Fix filename length error when using LLM-generated prompts

sam3/agent/agent_core.py

@@ -9,6 +9,7 @@
 
 from .client_llm import send_generate_request
 from .client_sam3 import call_sam_service
+from .helpers.filename_utils import sanitize_filename
 from .viz import visualize
 
 
@@ -319,12 +320,14 @@ def agent_inference(
                 print(f"🔍 Checking mask {i+1}/{num_masks}...")
                 image_w_mask_i, image_w_zoomed_in_mask_i = visualize(current_outputs, i)
 
+                # Sanitize the text prompt to create safe filenames
+                sanitized_prompt = sanitize_filename(LATEST_SAM3_TEXT_PROMPT)
                 image_w_zoomed_in_mask_i_path = os.path.join(
-                    sam_output_dir, rf"{LATEST_SAM3_TEXT_PROMPT}.png".replace("/", "_")
-                ).replace(".png", f"_zoom_in_mask_{i + 1}.png")
+                    sam_output_dir, f"{sanitized_prompt}_zoom_in_mask_{i + 1}.png"
+                )
                 image_w_mask_i_path = os.path.join(
-                    sam_output_dir, rf"{LATEST_SAM3_TEXT_PROMPT}.png".replace("/", "_")
-                ).replace(".png", f"_selected_mask_{i + 1}.png")
+                    sam_output_dir, f"{sanitized_prompt}_selected_mask_{i + 1}.png"
+                )
                 image_w_zoomed_in_mask_i.save(image_w_zoomed_in_mask_i_path)
                 image_w_mask_i.save(image_w_mask_i_path)
 
@@ -391,13 +394,11 @@ def agent_inference(
             }
 
             image_w_check_masks = visualize(updated_outputs)
+            # Sanitize the text prompt to create a safe filename
+            sanitized_prompt = sanitize_filename(LATEST_SAM3_TEXT_PROMPT)
             image_w_check_masks_path = os.path.join(
-                sam_output_dir, rf"{LATEST_SAM3_TEXT_PROMPT}.png"
-            ).replace(
-                ".png",
-                f"_selected_masks_{'-'.join(map(str, [i+1 for i in masks_to_keep]))}.png".replace(
-                    "/", "_"
-                ),
+                sam_output_dir,
+                f"{sanitized_prompt}_selected_masks_{'-'.join(map(str, [i+1 for i in masks_to_keep]))}.png"
             )
             image_w_check_masks.save(image_w_check_masks_path)
             # save the updated json outputs and append to message history

sam3/agent/client_sam3.py

@@ -9,6 +9,7 @@
 from sam3.model.box_ops import box_xyxy_to_xywh
 from sam3.train.masks_ops import rle_encode
 
+from .helpers.filename_utils import sanitize_filename
 from .helpers.mask_overlap_removal import remove_overlapping_masks
 from .viz import visualize
 
@@ -59,9 +60,8 @@ def call_sam_service(
     """
     print(f"📞 Loading image '{image_path}' and sending with prompt '{text_prompt}'...")
 
-    text_prompt_for_save_path = (
-        text_prompt.replace("/", "_") if "/" in text_prompt else text_prompt
-    )
+    # Sanitize the text prompt to create a safe filename
+    text_prompt_for_save_path = sanitize_filename(text_prompt)
 
     os.makedirs(
         os.path.join(output_folder_path, image_path.replace("/", "-")), exist_ok=True

sam3/agent/helpers/filename_utils.py

@@ -0,0 +1,38 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates. All Rights Reserved
+
+import hashlib
+import re
+
+
+def sanitize_filename(text_prompt: str, max_length: int = 200) -> str:
+    """
+    Sanitize a text prompt to be used as a filename.
+
+    Replaces invalid filesystem characters, truncates to max_length, and appends
+    a hash suffix to ensure uniqueness.
+
+    Args:
+        text_prompt: The text prompt to sanitize
+        max_length: Maximum length for the filename (default: 200)
+                   Leaves room for extensions and path components
+
+    Returns:
+        A sanitized filename-safe string
+    """
+    # Generate hash suffix for uniqueness (8 chars + 1 underscore = 9 chars)
+    prompt_hash = hashlib.md5(text_prompt.encode('utf-8')).hexdigest()[:8]
+    hash_suffix = f"_{prompt_hash}"
+
+    # Sanitize: replace invalid chars, collapse whitespace/underscores, trim
+    sanitized = re.sub(r'[<>:"/\\|?*\x00-\x1f\s]+', '_', text_prompt).strip('_.')
+
+    # Use default if empty after sanitization
+    sanitized = sanitized or "prompt"
+
+    # Truncate to fit max_length with hash suffix
+    available_length = max_length - len(hash_suffix)
+    if len(sanitized) > available_length:
+        sanitized = sanitized[:available_length]
+
+    return f"{sanitized}{hash_suffix}"
+

sam3/agent/inference.py

@@ -4,6 +4,7 @@
 import os
 
 from sam3.agent.agent_core import agent_inference
+from sam3.agent.helpers.filename_utils import sanitize_filename
 
 
 def run_single_image_inference(
@@ -27,8 +28,15 @@ def run_single_image_inference(
 
     # Generate output file names
     image_basename = os.path.splitext(os.path.basename(image_path))[0]
-    prompt_for_filename = text_prompt.replace("/", "_").replace(" ", "_")
-
+    # Sanitize the text prompt to create a safe filename
+    prompt_for_filename = sanitize_filename(text_prompt, max_length=150)
+
+    # Build base filename, truncating image_basename if needed to keep total length reasonable
+    # Leave room for separators and suffixes like "_pred.json"
+    max_base_length = 200
+    suffix_length = len(f"_{prompt_for_filename}_agent_{llm_name}")
+    if len(image_basename) + suffix_length > max_base_length:
+        image_basename = image_basename[:max(0, max_base_length - suffix_length)]
     base_filename = f"{image_basename}_{prompt_for_filename}_agent_{llm_name}"
     output_json_path = os.path.join(output_dir, f"{base_filename}_pred.json")
     output_image_path = os.path.join(output_dir, f"{base_filename}_pred.png")