diff --git a/src/mitm/modify.rs b/src/mitm/modify.rs
index f78158e..2cf3191 100644
--- a/src/mitm/modify.rs
+++ b/src/mitm/modify.rs
@@ -85,7 +85,17 @@ pub fn modify_request(body: &[u8], tool_ctx: Option<&ToolContext>) -> Option<Vec
         let before = contents.len();
 
         // Remove messages that are pure Antigravity context injection
+        // IMPORTANT: Never strip messages containing inlineData (images)
         contents.retain(|msg| {
+            // Always keep messages with image/binary data in any part
+            if let Some(parts) = msg["parts"].as_array() {
+                for part in parts {
+                    if part.get("inlineData").is_some() {
+                        return true;
+                    }
+                }
+            }
+
             if let Some(text) = msg["parts"][0]["text"].as_str() {
                 // Strip user_information (OS, workspace paths)
                 if text.starts_with("<user_information>") {
@@ -151,8 +161,16 @@ pub fn modify_request(body: &[u8], tool_ctx: Option<&ToolContext>) -> Option<Vec
             }
         }
 
-        // Remove now-empty messages
+        // Remove now-empty messages (but preserve messages with non-text parts like images)
         contents.retain(|msg| {
+            if let Some(parts) = msg["parts"].as_array() {
+                // Keep if any part has inlineData
+                for part in parts {
+                    if part.get("inlineData").is_some() {
+                        return true;
+                    }
+                }
+            }
             if let Some(text) = msg["parts"][0]["text"].as_str() {
                 !text.trim().is_empty()
             } else {
diff --git a/src/proto.rs b/src/proto.rs
index fcfdc48..8801284 100644
--- a/src/proto.rs
+++ b/src/proto.rs
@@ -111,34 +111,53 @@ pub fn build_init_metadata(
 
 /// Build the `SendUserCascadeMessageRequest` protobuf binary.
 ///
-/// Produces a byte-exact match to real Antigravity webview traffic.
-/// Verified against Chrome DevTools network capture 2026-02-12.
+/// See `build_request_with_image` for the full field layout documentation.
 ///
-/// Field layout:
+/// Field layout (key fields):
 ///   1: cascade_id (string)
-///   2: ChatMessage { 1: text, 6: Blob { 1: mime_type, 2: data } }
+///   2: items (ChatMessage { 1: text })
 ///   3: metadata { 1: client_name, 3: oauth_token, 4: "en", 7: version, 12: client_name }
-///   5: PlannerConfig { 1: inner_config, 7: { 1: 1 } }
-///      inner_config contains: f2 (conv mode), f13 (tool config), f15 (model), f21 (ephemeral), f32 (knowledge)
-///  11: conversation_history = true
+///   5: cascade_config (PlannerConfig)
+///   6: images (ImageData { 1: base64_data, 2: mime_type }, repeated)
+///  11: client_type
 #[allow(dead_code)]
 pub fn build_request(cascade_id: &str, text: &str, oauth_token: &str, model_enum: u32) -> Vec<u8> {
     build_request_with_image(cascade_id, text, oauth_token, model_enum, None)
 }
 
-/// Image data to embed in the ChatMessage protobuf.
+/// Image data to attach to a cascade message.
 pub struct ImageData {
     /// MIME type, e.g. "image/png", "image/jpeg", "image/webp", "image/gif"
     pub mime_type: String,
-    /// Raw image bytes (NOT base64 — already decoded)
+    /// Raw image bytes (NOT base64 — we base64-encode them into the proto)
     pub data: Vec<u8>,
 }
 
 /// Build `SendUserCascadeMessageRequest` with optional image attachment.
 ///
-/// When `image` is Some, the ChatMessage includes a Blob field (field 6)
-/// alongside the text (field 1). This matches how the real Antigravity
-/// webview sends images: `ChatMessage { text, blob: { mime_type, data } }`.
+/// Field layout (from JS protobuf-es registration in workbench bundle):
+///   1: cascade_id    (string)
+///   2: items         (ChatMessage, repeated) — contains the text
+///   3: metadata      (auto-populated by LS)
+///   4: experiment_config
+///   5: cascade_config (PlannerConfig)
+///   6: images        (ImageData, repeated) — TOP-LEVEL, NOT inside ChatMessage
+///   8: blocking      (bool)
+///   9: additional_steps (repeated)
+///  10: artifact_comments (repeated)
+///  11: client_type   (enum)
+///  12: file_diff_comments (repeated)
+///  13: file_comments (repeated)
+///  14: media         (repeated)
+///  15: agent_script_item
+///  16: propagate_error (bool)
+///  17: planner_response
+///
+/// ImageData proto (from codeium_common_pb):
+///   1: base64_data   (string) — BASE64 encoded, NOT raw bytes
+///   2: mime_type     (string)
+///   3: caption       (string)
+///   4: uri           (string)
 pub fn build_request_with_image(
     cascade_id: &str,
     text: &str,
@@ -151,13 +170,8 @@ pub fn build_request_with_image(
     // Field 1: cascade_id
     msg.extend(proto_string(1, cascade_id.as_bytes()));
 
-    // Field 2: ChatMessage { f1: text, f6?: Blob { f1: mime_type, f2: data } }
-    let mut chat_msg = proto_string(1, text.as_bytes());
-    if let Some(img) = image {
-        let mut blob = proto_string(1, img.mime_type.as_bytes());
-        blob.extend(proto_string(2, &img.data));
-        chat_msg.extend(proto_message(6, &blob));
-    }
+    // Field 2: items (ChatMessage) — text only, no blob
+    let chat_msg = proto_string(1, text.as_bytes());
     msg.extend(proto_message(2, &chat_msg));
 
     // Field 3: Metadata (Auth + Client ID)
@@ -169,7 +183,7 @@ pub fn build_request_with_image(
     meta.extend(proto_string(12, CLIENT_NAME.as_bytes()));
     msg.extend(proto_message(3, &meta));
 
-    // Field 5: PlannerConfig
+    // Field 5: PlannerConfig (cascade_config)
     let mut inner = Vec::new();
 
     // field 2: conversational mode { f4: 1, f14: 0 }
@@ -204,7 +218,17 @@ pub fn build_request_with_image(
     .concat();
     msg.extend(proto_message(5, &f5_payload));
 
-    // Field 11: conversation history flag
+    // Field 6: images (ImageData, repeated) — TOP-LEVEL field
+    // ImageData { 1: base64_data (string), 2: mime_type (string) }
+    if let Some(img) = image {
+        use base64::Engine;
+        let b64 = base64::engine::general_purpose::STANDARD.encode(&img.data);
+        let mut image_data = proto_string(1, b64.as_bytes());
+        image_data.extend(proto_string(2, img.mime_type.as_bytes()));
+        msg.extend(proto_message(6, &image_data));
+    }
+
+    // Field 11: client_type (but we use it for conversation history flag)
     msg.extend(bool_field(11, true));
 
     msg