csghub-server Files

v2.2.0-ce

✨ New Features

• AI Gateway & OpenAI-Compatible APIs
• Upstream Reliability: Added upstream health checks, circuit breaker state, fallback routing, and per-upstream availability filtering.
• Observability: Added Prometheus metrics for chat upstream attempts and optional LLM tracing for chat completions.

• Model Routing: Improved external model routing with relational upstreams, per-upstream provider/auth/model overrides, and session-aware routing.

• ClawHub & Skills

• Added ClawHub-compatible APIs for skill search, publish, resolve, download, whoami, and skill version lookup.
• Added skill version publishing from existing skill repositories through /skills/{namespace}/{name}/publish.

• Added persisted skill version records so published skills can expose installable versions.

• Platform Dataflow

• Added platform dataflow job APIs for creating, deleting, inspecting, and reading logs from Argo-based dataflow jobs.

• Added runner-side dataflow workflow support, including PVC creation, workflow status tracking, and log access.

• Activity Logs

• Added activity log collection for model, dataset, git, inference, fine-tune, serverless, and agent operations.

• Added admin activity log listing via /api/v1/admin/activity_logs.

• Code Repositories

• Added code package upload support with presigned S3 POST URLs.
• Added ZIP and tar.gz package import during code repository creation.

🚀 Enhancements & Bug Fixes

• AI Gateway Reliability
• Fixed model cache miss behavior and reduced stale routing issues.
• Improved chat fallback retry behavior for failed upstream attempts.
• Improved multimedia generation behavior, including Seedance video download handling.

• Added safer token/API-key handling in metering and proxy logs.

• Repository Search & Metadata

• Added model parameter range filters for model listing.
• Added dataset repository-size range filters for dataset listing.

• Fixed skill tag and skill mirror_from_saas route behavior..

• Resource Scheduling

• Added admin listing for all space resources.
• Improved resource availability checks by returning explicit errors when cluster resource status is uncertain.

• Made CPU-only scheduling to GPU/XPU nodes configurable.

• Finetune & Runner

• Allowed fine-tune logs to be read by either numeric job ID or task ID.

• Fixed runner service/workflow panic paths and cluster existence checks.

• Repository & LFS

• Improved LFS mirroring for small files.
• Added object size mismatch cleanup during LFS sync.
• Improved LFS worker lifecycle cleanup.

• Fixed XNet public repository download behavior when no access token is available.

• Proxy & Networking

• Sanitized unsafe forwarded headers.
• Used request host handling that preserves proper upstream Host behavior.

🛠 Maintenance

• Migrated AI Gateway upstream configuration into relational upstream, health state, and circuit state tables.
• Added persistence support for skill versions, activity logs, dataflow tasks, and repository range filters.
• Updated inference runtime configs and images, including AMD, HF inference toolkit, and vLLM variants.
• Added localized error codes for skill, git, agent, and task errors.
• Refreshed tests across AI Gateway, activity logs, ClawHub, dataflow, skill publishing, resource checks, and LFS sync.

Full Changelog: https://github.com/OpenCSGs/csghub-server/compare/v2.1.0-ce...v2.2.0-ce