Validates the request body — ensures imagePath, audioPath, resumeText (strings), and durationSeconds (number) are all provided.

Makes an authenticated request to the central Arcade API to fetch the user's current credit balance.

Calculates dynamic credit cost (50 base + 5 per 5s increment), checks balance, generates a job ID, computes target word count for script length, and builds public serve URLs for the uploaded audio and image files.

Sends the resume text to Claude Haiku with strict word-count constraints based on the selected duration (~2.5 words/second). Outputs only the spoken words — no markdown or labels.

Sends the voice sample to MiniMax Voice Cloning, which analyzes the audio and creates a custom voice model. Returns a voice_id for use with speech generation. Noise reduction and volume normalization are enabled.

Sends the script text and cloned voice_id to MiniMax Speech 2.8 Turbo, which generates natural-sounding speech audio in the cloned voice.

Creates a job record with status "processing", the generated script, and job ID. The frontend polls this record for completion.

Calls the Arcade transaction API to deduct the calculated credit cost from the user's balance.

Assembles the JSON response object with jobId, script, status, and remaining credits. Uses a function handler to avoid template escaping issues with script content.

Returns the job ID, generated script, status "processing", and updated credit balance. The client then polls GET /api/status for completion.