Host cloud MCP sessions on our own DO with the v2 stack

[RhysSullivan]

The big consolidation step: Cloudflare MCP session serving moves off McpAgent onto our own Durable Object hosting the SDK v2 sessionful stack. patches/[email protected] (911 lines) is deleted and agents is out of every MCP path.

• The session DO extends DurableObject directly: per-session v2 WebStandardStreamableHTTPServerTransport + the sessionful v2 assembly, direct stub.fetch response streaming (the McpAgent/PartyServer WebSocket bridge and its delivery-ack protocol are gone — stream resumability is the event store's job).
• do-event-store.ts: DO-storage EventStore carrying the patch's semantics as first-class code — 64 events / 2 MiB per stream, never evict the newest, >120 KiB events deliver live but don't persist, best-effort writes that never break the send path, replay strictly after Last-Event-ID.
• sse-response-rotation.ts: 30-minute stream rotation with graceful reconnect framing, aligned with the alarm policy.
• Priming: v2's transport only primes 2025-11-25+ clients (verified in the installed dist); older clients on slow tool calls still need one benign id-bearing event before the result, ported minimally as our code. e2e/cloud/mcp-priming-reconnect.test.ts is the CI arbiter.
• Native elicitation flows through the SDK legacy shim like the neutral hosts; model/browser pause-resume, approval leases, cross-DO model resume, and the alarm policy keep their semantics (alarm arming is now fully ours).
• Session addressing keeps DO-unique-id session ids (idFromString); class names and wrangler bindings unchanged, so no DO migrations. Deliberate cutover behavior: sessions created before this deploy get one clean 404 and clients re-initialize transparently — no storage-compat restore of agents-internal state.
• Patch-internal test suites rewritten as behavior-equivalent tests of our components; a real v1-SDK client drives the DO-hosted stack end to end incl. reconnect-mid-call replay.

Tests: 57 hosts/cloudflare + 204 hosts/mcp + 24 host-cloudflare + 244 cloud (3 known-environmental failures pre-exist on parent). Key acceptance signal is CI's cloud e2e: priming reconnect, SSE replay, oversize response, session survival.

Stacked on #1606. Next PR deletes the v1 assembly and the V2 naming.

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Preview URL	Updated (UTC)
✅ Deployment successful! View logs	executor-marketing	d62e666	Commit Preview URL Branch Preview URL	Aug 16 2026, 03:59 PM

[cloudflare-workers-and-pages]

Deploying with    Cloudflare Workers

The latest updates on your project. Learn more about integrating Git with Workers.

Status	Name	Latest Commit	Updated (UTC)
✅ Deployment successful! View logs	executor-cloud	d62e666	Aug 16 2026, 04:00 PM

[github-actions]

Cloudflare preview

Torn down — the PR is closed.

[pkg-pr-new]

Open in StackBlitz

@executor-js/cli

npm i https://pkg.pr.new/@executor-js/cli@1608

@executor-js/config

npm i https://pkg.pr.new/@executor-js/config@1608

@executor-js/execution

npm i https://pkg.pr.new/@executor-js/execution@1608

@executor-js/sdk

npm i https://pkg.pr.new/@executor-js/sdk@1608

@executor-js/codemode-core

npm i https://pkg.pr.new/@executor-js/codemode-core@1608

@executor-js/runtime-quickjs

npm i https://pkg.pr.new/@executor-js/runtime-quickjs@1608

@executor-js/plugin-file-secrets

npm i https://pkg.pr.new/@executor-js/plugin-file-secrets@1608

@executor-js/plugin-graphql

npm i https://pkg.pr.new/@executor-js/plugin-graphql@1608

@executor-js/plugin-keychain

npm i https://pkg.pr.new/@executor-js/plugin-keychain@1608

@executor-js/plugin-mcp

npm i https://pkg.pr.new/@executor-js/plugin-mcp@1608

@executor-js/plugin-onepassword

npm i https://pkg.pr.new/@executor-js/plugin-onepassword@1608

@executor-js/plugin-openapi

npm i https://pkg.pr.new/@executor-js/plugin-openapi@1608

executor

npm i https://pkg.pr.new/executor@1608

commit: d62e666

[RhysSullivan]

CI verdict: 3 cloud e2e shards red — the acceptance gate working as intended. Four distinct failures:

1. mcp-client-sessions (unknown session id must fail fast): likely idFromString throwing on arbitrary client-supplied session strings where the old path returned a clean 404. Needs a parse guard in the worker.
2. mcp-sse-replay (severed POST response replays on the next standalone GET): the patch replayed stranded POST-stream events onto the GET stream; the stock v2 eventStore only resumes the same stream via Last-Event-ID. The DO needs the cross-stream undelivered-response replay ported.
3. mcp-priming-reconnect (stock old client recovers a mid-call result): the minimal priming port is not effective under real workerd — re-verify against the severed-stream flow.
4. repro-transport-brick (concurrent GET+POST after idle-dispose): restore/dispose race in the new lifecycle.

Iterating with the narrow e2e scenarios locally; will push the fix commit to this branch.

[RhysSullivan]

Folded into #1611 (single consolidated PR per review preference).