result. Send `"view":"preview"` for data: when the result has a
   list, it is in `items` (the first `max_items`, default 10, up to 50, each

skills/beatapi-video/references/current.md

@@ -56,6 +56,7 @@ Each page is short and stands alone. Open one when its row applies.
 | 小红书选题与趋势 (keyword expansion → note search → comments → summary) | <https://beatapi.io/skill-refs/recipes/xiaohongshu-topic-research.md> |
 | Competitor accounts on 抖音 / TikTok / 小红书 | <https://beatapi.io/skill-refs/recipes/competitor-accounts.md> |
 | N 选 1 decisions with JEV | <https://beatapi.io/skill-refs/recipes/decide-with-jev.md> |
+| Edit or extend a video with Seedance 2.5 (`task_type`) | <https://beatapi.io/skill-refs/recipes/seedance-video-edit-and-drafts.md> |
 
 Direct APIs for developers (`/v1/responses`, `/v1/systemone`, OpenAPI):
 <https://docs.beatapi.io/>. Source: <https://github.com/BeatAPI/beatapi-skill>.

skills/beatapi-video/references/errors.md

@@ -13,7 +13,8 @@ user; when `true`, the same call may succeed later. A failed task carries the sa
 | 401 `missing_api_key` | The request had no key: add the `Authorization` header. |
 | 401 `invalid_api_key` | The key was rejected: ask the user to check it in their secure settings. Do not retry. |
 | 402 / `insufficient_credits` | Balance too low: send the user to <https://beatapi.io/dashboard/billing>. |
-| 400 `bad_request` | Inspect again and fix the named field. Not retryable. |
+| 400 `bad_request` | Fix the named field. If the message lists the values the model takes (e.g. `available: 16:9, 9:16`), use one of them. Not retryable: never resend unchanged. |
+| 400 naming an input file (e.g. `reference_videos[1] runs at 15 frames per second; this model takes 24 to 60.`) | The file is outside the model's limits. Send a file within the stated limit. Not retryable. |
 | 404 `not_found` | Use `suggestions`, or `GET /v1/models` for text models. Not retryable; do not loop. |
 | 429 | Wait `error.retry_after_seconds` (or `Retry-After`; MCP sees only the body). Free keys are rate limited before first top-up; batch calls. |
 | 5xx on a sync call (`retryable:true`) | It failed and was not charged. Retry once, then tell the user or try another capability. |

skills/beatapi-video/references/recipes/seedance-video-edit-and-drafts.md

@@ -0,0 +1,43 @@
+# Recipe: edit or extend a video with Seedance 2.5
+
+Seedance 2.5 makes a new video from references, edits one of them, or extends one. Inspect `model:seedance-2.5` before a paid run: its input schema and price are live. Full reference: <https://docs.beatapi.io/video-api/seedance-2-5>.
+
+Poll any task with the status call from the Run page (<https://beatapi.io/skill-refs/run.md>):
+
+```json
+{"reference":"model:seedance-2.5","operation":"status","task_id":"<id>"}
+```
+
+Repeat it every 8 seconds until `succeeded` or `failed`. The video is `data.output.media[0].url`.
+
+## Edit or extend a reference video
+
+Set `task_type` to say what `reference_videos` are for:
+
+- `reference`: makes a new video from them.
+- `edit`: edits one of them and keeps its length and aspect ratio. Leave `duration` and `aspect_ratio` out. The source must be 4 to 30 s.
+- `extend`: continues one. `duration` is the total length of the finished video, source included: a 5 s source extended by 5 s takes `duration: 10`.
+
+Without `task_type` the model decides from the prompt, and a prompt it reads as an edit fails if `duration` or `aspect_ratio` was set. Name the type.
+
+Refer to inputs by position: `@image1`, `@video1`, `@audio1`. For a new video from references, avoid "replace", "remove" or "extend" in the prompt, or set `task_type` to `reference`. For an edit, say exactly what changes and what stays.
+
+Edit:
+
+```json
+{"reference":"model:seedance-2.5","input":{"model":"seedance-2.5","task_type":"edit","prompt":"Make the jacket in @video1 red. Keep the person, the background and the camera move unchanged.","reference_videos":["https://example.com/clip.mp4"]}}
+```
+
+Extend a 5 s source to 10 s:
+
+```json
+{"reference":"model:seedance-2.5","input":{"model":"seedance-2.5","task_type":"extend","prompt":"@video1 continues: the car turns onto the coast road at sunset.","reference_videos":["https://example.com/clip-5s.mp4"],"duration":10}}
+```
+
+## Output format
+
+`output_format` is `mp4` (default) or `mov`, at the same price. Draft mode (`draft`) is not offered yet.
+
+## Inputs are checked first
+
+Before a task is created: images 300 to 6000 px per side, aspect 0.4 to 2.5, at most 30 MB; reference videos 2 to 30 s each, at most 30 s in total, 24 to 60 fps, at most 200 MB; audio 2 to 30 s each, at most 30 s in total, at most 15 MB. A file outside these limits gets a 400 that names the field, the index and the limit. Fix the file; never resend it unchanged.

skills/beatapi-video/references/run.md

@@ -19,6 +19,7 @@ Over MCP, pass the same JSON to `capabilities_run`.
   OpenAI's API does. Send a unique `idempotency_key` per task as a top-level field of
   the Run body, next to `reference` and `input` (or as an `Idempotency-Key`
   header), and reuse it only to retry the same task.
+- Reference media (images, videos, audio) is checked before a video task is created. A file outside the model's limits, or a reference image over 50 MB, gets a 400 that names the field, the index and the limit, and no task is created.
 - **Sync** capabilities (social data, web search/read/map, text models, JEV)
   return the result. Send `"view":"preview"` for data: when the result has a
   list, it is in `items` (the first `max_items`, default 10, up to 50, each