Version Packages

[github-actions]

This PR was opened by the Changesets release GitHub action. When you're ready to do a release, you can merge this and the packages will be published to npm automatically. If you're not ready to do a release yet, that's fine, whenever you add more changesets to main, this PR will be updated.

Releases

@cloudflare/worker-bundler@0.1.0

Minor Changes

• #1277 0cd0487 Thanks @zebp! - Introduce FileSystem abstraction for all bundler APIs.

  The files option on createWorker and createApp now accepts any FileSystem
  implementation in addition to a plain Record<string, string>. This lets callers
  back the virtual filesystem with persistent or custom storage — for example, a
  DurableObjectKVFileSystem that buffers writes in memory and flushes to Durable
  Object KV on demand, avoiding a KV write for every individual file operation.

  Three concrete implementations are exported from the package:

  • InMemoryFileSystem — a Map-backed filesystem suitable for tests and
    in-process pipelines. Accepts an optional seed object or Map of initial
    files.
  • DurableObjectKVFileSystem — a Durable Object KV-backed filesystem with a
    write-overlay. Writes accumulate in memory and are flushed to KV in one batch
    when flush() is called. Reads are served from the overlay first, so callers
    always observe their own writes immediately.
  • DurableObjectRawFileSystem — a thin Durable Object KV-backed filesystem
    with no buffering. Every write is committed to KV synchronously. Use when
    per-write durability is preferred over batching.

  createFileSystemSnapshot creates an InMemoryFileSystem from any sync or
  async iterable of [path, content] pairs, bridging async storage backends
  (e.g. Workspace from @cloudflare/shell) to the synchronous FileSystem
  interface.

  The FileSystem.read() method returns string | null (null = file does not
  exist) rather than an empty string, eliminating the need for a separate
  exists() check.

  Plain Record<string, string> objects continue to work unchanged — they are
  wrapped in an InMemoryFileSystem automatically.

• #1277 0cd0487 Thanks @zebp! - Export installDependencies, hasDependencies, and InstallResult so callers
  can pre-warm a FileSystem with npm packages independently of createWorker or
  createApp.

  When createWorker or createApp encounter a FileSystem that already contains
  a package under node_modules/, that package is skipped during installation,
  avoiding redundant network fetches. This makes a second call to
  installDependencies (or the internal call inside createWorker) a no-op for
  packages that were pre-installed into the same FileSystem.

• #1277 0cd0487 Thanks @zebp! - Add in-process TypeScript language service via createTypescriptLanguageService.

  createTypescriptLanguageService wraps any FileSystem in a
  TypescriptFileSystem that mirrors every write and delete into an underlying
  virtual TypeScript environment. Diagnostics returned by the language service
  always reflect the current state of the filesystem — an edit that fixes a type
  error immediately clears getSemanticDiagnostics.

  TypeScript is pre-bundled as a browser-safe artifact so it runs inside the
  Workers runtime without Node.js APIs. Lib declarations are fetched from the
  TypeScript npm tarball at runtime.

  Exposed under a separate ./typescript subpath export to keep the TypeScript
  bundle out of the main import path.

agents@0.10.1

Patch Changes

• #1286 d76f8b9 Thanks @threepointone! - Fix McpAgent.handleMcpMessage crashing with "Attempting to read .name before it was set" when the Durable Object wakes from hibernation via native DO RPC. The method now calls __unsafe_ensureInitialized() to hydrate this.name from storage and run onStart() before processing messages, matching the pattern used by _workflow_* RPC methods and alarm().

• #1278 8c7caab Thanks @threepointone! - Think now owns the inference loop with lifecycle hooks at every stage.

  Breaking: onChatMessage(), assembleContext(), and getMaxSteps() are removed. Use lifecycle hooks and the maxSteps property instead. If you need full custom inference, extend Agent directly.

  New lifecycle hooks: beforeTurn, beforeToolCall, afterToolCall, onStepFinish, onChunk — fire on every turn from all entry paths (WebSocket, chat(), saveMessages, auto-continuation).

  beforeTurn(ctx) receives the assembled system prompt, messages, tools, and model. Return a TurnConfig to override any part — model, system prompt, messages, tools, activeTools, toolChoice, maxSteps, providerOptions.

  maxSteps is now a property (default 10) instead of a method. Override per-turn via TurnConfig.maxSteps.

  MCP tools auto-merged — no need to manually merge this.mcp.getAITools() in getTools().

  Dynamic context blocks: Session.addContext() and Session.removeContext() allow adding/removing context blocks after session initialization (e.g., from extensions).

  Extension manifest expanded with context (namespaced context block declarations) and hooks fields.

@cloudflare/ai-chat@0.4.1

Patch Changes

• #1277 0cd0487 Thanks @zebp! - Fix race condition in messageConcurrency where rapid overlapping submits could bypass the latest/merge/debounce strategy. The concurrency decision checked queuedCount() before the turn was enqueued, but an intervening await persistMessages() allowed a second message handler to see a stale count of zero and skip supersede checks. A pending-enqueue counter now bridges this gap so overlapping submits are always detected.

• #1272 22da9b1 Thanks @threepointone! - Widen useAgentChat agent prop type to accept both typed and untyped useAgent connections. Previously, useAgent<MyAgent>() results could not be passed to useAgentChat due to incompatible call types. The agent prop now uses a structural type matching only the fields useAgentChat actually uses.

@cloudflare/think@0.2.1

Patch Changes

• #1275 37b2ce3 Thanks @threepointone! - Add built-in workspace to Think. Every Think instance now has this.workspace backed by the DO's SQLite storage, and workspace tools (read, write, edit, list, find, grep, delete) are automatically merged into every chat turn. Override workspace to add R2 spillover for large files. @cloudflare/shell is now a required peer dependency.

• #1278 8c7caab Thanks @threepointone! - Think now owns the inference loop with lifecycle hooks at every stage.

  Breaking: onChatMessage(), assembleContext(), and getMaxSteps() are removed. Use lifecycle hooks and the maxSteps property instead. If you need full custom inference, extend Agent directly.

  New lifecycle hooks: beforeTurn, beforeToolCall, afterToolCall, onStepFinish, onChunk — fire on every turn from all entry paths (WebSocket, chat(), saveMessages, auto-continuation).

  beforeTurn(ctx) receives the assembled system prompt, messages, tools, and model. Return a TurnConfig to override any part — model, system prompt, messages, tools, activeTools, toolChoice, maxSteps, providerOptions.

  maxSteps is now a property (default 10) instead of a method. Override per-turn via TurnConfig.maxSteps.

  MCP tools auto-merged — no need to manually merge this.mcp.getAITools() in getTools().

  Dynamic context blocks: Session.addContext() and Session.removeContext() allow adding/removing context blocks after session initialization (e.g., from extensions).

  Extension manifest expanded with context (namespaced context block declarations) and hooks fields.

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.