[codex] fix: prefer active IntelliJ workspace for relative paths

[yzlu0917]

Summary

• prefer the active editor's workspace directory when resolving ambiguous relative paths
• apply the same workspace prioritization to both existing-path lookup and new-file path inference
• add regression tests covering ambiguous multi-workspace resolution and the no-active-file fallback

Why

In IntelliJ multi-Maven setups, Continue can receive multiple workspace roots for related projects. When the same relative path exists in more than one workspace, the previous behavior effectively fell back to the first workspace root, which could cause ls and create_new_file to operate in a sibling project instead of the one containing the current file.

This change biases relative path resolution toward the workspace that contains the active file, while preserving the existing first-workspace fallback when there is no active file context.

Validation

• ran npm test -- --runInBand util/ideUtils.test.ts in core
• ran git diff --check

Closes #10644

---

Summary by cubic

Prefer the active editor’s workspace when resolving relative paths in IntelliJ multi-workspace projects to keep actions in the current project. Also adds dynamic provider model fetching with a simple UI hook to make model selection easier, and fixes a CLI install banner unmount update.

• Bug Fixes

  • Prioritize the current file’s workspace in resolveRelativePathInDir and inferResolvedUriFromRelativePath.
  • Fall back to the original first-workspace choice when there’s no active file or getCurrentFile isn’t available.
  • Add regression tests for ambiguous multi-workspace resolution, the no-active-file fallback, and the legacy-stub fallback.
  • GUI: prevent CliInstallBanner state updates after unmount during CLI detection.

• New Features

  • Add models/fetch core protocol and provider adapters to list models (Ollama, OpenRouter, Anthropic, Gemini, or generic via listModels).
  • Map model capabilities into config/types (tools, image); broaden tool/image support detection (Gemma 4, GPT-5+), and update OpenAI o-series checks.
  • UI: fetch provider models using the entered API key in Add Model; dynamic Ollama/OpenRouter lists; minor ModelSelectionListbox label tweak.
  • Add GPT-5.4 variants and Gemma 4; update docs; include OpenRouter request headers via @continuedev/openai-adapters.

^{Written for commit 198d7d8. Summary will update on new commits.}

[cubic-dev-ai]

No issues found across 2 files

[chatgpt-codex-connector]

💡 Codex Review

continue/gui/src/forms/AddModelForm.tsx

Lines 75 to 77 in 2f98da9

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

Ignore stale model-fetch responses after provider change

The provider check in setFetchedModelsList uses selectedProvider from the same closure as providerAtFetchTime, so it does not protect against in-flight requests completing after the user switches providers. In that flow, a late response for provider A can populate fetchedModelsList while provider B is selected, which surfaces and can save mismatched model IDs under the wrong provider. Use current provider state (or cancel previous requests) when applying async results.

---

continue/core/llm/fetchModels.ts

Lines 180 to 181 in 2f98da9

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

Preserve Gemini API version path in models endpoint URL

Building the endpoint with new URL("models", base) drops the final path segment when apiBase lacks a trailing slash. For example, a valid custom base like https://.../v1beta resolves to /models instead of /v1beta/models, so model discovery fails for those configurations. Normalize apiBase to a directory-style URL (trailing slash) before joining models.

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".