whoiskatrin

This commit **upgrades the `sandbox`'s dependency on `@cloudflare/containers` to version `0.2.2`** to properly support `sleepAfter` functionality for streamed commands. It introduces a **new feature** to **disconnect idle WebSocket transports** within the `sandbox`'s client, preventing premature termination of long-running `execStream` calls and improving resource management. This **dependency upgrade** and **feature implementation** ensure that `execStream` calls are not prematurely terminated while the container is active and that idle connections are properly cleaned up. Extensive **E2E tests** were added to validate the correct `sleepAfter` behavior for both quiet and idle streaming operations, ensuring stability and preventing regressions.

This commit **fixes critical bugs** in the **`ai-chat` package** related to stopping AI tool continuation streams and preventing orphaned server-side processes. It enhances the `useAgentChat` hook's `stop()` function to reliably call `abortActiveToolContinuation` in the `WebSocketChatTransport`, which now robustly handles pre-handshake aborts, manages request IDs for server cleanup, and prevents late `STREAM_RESUMING` messages. This **bug fix** ensures predictable stream termination and correct merging of continuation data, improving the stability and resource management of AI chat interactions.

This commit **fixes a critical bug** in the **`ai-chat` package** that led to **duplicate assistant messages** appearing during overlapping submits. It **refactors the streaming protection logic** within the `useAgentChat` hook, introducing an `anchorMessageId` to correctly restore in-progress assistant messages and consolidating local response ID handling for better state management. The `moveMessageById` function was renamed to `moveMessageToEnd` and simplified to always move a message to the end. This change includes **significant test improvements** in `use-agent-chat.test.tsx`, covering edge cases like multiple server broadcasts, mid-stream clears, and anchor-based restoration, ensuring the chat interface remains consistent and accurate even under complex interactions.

This commit introduces a **bug fix** within the **`ai-chat` package** to resolve an issue where duplicate assistant messages appeared during chat continuations. The core problem, tracked in #1229, stemmed from `messageId` being present in continuation start chunks, leading clients to create new messages instead of reusing existing ones. To address this, the `messageId` is now **stripped server-side** from these chunks within the `_handleWebSocketMessage` logic, ensuring proper message reuse. This change improves the reliability of AI chat interactions and is validated by new tests, including an SSE stub in the test agent, to simulate relevant stream data.

This commit introduces a new **GitHub issue template** specifically for **bug reports** within the `.github/ISSUE_TEMPLATE/` directory. This **maintenance** and **documentation improvement** effort aims to standardize the information collected when users report bugs. By providing a clear, structured template, it will streamline the issue submission process and ensure all necessary details are captured upfront, improving the efficiency of issue resolution.

This commit introduces **message concurrency controls** to the `AIChatAgent` within the `@cloudflare/ai-chat` package, allowing developers to define strategies like `queue`, `latest`, `merge`, `drop`, and `debounce` for overlapping `sendMessage()` submissions. It also **enhances** the `saveMessages()` method to accept a functional form for deriving messages and return `{ requestId, status }`, enabling callers to detect skipped turns. This **new capability** significantly improves the robustness and predictability of chat interactions by managing concurrent message flows, preventing race conditions and unexpected behavior. Comprehensive test coverage for all strategies, `onChatResponse` interactions, and programmatic turns ensures reliability. Documentation for `AIChatAgent.messageConcurrency` and the enhanced `saveMessages()` is also updated.

This commit **enhances the WebSocket transport client** by implementing a **retry mechanism for WebSocket upgrade requests** that fail with 503 responses during cold starts. Previously, these initial handshakes would fail fast, but now the `ws-transport` client will retry within the existing startup budget, ensuring each attempt uses a fresh request and abort signal. This **bug fix and feature enhancement** significantly improves the **reliability of initial WebSocket connections**, preventing client failures during server warm-up periods. The `base-transport` module was updated to expose retry timeout configurations, and new tests in `ws-transport.test.ts` validate this robust connection behavior.

This commit **enhances the reliability of WebSocket client connections** by implementing a retry mechanism for upgrade requests that fail with a 503 response during **cold starts** or deployments. The **WebSocket client transport** in `packages/sandbox/src/clients/transport/ws-transport.ts` now retries these initial handshakes, ensuring each attempt uses a fresh request to prevent abort signals from poisoning subsequent tries and adheres to the existing retry budget. This **bug fix** prevents premature connection failures, improving the overall stability of WebSocket-based services. New tests in `packages/sandbox/tests/ws-transport.test.ts` validate this improved retry behavior.

This commit **updates the GitHub Actions configuration** by adding the `actions: read` permission to the `e2e` job in the `.github/workflows/pr-privileged.yml` workflow. This **maintenance fix** addresses an issue where the **PR E2E workflow** would fail to execute due to insufficient permissions. By granting `actions: read`, the workflow can now successfully download required build artifacts, ensuring that end-to-end tests run properly for pull requests. This change **enables the E2E workflow to complete its checks**, preventing GitHub from rejecting it prematurely.

Improve structured logging across SDK and container (#456)

This commit implements a **bug fix** for the **AI chat persistence logic**, specifically ensuring that **Anthropic replay tokens** are correctly preserved. Previously, these tokens, vital for maintaining conversational context with Anthropic models, were not properly persisted. This **fix** modifies `packages/ai-chat/src/index.ts` by updating truncation methods and introducing new helper functions like `_truncateProviderExecutedToolPayloads` and `_isOpaqueReplayFieldKey` to handle these tokens. The change improves the reliability of **AI chat interactions** by preventing data loss related to Anthropic model state, and includes new **tests** in `packages/ai-chat/src/tests/sanitize-messages.test.ts` to validate the correct behavior.

This commit delivers a crucial **bug fix** to the **`ai-chat`** package, significantly improving the reliability and efficiency of its **auto-continuation** feature. It **coalesces rapid auto-continuation turns** to prevent duplicate model runs and redundant output, while also **preserving origin-tab continuation resume** to ensure stable handshakes and **queuing late follow-up results** to cover pre-stream race conditions. These changes, along with refined connection ID cleanup and synchronous queuing logic in `_enqueueAutoContinuation`, prevent redundant processing, enhance stream stability, and ensure proper resource management within the chat system. New regression tests have been added to validate these critical fixes.

This commit delivers a **bug fix** for the **AI chat agent** that resolves an issue where other chat tabs would stall or stop streaming during auto-continuations (e.g., tool results or approvals) initiated in a different tab. The `packages/ai-chat/src/index.ts` module was updated to correctly manage stream resume requests, ensuring only the originating connection waits for the continuation while others continue to receive updates, improving the multi-tab user experience. This fix introduces `_pendingAutoContinuationConnectionId` to track the active continuation. Additionally, new **tests** were added to `packages/ai-chat/src/tests/client-tools-continuation.test.ts` to stabilize and prevent regressions for multi-tab approval scenarios.

This commit provides a **fix** for the **AI chat Durable Object** to prevent it from hibernating prematurely during **tool continuation streaming**. It wraps the `_queueAutoContinuation` logic within a `keepAliveWhile` block, ensuring the Durable Object remains active from queueing through streaming. This resolves an issue where the DO could hibernate between receiving a tool result and starting the continuation stream, leading to **silently dropped LLM follow-ups** in the `ai-chat` module. Additionally, it **enhances test coverage** for various continuation edge cases, including empty bodies, chat clearing, and transport send failures, and improves test reliability by introducing event-based helpers to replace flaky sleep-based waits.

This commit delivers a crucial **bug fix** within the **`ai-chat` package** to **preserve assistant messages across chained continuations**, ensuring conversational context is maintained during multi-turn interactions. It addresses issues with auto-continuation logic, stream management, and tool continuations, preventing the loss or duplication of assistant messages. The changes impact the core chat logic (`index.ts`), React hooks (`react.tsx`), and WebSocket transport (`ws-chat-transport.ts`), alongside comprehensive test updates. This **enhances the reliability and consistency** of the AI chat system, providing a more robust and predictable user experience for complex conversational flows.

This commit introduces a **new cleanup mechanism** for **CI/CD workflows** to eagerly manage temporary Cloudflare resources. It adds a `cleanup-resources` job to `.github/workflows/pr-privileged.yml` that **deletes PR test workers, containers, and image tags immediately** after privileged PR jobs complete, preventing resource accumulation. A key refinement ensures that `pr-tagged` images are **preserved for failed reruns**, allowing GitHub's rerun functionality to reuse successful build outputs without requiring a full rebuild. Additionally, `.github/workflows/cleanup.yml` is **refactored** to use the `--json` flag for `wrangler containers list` commands and clarify comments, improving the overall robustness and clarity of the cleanup process. This **maintenance feature** significantly reduces resource waste and improves the efficiency of the PR testing environment.

This commit **updates the changelog for Agents SDK v0.8.0**, adding a new entry to `src/content/changelog/agents/2026-03-23-agents-sdk-v0.8.0.mdx`. It details significant new features such as readable state, idempotent schedules, a typed `AgentClient`, and the migration to Zod 4. The work involved initial creation, removal of an unreleased turn control API mention, and subsequent **documentation fixes** for bare class methods and unclosed code fences. This **maintenance** ensures users have an accurate and well-formatted record of changes in the **Agents SDK**, improving clarity on new capabilities and updates.

This commit introduces a significant **refactoring** and **API simplification** for **turn coordination** within the **`ai-chat` module**. The method `waitForPendingInteractionResolution()` has been renamed to `waitUntilStable()`, which now comprehensively waits for a fully stable conversation state, including any queued continuation turns, addressing previous race conditions. Public methods like `waitForIdle()` and `abortActiveTurn()` are demoted to private, with `resetTurnState()` exposed as the new public API for aborting active turns and invalidating continuations. These changes, along with **bug fixes** for pending interaction bookkeeping, aim to provide a more robust and intuitive API for managing chat agent lifecycle, requiring developers to **update their usage patterns** as detailed in the extensive **documentation updates** across the `README.md` and `docs/chat-agents.md`.

This commit **reduces the default `waitIntervalMS`** from 1000ms to 300ms for the **`Sandbox` component**. This **configuration adjustment** primarily affects the `packages/sandbox` module, making its internal waiting mechanism more responsive by default without requiring explicit configuration. The change is implemented in the `Sandbox` class's default options, with corresponding updates to the JSDoc for `waitIntervalMS` in `SandboxOptions` and relevant test cases. This aims to improve the default responsiveness and performance of sandbox operations.

This commit delivers a **bug fix** and **enhancements** to the **`ai-chat` package**, primarily focusing on **serializing chat turns** and improving the reliability of `saveMessages` operations. It introduces a monotonic `_chatEpoch` to prevent stale continuations and `saveMessages` calls from executing after a chat clear, ensuring **data consistency** and preventing race conditions within the `AIChatAgent`. New helper methods like `hasPendingInteraction` and `waitForIdleForTest` are added, along with comprehensive tests, to manage and verify chat turn states and tool-result continuations. This significantly improves the **robustness and predictability** of the AI chat agent's interaction flow, especially when dealing with rapid user input or chat resets.