A configuration per project, and smart routing for inline tiers 🧠
- A configuration per project and per agent. Each project and each agent you deploy on SLNG has its own configuration, with the model chosen in its Think section, and keeps your organization’s settings. A key created in a project answers with that project’s model.
- Your dashboard LLM keys in the Think section. LLM keys you save on the BYOK page are offered in your projects’ Think sections.
- Smart routing for inline tiers. Send two or more tiers in
slng_config, from cheapest to most capable, withslng/auto. The router answers each request with the cheapest tier it expects to handle it well, and thex-slng-modelheader tells you which model answered. One tier works exactly as before.
Breaking changes
- Two or more inline tiers no longer mean “preferred first, backup second”. If your
slng_configuses tiers that way, your backup now answers turns your preferred model used to answer: inline tiers are routed per turn, and turns that may need a tool call or look demanding go to your most capable tier. Withslng_pure_proxy, yourdefault_tieranswers every turn, and that is your most capable tier unless you set it. Put both models in one tier and give the backuppriority: 1. - Inline configuration keeps your settings. A
slng_configsent with a request no longer replaces the configuration your request uses, a project’s included. Each top-level field it sends replaces that field for the request, and saved settings it does not send, such as caching or a flow SLNG set up for your agent, now apply too. Sendingtiersalso sets aside the saveddefault_tier. - A stream that fails now ends with an error. When the model fails after a Chat Completions stream has started, the stream now ends with an error chunk and
data: [DONE], with nofinish_reasonand nousageblock, instead of a normalfinish_reason: "stop"chunk. The OpenAI SDKs raise anAPIErrorwhen they read it, so catch it and treat the text you already received as partial. On the/responsesendpoint the stream ends withresponse.failedinstead ofresponse.completed. slng_pure_proxyanswers from your default tier. With two or more tiers, aslng_pure_proxyrequest is now answered by yourdefault_tieron every turn, your most capable tier unless you set it, instead of by tier"1". If your tiers still mean “preferred first, backup second”, send"default_tier": 1or put both models in one tier.- Client-model
fallbacksis retired. Creating or updating a client model with a non-emptyfallbackslist now returns a422. To choose which key answers first, use Set as default on the BYOK page of the Vault (organization owners and admins). To add fallbacks for an agent, use Fallback routing in its Think section. - Client model responses no longer include
fallbacks. Listing, getting, creating and updating a client model returns nofallbacksfield, so code that reads it must stop doing so.
Hello, world! 👋
New day, new docs. We rebuilt this site from the ground up so it is easier to find your way around, and today is our first changelog entry. Welcome!Reorganized around how you actually work 🧭
- Understand the execution layer first. The concepts section explains what sits between your orchestrator and your models, and why, before you write a line of code.
- Follow the guides step by step to set things up. Each one walks you from zero to working.
- UI or API, your call. Every setup path gives you both options, so you can click through the dashboard or wire it up programmatically, whichever fits.