Retry without starting work twice

A lost response does not mean a lost request. Millwork may have accepted the run even though your application never saw the reply. Send the retry with the same request key, a value you choose and put in the Idempotency-Key header, and Millwork returns the run it already accepted instead of starting a new one. A run is one submitted task; the API calls it an execution. Goal: Send the same request twice and receive the same run. You are done when: both responses contain the same execution_id.

What the retry costs

Millwork charges the platform fee when it accepts a live run. A retry with the same request key and the same body returns that accepted run. No second run is created, so no second platform fee is charged. This recipe uses a test run (the API calls it Echo, mode: "echo"). It does not call a model, so it has no platform fee and no model cost. The same retry rule applies to a live run. See Model access and billing.

Before you begin

Set MILLWORK_API_URL and MILLWORK_API_KEY as shown in Make your first API call.

Choose a request key

Use one new key for each piece of work. Reuse that key only to retry the same request.

export REQUEST_KEY="echo-retry-$(date +%s)"

Send the same request twice

Both requests carry the same request key and the same body.

jq -n '{
  mode: "echo",
  task: {
    objective: "Check that a repeated request does not start another run."
  },
  policy: {
    data_classes: ["public"],
    budget: {
      max_cost_usd: 1,
      max_runtime_s: 30
    }
  }
}' > retry-request.json

for ATTEMPT in 1 2; do
  if ! curl --fail-with-body \
    --request POST \
    --header "Authorization: Bearer $MILLWORK_API_KEY" \
    --header "Idempotency-Key: $REQUEST_KEY" \
    --header "Content-Type: application/json" \
    --data @retry-request.json \
    --output "retry-$ATTEMPT.json" \
    "$MILLWORK_API_URL/executions"; then
    echo "Request $ATTEMPT failed. Read retry-$ATTEMPT.json before retrying." >&2
    exit 1
  fi
done

Expected result: both requests return 202 with an execution_id, "mode": "echo", and "status": "completed".

Compare the two run IDs

if ! jq -e -s '
  length == 2 and
  all(.[];
    (.execution_id | type == "string" and length > 0) and
    .mode == "echo" and .status == "completed"
  ) and
  .[0].execution_id == .[1].execution_id
' retry-1.json retry-2.json; then
  echo "The responses did not confirm the same completed test run." >&2
  exit 1
fi

Expected result: the command prints true.

If it fails

What happened	What to do
`idempotency_conflict`	The body changed between the two requests. Send the original body again, or use a new request key for new work.
No response after submitting	Send the exact same body with the same request key.
You do not know which body was sent	Stop. List your recent runs with `GET /v1/executions` and find the run before you create another request.

What this confirms

The same request key and body return the same run. This test run has no platform fee. Retrying a live run does not add another platform fee. This check confirms that the test run completed. For a live run, matching IDs confirm safe replay; follow its status separately as shown in Run lifecycle. A new request key starts new work, even with the same body. Millwork does not detect a duplicate by its body alone.

Output checks

Agent workflows

Review jobs

Operations recipes

Retry without starting work twice

What the retry costs

Before you begin

If it fails

What this confirms

​What the retry costs

​Before you begin

​If it fails

​What this confirms

What the retry costs

Before you begin

If it fails

What this confirms