Fix silent data loss in function_call arguments/output non-string JSON parsing

responsesInputToMessages unmarshaled function_call.arguments and
function_call_output.output as bare strings, silently dropping the
value when it arrived as an object or content-parts array. This
caused the model to lose tool-call context in multi-turn
conversations, increasing the likelihood of malformed tool-call JSON.

Add rawJSONToString (re-encodes non-string values as JSON strings)
and outputToString (extracts text from content-parts arrays, re-encodes
other non-string values). Add 3 regression tests covering object
arguments, array output, and object output.

internal/openai/responses.go

@@ -3,6 +3,7 @@ package openai
 import (
 	"encoding/json"
 	"errors"
+	"strings"
 )
 
 // --- Responses API request parsing & conversion ---
@@ -194,9 +195,7 @@ func responsesInputToMessages(input json.RawMessage) ([]map[string]any, error) {
 				inner["name"] = name
 			}
 			if v, ok := item["arguments"]; ok {
-				var args string
+				inner["arguments"] = rawJSONToString(v)
-				_ = json.Unmarshal(v, &args)
-				inner["arguments"] = args
 			}
 			if v, ok := item["call_id"]; ok {
 				var id string
@@ -215,9 +214,7 @@ func responsesInputToMessages(input json.RawMessage) ([]map[string]any, error) {
 				msg["tool_call_id"] = id
 			}
 			if v, ok := item["output"]; ok {
-				var out string
+				msg["content"] = outputToString(v)
-				_ = json.Unmarshal(v, &out)
-				msg["content"] = out
 			}
 			messages = append(messages, msg)
 		default:
@@ -279,6 +276,58 @@ func convertContentParts(raw json.RawMessage) any {
 	return result
 }
 
+// rawJSONToString converts a json.RawMessage to a string. If the value is
+// already a JSON string, it is used directly. Any other JSON value (object,
+// array, number, bool) is re-encoded as a JSON string so it can populate
+// fields that require a string, such as tool_calls[].function.arguments.
+// This prevents silent data loss when a field arrives as a non-string type.
+func rawJSONToString(v json.RawMessage) string {
+	var s string
+	if json.Unmarshal(v, &s) == nil {
+		return s
+	}
+	var anyValue any
+	if json.Unmarshal(v, &anyValue) == nil {
+		if b, err := json.Marshal(anyValue); err == nil {
+			return string(b)
+		}
+	}
+	return string(v) // last resort: raw bytes
+}
+
+// outputToString converts a function_call_output "output" value to a string
+// for the Chat Completions tool message content. The output may be:
+//   - a plain string (used directly)
+//   - an array of content parts (text extracted and concatenated)
+//   - any other JSON value (re-encoded as a JSON string)
+func outputToString(v json.RawMessage) string {
+	var s string
+	if json.Unmarshal(v, &s) == nil {
+		return s
+	}
+	// Try array of content parts — extract text from each part.
+	var parts []map[string]any
+	if json.Unmarshal(v, &parts) == nil {
+		var sb strings.Builder
+		for _, p := range parts {
+			if t, _ := p["text"].(string); t != "" {
+				sb.WriteString(t)
+			}
+		}
+		if sb.Len() > 0 {
+			return sb.String()
+		}
+	}
+	// Fallback: re-encode as a JSON string.
+	var anyValue any
+	if json.Unmarshal(v, &anyValue) == nil {
+		if b, err := json.Marshal(anyValue); err == nil {
+			return string(b)
+		}
+	}
+	return string(v)
+}
+
 // translateResponsesTools converts tools from the Responses API flat format
 // to the Chat Completions nested {type:"function",function:{…}} format.
 func translateResponsesTools(raw json.RawMessage) (json.RawMessage, error) {

internal/openai/responses_test.go

@@ -191,6 +191,90 @@ func TestParseResponsesRequest_FunctionCallInput(t *testing.T) {
 	}
 }
 
+// TestParseResponsesRequest_FunctionCallObjectArguments verifies that when
+// function_call.arguments arrives as a JSON object (not a string), it is
+// re-encoded as a JSON string instead of being silently dropped.
+func TestParseResponsesRequest_FunctionCallObjectArguments(t *testing.T) {
+	body := `{"model":"GLM-4.7","input":[
+		{"type":"function_call","call_id":"call_456","name":"task","arguments":{"operation":"create","summary":"test"}}
+	]}`
+	req, err := ParseResponsesRequest([]byte(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(req.Messages) != 1 {
+		t.Fatalf("messages = %d items", len(req.Messages))
+	}
+	tc, _ := req.Messages[0]["tool_calls"].([]any)
+	if len(tc) != 1 {
+		t.Fatalf("tool_calls = %v", req.Messages[0]["tool_calls"])
+	}
+	fn, _ := tc[0].(map[string]any)["function"].(map[string]any)
+	args, _ := fn["arguments"].(string)
+	if args == "" {
+		t.Fatalf("arguments was dropped (empty)")
+	}
+	// The re-encoded string must be valid JSON containing the original fields.
+	var parsed map[string]any
+	if err := json.Unmarshal([]byte(args), &parsed); err != nil {
+		t.Fatalf("arguments not valid JSON: %v", err)
+	}
+	if parsed["operation"] != "create" {
+		t.Fatalf("operation = %v", parsed["operation"])
+	}
+}
+
+// TestParseResponsesRequest_FunctionCallOutputArray verifies that when
+// function_call_output.output arrives as an array of content parts, the
+// text is extracted instead of being silently dropped.
+func TestParseResponsesRequest_FunctionCallOutputArray(t *testing.T) {
+	body := `{"model":"GLM-4.7","input":[
+		{"type":"function_call","call_id":"call_789","name":"get_weather","arguments":"{\"city\":\"NYC\"}"},
+		{"type":"function_call_output","call_id":"call_789","output":[{"type":"output_text","text":"sunny 72F"}]}
+	]}`
+	req, err := ParseResponsesRequest([]byte(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	if len(req.Messages) != 2 {
+		t.Fatalf("messages = %d items", len(req.Messages))
+	}
+	toolMsg := req.Messages[1]
+	if toolMsg["role"] != "tool" {
+		t.Fatalf("msg[1] role = %v", toolMsg["role"])
+	}
+	content, _ := toolMsg["content"].(string)
+	if content != "sunny 72F" {
+		t.Fatalf("content = %q, want %q", content, "sunny 72F")
+	}
+}
+
+// TestParseResponsesRequest_FunctionCallOutputObject verifies that when
+// function_call_output.output is a bare JSON object, it is re-encoded as a
+// JSON string instead of being silently dropped.
+func TestParseResponsesRequest_FunctionCallOutputObject(t *testing.T) {
+	body := `{"model":"GLM-4.7","input":[
+		{"type":"function_call","call_id":"call_obj","name":"run","arguments":"{}"},
+		{"type":"function_call_output","call_id":"call_obj","output":{"result":"success","code":200}}
+	]}`
+	req, err := ParseResponsesRequest([]byte(body))
+	if err != nil {
+		t.Fatal(err)
+	}
+	toolMsg := req.Messages[1]
+	content, _ := toolMsg["content"].(string)
+	if content == "" {
+		t.Fatal("content was dropped (empty)")
+	}
+	var parsed map[string]any
+	if err := json.Unmarshal([]byte(content), &parsed); err != nil {
+		t.Fatalf("content not valid JSON: %v", err)
+	}
+	if parsed["result"] != "success" {
+		t.Fatalf("result = %v", parsed["result"])
+	}
+}
+
 func TestUsageToResponses(t *testing.T) {
 	usage := map[string]any{
 		"prompt_tokens":             10,