X-Git-Url: https://git.giacomolli.uk/onechat/blobdiff_plain/6f7eda727e519ae43d7d35195cbe412a2725abd7..d647f33b3e70f9419f6c0740e1d01b210439a519:/responses.go

diff --git a/responses.go b/responses.go
index cdc1558..fddec73 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
@@ -64,13 +66,16 @@ type Client struct {
 }
 
 // Request is one model call. Input holds the input items in order, such as
-// userMessage values or output items of an earlier response.
+// userMessage values or output items of an earlier response. Node names the
+// node that a compaction builds, or is nil. It goes only to the usage file,
+// never to the endpoint.
 type Request struct {
 	Model        string
 	Effort       string
 	Instructions string
 	Tools        []Tool
 	Input        []any
+	Node         *nodeKey
 }
 
 // Tool is the definition of a function tool.
@@ -217,19 +222,31 @@ 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)
+	}
+
+	start := time.Now()
 	record := usageRecord{
-		Date:  time.Now().Format(time.RFC3339Nano),
+		Date:  start.Format(time.RFC3339Nano),
 		Model: request.Model,
 	}
 
+	if request.Node != nil {
+		record.Node = &[2]int{request.Node.level, request.Node.index}
+	}
+
 	response, err := c.send(httpRequest, &record)
 	if err != nil {
 		record.Error = err.Error()
 	}
 
+	record.Seconds = time.Since(start).Seconds()
 	record.InputTokens = response.Usage.Input
 	record.CachedTokens = response.Usage.Cached
 	record.OutputTokens = response.Usage.Output
+	record.TextBytes = len(response.Text)
 
 	return response, errors.Join(err, c.usage.Write(record))
 }
@@ -265,16 +282,21 @@ type usageLog struct {
 	path  string
 }
 
-// usageRecord is one line of usage.jsonl. Status is the HTTP status, or 0 when
-// no response came. Quota holds the x-codex-* headers of the response, without
-// the turn state.
+// usageRecord is one line of usage.jsonl. Date is the start of the call, and
+// Seconds its length. Node is the node [l, i] of a compaction. Status is the
+// HTTP status, or 0 when no response came. TextBytes is the size of the text of
+// the reply. Quota holds the x-codex-* headers of the response, without the
+// turn state.
 type usageRecord struct {
 	Date         string            `json:"date"`
+	Seconds      float64           `json:"seconds"`
 	Model        string            `json:"model"`
+	Node         *[2]int           `json:"node,omitempty"`
 	Status       int               `json:"status"`
 	InputTokens  int               `json:"input_tokens"`
 	CachedTokens int               `json:"cached_tokens"`
 	OutputTokens int               `json:"output_tokens"`
+	TextBytes    int               `json:"text_bytes"`
 	Quota        map[string]string `json:"quota,omitempty"`
 	Error        string            `json:"error,omitempty"`
 }