diff --git a/agent/error_classifier.py b/agent/error_classifier.py
index 7ce749f594..9add7c84ae 100644
--- a/agent/error_classifier.py
+++ b/agent/error_classifier.py
@@ -169,7 +169,8 @@ _IMAGE_CORRUPT_PATTERNS = (
 _MULTIMODAL_TOOL_CONTENT_PATTERNS = (
     "text is not set", "tool message content must be a string", "tool content must be a string",
     "tool message must be a string", "expected string, got list", "expected string, got array",
-    "tool_call.content must be string",
+    "tool_call.content must be string", "tool.content.str", "tool.content",
+    "input should be a valid string",
 )
 
 # Local-inference memory/resource-ceiling rejections (oMLX/MLX memory guard,
@@ -763,6 +764,7 @@ def _classify_400(c: _Ctx) -> Verdict:
 _STATUS_HANDLERS: Dict[int, Callable[[_Ctx], Verdict]] = {
     400: _classify_400, 401: lambda c: _V_AUTH_ROTATE, 402: lambda c: _classify_402(c.msg, dict),
     403: _status_403, 404: _status_404, 408: lambda c: _V_TIMEOUT, 413: lambda c: _V_PAYLOAD_TOO_LARGE,
+    422: lambda c: _first_match(c.msg, _IMAGE_TOOL_RULES) or _V_FORMAT_ERROR,
     429: _status_429, 500: _status_5xx, 502: _status_5xx,
     503: lambda c: _first_match(c.msg, _OVERFLOW_AS_5XX_RULES) or _V_OVERLOADED,
     529: lambda c: _first_match(c.msg, _OVERFLOW_AS_5XX_RULES) or _V_OVERLOADED,
diff --git a/plugins/model-providers/opencode-zen/__init__.py b/plugins/model-providers/opencode-zen/__init__.py
index 1a3be51829..a2ee868dc3 100644
--- a/plugins/model-providers/opencode-zen/__init__.py
+++ b/plugins/model-providers/opencode-zen/__init__.py
@@ -115,10 +115,11 @@ opencode_go = OpenCodeGoProfile(
     name="opencode-go", aliases=("opencode_go", "go", "opencode-go-sub"), env_vars=("OPENCODE_GO_API_KEY",),
     base_url="https://opencode.ai/zen/go/v1", default_headers=dict(_ATTRIBUTION_HEADERS),
     default_aux_model="glm-5",
-    # opencode-go proxies to Xiaomi MiMo for mimo-* models; MiMo rejects
-    # list-type tool content with "text is not set" (400).  The direct
-    # xiaomi profile already sets supports_vision_tool_messages=False;
-    # propagate the same safety here for the relay path.
+    # The Go relay's upstream validates tool content as a strict string: list-type tool
+    # content (native vision embeds) 422s with ``messages.N.tool.content.str Input should
+    # be a valid string`` (Console Go, #104731) or 400s ``text is not set`` (MiMo, #47026),
+    # and the rejected row stays in history so every later call dies too. Images in user
+    # messages are fine, so vision itself keeps working via the text-summary downgrade.
     supports_vision_tool_messages=False,
 )
 
diff --git a/tests/plugins/model_providers/test_opencode_go_profile.py b/tests/plugins/model_providers/test_opencode_go_profile.py
index 54fae9b15f..2accf6bf3d 100644
--- a/tests/plugins/model_providers/test_opencode_go_profile.py
+++ b/tests/plugins/model_providers/test_opencode_go_profile.py
@@ -253,3 +253,12 @@ class TestOpenCodeGoFullKwargsIntegration:
         )
         assert "extra_body" not in kwargs
         assert kwargs["reasoning_effort"] == "high"
+
+
+class TestOpenCodeGoVisionToolMessages:
+    """OpenCode Go provider profile declares supports_vision_tool_messages=False (#104731)."""
+
+    def test_supports_vision_true_and_tool_messages_false(self, opencode_go_profile):
+        assert opencode_go_profile.supports_vision is True
+        assert opencode_go_profile.supports_vision_tool_messages is False
+
diff --git a/tests/run_agent/test_multimodal_tool_content_recovery.py b/tests/run_agent/test_multimodal_tool_content_recovery.py
index e2c91e3fb2..7119310e80 100644
--- a/tests/run_agent/test_multimodal_tool_content_recovery.py
+++ b/tests/run_agent/test_multimodal_tool_content_recovery.py
@@ -196,3 +196,42 @@ class TestRecoveryEndToEndClassification:
         )
         result = classify_api_error(err, provider="alibaba", model="qwen3.5-plus")
         assert result.reason == FailoverReason.multimodal_tool_content_unsupported
+
+    def test_opencode_go_422_console_go_classifies(self):
+        """Regression test for #104731: Console Go relay HTTP 422 on list-type tool content."""
+        err = _FakeApiError(
+            status_code=422,
+            message=(
+                "Error code: 422 - {'error': {'param': 'messages.67.tool.content.str', "
+                "'type': 'invalid_request_error', 'message': 'Error from provider (Console Go): "
+                "Upstream request failed: [invalid_request_error] Input should be a valid string'}}"
+            ),
+        )
+        result = classify_api_error(err, provider="opencode-go", model="deepseek-v4-flash-vision-exp")
+        assert result.reason == FailoverReason.multimodal_tool_content_unsupported
+        assert result.retryable is True
+
+
+class TestOpenCodeGoProactiveToolResultDowngrade:
+    def _multimodal_result(self, png_b64: str = "iVBORw0KGgoAAAA"):
+        return {
+            "_multimodal": True,
+            "content": [
+                {"type": "text", "text": "detected 2 objects"},
+                {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{png_b64}"}},
+            ],
+            "text_summary": "detected 2 objects",
+            "meta": {"png_bytes": 1024},
+        }
+
+    def test_returns_text_summary_for_opencode_go_proactively(self, monkeypatch):
+        """OpenCode Go rejects list-type tool content, so _tool_result_content_for_active_model
+        should proactively downgrade to a text summary (#104731)."""
+        agent = _make_agent(provider="opencode-go", model="deepseek-v4-flash-vision-exp")
+        agent._no_list_tool_content_models = set()
+        monkeypatch.setattr(agent, "_model_supports_vision", lambda: True)
+        out = agent._tool_result_content_for_active_model("vision_analyze", self._multimodal_result())
+        assert isinstance(out, str)
+        assert "data:image" not in out
+        assert "image_url" not in out
+