Render a video clip from YOUR prompt on the model YOU choose, in one of three MODES that follow from the fields you pass: TEXT (prompt only), IMAGE-TO-VIDEO (first_frame: an https still that becomes frame one, optionally last_frame to end on; the clip animates the still) or REFERENCE (refs: https images such as a portrait, a character sheet from list_characters or a product photo whose identity/look is kept; video_refs: clips whose motion or framing is followed; audio_refs: a voice or sound to carry; address them in the prompt as @image1, @video1, @audio1). Frames and refs cannot be mixed. Write the shot like a director: who is in frame, where, what happens, camera, and the exact spoken words in quotes if anyone talks. HEBREW: the video models cannot speak Hebrew, so Hebrew words in quotes are never left to them: Gemini speaks the line and OmniHuman 1.5 lip-syncs a front-facing still to it (your first_frame, or a keyframe made from the prompt and refs). 9:16 or 1:1, up to 35 s, no video_refs or audio_refs; priced as the voice + the keyframe + 26 credits per lip-sync second instead of the model's rate (call quote); voiceover sets the voice (sulafat, achird) or the gender. Models: seedance-2.0, seedance-2.5, wan-3.0-prime (default seedance-2.5; call list_models for each model's modes, limits, prices and recent results; or pass model:"auto"). Aspect: 9:16 default, also 16:9, 1:1, 4:3, 3:4, 3:2, 2:3, 21:9, adaptive where the selected model supports it. Quality: 480p, 720p, 1080p, or 4k where the selected model supports it; omit for the selected model's default. The price per second follows the model and quality (seedance-2.5: 60 credits/s at 480p, 125 credits/s at 720p, 225 credits/s at 1080p; call list_models for the full table). Reference clip seconds are billed like output seconds. Call quote first if the user cares about cost. Returns a job id, then call get_run_status until it is done and hand the user the URL.

When the user wants to provide a product photo, portrait, screenshot, or image reference, check tools/list. If open_upload_panel is available, offer it first: call open_upload_panel with {}, show its inline panel or returned browser link, and wait for the user to finish. Then call get_uploads with the returned session_id and use the exact ready image_url in refs, first_frame, last_frame, or the matching skill image field. Do not start a reference-dependent generation until the required images are ready. The panel accepts still PNG, JPEG, and WebP, up to 10 images and 25 MB each. Images expire 24 hours after the panel is created; show the returned expiry, request a new upload if expired, and never make a permanent copy. Uploading uses no generation credits. Do not ask the user for base64, shell commands, or a public hosting service; do not continuously poll while waiting for them. Only call these tool names when discovered. If they are missing from the connected tool catalog, call upload_image with {} to open the same panel; use the returned panel: upload_key with upload_image to retrieve images after the user finishes. If that reports the panel is unavailable, use upload_image only for an already accessible file or URL. Do not build an upload page or request a local folder as a substitute for the existing panel. Never invent access to a chat attachment: if you cannot read it, explain that and offer the panel when available. Browser fallback works without inline UI support; do not promise an inline panel in every client. Inspect the native image previews returned with retrieved uploads before describing them or writing image-specific prompts. Never infer what the image shows from an email domain, account metadata, or filename. Use the original URLs for generation, not preview bytes. If the session ID was lost, call get_uploads with {} to list recent sessions, or upload_image with {"upload_key":"panel:recent"} for cached catalogs; retrieve the matching session before asking for re-upload. Account-wide recent uploads may belong to other conversations, so clarify ambiguous selection. Uploading stores an image; it does not attach it to a generation automatically. If the user already requested a generation with these images, continue that request using the returned URLs; do not stop at "upload complete" or ask again whether to use them. If make_ugc fits the requested outcome, pass a person photo as image or a product photo as product_image, then call quote_ugc before spending. Otherwise compose the direct generators: for generate_image, put the relevant image_url values in refs; for generate_video, use refs for product/person/appearance references, or first_frame for animating a still (last_frame only for an explicitly requested ending frame), and never combine refs with frame fields. Preserve the user's intended image roles and model limits; ask only if the role or selection is unclear. Include the URL in the tool arguments, not only in the prompt. Do not substitute a newly generated image for the uploaded reference. If the request was upload-only, report readiness and wait for a generation request; uploading alone does not authorize spending credits. After submitting, poll get_run_status and return the result.

IMAGES: pass an https URL. If you only have raw bytes or a data: URL, call upload_image FIRST and pass the URL it returns. Never paste base64 into these arguments, the client prints tool arguments in the conversation, so a base64 image becomes a wall of text for the user and is re-sent on every retry.

RETRY SAFETY: set request_id to a fresh UUID for a new generation. Reuse the exact same request_id and inputs only to recover a lost response; that returns the original job without another charge. A deliberately new generation needs a new request_id. Keep the returned job_id and poll get_run_status. Never invent a refund.