X-Git-Url: http://git.giacomolli.uk/onechat/blobdiff_plain/6f7eda727e519ae43d7d35195cbe412a2725abd7..ba9faee0cd6c35020997794ef7bf134e8db12e96:/responses.go?ds=inline

diff --git a/responses.go b/responses.go
index cdc1558..096439f 100644
--- a/responses.go
+++ b/responses.go
@@ -48,8 +48,10 @@ type apiKey string
 // Client sends model calls to a Responses endpoint for one agent. The ChatGPT
 // backend and the OpenAI API take the same request body and differ only in the
 // URL and the credentials. Every call sends the fixed cache key of the agent,
-// so the turns and the compactions of the agent share one prompt cache. Every
-// call is recorded in the usage file.
+// so the turns and the compactions of the agent share one prompt cache. The
+// ChatGPT backend caches only with the key in the session_id header, so a call
+// to it sends the key there and in the body. Every call is recorded in the
+// usage file.
 //
 // The request never holds max_output_tokens or metadata, because the ChatGPT
 // backend refuses them. It sets store to false, so the client sends the whole
@@ -217,6 +219,11 @@ func (c *Client) Create(ctx context.Context, request Request) (Response, error)
 		return Response{}, fmt.Errorf("authorize model call: %w", err)
 	}
 
+	_, chatgpt := c.auth.(*chatgptAuth)
+	if chatgpt {
+		httpRequest.Header.Set("session_id", c.cacheKey)
+	}
+
 	record := usageRecord{
 		Date:  time.Now().Format(time.RFC3339Nano),
 		Model: request.Model,