exoclaw-provider-openai

Direct-httpx OpenAI-compatible provider for exoclaw.

Uses httpx directly (no OpenAI SDK, no LiteLLM) so the request body can be streamed as the JSON is built rather than materialized as one contiguous string before the POST. On long conversations this drops peak per-turn RAM from ≈3× prompt-size (list + JSON dump + httpx buffer) to ≈1× streaming buffer — see exoclaw/docs/memory-model.md Step B.

Shape

from exoclaw_provider_openai import Deployment, OpenAIStreamingProvider

provider = OpenAIStreamingProvider(
    default_model="zai/glm-4.7",
    deployments={
        "zai/glm-4.7":           Deployment(base_url="https://openrouter.ai/api/v1", api_key=OPENROUTER_KEY),
        "minimax/minimax-m2.7":  Deployment(base_url="https://openrouter.ai/api/v1", api_key=OPENROUTER_KEY),
        "openai/gpt-5.4":        Deployment(base_url="https://api.openai.com/v1",    api_key=OPENAI_KEY),
        "zai-direct/glm-5.1":    Deployment(base_url="https://api.z.ai/api/coding/paas/v4", api_key=ZAI_KEY),
    },
    fallbacks={
        "zai/glm-4.7":   ["minimax/minimax-m2.7"],
        "zai/glm-5.1":   ["minimax/minimax-m2.7"],
    },
)

Each model name maps to exactly one deployment (base URL + API key + optional extra headers). fallbacks is a per-model list — when the primary raises a retryable error (429, 5xx, timeout) the provider walks the fallback list until one succeeds or all are exhausted.

Routing policy

• Each model → exactly one deployment. No load balancing across deployments for a single model.
• Fallbacks are strict per-model lists. No automatic cross-provider fallback.
• SSE response streaming is always on (the streaming-request-body path requires it as the codec; non-streaming responses aren't supported).
• TTFT timeout: if the first response byte doesn't arrive inside stream_ttft_timeout seconds, the request is abandoned and the fallback list is tried. Default 15 s.

What's intentionally skipped vs. LiteLLM

• Multi-provider cost tracking / usage normalization (OpenRouter returns the OpenAI shape directly; openclaw doesn't use cost tracking).
• Automatic Anthropic cache_control tagging. If your deployment needs it, stamp it on the messages before handing them to chat() — this provider is OpenAI-schema only.
• Deep parameter remapping across providers. Everything here is the OpenAI chat-completions schema; it's the caller's responsibility to send messages in that shape.