Respect adapter endpoint routing in CLI stream retries

[lyonsno]

Summary

Fixes #10474.

The CLI retry helper was routing Responses-capable model names directly to llmApi.responsesStream() whenever that optional method existed. That bypassed the OpenAI adapter's apiBase guard, so OpenAI-compatible providers and proxies could be sent to /v1/responses even when they only support /v1/chat/completions.

This change keeps endpoint selection inside the adapter by having chatCompletionStreamWithBackoff() delegate streaming calls to the required llmApi.chatCompletionStream() contract. The OpenAI adapter still uses the Responses API for official OpenAI endpoints when appropriate, but custom apiBase values stay on the chat-completions path.

Changes

• Remove the CLI helper's direct responsesStream() routing branch.
• Add a regression test proving a Responses-capable request with both adapter methods present still goes through chatCompletionStream().
• Update the streaming/tool-call preservation test so it fails if the CLI calls responsesStream() directly.

Verification

• npm --prefix extensions/cli test
• npm --prefix packages/openai-adapters test -- --run
• npm --prefix extensions/cli run build
• npm --prefix extensions/cli run test:e2e
• npm --prefix extensions/cli run test:smoke
• npm --prefix extensions/cli run typecheck
• git diff --check

I also ran two manual regression smokes:

• a local OpenAI-compatible proxy with a Responses-capable model name and custom apiBase, which hit /v1/chat/completions once and /v1/responses zero times;
• the official OpenAI path with a Responses-capable model, to confirm adapter-owned Responses routing still streams successfully.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[cubic-dev-ai]

No issues found across 3 files

_{Re-trigger cubic}

[chatgpt-codex-connector]

💡 Codex Review

continue/gui/src/forms/AddModelForm.tsx

Lines 75 to 77 in b3af874

	setFetchedModelsList((prev) =>
	selectedProvider.provider === providerAtFetchTime ? models : prev,
	);

Guard fetched models with current provider state

If the user clicks the refresh icon and then switches providers before the request completes, this closure still compares against the provider value captured when the request started, so the condition is always true and the old provider's fetched models are inserted into the newly selected provider's model list. This can make the form offer models with the wrong providerOptions until another selection clears the list.

---

continue/core/llm/fetchModels.ts

Lines 180 to 181 in b3af874

	const base = apiBase || "https://generativelanguage.googleapis.com/v1beta/";
	const url = new URL("models", base);

Normalize Gemini API base before appending models

When apiBase is provided without a trailing slash, new URL("models", base) replaces the final path segment instead of appending to it; for example https://generativelanguage.googleapis.com/v1beta becomes https://generativelanguage.googleapis.com/models?... rather than /v1beta/models. That makes model fetching fail for the common custom-base spelling without a trailing slash.

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[lyonsno]

I have read the CLA Document and I hereby sign the CLA

Recheck