fix(proxy): use the attached marker for orphan tool result media (#8157)

src-tauri/src/proxy/providers/transform_codex_chat.rs

@@ -20,7 +20,8 @@ use crate::proxy::{
     },
     tool_media::{
         chat_file_from_input_file, flush_pending_chat_tool_media, plan_chat_tool_output_media,
-        queue_chat_tool_output_media, strip_and_clamp_media_from_tool_value, ToolMediaScope,
+        plan_chat_tool_output_media_with_marker, queue_chat_tool_output_media,
+        strip_and_clamp_media_from_tool_value, ToolMediaScope, TOOL_RESULT_MEDIA_ATTACHED_MARKER,
         TOOL_RESULT_MEDIA_MOVED_MARKER,
     },
 };
@@ -1213,9 +1214,10 @@ fn flush_pending_orphan_results(
 /// 上下文，但不能以 role=tool 的形式发给严格上游。
 ///
 /// 输出里的图片/文件媒体块复用配对路径的提取逻辑：文本 body 内联
-/// `TOOL_RESULT_MEDIA_MOVED_MARKER`，媒体本体作为同一条 user 消息的
-/// content parts 跟随，避免图片被 base64 序列化成纯文本——模型既看不到
-/// 图，token 还会被瞬间撑爆。
+/// `TOOL_RESULT_MEDIA_ATTACHED_MARKER` 占位，媒体本体作为同一条 user
+/// 消息的 content parts 跟随，避免图片被 base64 序列化成纯文本——模型既
+/// 看不到图，token 还会被瞬间撑爆。配对路径的 "moved to the following
+/// user message" 占位在本路径会指向一条不存在的消息，不能沿用。
 fn append_orphan_tool_output_as_user_message(messages: &mut Vec<Value>, item: &Value) {
     let call_id = item.get("call_id").and_then(|v| v.as_str()).unwrap_or("");
     let (body, media_parts) = plan_orphan_tool_output_body(item.get("output"));
@@ -1239,10 +1241,16 @@ fn append_orphan_tool_output_as_user_message(messages: &mut Vec<Value>, item: &V
 }
 
 /// 计算孤儿工具结果的文本 body 与媒体 parts。媒体识别复用配对路径的
-/// `plan_chat_tool_output_media`；没有媒体时保持旧版逐字段的纯文本序列化，
-/// 不给无媒体输出多加一层 JSON 包装。
+/// `plan_chat_tool_output_media_with_marker`；没有媒体时保持旧版逐字段的
+/// 纯文本序列化，不给无媒体输出多加一层 JSON 包装。
+///
+/// 孤儿路径的媒体就挂在这条 user 消息自己的后续 parts 里，占位符用
+/// `TOOL_RESULT_MEDIA_ATTACHED_MARKER`（与 Anthropic/Gemini/Responses
+/// bridge 的"媒体已作为原生媒体挂载"用词一致）。
 fn plan_orphan_tool_output_body(output: Option<&Value>) -> (String, Vec<Value>) {
-    if let Some(plan) = output.cloned().and_then(plan_chat_tool_output_media) {
+    if let Some(plan) = output.cloned().and_then(|output| {
+        plan_chat_tool_output_media_with_marker(output, TOOL_RESULT_MEDIA_ATTACHED_MARKER)
+    }) {
         return (plan.tool_content, plan.media_parts);
     }
     let body = match output {
@@ -2996,6 +3004,14 @@ mod tests {
         let text = parts[0]["text"].as_str().unwrap();
         assert!(text.starts_with(ORPHAN_TOOL_OUTPUT_MARKER), "{text}");
         assert!(text.contains("screenshot follows"), "{text}");
+        assert!(
+            text.contains(TOOL_RESULT_MEDIA_ATTACHED_MARKER),
+            "inline media placeholder must match where the media lands: {text}"
+        );
+        assert!(
+            !text.contains("following user message"),
+            "media travels in this same message: {text}"
+        );
         assert!(
             !text.contains(&data_url),
             "base64 image must not stay in the text body: {text}"

src-tauri/src/proxy/tool_media.rs

@@ -57,11 +57,25 @@ impl ToolMediaScope {
 /// Scalar strings remain scalar strings after replacement. This matters for a
 /// raw image data URL and for JSON encoded inside a tool-output string: adding
 /// another layer of JSON string quotes would change what the model sees.
-pub(crate) fn plan_chat_tool_output_media(mut output: Value) -> Option<ChatToolOutputMediaPlan> {
+///
+/// The placeholder says the media was "moved to the following user message".
+/// Callers that re-attach the extracted media inside the same message must use
+/// [`plan_chat_tool_output_media_with_marker`] so the placeholder matches where
+/// the media actually lands.
+pub(crate) fn plan_chat_tool_output_media(output: Value) -> Option<ChatToolOutputMediaPlan> {
+    plan_chat_tool_output_media_with_marker(output, TOOL_RESULT_MEDIA_MOVED_MARKER)
+}
+
+/// Same extraction as [`plan_chat_tool_output_media`], with a caller-chosen
+/// placeholder text for the stripped media blocks.
+pub(crate) fn plan_chat_tool_output_media_with_marker(
+    mut output: Value,
+    replacement_marker: &str,
+) -> Option<ChatToolOutputMediaPlan> {
     let output_was_string = output.is_string();
     let replacement_block = json!({
         "type": "text",
-        "text": TOOL_RESULT_MEDIA_MOVED_MARKER
+        "text": replacement_marker
     });
     let mut media_parts = Vec::new();
     let replaced = strip_and_clamp_media_from_tool_value(
@@ -69,7 +83,7 @@ pub(crate) fn plan_chat_tool_output_media(mut output: Value) -> Option<ChatToolO
         &mut media_parts,
         ToolMediaScope::AllSupported,
         &replacement_block,
-        TOOL_RESULT_MEDIA_MOVED_MARKER,
+        replacement_marker,
     );
     if replaced == 0 {
         return None;