fix(agent): recover 'not a multimodal model' image rejections from text-only endpoints (port from block/buzz#5318)

agent/conversation_loop.py

@@ -4098,6 +4098,17 @@ def _perform_api_call(next_api_kwargs):
                     # 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",
+                    # Text-only serving endpoints (Crusoe serverless, vLLM
+                    # deployments of text-only checkpoints, various
+                    # OpenAI-compatible hosts) reject any request whose
+                    # history contains an image with HTTP 400
+                    # "<model-id> is not a multimodal model". Without this
+                    # phrase the image stays in history, every retry fails
+                    # identically, and the session is poisoned until manual
+                    # intervention. Observed live by block/buzz on
+                    # crusoeai/GLM-5.2-NVFP4 (block/buzz#5318: 8 wedged
+                    # trials, 12.7h aggregate idle-after-poison).
+                    "not a multimodal model",
                 )
                 _err_lower = _err_body.lower()
                 _looks_like_image_rejection = any(

tests/run_agent/test_image_rejection_fallback.py

@@ -139,6 +139,7 @@ class TestImageRejectionPhraseIsolation:
         "model does not support image",
         "image_url'. expected",
         "no endpoints found that support image input",
+        "not a multimodal model",
     )
 
     def _matches(self, body: str) -> bool:
@@ -158,6 +159,33 @@ def test_anthropic_image_too_large_does_not_trip(self):
         for body in bodies:
             assert self._matches(body) is False, f"false positive on: {body}"
 
+    def test_text_only_endpoint_multimodal_rejection_trips(self):
+        """Crusoe/vLLM text-only endpoints reject imaged history with
+        "<model-id> is not a multimodal model" (HTTP 400). This must route
+        to the strip-images fallback or the session poisons permanently —
+        observed live by block/buzz on crusoeai/GLM-5.2-NVFP4
+        (block/buzz#5318: 8 wedged benchmark trials).
+        """
+        bodies = [
+            "400: crusoeai/GLM-5.2-NVFP4 is not a multimodal model",
+            "Error: my-org/text-model-v2 is not a multimodal model.",
+        ]
+        for body in bodies:
+            assert self._matches(body) is True, f"missed rejection: {body}"
+
+    def test_multimodal_tool_content_shapes_do_not_trip(self):
+        # From agent/error_classifier.py _MULTIMODAL_TOOL_CONTENT_PATTERNS —
+        # these mean "tool message content must be a string", recovered by
+        # downgrading tool content to text, NOT by stripping all images
+        # session-wide.
+        bodies = [
+            "tool message content must be a string",
+            "expected string, got list",
+            "text is not set",
+        ]
+        for body in bodies:
+            assert self._matches(body) is False, f"false positive on: {body}"
+