From 8551e939d582ab6fa7c55d7f20afe092325c1611 Mon Sep 17 00:00:00 2001
From: Cursor Agent <cursoragent@cursor.com>
Date: Wed, 27 May 2026 13:54:11 +0000
Subject: [PATCH] fix(anthropic): plug content drop, compaction SSE shape, and
 compaction leak

- Sync streaming __next__ no longer drops a buffered holding_chunk when
  the usage-merge path has already fired. Restoring the prior unconditional
  flush behavior preserves provider-emitted content (the SSE-ordering nit
  of a trailing content delta is preferable to silent content loss).
- compaction content_block_start now carries the full block shape
  ({"type": "compaction", "content": ""}) to match the text-block
  pattern and Anthropic's native streaming shape, so clients that key off
  content_block_start see the field.
- apply_compact_20260112 now slices around / strips compaction blocks
  before the opt-in gate check. Previously, when summary_model was not
  configured the editor returned the raw messages, leaking Anthropic-only
  compaction content blocks to non-Anthropic providers that reject them.

Co-authored-by: Yassin Kortam <yassin@berri.ai>
---
 .../adapters/streaming_iterator.py            | 18 ++++++++++---
 .../context_management/editors/compact.py     | 26 +++++++++++--------
 2 files changed, 29 insertions(+), 15 deletions(-)

diff --git a/litellm/llms/anthropic/experimental_pass_through/adapters/streaming_iterator.py b/litellm/llms/anthropic/experimental_pass_through/adapters/streaming_iterator.py
index bec40c8c7a5..2eca78b5519 100644
--- a/litellm/llms/anthropic/experimental_pass_through/adapters/streaming_iterator.py
+++ b/litellm/llms/anthropic/experimental_pass_through/adapters/streaming_iterator.py
@@ -175,7 +175,12 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
             {
                 "type": "content_block_start",
                 "index": compaction_index,
-                "content_block": {"type": "compaction"},
+                # Mirror the text-block shape ({"type": "text", "text": ""}):
+                # send an empty ``content`` field so clients that introspect
+                # ``content_block_start`` see the full block schema. The
+                # actual summary text arrives via the ``content_block_delta``
+                # below.
+                "content_block": {"type": "compaction", "content": ""},
             }
         )
         self.chunk_queue.append(
@@ -394,9 +399,14 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
                     )
                     self.holding_stop_reason_chunk = None
 
-                if self.holding_chunk is not None:
-                    self.chunk_queue.append(self.holding_chunk)
-                    self.holding_chunk = None
+            # Always flush any buffered content delta, even when usage has
+            # already been merged + emitted: dropping it would silently lose
+            # provider-emitted content, which is worse than the SSE ordering
+            # nit of trailing a content chunk after the final message_delta
+            # (the prior sync ``__next__`` behavior).
+            if self.holding_chunk is not None:
+                self.chunk_queue.append(self.holding_chunk)
+                self.holding_chunk = None
 
             if not self.sent_last_message:
                 self.sent_last_message = True
diff --git a/litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py b/litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py
index d1ef9ebc15e..c23461bcd9b 100644
--- a/litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py
+++ b/litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py
@@ -486,17 +486,10 @@ async def apply_compact_20260112(
     if warnings:
         applied["warnings"] = warnings
 
-    # Opt-in gate: no summary model configured → no-op.
-    summary_model = _read_summary_model_setting()
-    if summary_model is None:
-        applied["error"] = "summary_model_not_configured"
-        return PolyfillResult(
-            messages=messages,
-            system=system,
-            applied_edits=[applied],
-        )
-
-    # Phase A: slice around any existing compaction block.
+    # Phase A: slice around any existing compaction block. Runs before the
+    # opt-in gate below so that even when summarization is disabled we still
+    # strip Anthropic-only ``compaction`` blocks from messages going to
+    # non-Anthropic backends (which would reject them).
     effective_messages, prior_compaction_block = _slice_around_compaction_block(
         messages
     )
@@ -513,6 +506,17 @@ async def apply_compact_20260112(
 
     downstream_messages = _strip_compaction_blocks(effective_messages)
 
+    # Opt-in gate: no summary model configured → no-op (but still return the
+    # Phase A-sliced/stripped messages so compaction blocks don't leak).
+    summary_model = _read_summary_model_setting()
+    if summary_model is None:
+        applied["error"] = "summary_model_not_configured"
+        return PolyfillResult(
+            messages=downstream_messages,
+            system=augmented_system,
+            applied_edits=[applied],
+        )
+
     # Phase B: threshold check.
     try:
         current_tokens = _count_effective_tokens(