Spring cleaning: upstream Rook seams, fix the transformed-await gate, harden the OPFS pool

.changeset/spring-cleaning-browser-host.md

@@ -0,0 +1,42 @@
+---
+"@mcp-b/do-runtime": minor
+---
+
+Ship the Worker half of the browser alarm protocol. `createBrowserAlarmProjector()`
+supplies the `AlarmScheduler`'s `projectWake` and an `acknowledge()` for the
+`BrowserAlarmCoordinator`'s `deliver()`. Projections leave one at a time, and each draws
+its generation only after the previous one was sent. A consumed wake is acknowledged only
+after the latest projection is accepted, no delivery or cleanup is active, and the next
+wake is absent or later. If the latest projection failed, `acknowledge()` sends it again
+first, so a scheduler with nothing new to project cannot leave the wake retrying forever.
+`nextGeneration()` must be durable across Worker restarts: the coordinator silently drops
+any projection older than the generation it journaled, so an in-memory counter would stall
+every wake after a restart. `parseBrowserAlarmProjection()` is now exported.
+
+Add `connectMessagePortWebSocket()` to `@mcp-b/do-runtime/browser`. It routes one
+MessagePort socket through a Workers-style `fetch` such as the Agents SDK's
+`routeAgentRequest()`. A socket nothing routes closes with 1011, and a refused upgrade
+closes with 1008 and the refusal's text when it is printable ASCII of at most 123 bytes.
+Previously every failure closed with a generic 1011 that dropped the Agent's reason.
+`serveMessagePortWebSockets()` now takes `(bridge, url) => Promise<void>`, such as
+`(bridge, url) => connectMessagePortWebSocket(bridge, url, route)`, instead of a function
+resolving a URL to a socket. It reports a connection failure after closing the client,
+and a socket that finishes connecting after `stop()` now closes with
+"MessagePort transport closed" instead of "host stopped".
+
+Gate a hibernatable socket that is a host transport, such as a `MessagePortWebSocket`
+rehydrated after a Worker restart. The actor used to receive the transport itself, so its
+`send()` and `close()` could leave before a preceding storage write was confirmed. It now
+receives one stable socket per transport that waits for the output gate like a
+`WebSocketPair` half; hibernation hosts still see the transport.
+
+`OffscreenDocumentAdapter` gains optional `ready()` and `replaceUnready()` hooks. The
+coordinator runs readiness in the same single flight as creation, for new and existing
+documents, and replaces a document that fails it at most once.
+
+Fix two `MessagePortWebSocket` hangs. A throwing `onmessage`, `onopen` or `onclose`
+handler skipped the `addEventListener` listeners behind it; during the open flush it also
+dropped the queued frames and held every later frame in the queue. Handler errors are now
+reported the way `EventTarget` reports a throwing listener. A throw while bridging a
+connected socket, such as a second `accept()`, left the brokered client connecting; it now
+closes with 1011.

.changeset/spring-cleaning-maintenance.md

@@ -0,0 +1,24 @@
+---
+"@mcp-b/do-runtime": patch
+---
+
+Validate both `withEnvAndExports()` scopes before installing either. A non-object `exports`
+argument previously left the `env` scope installed for the whole realm, so every later `env`
+read resolved against it.
+
+Treat a `Timer.afterDelay` that rejects on abort, as `node:timers/promises` does, as
+cancellation. Replacing or deleting a waiting alarm previously raised an unhandled
+`AbortError`, which terminates a Node host by default. The `Timer` contract now states that
+an aborted wait may stay pending or reject, but must not resolve.
+
+Close the databases `createActorContainer()` opened when an open-time check refuses, such as
+storage written by a newer release. Each refused attempt previously leaked its handles: a
+retry opened another connection to the same OPFS file, and on Node the provider's
+`exportSnapshot()` refused from then on. A `FacetHost.abort` that throws is now recorded like a
+failed facet deletion instead of becoming an unhandled rejection.
+
+Report `rowsWritten` as 0 for statements without result columns that write nothing, in both
+SQLite backends. DDL previously repeated the last write's count, because SQLite does not
+reset `sqlite3_changes()` for it. `SqliteWasmActorStorage.copyFrom()` now reads every source
+file in the same task as its recovery-sidecar check, so a source that is still running cannot
+open a write transaction between the two.

.changeset/spring-cleaning-opfs-pool.md

@@ -0,0 +1,15 @@
+---
+"@mcp-b/do-runtime": minor
+---
+
+Add `installSqliteWasmHost()` to `@mcp-b/do-runtime/backends/sqlite-wasm`. It installs an OPFS SAH
+pool through sqlite-wasm's `installOpfsSAHPoolVfs()` with the `name`, `directory`, `clearOnInit`
+and `initialCapacity` options, and returns the `{ pool, capi }` host the sqlite-wasm provider
+takes. Concurrent calls for one pool share a single install. Install pools through it:
+
+- It waits up to 10 seconds for a terminated worker to release the pool before installing, then
+  rejects with a `NoModificationAllowedError`. Installing while the previous worker was still
+  shutting down could delete the pool's directory and every database in it.
+- A transaction interrupted by worker termination is now rolled back when its database is
+  reopened. Previously, reopening could expose the interrupted transaction's writes in place of
+  committed rows and left the recovery journal behind, so snapshot export stayed refused.

.changeset/spring-cleaning-scope-and-vite.md

@@ -0,0 +1,14 @@
+---
+"@mcp-b/do-runtime": minor
+---
+
+Export `ACTOR_SCOPE_GLOBALS` from `@mcp-b/do-runtime` and `facetScopeBanner()` from
+`@mcp-b/do-runtime/vite`. The banner binds every name `installActorScope()` writes to a
+facet bundle's own scope. Banners that bound only timers, `fetch` and `crypto` left
+`WebSocketPair` on the root actor's global, so a facet's socket frames waited on the root's
+output gate and could leave before the facet's own write committed.
+
+`doRuntimeAwaitTransform()` now resolves the imports it injects to the package's own files,
+ahead of any host alias for them, so hosts no longer alias `@mcp-b/do-runtime/gate` or
+`@mcp-b/do-runtime/browser/async-hooks`. Add a `@mcp-b/do-runtime/cloudflare-email` export,
+a data-only `EmailMessage` for hosts to alias `cloudflare:email` to.

.changeset/spring-cleaning-transformed-await-gate.md

@@ -0,0 +1,21 @@
+---
+"@mcp-b/do-runtime": patch
+---
+
+Keep the input lock and the implicit transaction across transformed awaits that settle while
+the actor still holds its lock. `doRuntimeAwaitTransform()` previously resumed every await
+through a macrotask hop and a fresh input lock queued behind waiting events, including awaits
+of storage calls and plain values. Another event could then run between `await
+storage.get()` and the following `put()`, so two concurrent read-modify-writes lost an update.
+The implicit transaction also committed at the await, so writes survived an abort that workerd
+rolls back. A transformed `await ctx.blockConcurrencyWhile()` let queued events run before its
+caller resumed. Every consumer that transforms the Agents SDK was exposed on its storage
+sequences.
+
+A transformed await now continues in the same checkpoint when its promise settles while the
+actor holds its lock: storage calls, plain values, and resumptions the runtime already admitted.
+Settled awaits no longer pay a macrotask each. An await on foreign I/O still re-enters through
+a fresh lock. When another actor's continuation owns the checkpoint, the await waits for a later
+task without releasing its lock, and its implicit transaction commits at that hand-off. Code that
+relied on a storage-only await to let other events in now blocks them, as on workerd. The Node
+and browser conformance lanes now also run the suite with the probe compiled by the transform.

.github/dependabot.yml

@@ -18,10 +18,5 @@ updates:
       actions:
         patterns:
           - "*"
-    ignore:
-      # changesets/action v2 requires Changesets CLI v3; we are on CLI v2.
-      # Lift this when the CLI is upgraded.
-      - dependency-name: changesets/action
-        update-types: ["version-update:semver-major"]
     commit-message:
       prefix: "ci(deps)"

.github/workflows/ci.yml

@@ -69,8 +69,15 @@ jobs:
       - name: Test browser conformance
         run: pnpm test:conformance-browser
 
-      - name: Build examples
-        run: pnpm examples:build
+      - name: Test transformed Node conformance
+        run: pnpm test:conformance-node-transformed
+
+      - name: Test transformed browser conformance
+        run: pnpm test:conformance-browser-transformed
+
+      # The extension e2e builds the extension itself; vibe's e2e uses the dev server.
+      - name: Build vibe-platform for production
+        run: pnpm --filter do-runtime-example-vibe-platform build
 
       - name: Test examples
         run: pnpm test:examples

.gitignore

@@ -1,5 +1,6 @@
 node_modules/
 dist/
+.sdk-pack/
 .tsbuild/
 .wrangler/
 __screenshots__/