Sync the self-hosting stack

CHANGELOG.md

@@ -9,6 +9,38 @@ one of them changed. An entry says which artefact and which version, so `web 0.1
 Every version published to a channel appears here. Nothing is written from memory afterwards: a build
 cannot be published without its entry.
 
+## 2026-08-28 — Models other than OpenAI's work again `api 0.96.0`
+
+**Read this if you point Jarvis at anything other than OpenAI.** Self-hosted models, Qwen, Llama,
+Mistral, anything behind LiteLLM or a similar gateway. If you use OpenAI, nothing here changes for
+you.
+
+A conversation in which the assistant had used a tool would fail from that point on, with the
+provider answering:
+
+```text
+400 System message must be at the beginning
+```
+
+and failing the same way on every retry and every later message in that conversation. There was no
+way round it except starting a new conversation and avoiding tools, which is most of what Jarvis
+does.
+
+The cause was on our side. Jarvis records what a tool returned as its own entry in the context,
+placed immediately after the assistant turn it belongs to — the position is what says "these are the
+results of the turn above". That entry was labelled as a system instruction, which OpenAI accepts
+anywhere in a conversation and almost every other provider rejects outright.
+
+It is now labelled in a way every provider accepts at any position, and it still says in its own
+first sentence that the platform wrote it and the assistant did not — which is what stops a model
+learning to write tool results instead of asking for them. Nothing changes about what the model is
+told; only the envelope it arrives in.
+
+A guard now also sits on the last line before any request leaves for your provider, so a
+mislabelled entry is corrected and logged rather than costing you a conversation.
+
+**Nothing to do.** Take the build; conversations that were failing will work on their next message.
+
 ## 2026-08-27 — A machine could show Offline while you were using it `api 0.95.0` `web 0.121.0`
 
 **Only affects deployments running more than one copy of the api.** On Docker Compose this cannot