Skip to content

HAI Proxy onboarding: set up the OpenAI and Gemini APIs too - #322

Merged
stippi merged 2 commits into
mainfrom
feature/hai-proxy-providers
Nov 18, 2026
Merged

stippi merged 2 commits into
mainfrom
feature/hai-proxy-providers

Conversation

@stippi

@stippi stippi commented Nov 17, 2026

Copy link
Copy Markdown
Owner

Part of #184.

Onboarding

The SAP Hyperspace AI Proxy suggestion now creates one provider per API the proxy offers, all with the entered key:

Provider Type Base URL Model offered
hai-proxy anthropic …/anthropic/v1 Claude Opus 4.8, Sonnet 4.6 (as before)
hai-proxy-openai openai-responses …/openai/v1 GPT-6.1 Sol
hai-proxy-vertex vertex (bearer_auth) …/gemini/v1beta Gemini 3.8 Flash
  • A suggestion now holds a list of providers, each with its own models; a model can carry its own config. Gemini through the proxy gets a tool_config without streamed function-call arguments, which the proxy rejects.
  • ProviderSuggestion::apply_to works on in-memory config, so it is tested without files; apply_suggestion only reads and writes.
  • The LiteLLM endpoint from the issue is left out for now; whether it is needed is still to be clarified.

GPT reasoning defaults

The openai-responses reasoning style now sends summary: "concise" and context: "all_turns" with every effort (instead of summary: "auto"), which until now every GPT model needed in its own config. This applies to every Responses API provider, including the ChatGPT subscription.

Not verified yet

  • Not tried in the app or against the real endpoints.
  • Whether the ChatGPT subscription (Codex backend) accepts reasoning.context.

The HAI Proxy suggestion now creates one provider per API the proxy
offers, all with the entered key: hai-proxy (Anthropic, as before),
hai-proxy-openai (Responses API) and hai-proxy-vertex (Gemini with bearer
auth). A suggestion therefore holds a list of providers, each with its
own models; apply_to does the work on in-memory config so it can be
tested without files. The LiteLLM endpoint is left out for now.

Refs #184
Requests on the Responses API now ask for `summary: "concise"` and
`context: "all_turns"` with every reasoning effort, which until now every
GPT model needed in its own config.

The HAI Proxy suggestion can give a model its own config: Gemini there
gets the tool_config without streamed function-call arguments, which the
proxy rejects.

Refs #184
@stippi
stippi merged commit 81d0f69 into main Nov 18, 2026
9 checks passed
@stippi
stippi deleted the feature/hai-proxy-providers branch November 18, 2026 14:55
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant