Openrouter Research

Capture:

• Base URL: https://openrouter.ai/api/v1
• Auth header format (Authorization: Bearer vs API-Key)
• Required request headers: HTTP-Referer, X-Title
• Chat completions endpoint: POST /v1/chat/completions
• Request body schema: model, messages, temperature, max_tokens, stream
• Response schema: choices[0].message.content
• Error response format and status codes (429, 402, 503, 500)

Step 2: Available Grok Models

1. Fetch: https://openrouter.ai/models?q=grok
2. Search: OpenRouter xAI Grok models list 2025
3. Fetch: https://openrouter.ai/x-ai if available

For each available Grok model, capture:

• Full model ID string (e.g. x-ai/grok-beta, x-ai/grok-vision-beta)
• Context window size (tokens)
• Supports image input (vision)? — critical for judge_service
• Cost per 1M input/output tokens

Determine the best model for each use case:

• prompt_gen: text-only, large context
• prompt_revise: text-only
• judge: MUST support vision/image input

Step 3: Image Input Format for Vision Models

Since judge_service.py sends a generated image to Grok for evaluation:

1. Search: OpenRouter vision model image input base64 format 2025
2. Fetch: https://openrouter.ai/docs/features/vision
3. Search: OpenAI-compatible vision API image_url content type format

Capture:

• Message content format for image:

  {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,..."}}
  

  vs URL-based image reference
• Max image size limits
• Supported formats (JPEG, PNG, WebP)
• Whether OpenRouter passes base64 to xAI directly or requires URL

Step 4: Error Handling and Retry Strategy

1. Search: OpenRouter API error handling retry 429 503 2025
2. Fetch: https://openrouter.ai/docs/api-reference/errors if available

Capture:

• 429 rate limit: retry-after header? fixed backoff interval?
• 402 payment/balance: raise immediately, no retry
• 503 model unavailable: retry with backoff
• Recommended timeout values for text calls vs vision calls
• Max retries pattern

Output Format

Section 1: openrouter_tool.py Implementation Blueprint

# Async httpx client structure
BASE_URL = "https://openrouter.ai/api/v1"

# Required headers
headers = {
    "Authorization": f"Bearer {api_key}",
    "HTTP-Referer": "...",
    "X-Title": "SFUMATO",
    "Content-Type": "application/json",
}

# chat_completion(model, messages, temperature, max_tokens) -> str
# vision_completion(model, text_messages_plus_image_message) -> str
# Retry pattern for 429/503 (show how many retries, backoff)
# Error handling: which codes to retry vs raise immediately

Section 2: Model Selection Table

Section 3: Image Message Format

Exact Python dict structure to use when sending image to judge:

{
    "role": "user",
    "content": [
        {"type": "text", "text": "...judge prompt..."},
        {"type": "image_url", "image_url": {"url": "data:image/jpeg;base64,..."}}
    ]
}

Section 4: Recommended Constants for config.py

OPENROUTER_BASE_URL = "https://openrouter.ai/api/v1"
OPENROUTER_MODEL_PROMPT_GEN = "x-ai/grok-..."
OPENROUTER_MODEL_PROMPT_REVISE = "x-ai/grok-..."
OPENROUTER_MODEL_JUDGE = "x-ai/grok-vision-..."  # must support vision
OPENROUTER_TIMEOUT_TEXT = 30
OPENROUTER_TIMEOUT_VISION = 45
OPENROUTER_MAX_RETRIES = 2

Notes

• Never log API key, response.text for large vision calls, or base64 image data
• Image for judge: read from data/sessions/<id>/iter_<n>.jpg, encode as base64
• The judge call is the only one that uses vision; wrap it separately in vision_completion()
• HTTP-Referer should be "http://localhost:5000" for local dev