Skip to content

v0.7.29

Choose a tag to compare

View all tags
@SilenNaihin SilenNaihin released this 04 Sep 02:29
v0.7.29
d2c44c3
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Gateway engine release 0.7.29. Native data plane: exp-gateway-native 0.3.29.

Highlights

  • Hunyuan (hy4-preview) preserved-thinking round-trip. Tencent-native plaintext reasoning_content is returned for display and round-tripped across tool-loop turns via a domain-separated, gateway-issued opaque carrier (AES-256-GCM, key bound to the provider credential + route context) — a Harbor/Terminus-2 tool loop works without BYOK TokenHub. Enable-thinking inputs (thinking:{type:"enabled"}, nested reasoning:{effort}, chat_template_kwargs:{enable_thinking:true}) are translated to think mode instead of 400'd.
  • Per-rung reasoning exposure capability. ModelCapabilities.reasoning_output_exposed gates plaintext reasoning output per rung (default false), required together with a resolved reasoning-carrier route — an absent capability fails closed and stays stripped even on an exposable endpoint. Exposure is per-rung, never inferred from the base URL.

Also included since 0.7.28

  • Pricing: account for cache_write_input_tokens when cached_input_tokens is none (#697).
  • Chat Completions: accept store:false (no-op retention opt-out) (#761).
  • Responses: an output-less turn stays continuable from its id (#763).
  • Batch lane: failures name the provider's reason on the batch object (#760); an invalid key gets the uniform 401 (#762).
Assets 2
Loading