Download Latest Version v2.5.0-ce source code.zip (9.7 MB) Google Add to Preferred Sources
Home / v2.5.0-ce
Name Modified Size InfoDownloads / Week
Parent folder
README.md 2026-09-02 4.6 kB
0
v2.5.0-ce source code.tar.gz 2026-09-02 9.1 MB
0
v2.5.0-ce source code.zip 2026-09-02 9.7 MB
0
Totals: 3 Items   18.8 MB 0

v2.5.0-ce

New Features

  • AI Gateway
  • Added OpenAI-compatible /v1/audio/translations support.
  • Added Anthropic Messages API support through /v1/messages.
  • Added hosted CSGHub Responses API support and chat adapter tool handling.
  • Added protocol-wide metrics reporting, dashboard data pipelines, and health checks for chat, responses, embedding, rerank, audio, image, video, OCR, and messages APIs.

  • Added PaddleOCR-VL PaddleX document parsing runtime and improved OCR preprocessing with embedded image support.

  • Agents & Sandbox

  • Added Agent Registry support and the CSGClaw agent type.
  • Added ms-swift agentic job runtime.
  • Integrated CSGClaw instances with Sandbox V2 deploy management.

  • Added configurable CSGClaw sandbox readiness probes and agent names in shared instance responses.

  • Evaluation & Finetune

  • Added staged finetune workflows to reduce runtime cost.
  • Added Claw Eval Codex configuration for OpenAI-compatible API evaluation.

  • Upgraded EvalScope and synchronized dataset rules.

  • Analytics & Governance

  • Added PostHog server-side analytics.
  • Added Redis-backed API rate limiting middleware.
  • Added user API authorization.
  • Added history archive support.

Enhancements & Bug Fixes

  • AI Gateway Reliability
  • Migrated protocol handlers to the new pipeline architecture.
  • Fixed billing skips for failed chat, responses, embedding, and rerank requests.
  • Fixed insufficient-balance response handling.
  • Fixed upstream unknown state so it no longer blocks traffic or marks models unavailable.
  • Fixed protocol-specific health-check defaults and Swagger generation ambiguity.
  • Added audio/video/image moderation support.
  • Added generated audio duration capture for TTS responses.

  • Added cached prompt token accounting and usage metering.

  • Repository, Mirror & Import

  • Added organization tag filters.
  • Optimized model index and repository list count queries.
  • Fixed deterministic namespace path lookup.
  • Fixed mirror deletion when the repository was already deleted.
  • Added MCP and Skill repository synchronization.
  • Rejected invalid mirror metadata source URLs.
  • Safely handled large HF NDJSON commit payloads.

  • Optimized skill/code package import and S3 package sync.

  • Deploy, Resource & Runtime

  • Checked cluster status for serverless detail queries and added force-stop support.
  • Improved deploy updates when clusters are offline.
  • Added serverless status filtering and multi-status query support.
  • Added XPU memory size support for physical XPU and MIG resources.
  • Added APIs to update space resource scenario constraints.
  • Added Ascend NPU runtime images and Ascend 910 HAMI support.

  • Added TGI ROCm inference image and updated llama-factory AMD engine support.

  • Security & Access Control

  • Fixed API key leakage in logs.
  • Protected /events from unauthorized clients exhausting database resources.
  • Added private endpoint permission checks for inference and version access.
  • Added feature-gating groundwork with OpenFeature and license integration.

  • Added sensitive-check gating for agent templates.

  • Accounting & Reporting

  • Added token usage information to evaluation summaries.
  • Added LLM log curated dataflow email notifications.
  • Reported source line and dataset row counts in LLM log sync emails.
  • Applied global timezone handling to admin queries, export data, recharge ranges, and accounting bills.
  • Capped accounting export page size to the upstream per-page limit.
  • Used recorded_at for AccountStatements and AccountMeterings partitioning.

Maintenance

  • Updated ms-swift configuration and Dockerfiles.
  • Upgraded agent-sandbox to v0.5.4.
  • Bumped GitPython to 3.1.58 and 3.1.61 in relevant runtime requirements.
  • Updated GitHub Actions checkout and setup-go workflow versions, pinning Go 1.26.
  • Fixed lint and stabilized database/user-like tests.
  • Regenerated mocks for exported availability interfaces only.
  • Added agent and merge request contract documentation.

New Contributors

Source: README.md, updated 2026-09-02