Merge current dev and refresh PR #6738 from the 7011 integration

src/context_compactor.py

@@ -13,6 +13,7 @@ from typing import Any, Dict, List, Optional
 from src.model_context import estimate_text_tokens, get_context_length, estimate_tokens
 from src.llm_core import llm_call_async
 from src.endpoint_resolver import resolve_endpoint
+from src.utility_effort import effort_for_call
 from src.settings import get_setting
 from core.models import ChatMessage
 
@@ -91,6 +92,83 @@ def prune_multimodal_images(
     return pruned
 
 
+_IMAGE_TURN_LABEL_RE = re.compile(r"\[Image(?:s)?(?:\s+attached)?:\s*([^\]]+)\]", re.IGNORECASE)
+
+
+def _image_turn_label(message: Dict) -> str:
+    """Best-effort '[Image: name]' label for a message whose image parts are dropped."""
+    name = ""
+    content = message.get("content") if isinstance(message, dict) else None
+    if isinstance(content, list):
+        for item in content:
+            if isinstance(item, dict) and item.get("type") in _MULTIMODAL_IMAGE_TYPES:
+                candidate = item.get("name") or (item.get("image_url") or {}).get("name")
+                if candidate:
+                    name = str(candidate)
+                    break
+        if not name:
+            text = _content_as_text(content)
+            match = _IMAGE_TURN_LABEL_RE.search(text)
+            if match:
+                name = match.group(1).strip()
+    return f"[Image: {name}]" if name else "[Image]"
+
+
+def keep_recent_image_turns(messages: List[Dict], *, keep: int = 2) -> List[Dict]:
+    """Keep image parts only for the last `keep` messages that carry an image.
+
+    Older image blocks are dropped and replaced with a short ``[Image: name]``
+    text label so the model still knows a picture was shared (and its name)
+    without paying the image tokens on every later turn. Non-text, non-image
+    multimodal blocks (e.g. audio) are left untouched.
+    """
+    limit = max(0, int(keep))
+    image_messages = [
+        index for index, message in enumerate(messages)
+        if isinstance(message, dict)
+        and isinstance(message.get("content"), list)
+        and any(
+            isinstance(item, dict) and item.get("type") in _MULTIMODAL_IMAGE_TYPES
+            for item in message["content"]
+        )
+    ]
+    if len(image_messages) <= limit:
+        return list(messages)
+    selected = set(image_messages[len(image_messages) - limit:]) if limit else set()
+
+    trimmed: List[Dict] = []
+    for message_index, message in enumerate(messages):
+        if message_index in selected or not isinstance(message, dict):
+            trimmed.append(message)
+            continue
+        content = message.get("content")
+        if not isinstance(content, list):
+            trimmed.append(message)
+            continue
+        kept_blocks = [
+            item for item in content
+            if not (isinstance(item, dict) and item.get("type") in _MULTIMODAL_IMAGE_TYPES)
+        ]
+        if len(kept_blocks) == len(content):
+            trimmed.append(message)
+            continue
+        label = _image_turn_label(message)
+        # Attach the label to the trailing text block when possible so we do not
+        # introduce an extra content block; otherwise append a fresh text block.
+        labelled = False
+        for item in reversed(kept_blocks):
+            if isinstance(item, dict) and item.get("type") == "text":
+                item["text"] = (item.get("text") or "").rstrip() + f"\n{label}"
+                labelled = True
+                break
+        if not labelled:
+            kept_blocks.append({"type": "text", "text": label})
+        cloned = dict(message)
+        cloned["content"] = kept_blocks
+        trimmed.append(cloned)
+    return trimmed
+
+
 COMPACT_THRESHOLD = 0.85  # Trigger compaction at 85% of context window
 SUMMARY_MAX_TOKENS = 1024
 SMALL_CONTEXT_LIMIT = 8192  # Models with context <= this get aggressive trimming
@@ -556,10 +634,8 @@ async def maybe_compact(
     )
 
     # Use utility model if configured, otherwise fall back to session model
-    util_url, util_model, util_headers = resolve_endpoint("utility", owner=owner)
-    compact_url = util_url or endpoint_url
-    compact_model = util_model or model
-    compact_headers = util_headers if util_url else headers
+    # Compact with the chat model itself (the one being evaluated), never a separate utility model.
+    compact_url, compact_model, compact_headers = endpoint_url, model, headers
 
     prompt = SELF_SUMMARY_SYSTEM_PROMPT.replace(
         "{count}", str(len(older))
@@ -580,6 +656,7 @@ async def maybe_compact(
             max_tokens=SUMMARY_MAX_TOKENS,
             headers=compact_headers,
             timeout=30,
+            reasoning_effort=effort_for_call(compact_url, compact_model, owner, session),
         )
     except Exception as e:
         logger.error(f"Compaction summary failed: {e}")