Have a person check an email before the relay sends it to a customer. This workflow uses the downloadable relay’s local queue. It does not create a Voiceflow inbox or a managed human-agent hand-off connection.
Before you begin
Complete a connection manual and leave AUTO_SEND=false. The reviewer needs private operator access to the host, the original provider mail and an approved reply policy. Run commands inside the extracted email-relay folder. Locally prefix each command with node --env-file=.env; on a running Docker host use docker compose exec relay node.
Review one customer reply
Find the held work
Replace JOB_ID with the complete ID printed by list. show includes the sender, original text, candidate, thread and state. Its output can contain customer data; use a private terminal. Later messages in a held thread wait so a review cannot be bypassed by a follow-up.
Correct or discard the reply
To replace the candidate, write a UTF-8 plain-text file containing the exact customer reply, then run:
For Docker, copy that file into the container before using its path:
A manual reply changes held input to review; it does not run an agent turn or extract an attachment. Remove [HUMAN_REVIEW] from the final customer text when replacing that candidate. To suppress a message entirely, use cli.mjs discard JOB_ID. Discard does not send an explanation or delete the provider’s original mail.
Approve once
The next worker pass sends the cached reply and records provider acceptance. Approval does not run the agent again. Inspect the customer inbox and the provider activity record before calling the request resolved. A later bounce needs operational follow-up even if the job says sent.
Verify the review boundary
Test a normal candidate
Send a pilot policy question. Confirm the customer receives nothing before approval. Replace its candidate, approve once and verify exactly the reviewed text arrives. Try approving the same sent job again; the CLI must reject that state.
Test a held conversation
Send an attachment or an out-of-scope question that returns the example’s [HUMAN_REVIEW] marker. Confirm held work and no automatic send. Send a follow-up in that thread; it should wait. Resolve the earlier item by sending a reviewed reply or discarding, then verify the follow-up can proceed.
An email may already have been sent
Resolve remote effects before using a recovery command.
Stop the worker. Inspect provider sent/activity logs and the customer inbox. For uncertain-send, cli.mjs mark-sent JOB_ID records known acceptance. Only after proving the provider did not accept it, use cli.mjs retry-send JOB_ID --confirmed-not-sent to reuse the cached reply. If the outcome remains unknown, leave it held and resolve manually. A timeout is not proof of failure.
For uncertain-turn, inspect the runtime transcript and any business-system effects first. cli.mjs restart JOB_ID --reviewed-side-effects begins a new session and can repeat actions; use it only after that review. While Docker is stopped use docker compose run --rm relay node cli.mjs ..., then restart the service.
Mail sends while the reviewer expects it to wait
Inspect automatic sending and already approved work.
Check AUTO_SEND and job states. Setting it to false affects future candidates; it does not revoke a prior approval or recall accepted mail. Stop the worker and reconcile queued ready jobs before resuming. Keep operator access restricted: this CLI is the approval authority.
For help desk connections, reassign the native ticket/conversation to a person before replying manually. Approval rechecks native ownership and latest content; a changed job becomes ignored. Resolve or discard old held work before returning assignment to automation. These supervised examples require review before sending.
Help desk messages arriving during review can overlap questions already run through the agent. Those newest jobs are held as review-input, with helpdesk.priorJobs in show. Reconcile the cached answers and newest questions, discard stale jobs and write one complete reply for the newest held job. This avoids repeating agent/tool execution. A manually edited or written reply does not rewrite the runtime’s conversation history; keep the thread with a person if a material correction or human exchange makes that saved context unsafe to reuse.