diff --git a/agent/conversation_loop.py b/agent/conversation_loop.py
index 7520a1fcc4..99da6974ca 100644
--- a/agent/conversation_loop.py
+++ b/agent/conversation_loop.py
@@ -4686,6 +4686,17 @@ def run_conversation(
                     # every subsequent message queued behind the stuck turn —
                     # the P1 in issue #21160. The 404 passes the 4xx gate below.
                     "no endpoints found that support image input",
+                    # Kimi / Moonshot / other OpenAI-compatible Chinese
+                    # providers reject truncated or corrupt image bytes with
+                    # HTTP 400 "Invalid request: prepare image failed ...
+                    # failed to decode image: invalid or unsupported image
+                    # format". Like the Codex case above, the bad bytes are
+                    # baked into immutable conversation history and re-sent on
+                    # every retry, wedging the session. Strip the images so the
+                    # turn recovers instead of exhausting retries. (issue
+                    # #76884; complements the proactive full-decode validation
+                    # in tools/vision_tools._normalize_to_supported_image)
+                    "failed to decode image",
                 )
                 _err_lower = _err_body.lower()
                 _looks_like_image_rejection = any(
diff --git a/tests/run_agent/test_image_rejection_fallback.py b/tests/run_agent/test_image_rejection_fallback.py
index cd254ee6aa..d4a643e7dd 100644
--- a/tests/run_agent/test_image_rejection_fallback.py
+++ b/tests/run_agent/test_image_rejection_fallback.py
@@ -139,12 +139,21 @@ class TestImageRejectionPhraseIsolation:
         "model does not support image",
         "image_url'. expected",
         "no endpoints found that support image input",
+        "failed to decode image",
     )
 
     def _matches(self, body: str) -> bool:
         low = body.lower()
         return any(p in low for p in self._REJECTION_PHRASES)
 
+    def test_kimi_truncated_image_trips_recovery(self):
+        # Kimi/Moonshot reject truncated image bytes with this 400; the
+        # bad bytes are in immutable history so stripping must fire.
+        body = ("HTTP 400: Invalid request: prepare image failed error, "
+                "status code: 400, message: failed to decode image: invalid "
+                "or unsupported image format")
+        assert self._matches(body) is True
+
     def test_anthropic_image_too_large_does_not_trip(self):
         # From agent/error_classifier.py _IMAGE_TOO_LARGE_PATTERNS —
         # these must route to image_too_large / _try_shrink_image_parts_in_messages,