[mattzcarey]

A failed MCP connect or discovery now schedules a durable connect:<serverId> Lifecycle job. It retries after 30s, 1m, 2m, 5m and 10m, then every 30m, and after 24h records a reconnect-exhausted failure. An auth-expiry:<serverId> job moves a server whose authorization link expired unused to failed, so a dead link is no longer shown and no new one gets minted.

Stacked on #2661, and part of the same work on #2658. Every retry used to live in memory. _connectWithRetry made 3 quick attempts, and then the server sat in failed until the isolate died, which with live sockets could take hours. Nothing woke the object to try again, and the connect after an OAuth callback was a fire-and-forget promise.

Things callers will notice:

• retries stop when authorization was refused, since every attempt would get the same answer: a 401 the agent can't answer with a link (a revoked static token, no OAuth callback), a 403 left after any step-up, or an OAuth error such as invalid_target or access_denied. The server records an authorization-failed failure. Network errors, 5xx, timeouts, server_error and temporarily_unavailable still back off.
• these retries add alarm wakes. An object with an unreachable MCP server wakes about 50 times in its first day, until it connects or gives up.
• recorded failures survive wakes, and only an explicit reconnect clears them: addMcpServer(), connectToServer(), establishConnection() or registerServer()
• addMcpServer() on a failed server reconnects. It used to throw.
• a server with a reconnect already scheduled stays failed when the object wakes for another reason, and waits for that job
• standalone managers take new MCPClientManager(name, version, { reconnect: { delaysMs, giveUpAfterMs } })
• the manager adds a failure column to cf_agents_mcp_servers the first time it records a failure

pnpm run check, the agents workers suite and the conformance lanes pass. New tests run both jobs through real Lifecycle alarms on a plain Durable Object, including staying failed after eviction, giving up and then reconnecting explicitly, and moving jobs when migrateServerId() renames a server.

The durable reconnect schedule and the link-expiry job here are removed in #2778, which records failures and leaves retrying to reconnectMcpServer(), as Claude Code does. Review refusal and failure recording here, and the rest there.

[agent-think]

🟡 agents import sizes: 4 entry points grew

	Entry point	Exports	Largest gzip change	Size now
🟡	agents/mcp	15 resized	up to +1.8 KiB (+0.49%)	361.4 KiB
🟡	agents	5 resized	up to +1.8 KiB (+0.66%)	271.7 KiB
🟡	agents/chat-sdk	3 resized	up to +1.8 KiB (+0.65%)	273.5 KiB
🟡	agents/mcp/client	1 resized	+1.7 KiB (+1.02%)	167.2 KiB

Changed exports (24)

	Import	Gzip change	Size now
🟡	agents/mcp#RPCClientTransport	+1.8 KiB (+0.49%)	361.4 KiB
🟡	agents/mcp#normalizeServerId	+1.8 KiB (+0.49%)	361.4 KiB
🟡	agents#getCurrentAgent	+1.8 KiB (+0.66%)	271.7 KiB
🟡	agents/mcp#DurableObjectEventStore	+1.8 KiB (+0.49%)	361.4 KiB
🟡	agents/chat-sdk#ChatSdkStateAgent	+1.8 KiB (+0.65%)	273.5 KiB
🟡	agents#DEFAULT_AGENT_STATIC_OPTIONS	+1.8 KiB (+0.65%)	271.7 KiB
🟡	agents#Agent	+1.8 KiB (+0.65%)	271.7 KiB
🟡	agents#StreamingResponse	+1.8 KiB (+0.65%)	271.7 KiB
🟡	agents/mcp#RPC_DO_PREFIX	+1.8 KiB (+0.49%)	361.4 KiB
🟡	agents/mcp#ElicitRequestSchema	+1.8 KiB (+0.49%)	361.4 KiB
🟡	agents/mcp#RPCServerTransport	+1.8 KiB (+0.49%)	361.4 KiB
🟡	agents#routeAgentEmail	+1.8 KiB (+0.65%)	272 KiB
🟡	agents/mcp#getMcpAuthContext	+1.8 KiB (+0.49%)	361.4 KiB
🟡	agents/mcp#StreamableHTTPEdgeClientTransport	+1.7 KiB (+0.48%)	361.5 KiB
🟡	agents/mcp#McpAgent	+1.7 KiB (+0.48%)	361.4 KiB
🟡	agents/chat-sdk#ChatSdkStateAdapter	+1.7 KiB (+0.64%)	274.3 KiB
🟡	agents/chat-sdk#createChatSdkState	+1.7 KiB (+0.63%)	274.3 KiB
🟡	agents/mcp#SSEEdgeClientTransport	+1.7 KiB (+0.48%)	361.4 KiB
🟡	agents/mcp#createLegacyMcpHandler	+1.7 KiB (+0.44%)	395.6 KiB
🟡	agents/mcp#experimental_createMcpHandler	+1.7 KiB (+0.43%)	395.8 KiB
🟡	agents/mcp/client#MCPClientManager	+1.7 KiB (+1.02%)	167.2 KiB
🟡	agents/mcp#MCP_SERVER_ID_MAX_LENGTH	+1.7 KiB (+0.47%)	361.3 KiB
🟡	agents/mcp#WorkerTransport	+1.7 KiB (+0.45%)	365.1 KiB
🟡	agents/mcp#createMcpHandler	+1.6 KiB (+0.4%)	408.5 KiB

How this works

Each runtime export is bundled on its own, minified, and gzipped. Changes smaller than 100 B, or smaller than 1% and 1 KiB, are ignored. Growth over 10% or 5 KiB is marked 🔴. This report is informational and does not fail CI. The workflow artifact contains every measurement.

_{Compared 0406eaf1 → 446f3588 · workflow run · reported by agent-think[bot]}

[pkg-pr-new]

Open in StackBlitz

agents

npm i https://pkg.pr.new/agents@2662

@cloudflare/ai-chat

npm i https://pkg.pr.new/@cloudflare/ai-chat@2662

@cloudflare/codemode

npm i https://pkg.pr.new/@cloudflare/codemode@2662

hono-agents

npm i https://pkg.pr.new/hono-agents@2662

@cloudflare/shell

npm i https://pkg.pr.new/@cloudflare/shell@2662

@cloudflare/think

npm i https://pkg.pr.new/@cloudflare/think@2662

@cloudflare/voice

npm i https://pkg.pr.new/@cloudflare/voice@2662

@cloudflare/worker-bundler

npm i https://pkg.pr.new/@cloudflare/worker-bundler@2662

commit: 446f358

[changeset-bot]

⚠️ No Changeset found

Latest commit: 446f358

Merging this PR will not cause a version bump for any packages. If these changes should not result in a new version, you're good to go. If these changes should result in a version bump, you need to add a changeset.

This PR includes no changesets

When changesets are added to this PR, you'll see the packages that this PR includes changesets for and the associated semver types

Click here to learn what changesets are, and how to add one.

Click here if you're a maintainer who wants to add a changeset to this PR

[devin-ai-integration]

Devin Review found 1 new potential issue.

⚠️ 1 issue in files not directly in the diff

⚠️ MCP client release lacks a changeset

When this feature reaches main without another changeset, agents receives no release entry. The release workflow versions packages through changeset version, so the new public reconnect option remains unpublished.