mnfst/llm-gateway manifest@6.27.0

manifest v6.27.0

on GitHub

✨ Minor Changes

• 9c773fc: Custom providers take several API keys, each a separate connection managed like a native provider's keys.

🐛 Patch Changes

• 60e0365: Forward tool_choice and parallel_tool_calls to Anthropic when translating Chat Completions or Responses requests. They were dropped, so a forced or disabled tool call was left to the model's own choice.
• be73101: Route Bedrock GPT-6 models (and later numbered GPT families) through the namespaced OpenAI Responses API path, which Bedrock requires for them.
• ac51ed8: Keep system prompts sent as content-part arrays when routing to Google Gemini. They were dropped from systemInstruction, so Gemini answered without the system prompt (including Responses API developer instructions).
• cbb96fa: Send developer messages to Gemini as the system instruction instead of a user turn.
• 74d8d04: Report a truncated or filtered non-streaming Responses API reply as finish_reason: "length" / "content_filter" instead of "stop", matching the streaming path.
• 20208e5: Keep cached-input and reasoning token counts on /v1/responses replies served by OpenAI-compatible Chat Completions upstreams, so clients see them and Manifest records and prices cache reads correctly.