whoiskatrin

WIP - Add Telnyx voice provider (#1461) * Add Telnyx voice provider * Add empty changeset for Telnyx provider * Address Telnyx provider review feedback * Fix Telnyx example typecheck * Update voice-providers/telnyx/src/transport/phone-transport.ts Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com> * Refocus Telnyx example on phone bridge * dependency version pinning * port telnyx tests * Bump deps, add aria-label, adjust tests Update dependency versions in examples/telnyx-voice-agent/package.json (bumping @cloudflare/kumo, ai, react, and several devDependencies) to newer releases for compatibility. Add an aria-label ("Text message") to the text input in the Telnyx voice agent UI to improve accessibility. Harden durable-chat recovery tests by checking scheduled callback counts before running retry/continue tasks and waiting for the agent to become idle, reducing flakiness in CI. * Harden Telnyx voice provider lifecycle and exports Keep the Telnyx package root server-safe by moving browser WebRTC helpers to the /browser entrypoint, and update the example/docs to use the clearer root-plus-browser import model. This avoids pulling @telnyx/webrtc into Worker bundles while keeping STT, TTS, and JWT helpers available from server-safe paths. Tighten the phone bridge and credential lifecycle so reconnects preserve the preferred audio format, stale async bridge setup is cancelled without leaking browser audio resources, in-flight starts settle on stop, late STT WebSockets are closed, and failed JWT token creation cleans up orphaned Telnyx credentials. The example now warns about local-only unauthenticated token minting, handles overlapping connect/disconnect attempts, and documents the live browser bridge requirement. Co-authored-by: Cursor <cursoragent@cursor.com> * Notify on websocket close error Ensure the TTS stream is unblocked if websocket.close() throws by calling notify() in the close handler catch. Add a test that simulates close() throwing (via MockWebSocket) and verifies aborting the request immediately resolves the synthesize promise to null. --------- Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com> Co-authored-by: Christopher Little-Savage <clittle-savage@cloudflare.com> Co-authored-by: Sunil Pai <spai@cloudflare.com> Co-authored-by: Cursor <cursoragent@cursor.com>

fix: stop oversized sessions from bricking the DO with SQLITE_NOMEM on wake (#1724) * fix: stop oversized sessions from bricking the DO with SQLITE_NOMEM on wake (#1710) Four coordinated changes across agents + @cloudflare/think: 1. AgentSessionProvider.getHistory() no longer carries message content through the recursive CTE and its ORDER BY sorter (2-3 transient copies of the whole transcript inside SQLite's allocator); content is fetched in bounded chunks via json_each. 2. Think.onStart degrades instead of throwing when a data-driven step fails (transcript hydration, declared-task reconcile, durable-work recovery) — a throw there is re-run on every wake, including alarm retries, permanently bricking the DO. 3. hydrationByteBudget (default 24MB): oversized transcripts hydrate as a bounded recent window instead of materializing fully in memory. 4. mediaEviction (default on): aged inline media (data-URL file parts, large strings in tool outputs) is evicted from stored messages in background passes, preserved as workspace files by default. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> * docs(think): document budget-exemption contract on _readMessagesFromStorage (#1710) * fix: close remaining full-history reads and harden memory bounds (#1710) Deep-review follow-up to the SQLITE_NOMEM fix. The original change bounded onStart hydration, but three helper paths still materialized the full transcript, undermining the budget; this closes them and hardens the new mechanisms. Session layer (agents): - Skill restore no longer bypasses budgeted hydration. The init-time loaded-skill scan is skipped entirely when no skill-capable context provider is configured, and when one is, it enumerates rows via getHistoryRowStats and fetches assistant messages ONE AT A TIME instead of reading the full history (full-read fallback for providers without row stats). A skill block added later via addContext() triggers the scan at that point. HistoryRowStat gains `role` to support the filter. - New Session.internal_rewriteMessage(): maintenance write path that skips the public updateMessage side effects (status broadcast + its FULL-history token estimate) while still notifying the message-change listener. Media eviction rewrites rows through it, so a pass no longer triggers a full-history read per rewritten row. - getRecentHistory(maxContentBytes, minRecentMessages?) gains a window floor: the most recent N rows are always included even when they exceed the byte budget (rows are write-capped, so the floor stays bounded). - Honest fallback: providers without getRecentHistory now report the real serialized size (not 0) and warn once that the budget is unenforced. - Content hydration chunks are bounded by cumulative stored bytes (4MB) as well as row count, removing the 50-near-cap-rows (~90MB) worst case. Think: - Budgeted hydration passes MODEL_RECENT_WINDOW (4 — the truncateOlderMessages default) as the floor, so windowing can never shrink this.messages below the span the model replays at full fidelity. - Media eviction: keepRecentMessages is clamped to that same window (a misconfigured low value can never strip content the model still sees); a pass that stops at maxRowsPerPass with eligible rows remaining schedules the next pass itself so backlogs drain; providers without row-stats support warn once instead of silently no-opping. - chat:hydration:windowed emits on change rather than on every safe- boundary sync (a chronically oversized session syncs many times per turn and would spam identical events). - Public getOnStartDegradations() accessor; stale onStart/JSDoc comments updated to describe the budgeted behavior. Tests (~30 new): restore-scan gating and bounded-restore call counts via stub providers, addContext late-skill scan, honest fallback metadata, silent-rewrite no-broadcast contract, floor semantics and corrupt-leaf behavior at the provider, role in row stats, 6MB multi-chunk round-trip, pure-function coverage for media-eviction.ts (markers, shape preservation, depth limit, no-mutation), eviction clamp and automatic pass chaining, and observability event assertions for chat:onstart:degraded / chat:hydration:windowed / chat:media:evicted. Co-authored-by: Cursor <cursoragent@cursor.com> --------- Co-authored-by: Claude Fable 5 <noreply@anthropic.com> Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com> Co-authored-by: Sunil Pai <spai@cloudflare.com> Co-authored-by: Cursor <cursoragent@cursor.com>

feat: add file watching capabilities with inotify support (#324) * feat: add file watching capabilities with inotify support * Create empty-poets-serve.md * Potential fix for code scanning alert no. 42: Incomplete string escaping or encoding Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com> * fixes for claude review * update tests to verify regex format for default and custom excludes * Fix error handling and type safety in WatchService and FileWatch classes. Update tests to validate new event parsing logic and ensure proper handling of inotifywait output. * Refactor WatchService tests to validate combined regex patterns for default and custom excludes * Added timeouts for event propagation * Refactored and cleaned * Small ws transport related fixes * Timing changes help account for the additional buffering * Fix WebSocket blocking issue for SSE streaming responses Streaming responses (like file watch events) were blocking the WebSocket message handler because handleStreamingResponse was awaited. This prevented other messages from being processed while a stream was active. Run streaming response handlers in the background with error logging, allowing the message handler to return immediately and process subsequent messages. * Add debug logging and stream tracking for WebSocket streaming Add detailed logging to trace streaming response handling and track active streams to prevent potential garbage collection of Response objects. * Acquire stream reader synchronously before async execution The WebSocket message handler needs to capture the Response body reader before any await points. When getReader() was called inside the async handleStreamingResponse method, Bun's WebSocket handler would return before the reader was acquired, potentially invalidating the Response body stream. By getting the reader synchronously in handleRequest before the promise starts executing, we ensure the stream remains valid throughout the async streaming loop. * Wait for inotifywait watches to be established before signaling ready The watching SSE event was sent immediately when the stream started, before inotifywait finished setting up watches. This caused flaky tests because file operations could occur before the watch was truly ready. Now we read stderr and wait for the 'Watches established' message from inotifywait before sending the watching event to clients. * Add timeout to waitForWatchesEstablished to prevent hanging If inotifywait fails to output 'Watches established' within 10 seconds, the function will return and allow the stream to proceed. This prevents indefinite hangs if inotifywait behaves unexpectedly. * Wait for first message before returning WebSocket stream For WebSocket streaming, errors were deferred until stream consumption. This caused issues where watchStream() would return successfully even when the server returned an error response. Now requestStream() waits for the first message before returning: - If it's a stream chunk, return the stream (success case) - If it's an error response, throw immediately (error case) This makes WebSocket streaming behavior match HTTP streaming, where errors are thrown immediately rather than deferred. * Address code review issues in file watching Replace empty catch blocks with debug/warn logging throughout WatchService and FileWatch to make failures visible. Fix FileWatch.established() to reject on AbortSignal during establishment, preventing indefinite hangs. Strengthen the SSE event type guard to validate required fields per event type. Add WATCH_STOP_ERROR code, integrate watch cleanup into server shutdown, and rewrite changeset for end users. * Fix file watch E2E test timeout race condition The watchWithActions timeout started at stream creation but blocked for ~6.5s inside the event handler (pre-action delay + file ops + post-action delay), leaving insufficient time to read events in CI. Reset the timeout after actions complete so the full window is available for event collection. * Fix exclude test timeout in file watch E2E The combined wait time (~18s) approached the 30s Vitest timeout. With excludes filtering most events, only 2-4 arrive, so the high stopAfterEvents threshold was never reached and the test always fell through to the full 12s reader timeout. * Remove low-level watch API to simplify public interface Remove watchStream(), stopWatch(), and listWatches() methods from Sandbox class. The handle-based API via watch() is sufficient for all use cases and prevents resource management confusion. Keep the internal WatchService methods for container use but don't expose them through the SDK public API. * Harden file watch stream lifecycle * Stabilize explicit watch stop e2e test * Isolate file-watch e2e sessions per test * Stabilize file-watch error and stop e2e cases * Stabilize websocket e2e flake handling Treat watch stop as idempotent when a watcher is already gone and make OpenCode proxy health checks resilient to transient startup failures. Relax the foreground timing threshold to reduce transport-related CI jitter without masking blocking behavior. * Treat ESRCH as success when stopping already-gone watch process Handle the race where a watch process exits before stopWatch() is called. When process.kill() throws ESRCH (no such process), clean up the watch entry and return success instead of an error. * Isolate file watch e2e sandbox * Route watch E2E through SDK surface * Use SDK watch bridge without new public API * Remove unused watch stop endpoint and reduce stream logging verbosity Remove /api/watch/stop endpoint and WatchClient.stop() method as watch lifecycle is now managed through handle-based API. Clean up verbose debug logging in WebSocket stream handler, keeping only completion summary. Clear setup timeout in watch service to prevent leak. * Scope stream cancellation to owning WebSocket connection Track connectionId for each active stream and only allow cancellation from the connection that initiated it. Prevent cross-connection cancel messages and ensure onClose only cancels streams owned by the closing connection. * Address review: parseSSEStream abort, watch readiness, include/exclude validation - Fix parseSSEStream to register abort listener that calls reader.cancel(), unblocking idle streams instead of polling signal.aborted between reads - Make watch() block until the watching event is received so callers can immediately perform actions that depend on the watcher being active - Reject requests that specify both include and exclude with a clear validation error since inotifywait does not support both simultaneously - Add timeout-race tests for WebSocket transport stream requests - Simplify E2E watch helper: use parseSSEStream + AbortSignal.timeout, remove all hardcoded delays - Remove test worker workaround that silently dropped exclude when include was present * Simplify non-existent path error test watch() now throws on establishment failure, so the test worker returns a clear error response. The 3-path defensive logic is no longer necessary. * Simplify stop-watch test: remove manual watching detection watch() now blocks until established, so the manual loop scanning for the watching event is redundant. Read one chunk to confirm the stream is live, then cancel. * chore: retrigger CI * Update .changeset/empty-poets-serve.md Co-authored-by: Naresh <ghostwriternr@gmail.com> * FIxed small review comments --------- Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com> Co-authored-by: opencode-agent[bot] <opencode-agent[bot]@users.noreply.github.com> Co-authored-by: Naresh <ghostwriternr@gmail.com>

fix(ai-chat): preserve assistant messages across chained continuations (#1162) * Bump core deps: agents, hono, AI providers Upgrade various dependencies across the repo: bump @openai/agents and @openai/agents-extensions to ^0.8.0, workers-ai-provider to ^3.1.7, hono to ^4.12.9, and ai-gateway-provider to ^3.1.2. Update package-lock.json to reflect these new versions (including updated optional provider sub-dependencies). Also apply minor formatting/whitespace cleanup in ai-chat and codemode CHANGELOGs and update example and site package.json files to use the new versions. * fix(ai-chat): preserve assistant messages across chained continuations * Clear _pendingAutoContinuation when continuation response has no body * fix(ai-chat): reuse resume handlers for tool continuations * Update packages/ai-chat/src/index.ts Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com> * fix(ai-chat): close stream controller when resume request send fails --------- Co-authored-by: Sunil Pai <spai@cloudflare.com> Co-authored-by: devin-ai-integration[bot] <158243242+devin-ai-integration[bot]@users.noreply.github.com>

add microsite to the repository (#585) * add microsite to the repository * Update .gitignore * cleanup * revert cached change from another PR * revert cached change from another pr * fix all the any types * config fixes * Update package-lock.json * github lightningcss --------- Co-authored-by: Sunil Pai <spai@cloudflare.com>

Add PTY terminal passthrough for browser clients (#310) * Add PTY types to shared package * Add PtyManager for container PTY lifecycle * Add PTY handler and route registration * Add PTY message handling to WebSocket adapter * Add PTY methods to transport interface * Add PtyClient for SDK PTY operations * Add pty namespace to Sandbox class * Add E2E tests for PTY workflow * Add changeset for PTY support * Skip PTY tests when PTY allocation fails * Fix pty manager tests * fix any types, logger * fix silent logging * fix pty tests for resizing * update claude review yml * revert review change * update http tests * more test updates * remove the plugin for review * Add error handling to PTY callbacks and terminal operations * Improve PTY error handling based on code review * add structured exit codes * change fire and forget strategy * add more e2e tests * more fixes and tests * fix ws * update error propagation * update resizing tests * add collab terminal example * Potential fix for code scanning alert no. 40: Insecure randomness Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com> * Potential fix for code scanning alert no. 41: Insecure randomness Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com> * Update dependency in examples * Add PTY listeners cleanup * minor nits * update tests setup * Add logging for PTY listener registration errors and improve error handling * Enhance error handling and logging in WebSocketTransport and PtyHandler; add tests for PTY listener registration and cleanup behavior * Add connection-specific PTY listener cleanup on WebSocket close * Remove outdated comment regarding connection cleanup functions in WebSocketAdapter * Fix error handling in PTY management by updating kill method to return success status and error messages * implement signal handling for Ctrl+C, Ctrl+Z, and Ctrl+\ in the PTY manager * Potential fix for code scanning alert no. 43: Insecure randomness Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com> * Changes based on review comments * Update dependencies and improve PTY handling in collaborative terminal example * Update PTY workflow tests to expect correct HTTP status codes for error responses * extractPtyId method to retrieve PTY IDs from responses, and update handleRegularResponse to return parsed body for further processing * Update PTY workflow tests to expect 'message' field in error responses instead of 'error' * Fixed handlers for tests * Use PR-specific Docker build cache scope to avoid cross-PR cache pollution * Add debug logging to router for route registration and matching * Add INFO-level route logging to debug container caching issues * Add retry logic for WebSocket server readiness in e2e tests The WebSocket connect tests were flaky because they didn't wait for the echo server to be ready after /api/init. Added a helper function that retries WebSocket connection with backoff before running tests. * Remove debug logging added during PTY route investigation The 404 issues were caused by stale container instances, not route registration problems. Reverting the debug logging changes: - Remove INFO-level route logging from router - Remove logRegisteredRoutes() method - Revert PR-specific Docker cache scope (not needed) * Fix sync-docs workflow to handle PR bodies with special characters Use quoted heredoc and printf to safely handle PR description content that may contain backticks, code blocks, or other shell-sensitive characters. Pass PR body via environment variable to prevent shell interpretation during prompt construction. * Fix sync-docs workflow shell escaping for opencode run Use environment variable to pass prompt to opencode run, avoiding shell interpretation of special characters like parentheses, backticks, and dollar signs that appear in PR descriptions with code examples. The prompt is stored in OPENCODE_PROMPT env var which GitHub Actions sets safely, then referenced with double quotes in the shell command. * Fix lint errors and align env type signatures The recent env var changes in 7da85c0 introduced Record<string, string | undefined> but missed updating getInitialEnv return type and getSessionInfo. Also aligns vite-plugin versions across examples. * send heartbeat events to keep container alive * Add PTY terminal passthrough for browser clients Enables browser-based terminal UIs to connect to sandbox shells via WebSocket. The terminal() method proxies connections to the container's PTY endpoint with output buffering for replay on reconnect. * Add tests and infrastructure for PTY terminal (#375) * Add tests and infrastructure for PTY terminal Unit tests for ring buffer, PTY spawning, and WebSocket handler. E2E tests for PTY workflow and browser terminal addon integration. Updates CI workflows and documentation for new test patterns. * Add collaborative terminal example and refine xterm addon (#376) * Refactor SandboxAddon to use connect(target) * Add collaborative terminal example Demonstrates the SandboxAddon connect() API with real-time room switching, presence tracking, and session isolation across multiple terminal rooms. Co-authored-by: Naresh <naresh@cloudflare.com> * Clear bash startup warning in container PTY Bash emits 'cannot set terminal process group' warnings in containers where the shell isn't a session leader. This clutters the first lines of terminal output for users. * Add build/** to turbo output cache React Router outputs to build/ instead of dist/, causing turbo to warn about missing outputs for the collaborative-terminal example. * Update collaborative-terminal README for new API The previous README documented the old PTY API (sandbox.pty.create() with JSON WebSocket protocol) that no longer exists. Updated to reflect the current implementation using session.terminal(request) with direct WebSocket passthrough and SandboxAddon for terminal integration. --------- Co-authored-by: katereznykova <kreznykova@cloudflare.com> --------- Co-authored-by: katereznykova <kreznykova@cloudflare.com> --------- Co-authored-by: Copilot Autofix powered by AI <62310815+github-advanced-security[bot]@users.noreply.github.com> Co-authored-by: Naresh <naresh@cloudflare.com> Co-authored-by: Steve James <sjames@cloudflare.com> Co-authored-by: Naresh <ghostwriternr@gmail.com>

add client-defined tools and prepareSendMessagesRequest options (#729) * add client-defined tools and prepareSendMessagesRequest options * fix client tools execution * minor nits * some claude review feedback * refactor * Refactor useAgentChat * more nits * take a different approach * cleanup after the refactor * fix param convertion * improve tool execution handling in useAgentChat * fix request body structure in useAgentChat to include id, messages, and trigger * review comments on types * filter broadcast id's out * fix types * fix msg duplication * fix the duplication issue in client tools * fix types * fix types * remove example from the git * Send tool results to server as source of truth * remove automatic message sending after tool results * cleanup the tests * fix OpenAI item ID stripping to prevent duplicate errors in persisted messages * minor nits * add server-side auto-continuation after tool results for improved UX; update message handling to merge tool results into existing assistant messages * update useAgentChat to include autoContinueAfterToolResult in message handling * fix for non human in the loop tools * update the tests * Update slow-jobs-boil.md with new options Add client-defined tools and options for message requests.

fix(ai-chat): serialize chat turns and saveMessages (#1142) * fix(ai-chat): serialize chat turns and saveMessages * fix(ai-chat): add pending interaction helpers * Skip stale continuations after chat clear Introduce a monotonic _chatEpoch (incremented on CF_AGENT_CHAT_CLEAR) so queued continuations or save operations enqueued under an older epoch are skipped after a chat clear. Wire apply/approval promises as prerequisites for auto-continuations (catching rejections) so continuations only run if the tool result/approval was applied, and check epoch before running exclusive chat turns. Snap clientTools/body in saveMessages and avoid running them if the epoch changed. Minor cleanup: remove an unnecessary _getAbortSignal call in cancel flow. Add tests and test helpers (isChatTurnActiveForTest, waitForIdleForTest, persistToolCallMessage, getMessageCount) to cover tool-result continuations, chat-clear skipping, and saveMessages behavior; replace some fixed delays with waitForIdleForTest. --------- Co-authored-by: Sunil Pai <spai@cloudflare.com>

feat: Add backup and restore API (#396) * Add backup and restore API for directory snapshots Introduces createBackup() and restoreBackup() methods to the Sandbox class, enabling point-in-time snapshots of directories. Backups are stored as compressed squashfs archives in R2, with instant restore via FUSE overlayfs (copy-on-write). Key features: - Compression with mksquashfs for efficient storage - Instant restore without extracting archives - Copy-on-write semantics preserve original backup - TTL-based automatic cleanup - Includes time-machine example and comprehensive E2E tests * Fix backup restore and stale E2E test expectations The doRestoreBackup finally block deleted the .sqsh archive after every restore, but squashfuse needs it as backing storage for the mount lifetime. Changed to catch-only cleanup so the file persists on success. Also use lazy unmount (-uz) in the container to handle stale FUSE mounts from previous restores. Updated E2E tests to expect correct HTTP status codes from the centralized error handling system (404 for not found, 408 for timeout, 400 for bad request) instead of blanket 500s. * Unmount FUSE before re-downloading backup archive Restoring the same backup a second time overwrote the .sqsh file while squashfuse still held it open, corrupting the backing store. Tear down existing FUSE mounts in the DO before writing the new archive. Also revert two E2E status expectations to 500 since IS_DIRECTORY and NOT_DIRECTORY errors are indistinguishable from generic FileSystemError across the RPC boundary. * Ensure clean upper layer on repeated backup restores When rm -rf on the mount base fails due to a lingering squashfuse mount point, the overlayfs upper directory survives with stale writes. Explicitly remove the upper and work directories first (never mount points) so the new overlay starts with a clean writable layer. * Add diagnostic logging to backup COW test * Use unique mount base per restore to avoid stale upper layer Each overlayfs restore now creates a fresh mount base with a unique suffix so writes from a previous overlay session cannot leak into subsequent restores. Old mount bases are torn down best-effort. * Skip archive re-download when file already exists A lazily-unmounted squashfuse may still hold the .sqsh file open. Overwriting it corrupts the backing store for the new mount. Compare the existing file size with the R2 object size and skip the download when they match. * Add targeted diagnostic for backup COW test * Remove diagnostic logging from backup COW test * Retrigger CI with clean container state * Use id in restore backup result * Revert chunked-read plumbing from file API The offset/length parameters on readFile were added for chunked base64 backup reads. With presigned URLs handling all transfers, chunked reads are unnecessary. * Replace base64 backup transfer with presigned R2 URLs Presigned URLs are the sole transfer mechanism for backup archives. The container curls directly to/from R2, bypassing the DO. This achieves 40-155x faster transfers (~24 MB/s up, ~93 MB/s down). Remove all base64 transfer code (uploadBackupSingle, Chunked, downloadBackupSingle, Chunked, uint8ArrayToBase64, base64ToUint8Array), the 500MB size cap, and the 50MB chunk constant. R2 credentials (R2_ACCESS_KEY_ID, R2_SECRET_ACCESS_KEY, CLOUDFLARE_ACCOUNT_ID, BACKUP_BUCKET_NAME) are now required for backup to function. * Add presigned URL config to test and example wrangler configs Add BACKUP_BUCKET_NAME var and R2 credential types for presigned URL backup transfers. Update time-machine example with required vars and secret instructions. * Update changeset to reflect presigned URL transfer approach * Harden presigned URL backup implementation Isolate backup shell ops in a dedicated session to prevent interference with user exec() calls. Encode URL path segments, use -sSf and --retry for curl, download to .tmp with atomic mv, and change TTL default to 3 days with no upper cap. * Use backup session for container API calls too * Add R2 credentials as wrangler secrets for backup E2E --------- Co-authored-by: Naresh <naresh@cloudflare.com>

Add message concurrency controls to AIChatAgent (#1192) Add `AIChatAgent.messageConcurrency` with queue, latest, merge, drop, and debounce strategies for overlapping sendMessage() submits. Enhance `saveMessages()` to accept a functional form for deriving messages from the latest transcript, and return { requestId, status } so callers can detect skipped turns. Test coverage for all strategies, onChatResponse interaction with each strategy, queued programmatic turns, and clear/epoch behavior. Made-with: Cursor Co-authored-by: Sunil Pai <spai@cloudflare.com>