mnfst/llm-gateway manifest@6.28.7

manifest v6.28.7

on GitHub

🐛 Patch Changes

• 52a921c: Surface Claude thinking text as reasoning_content when an Anthropic Messages reply is translated to a Chat Completions reply, both streaming and non-streaming.
• f93cde5: Report Anthropic refusal stops as content_filter and model_context_window_exceeded stops as length on chat completions routes. Both were reported as stop, so a refused or truncated Claude reply looked like a finished turn.
• b3f1251: Fix Chat Completions requests with function tools to Amazon Bedrock GPT cross-Region inference profiles (such as us.openai.gpt-6-luna). Bedrock Runtime defaults these models to reasoning and rejects that next to function tools, so a request with tools and no reasoning_effort now sends none, the OpenAI default. A reasoning_effort the caller sets is now forwarded instead of dropped.
• 78780f9: Speed up harness-scoped request logs by reusing the resolved harness ID when querying unlinked legacy attempts.
• ca55315: Send provider-required request headers from the Playground. Runs against OpenCode Go and OpenCode Zen now carry x-opencode-session, so the provider no longer rejects them with MissingSessionID. Headers set in the request editor still take precedence.
• 352618a: Keep token-limited and refused replies visible on /v1/responses when the upstream speaks Chat Completions. A length or content_filter finish now yields status: "incomplete" (streams end with response.incomplete), and message.refusal / delta.refusal become refusal content parts instead of an empty completed response. Non-stream requests answered by a native Responses upstream over SSE now keep the response.incomplete and response.failed snapshots instead of reporting them as completed.
• df977e5: Reuse resolved harness IDs across analytics queries and keep cached request filters isolated when harness names are reused.