v1.3.35-vscode
ContinueSummary
This release introduces a new option to opt out of the Responses API and enhances provider compatibility by adding necessary headers for OpenRouter identification. Numerous bug fixes address issues with model configuration, message ordering in Gemini, tool call handling, and resource management across various providers like Ollama and vLLM.
New Features
- Added `useResponsesApi` option to allow users to opt out of the Responses API.
- Enabled OpenRouter to send HTTP-Referer and X-Title headers to identify the application when making requests.
Bug Fixes
- Removed Llama 3.1 405B from the Groq provider.
- Fixed an issue where Gemini models merged consecutive same-role messages, causing ordering errors.
- Fixed an issue where tool arguments (MCP tool args) were not being coerced to match schema string types.
- Fixed mapping of `reasoning-delta` to `reasoning_content` instead of `content` for certain models.
- Fixed an issue preventing multiple context providers of the same type from being configured in `config.yaml`.
- Stopped CLI free-trial polling for models that are not in a free-trial state.
- Removed inline backtick fences from tool instruction prose.
- Fixed handling of multiple zip files during the JetBrains release artifact creation step.
- Fixed hiding the thinking indicator when the thinking content is empty.
- Fixed OpenRouter support for Gemini 3, including suffix stripping, `thought_signature`, and the autocomplete endpoint.
- Fixed listener leaks and redundant file reads occurring during autocomplete operations.
- Preserved tool calls when thinking models return no text content.
- Fixed an issue where users could not correct an API key after entering an invalid one for xAI/Gemini.
- Added actionable error reporting when Ollama fails to parse tool calls.
- Ensured the vLLM provider respects the user-configured `contextLength` and model settings.
Improvements
- Added `keepAlive` configuration support to YAML completion options schema.
- Included the model name in the completion request body for `llama.cpp` providers.
- Added support for the `reasoning_content` field for Kimi models in the Moonshot provider.
- Ensured the context length specified in the YAML model configuration is respected.
- Lazy-loaded the Ollama `/api/show` endpoint to reduce unnecessary initial requests.
- Ensured installation steps are not skipped by default and the lock file is synchronized.
- Added documentation clarifying where secrets can be templated from.
- Added troubleshooting documentation specifically for Ollama memory errors.