Repository navigation
Description
Before You Submit
- I'm using the latest version of Salesforce CLI (reproduced on 2.157.8)
- I've searched the existing issues (related: Unhandled
SocketError: other side closedfrom undici HTTP/2 client crashes long-running data commands #3622, which is about HTTP/2 crashes in data commands; this one is HTTP/1.1 and handled, but not retried)
Command
sf project deploy start (also deploy validate, deploy quick: any command that waits with MetadataApiDeploy.pollStatus)
Org Type
Sandbox
Did this work before?
Not sure
Summary
When Salesforce closes the HTTP connection between two deploy status checks, sf project deploy start --wait N exits 10 with MetadataTransferError: Metadata API request failed: fetch failed, while the deploy keeps running and later succeeds in the org. In CI this marks a successful deploy as failed.
Cause, as far as we can read the code:
- jsforce-node's
fetchWithRetriesretries onlyGET/PUT/HEAD/OPTIONS/DELETE;checkDeployStatusis a SOAPPOST, so the request is not retried. - source-deploy-retrieve
MetadataTransfer.isRetryableErrortolerates network errors by matchingerror.messageagainstECONNRESET,socket hang up, etc. With undici the message isfetch failed(the cause,SocketError: other side closed,UND_ERR_SOCKET, is inerror.cause), so none match andpoll()rethrows. Same list in SDR 13.6.0 and 13.12.0.
We saw it 4 times in 20 minutes, from two different machines (GitHub Actions runners and a cloud VM), against the same sandbox, with no Trust incident.
Steps to reproduce
The closure is intermittent; we reproduce it deterministically with a local CONNECT proxy that closes the tunnel of one status check:
- Run a CONNECT proxy on 127.0.0.1:18081 that, after ~20 s, closes the next tunnel right after the client sends its request (about 20 lines of Node
net; happy to share). HTTPS_PROXY=http://127.0.0.1:18081 sf project deploy start --dry-run --source-dir force-app/main/default/classes -o <sandbox> --test-level RunSpecifiedTests --tests SomeTest --wait 10 --json
Expected Result
The CLI treats the closed connection as a transient network error (as it does for ECONNRESET), keeps polling, and returns the deploy's final result (exit 0 here).
Actual Result
Exit 10 after the first closed connection:
{"status":10,"name":"MetadataTransferError","message":"Metadata API request failed: fetch failed","data":{"id":"0Af..."}}The CLI log (~/.sf/sf-YYYY-MM-DD.log) has:
TypeError: fetch failed
at Object.processResponse (undici/lib/web/fetch/index.js:272:16)
at async fetchWithRetries (@jsforce/jsforce-node/lib/request.js:111:25)
caused by: SocketError: other side closed
at TLSSocket.onHttpSocketEnd (undici/lib/dispatcher/client-h1.js:1024:22)
sf project deploy report --job-id <id> afterwards shows the job Succeeded.
Suggested fix
In isRetryableError, also inspect error.cause (and error.cause.cause) for code in UND_ERR_SOCKET, ECONNRESET, ETIMEDOUT, …, or add fetch failed / other side closed to retryableErrors. A status check is read-only, so retrying it is safe.
Our workaround: when the command exits 10 with this error and data.id, we follow the same job with sf project deploy report --job-id <id> --wait <remaining>.
System Information
@salesforce/cli/2.157.8 linux-x64 node-v24.22.2
@salesforce/source-deploy-retrieve 13.11.0, @jsforce/jsforce-node 3.12.0, undici 8.12.3
(also @salesforce/cli 2.156.5: SDR 13.6.0, undici 8.12.2)
Activity
git2gus commentedon Nov 9, 2026
This issue has been linked to a new work item: W-24844126
k80bowman commentedon Nov 9, 2026
Thank you for reporting this, we will take a look.
thvd commentedon Nov 13, 2026
Independently observed the same error during Metadata API deployment status polling. The CLI debug log captured:
Successful
InProgressstatus polls immediately preceded the error. The serialised log retained the cause in the stack, but did not expose a structured cause code, so I cannot independently confirmUND_ERR_SOCKET.Would you welcome a focused PR to source-deploy-retrieve that recognises transient nested network causes and uses the existing bounded polling retries, preserving the same deployment ID rather than resubmitting the deployment?