Spaces:
Running
Running
Commit History
8 days ago
-
feat: registry-based model family, 64K Gemini cap, disabled-model filter
Hermes Agentd11d528
-
Resolve Gemini grounding redirect citation URLs
farooquiowais +1: CursorCo-authored-by: Cursor <cursoragent@cursor.com>cd036d7 -
feat: surface Gemini search grounding as OpenAI url_citation annotations
proxy-editCapture candidate.groundingMetadata (queries, source URLs, per-segment supports) in both streaming (sse-streamer) and non-streaming (sse-parser -> response-converter) paths, normalize via new format/grounding.js, and expose on the OpenAI endpoint as message.annotations (url_citation) plus a richer message.grounding object. Additive only; existing behavior unchanged.d059023
-
feat: raise proxy output cap and defaults to 32768
farooquiowais- GEMINI_MAX_OUTPUT_TOKENS 16384 -> 32768 - OpenAI compat default fallback 16384 -> 32768 - Anthropic /v1/messages default fallback 16384 -> 32768 This supports single-shot tabular extraction with larger column counts while keeping explicit max_tokens respected.1b034a4 -
feat: raise default max_tokens fallback from 4096 to 16384
farooquiowaisOpenAI-compatible (/v1/chat/completions) and Anthropic-compatible (/v1/messages) endpoints now default to 16384 output tokens when the client omits max_tokens. This matches the existing Gemini hard cap and mitigates truncation for streaming clients (e.g. Mike tabular batch extraction) until they send explicit per-request budgets.2be6e7d
-
fix: map thinking blocks to reasoning_content in OpenAI endpoints
farooquiowais64a3efa
-
Fix gemini-3.1-pro-high by routing to pro-low with thinkingLevel high.
The upstream API rejects gemini-3.1-pro-high as a direct model id. Map pro-high requests to gemini-3.1-pro-low with high thinking, skip duplicate thinking on tier SKUs, and add legacy Antigravity client headers. Co-authored-by: Cursor <cursoragent@cursor.com>4badc3b