Skip to content

v2.7.0-ce

Pre-release
Pre-release

Choose a tag to compare

@Rader Rader released this 05 Nov 07:24

v2.7.0-ce

✨ New Features

  • AI Gateway

    • Added upstream priority routing and user routing tags. Model catalog filtering and pagination now run in the database. [#1623](#1623)
  • Inference & Model Support

    • Upgraded vLLM to 0.31, SGLang to 0.5.22, and llama.cpp to b11785.
    • Added the typed-decisions model task and video API compatibility for SGLang Diffusion. [#1623](#1623)
  • Authentication & Administration

    • Added Portico OIDC sign-in, configurable CORS origins, and administrative APIs for user tags. [#1623](#1623)

🚀 Enhancements & Bug Fixes

  • AI Gateway Reliability

    • When an upstream’s /models health probe returns HTTP 429 or 499, the gateway checks the inference endpoint before marking the upstream unhealthy. Upstreams with the same endpoint and authentication settings now share a probe. [#1605](#1605)
    • Improved OpenAI-compatible error responses and multimodal model discovery. Added a tokenizer fallback, stopped forwarding Anthropic user metadata to upstreams, and corrected token metering for chat requests with images. [#1623](#1623)
  • Responses & Anthropic Messages Compatibility

    • Report truncated Responses output as incomplete, keep parallel tool calls separate when streamed chunks omit indexes, and report tool_use when either a Chat Completions or Responses upstream emits tool calls. [#1607](#1607) · [#1603](#1603) · [#1606](#1606) · [#1618](#1618)
    • Map Chat Completions’ content_filter finish reason to Anthropic’s refusal stop reason. [#1621](#1621)
    • Preserve cached and reasoning token details in Responses usage, exclude cache reads from Anthropic input_tokens, and count streamed delta.reasoning in fallback token estimates. [#1614](#1614) · [#1615](#1615) · [#1622](#1622)
  • Repositories, Mirrors & Storage

    • Preserve repository visibility during synchronization, validate mirror source paths, support case-insensitive resource lookup, and preserve normal job order when urgent jobs preempt the queue. [#1623](#1623)
    • Fix completion of large LFS multipart uploads and align S3 signing with the forwarded host. Keep platform authorization headers out of presigned LFS upload requests and URLs. [#1623](#1623) · [#1608](#1608)
  • Deployments, Evaluation & Security

    • Correct multi-node vLLM topology and quantized model parameter counts, skip GGUF projector files during metadata scans, recognize video task aliases, and return stable evaluation framework IDs. [#1623](#1623)
    • Restrict sign-in redirects to registered callbacks and prevent stale webhook events from changing the state of a deleted sandbox. [#1623](#1623)
  • Accounting

    • Record token IDs in usage events, correct model usage attribution, and exclude API usage from compute consumption reports. [#1623](#1623)

🛠 Maintenance

  • Updated CSGClaw runtime profiles and shared agent template and execution contracts. [#1610](#1610) · [#1623](#1623)
  • Updated shared accounting schemas and statement indexes, including quota record mappings. [#1601](#1601) · [#1604](#1604) · [#1623](#1623)

Upgrade Notes

  • Ascend vLLM profiles are no longer registered automatically at startup [#1611](#1611)
  • If a browser client calls the server from another origin, review the CORS allowlist before upgrading. [#1623](#1623)

New Contributors

Full Changelog [v2.6.0-ce…v2.7.0-ce](v2.6.0-ce...v2.7.0-ce)