Inja LLM Gateway
Get started

M6 surface notes

Operator notes for regional providers, media passthrough, completions policy, and the Realtime↔Live bridge decision.

Companion to the shipped OAuth, Realtime, platform, and fidelity guides. Documents passthrough-complete or policy-complete behavior.

Regional openai_compat

Provider Notes
xAI Chat/Responses; deferred completion route; Imagine needs image_genxAI
Moonshot/Kimi API keys, K3, Kimi Code CLI, helpers — Moonshot / Kimi; regional base_url
Z.AI/GLM Regional bases — Z.AI
DeepSeek Completions/FIM — DeepSeek FIM
Groq STT-first — Groq STT

Rule: passthrough never strips unknown keys. Regional quirks are base_url + capability flags.

Completions multi-dialect

Ingress Behavior
OpenAI /v1/completions openai / openai_compat passthrough (DeepSeek FIM on /beta)
Translate to Anthropic/Google Not supported (fail closed)

Anthropic beta & content blocks

  • anthropic-beta is forwarded (not allowlisted) on passthrough.
  • Skills, tunnels, memory, agents/sessions/environments are proxied.
  • Unknown content blocks: forwarded on PT; only IR types on translate.

Media / files

Surface Behavior
Image streaming Passthrough when upstream streams
STT diarization / formats Multipart fields forwarded on openai family
Moderations multimodal /v1/moderations body passthrough
Large uploads max_body_bytes (default 32MiB; set 536870912 for 512MiB) + Uploads API
max_body_bytes: 536870912  # 512 MiB

Realtime ↔ Live bridge

Shipped decision: fail-closed with unsupported_realtime_bridge. Same-protocol TLS WebSocket is production-ready. Full event mappers are a future milestone — IR placeholders live in canonical/realtime.go.

Type to search…

↑↓ navigate↵ selectEsc close