Sunil Pai

This commit performs a **documentation cleanup** within the **`ai-chat` package**, specifically addressing **formatting issues** in the `CHANGELOG.md` file. It removes extraneous blank lines to improve the readability and consistency of the release notes. This **maintenance** task ensures the changelog is well-formatted, providing a clearer history of changes for users and contributors.

This commit performs a **style fix** within the **AI Chat package**, specifically addressing formatting inconsistencies in its documentation. It removes extraneous blank lines from the `packages/ai-chat/CHANGELOG.md` file to ensure better readability and adherence to project style guidelines. This is a **maintenance** task focused on improving the consistency and presentation of the changelog. The scope of this change is limited to the visual formatting of the `ai-chat` module's release history.

This commit **fixes a critical cache invalidation bug** in the **`useAgentChat` hook** within the **`@ai-chat` package**. The `initialMessagesCacheKey` calculation is now updated to **ignore URL query parameters**, preventing unnecessary cache misses and broken stream resumes that occurred when dynamic parameters like authentication tokens rotated. By focusing the cache key on the agent's core identity (origin, pathname, agent, name), this change significantly **improves the stability and performance** of agent chat interactions. **New tests** have been added to confirm the cache's stability when only query parameters change and to ensure proper refetching when the agent's identity genuinely shifts.

This commit introduces **server-driven messaging capabilities** to the **`ai-chat`** package, primarily by adding an `onChatResponse` lifecycle hook to the `AIChatAgent` for agents to react after a chat turn completes. It also provides new `isServerStreaming` and `isStreaming` flags within the `useAgentChat` hook, offering **client-side indicators** for server-initiated streams and unified streaming states. This **new capability** enables more dynamic chat experiences, such as proactive messages and chained reasoning, and is supported by comprehensive documentation, examples, and end-to-end tests. The work also includes **refactoring** to rename related types and hooks for clarity and a **bug fix** to ensure pending responses are cleared on stop, enhancing overall stability.

This commit **introduces an `auto` transport option for MCP endpoints**, enabling **auto-negotiation** between Streamable HTTP and legacy SSE protocols on the same endpoint. This **new capability** adds a `createAutoHandler` function with routing rules to dynamically serve the appropriate protocol based on the request, and integrates it into the `McpAgent`. The **`packages/agents` module**, specifically the **MCP subsystem**, is primarily affected, with updates to `ServeOptions`, comprehensive new tests, and documentation to support this more flexible communication method. This allows for more robust and simplified client-server communication by consolidating different protocol handling onto a single endpoint.

This commit performs a **maintenance update** by bumping numerous development and runtime dependencies across the project's configuration. It updates critical tools like **`vite`**, **`wrangler`**, **`ai`**, and **`partyserver`**, along with related packages such as `@ai-sdk/react` and `@cloudflare/vite-plugin`. These updates affect a wide range of **example applications** and several **core packages** (e.g., `agents`, `ai-chat`), ensuring the project benefits from the latest bug fixes, performance enhancements, and features from these upstream libraries. A new changeset was added, indicating a planned patch release for the `agents` package as a result of these dependency changes, and `package-lock.json` was updated accordingly.

This commit introduces **new end-to-end Playwright tests** for the `codemode` package, adding `executor.spec.ts` and `llm-codemode.spec.ts` to cover direct executor flows, multi-provider namespaces, and LLM-driven code generation. To support these tests, the **test worker (`packages/codemode/e2e/worker.ts`) is significantly extended** with new endpoints like `/execute`, `/mcp/execute`, `/run-multi`, and `/types/multi`, incorporating various tool sets and presets. This **new capability** provides **deterministic e2e coverage** of the entire HTTP → Worker → sandbox → tool RPC → result pipeline, including LLM integration. The changes ensure robust validation of core `codemode` functionalities, from HTTP requests to sandbox execution and tool RPC.

This commit introduces a comprehensive suite of **end-to-end Playwright tests** for the **`ai-chat` module**, significantly enhancing its **testing coverage** and validating complex behaviors. The new tests cover critical functionalities including multi-tab continuation, `data-*` parts handling, stream error recovery, and persistence features like `sanitizeMessageForPersistence` and `maxPersistedMessages` trimming. To enable these validations, supporting **agent implementations** such as `SanitizeAgent`, `MaxPersistedAgent`, and `DataPartsAgent` were added to `packages/ai-chat/e2e/worker.ts`, alongside necessary **Durable Object registrations** in `wrangler.jsonc`. This work provides crucial **validation** for the `ai-chat` system's resilience, data integrity, and streaming capabilities.

This commit performs **maintenance** on the **AI Chat feature's changelog** by **removing stray blank lines** from `packages/ai-chat/CHANGELOG.md`. This **formatting cleanup** specifically targets list spacing within the changelog entries. The primary goal is to **improve the readability** and overall presentation of the release notes for the **AI Chat module**.

This commit **enhances the `ai-chat` module** by implementing the **stream resume handshake protocol**, introducing `CF_AGENT_STREAM_RESUMING` and `CF_AGENT_STREAM_RESUME_ACK` for robust continuation. It also **fixes a bug** in the `AIChatAgent`'s error handling by ensuring `_pendingResumeConnections` are cleared within `_markStreamError` when a stream fails, preventing stale resume states. These changes improve the reliability of stream resumption and error recovery, with corresponding updates to **e2e tests** in `client-tools.spec.ts` to validate the new handshake and continuation logic. The overall impact is a more stable and predictable streaming experience for users of the `ai-chat` agent.

This commit **improves the developer experience** for the **push notifications example** by modifying the `generate-vapid-keys` script. Instead of printing VAPID keys to the console, the script now **automatically writes `VAPID_PUBLIC_KEY` and `VAPID_PRIVATE_KEY` directly into the `.env` file**, creating or updating entries as needed. This **streamlines setup** by eliminating a manual copy/paste step for configuring VAPID keys. Additionally, the `examples/push-notifications/README.md` has been updated to reflect this automated process and remind users to configure `VAPID_SUBJECT`.

This commit significantly **improves the build and development consistency** across numerous **example projects** by addressing critical configuration issues. It introduces per-example `.gitignore` files to standardize ignored artifacts and, more importantly, **fixes Tailwind CSS scanning** by updating `src/styles.css` files with correct `@source` entries and import orders. These **maintenance updates** ensure Tailwind properly identifies and applies styles from `Kumo` and `streamdown` components in both monorepo and standalone setups, resolving a **styling bug**. Additionally, a `deploy` script was added to the `cross-domain` example's `package.json` for streamlined deployment, enhancing the overall **developer experience** for the `examples` subsystem.

This commit introduces a **new example** in `examples/structured-input` that demonstrates **LLM-driven interactive forms** via client-side tools. It features a React client (`client.tsx`) capable of rendering dynamic input forms (multiple choice, yes/no, free text, rating) based on tool calls from an `AIChatAgent` (`server.ts`) powered by Workers AI. The example defines tool schemas using Zod (`tools.ts`) and illustrates the full workflow from the LLM requesting structured input to the client returning user responses. This **new capability example** also includes necessary build configurations (Vite, Wrangler), type fixes for tool input handling, and consolidates entry points for adherence to project conventions. Ultimately, it provides a comprehensive blueprint for developers to implement **interactive user experiences** where the LLM guides and collects structured data from the user.

This commit performs a **visual improvement** to the `examples/elevenlabs-starter/README.md` file. It **increases the height of badge images** from 28 to 32 pixels and **removes the '×' separator** that previously existed between the ElevenLabs and Cloudflare Agents badges. This **formatting tweak** consolidates spacing, resulting in **cleaner alignment** and a more consistent header layout within the example's documentation. The change is a **maintenance** task focused on enhancing the aesthetic presentation of the project's examples.

This commit introduces a **new example application**, `examples/elevenlabs-starter`, demonstrating the integration of **ElevenLabs APIs with Cloudflare Agents**. This **feature addition** provides a comprehensive full-stack starter demo, including server-side agent implementations for voice chat, soundscapes, character creation, and music generation, alongside a React client, shared utilities, and configuration files. It showcases diverse AI audio capabilities within the `examples` directory. Additionally, `examples/dynamic-workers-playground/package.json` is updated to include the `agents` dependency, enabling other examples to reference the Agents SDK.

This commit performs a significant **refactoring** by **removing the shared `@cloudflare/agents-ui` dependency** and its entire `packages/agents-ui` workspace. It **inlines common UI components** such as `ModeToggle` and `ConnectionIndicator` directly into numerous **example and experimental applications**, including `a2a`, `ai-chat`, `playground`, and `gadgets-chat`. The **theme management system** is also refactored, switching all affected applications to Kumo's default theme, eliminating `ThemeProvider` wrappers, and adopting a simpler `localStorage` and `data-mode` approach for theme toggling. This change **reduces external UI dependencies** across the project's examples, standardizes their styling, and replaces `PoweredByAgents` with `PoweredByCloudflare`.

This commit introduces a **new capability** to **normalize `CallTool` results** within the **`codeMcpServer`** by implementing `unwrapMcpResult`. This utility processes diverse `CallTool` outputs, ensuring **sandbox code receives plain, consistent values** by prioritizing `toolResult` or `structuredContent`, parsing single-text JSON, concatenating plain text, or throwing on `isError` for catchable exceptions. The change significantly improves the predictability and usability of `CallTool` outputs, with updated tests in `mcp.test.ts` validating the new unwrapping logic and error handling. A changeset note is included for this important patch release.

This commit performs a significant **dependency upgrade**, bumping the `@modelcontextprotocol/sdk` from version 1.26.0 to 1.28.0 across all examples and core packages. A key **new capability** introduced is the population of the `RequestInfo.url` field within the `StreamableHTTPServerTransport` and `WorkerTransport` modules. This enhancement directly benefits **tool handlers** by providing them with access to the full request URL and its query parameters, enabling more context-aware operations within the **Model Context Protocol (MCP) transports**. The change improves the overall utility of the SDK by exposing vital request details, alongside minor **maintenance** by removing an unused script.

This commit **integrates the `@streamdown/code` plugin** across numerous example and experimental applications, enabling enhanced code rendering capabilities within **Streamdown components**. It involves **updating package dependencies** for `streamdown` to `^2.5.0` and `@streamdown/code` to `^1.1.1` in all affected `package.json` files, along with corresponding `package-lock.json` updates. Furthermore, **CSS styles are updated** in modules like `examples/*/src/styles.css` and `site/ai-playground/src/index.css` to correctly source the new distribution files from `streamdown` and `@streamdown/code`. This **feature enhancement** significantly expands the display functionality of Streamdown-based applications by providing dedicated code block support, and also adjusts the `scripts/check-updates.ts` script to no longer reject these specific package updates.

This commit **improves the stability and reliability of `useAgent` React tests** by addressing a **race condition**. It modifies the tests to **wait for the server's initial state broadcast** before executing local `setState` operations, preventing the server's state from inadvertently overwriting test-driven changes. Specifically, `vi.waitFor` blocks now query `[data-testid="agent-state"]` and assert that `textContent` is not "undefined" and `capturedAgent.identified` is true, ensuring the server's state is properly established. This **test fix** prevents flaky failures and ensures accurate testing of the **`useAgent` hook's behavior** within the `agents` package.