<!-- CURSOR_SUMMARY -->
> [!NOTE]
> Adds a context-limit banner with one-click “summarize into new chat,”
refactors token counting with react-query, and persists per-message max
token usage.
>
> - **Chat UX**
> - **Context limit banner** (`ContextLimitBanner.tsx`,
`MessagesList.tsx`): shows when within 40k tokens of `contextWindow`,
with tooltip and action to summarize into a new chat.
> - **Summarize flow**: extracted to `useSummarizeInNewChat` and used in
chat input and banner; new summarize system prompt
(`summarize_chat_system_prompt.ts`).
> - **Token usage & counting**
> - **Persist max tokens used per assistant message**: DB migration
(`messages.max_tokens_used`), schema updates, and saving usage during
streaming (`chat_stream_handlers.ts`).
> - **Token counting refactor** (`useCountTokens.ts`): react-query with
debounce; returns `estimatedTotalTokens` and `actualMaxTokens`;
invalidated on model change and stream end; `TokenBar` updated.
> - **Surfacing usage**: tooltip on latest assistant message shows total
tokens (`ChatMessage.tsx`).
> - **Model/config tweaks**
> - Set `auto` model `contextWindow` to `200_000`
(`language_model_constants.ts`).
> - Improve chat auto-scroll dependency (`ChatPanel.tsx`).
> - Fix app path validation regex (`app_handlers.ts`).
> - **Testing & dev server**
> - E2E tests for banner and summarize
(`e2e-tests/context_limit_banner.spec.ts` + fixtures/snapshot).
> - Fake LLM server streams usage to simulate high token scenarios
(`testing/fake-llm-server/*`).
>
> <sup>Written by [Cursor
Bugbot](https://cursor.com/dashboard?tab=bugbot) for commit
2ae16a14d50699cc772407426419192c2fdf2ec3. This will update automatically
on new commits. Configure
[here](https://cursor.com/dashboard?tab=bugbot).</sup>
<!-- /CURSOR_SUMMARY -->
<!-- This is an auto-generated description by cubic. -->
---
## Summary by cubic
Adds a “Summarize into new chat” trigger and a context limit banner to
help keep conversations focused and avoid hitting model limits. Also
tracks and surfaces actual token usage per assistant message, with a
token counting refactor for reliability.
- **New Features**
- Summarize into new chat from the input or banner; improved system
prompt with clear output format.
- Context limit banner shows when within 40k tokens of the model’s
context window and offers a one-click summarize action.
- Tooltip on the latest assistant message shows total tokens used.
- **Refactors**
- Token counting now uses react-query and returns estimatedTotalTokens
and actualMaxTokens; counts are invalidated on model change and when
streaming settles.
- Persist per-message max_tokens_used in the messages table; backend
aggregates model usage during streaming and saves it.
- Adjusted default “Auto” model contextWindow to 200k for more realistic
limits.
- Improved chat scrolling while streaming; fixed app path validation
regex.
<sup>Written for commit 2ae16a14d50699cc772407426419192c2fdf2ec3.
Summary will update automatically on new commits.</sup>
<!-- End of auto-generated description by cubic. -->
Fixes#710
This PR implements comprehensive Azure OpenAI integration for Dyad,
enabling users to leverage Azure
OpenAI models through proper environment variable configuration. The
implementation adds Azure as a
supported provider with full integration into the existing language
model architecture, including support
for GPT-5 models. Key features include environment-based
configuration using `AZURE_API_KEY` and `AZURE_RESOURCE_NAME`,
specialized UI components that provide clear
setup instructions and status indicators, and seamless integration with
Dyad's existing provider system.
The Azure provider leverages the @ai-sdk/azure package (v1.3.25) for
compatibility with the current
TypeScript language model interfaces.
The implementation includes robust error handling for missing
configuration, comprehensive test coverage
with 9 new unit tests covering critical functionality like model client
creation and error scenarios, and
an E2E test for the Azure-specific settings UI.
<img width="1510" height="908" alt="Screenshot 2025-08-18 at 9 14 32 PM"
src="https://github.com/user-attachments/assets/04aa99e1-1590-4bb0-86c9-a67b97bc7500"
/>
---------
Co-authored-by: graphite-app[bot] <96075541+graphite-app[bot]@users.noreply.github.com>
Co-authored-by: Will Chen <willchen90@gmail.com>
Fixes#348Fixes#274Fixes#149
- Connect to existing repos
- Push to other branches on GitHub besides main
- Allows force push (with confirmation) dialog
---------
Co-authored-by: graphite-app[bot] <96075541+graphite-app[bot]@users.noreply.github.com>
This uses Gemini's native [thinking
summaries](https://cloud.google.com/vertex-ai/generative-ai/docs/thinking#thought-summaries)
which were recently added to the API.
Why? The grafted thinking would sometimes cause weird issues where the
model, especially Gemini 2.5 Flash, got confused and put dyad tags like
`<dyad-write>` inside the `<think>` tags.
This also improves the UX because you can see the native thoughts rather
than having the Gemini response load for a while without any feedback.
I tried adding Anthropic extended thinking, however it requires temp to
be set at 1, which isn't ideal for Dyad's use case where we need precise
syntax following.
