A stronger foundation for every provider
ChatGPT MD 3.2.0 is a reliability-focused release that modernizes the plugin's AI stack while tightening the boundaries around provider routing, streaming, and human-approved tool execution.
The headline change is the migration to AI SDK 7 and current provider SDKs. Underneath, provider configuration now flows through a centralized registry, making model selection, credentials, base URLs, defaults, and frontmatter behavior more consistent across OpenAI, OpenRouter, Anthropic, Gemini, Ollama, LM Studio, and Z.AI.
Highlights
- Correct provider routing everywhere. Title inference now derives its provider from the configured model, so non-OpenAI models use the right API key, base URL, and model-name handling.
- No stale provider clients. Provider instances are rebuilt when the provider, URL, or API key changes, preventing credentials or routing from leaking across requests.
- Safer tool calling. Models receive tool declarations, but execution remains inside ChatGPT MD's explicit approval flow. Tool calls now pass through stricter normalization, schema validation, and deterministic result formatting.
- More resilient streaming. Request handling, stream consumption, cancellation, and error reporting have been hardened across desktop and mobile environments.
- Current model coverage. The default tool-capable model list has been refreshed for OpenAI, Anthropic, Gemini, OpenRouter, and Z.AI.
Under the hood
This release introduces a provider registry as the operational source of truth for provider metadata and extracts smaller components for runtime provider construction, stream consumption, settings rendering, tool approval, and result formatting. Provider API responses are parsed more defensively, and OpenAI prompt/search model filtering has been corrected.
The test suite now covers 173 tests across 22 suites, including provider routing, URL construction, settings migrations, tool approval behavior, streaming, cancellation, and editor/file interactions. CI also validates formatting, linting, type safety, tests, production builds, and runtime dependency audits.
Privacy remains the boundary
The SDK upgrade does not bypass ChatGPT MD's human-in-the-loop design. Vault and web tools still require approval before execution and before selected results are shared with a model.
Upgrade notes
This is a drop-in upgrade from 3.1.x with no user configuration changes required. Existing agents, provider settings, frontmatter, and tool approval workflows continue to work as before.
Full diff: 3.1.0...3.2.0