You are creating an importable template package for Creavax Template Builder.

CANONICAL LIVE CONTEXT
If web access is available, read the current Creavax rules, schema and the current visible admin-curated Template Builder models with their real capabilities first:
https://creavax.com/developers/templates/ai-context.json

The models exposed for Template Builder discovery come from the exact visible Admin Curated Models category memberships. Curated membership controls which real models are offered in the builder; it does not change a model's real capability contract. Keep template model slots portable and requirement-based; do not hard-code curated category IDs, provider IDs or model_key values inside template.models.

Your final answer must be one valid .creavax-template.json document. Do not create application code, React components, Laravel/PHP code, ZIP patches, explanations, or installation steps.

Required root structure:
{
  "schema_version": "1.0",
  "type": "creavax_template",
  "template": { ... }
}

Creavax workflow architecture:
INPUT -> AGENT
AGENT -> MODEL
AGENT -> AGENT
AGENT -> OUTPUT

The Model is an execution dependency of an Agent. The Model is NOT the producer node of the final Template Output.

Correct:
{ "from": "product_scene_creator", "to": "image_model" }
{ "from": "product_scene_creator", "to": "final_image" }

Forbidden:
{ "from": "image_model", "to": "final_image" }

Never connect a Model directly to an Output. Creavax may reject that graph with:
Output must be connected to an Agent, not directly to a Model.

INPUTS
- Every user-configurable value must be an Input.
- Use meaningful unique snake_case keys.
- Canonical buyer input types include text, select, number, image, video, audio, three_d and file.
- Select inputs should include a valid default and unique label/value options when appropriate. Numeric inputs use type "number" and may include a numeric default.

AGENTS
- Each Agent must have a unique key, name and clear system_prompt.
- The system_prompt should explain how to use inputs, what to preserve, what to generate, the expected result and important restrictions.
- Complex templates may use multiple Agents connected as a directed workflow. Agent -> Agent is supported.
- An upstream Agent's result becomes runtime input for the downstream Agent. Chains and branches are allowed, but Agent cycles/loops are not.
- Each Agent may use up to three Model slots.
- Each Model slot/node belongs to exactly one Agent. Never connect the same Model slot to multiple Agents. If two Agents need the same capability, create two separate Model slots; Creavax may resolve both slots to the same real provider model.
- Do not add extra Models only to fill all three slots. Every connected Model is an executable step and can add latency and credit cost.
- When an Agent genuinely benefits from multiple Model stages, connect them explicitly with target_port values chatModel1, chatModel2 and chatModel3 to preserve execution order.
- A Model slot may define its own instruction and prompt_composition. Use these for stage-specific work instead of forcing every Model in the Agent to share one generic prompt.
- prompt_composition is optional, but when provided it MUST be non-empty. {{prompt}} is optional: use it only to choose where structured connected text inputs are inserted. Without {{prompt}}, Creavax appends those structured inputs after the authored prompt text.
- Image/file inputs remain real media/reference bindings and must not be replaced by filenames in prompt text. If no custom final Prompt text is required, omit prompt_composition.
- A later Model in the same Agent may declare depends_on with one or more earlier Model keys. Creavax passes those earlier outputs into the later Model. A dependency must point to an earlier slot in the same Agent.
- Agent -> Agent is also a data connection: the downstream Agent can receive the upstream Agent's final result when its Model accepts that result type.
- Execution follows the Agent graph automatically: independent Agents may run in parallel; a downstream Agent waits only for its connected upstream Agent dependencies; mixed graphs combine parallel branches with sequential hand-offs.
- Do not hardcode provider-specific model names unless the user explicitly asks for one.

MODELS
- Template Builder model discovery uses the exact visible model memberships configured in Admin > Curated Models.
- Curated category membership controls discovery/availability only. It must never be used to invent or rewrite a model's inputs, outputs, operations, features or runtime controls.
- Define a Model slot by capability requirements, not by curated category, provider ID or model_key.
- Describe only the inputs, outputs, operations and optional features genuinely required by the Agent.
- Provider request fields are not automatically semantic input modalities. A prompt/instruction field, JSON/config value, reference URL, task ID, mask, bounding box or provider-specific control must not create a new semantic input or generation operation unless that endpoint independently supports it.
- The live AI Context lists only the currently visible admin-curated models available to Template Builder, together with their real capability contracts. Use that list to understand what can currently be resolved; do not bind the portable JSON to a specific listed model.
- If a preferred or referenced real model is unavailable or unregistered, Creavax may replace it only with an active model that satisfies the complete inputs + outputs + operations + features + required runtime-controls intersection.
- Example for image editing:
  "requirements": {
    "inputs": ["image", "text"],
    "outputs": ["image"],
    "operations": ["image_edit"],
    "features": []
  }

MULTI-MODEL AGENTS
- Use multiple Models only when each slot has a distinct job. Connected Model slots execute in chatModel1 -> chatModel2 -> chatModel3 order.
- Give each Model a stage-specific instruction. prompt_composition is optional; if defined it must be non-empty. {{prompt}} is only an optional placement token for connected structured text inputs.
- Use depends_on to pass an earlier Model result into a later Model in the same Agent.
- Example:
  "models": [
    {
      "key": "draft_image_model",
      "label": "Draft Image Model",
      "requirements": { "inputs": ["image", "text"], "outputs": ["image"], "operations": ["image_edit"], "features": [] },
      "instruction": "Create the first advertising composition."
    },
    {
      "key": "refine_image_model",
      "label": "Refine Image Model",
      "requirements": { "inputs": ["image", "text"], "outputs": ["image"], "operations": ["image_edit"], "features": [] },
      "instruction": "Refine the previous result while preserving the product exactly.",
      "depends_on": ["draft_image_model"]
    }
  ]
  "connections": [
    { "from": "product_agent", "to": "draft_image_model", "target_port": "chatModel1" },
    { "from": "product_agent", "to": "refine_image_model", "target_port": "chatModel2" }
  ]

RUNTIME MODEL SETTINGS
- When a user-facing Input must control a real model setting, use model.setting_bindings.
- Model compatibility is strict intersection: every required input, output, operation, feature and bound runtime setting must be supported by the resolved real model.
- Every setting_bindings[].setting is promoted to a required runtime control before model selection; candidates missing that control are excluded before execution/provider invocation.
- This is intended for portable runtime controls such as video duration, aspect ratio, resolution, quality, FPS or output count.
- Every setting_binding must reference an existing Input whose schema type is valid for that setting. The setting must be a canonical supported runtime control, and the binding must not conflict with the Model slot requirements.
- Use only Input types supported by the live Creavax Template schema. The canonical numeric buyer input type is "number"; use it for numeric values and setting bindings.
- Use cast to convert the selected value to the real runtime type: string, integer, number or boolean.
- Example for dynamic video duration:
  {
    "key": "video_model",
    "label": "Compatible Video Model",
    "requirements": {
      "inputs": ["image", "text"],
      "outputs": ["video"],
      "operations": ["video_generation"],
      "features": []
    },
    "setting_bindings": [
      { "setting": "duration_seconds", "input": "video_duration", "cast": "integer" }
    ]
  }
- The runtime value is applied before model routing and pricing, so a longer duration can produce a higher credit cost when the selected provider model is billed by duration.
- Never bind Inputs to provider identity, model identity, billing internals, workflow ids or other protected settings. Creavax rejects unsafe bindings.

IMAGE / REFERENCE INPUTS
- Connect the image Input to the Agent.
- Creavax runtime can pass an Agent's connected image/reference Inputs to that Agent's Model.
- Do not add a visible Image -> Model connection merely to provide the reference image.

OUTPUTS
- Outputs represent final workflow results.
- Every final Output must be connected from the Agent that produces it, never directly from its Model.

CONNECTIONS
Before returning JSON, verify:
1. Every from/to key exists.
2. Every required Input is connected to the Agent that uses it.
3. Every Model is connected from exactly one Agent that owns/uses it. A Model slot must never be shared by multiple Agents.
4. Agent -> Agent connections may be used for multi-step workflows; they create both data and execution dependencies.
5. Every final Output is connected from an Agent.
6. No Model -> Output connection exists.
7. No Agent cycle/loop exists.
8. No duplicate or meaningless connections exist.
9. All keys are unique.
10. Every defined prompt_composition is non-empty. {{prompt}} is optional; without it, Creavax appends structured connected text inputs automatically. Media inputs remain separate runtime bindings.

COMMERCE
- seller_markup is supported in template JSON and represents the creator markup percentage.
- Example: "seller_markup": 25

TRANSLATIONS AND SEO
- Include useful English copy and Persian and Arabic translations when appropriate.
- Use template.translations.fa and template.translations.ar for localized title/description.
- Keep internal machine keys untranslated and stable.
- SEO supports title, description and keywords.
- SEO keywords are limited to a maximum of 10 unique items.
- Example: "seo": { "title": "...", "description": "...", "keywords": ["..."] }

DO NOT
- invent or change Creavax Model Capability rules;
- create Curated Mapping or Classification rules;
- create source code or a ZIP;
- return Markdown commentary around the JSON.

FINAL CHECK
- schema_version is "1.0"
- type is "creavax_template"
- template exists
- JSON syntax is valid
- no comments inside JSON
- no trailing commas
- all connection endpoints exist
- INPUT -> AGENT
- AGENT -> MODEL
- every MODEL slot has exactly one AGENT owner
- AGENT -> AGENT when a multi-step workflow is needed
- independent AGENTS may run in parallel; dependent AGENTS wait for upstream AGENTS
- AGENT -> OUTPUT
- no Agent cycles
- never MODEL -> OUTPUT
- no more than 10 SEO keywords
- provider request fields do not create semantic input modalities or operations unless the endpoint independently supports them
- unavailable/unregistered preferred models may be replaced only by a model satisfying the full inputs + outputs + operations + features + required runtime-controls intersection
- every setting_bindings entry references an existing compatible Input, uses a canonical supported runtime control and does not conflict with Model requirements
- duration/resolution/etc. controls use setting_bindings so pricing and provider settings receive the real runtime value

Return only the final valid JSON, ready to save as:
template-name.creavax-template.json