Merge pull request #1408 from NousResearch/hermes/hermes-daa73839

fix: make Claude image handling work end-to-end

agent/anthropic_adapter.py

@@ -497,6 +497,66 @@ def convert_tools_to_anthropic(tools: List[Dict]) -> List[Dict]:
     return result
 
 
+def _image_source_from_openai_url(url: str) -> Dict[str, str]:
+    """Convert an OpenAI-style image URL/data URL into Anthropic image source."""
+    url = str(url or "").strip()
+    if not url:
+        return {"type": "url", "url": ""}
+
+    if url.startswith("data:"):
+        header, _, data = url.partition(",")
+        media_type = "image/jpeg"
+        if header.startswith("data:"):
+            mime_part = header[len("data:"):].split(";", 1)[0].strip()
+            if mime_part.startswith("image/"):
+                media_type = mime_part
+        return {
+            "type": "base64",
+            "media_type": media_type,
+            "data": data,
+        }
+
+    return {"type": "url", "url": url}
+
+
+def _convert_content_part_to_anthropic(part: Any) -> Optional[Dict[str, Any]]:
+    """Convert a single OpenAI-style content part to Anthropic format."""
+    if part is None:
+        return None
+    if isinstance(part, str):
+        return {"type": "text", "text": part}
+    if not isinstance(part, dict):
+        return {"type": "text", "text": str(part)}
+
+    ptype = part.get("type")
+
+    if ptype == "input_text":
+        block: Dict[str, Any] = {"type": "text", "text": part.get("text", "")}
+    elif ptype in {"image_url", "input_image"}:
+        image_value = part.get("image_url", {})
+        url = image_value.get("url", "") if isinstance(image_value, dict) else str(image_value or "")
+        block = {"type": "image", "source": _image_source_from_openai_url(url)}
+    else:
+        block = dict(part)
+
+    if isinstance(part.get("cache_control"), dict) and "cache_control" not in block:
+        block["cache_control"] = dict(part["cache_control"])
+    return block
+
+
+def _convert_content_to_anthropic(content: Any) -> Any:
+    """Convert OpenAI-style multimodal content arrays to Anthropic blocks."""
+    if not isinstance(content, list):
+        return content
+
+    converted = []
+    for part in content:
+        block = _convert_content_part_to_anthropic(part)
+        if block is not None:
+            converted.append(block)
+    return converted
+
+
 def convert_messages_to_anthropic(
     messages: List[Dict],
 ) -> Tuple[Optional[Any], List[Dict]]:
@@ -533,11 +593,9 @@ def convert_messages_to_anthropic(
             blocks = []
             if content:
                 if isinstance(content, list):
-                    for part in content:
+                    converted_content = _convert_content_to_anthropic(content)
-                        if isinstance(part, dict):
+                    if isinstance(converted_content, list):
-                            blocks.append(dict(part))
+                        blocks.extend(converted_content)
-                        elif part is not None:
-                            blocks.append({"type": "text", "text": str(part)})
                 else:
                     blocks.append({"type": "text", "text": str(content)})
             for tc in m.get("tool_calls", []):
@@ -587,12 +645,11 @@ def convert_messages_to_anthropic(
 
         # Regular user message
         if isinstance(content, list):
-            converted_blocks = []
+            converted_blocks = _convert_content_to_anthropic(content)
-            for part in content:
+            result.append({
-                converted = _convert_user_content_part_to_anthropic(part)
+                "role": "user",
-                if converted is not None:
+                "content": converted_blocks or [{"type": "text", "text": ""}],
-                    converted_blocks.append(converted)
+            })
-            result.append({"role": "user", "content": converted_blocks or [{"type": "text", "text": ""}]})
         else:
             result.append({"role": "user", "content": content})
 

agent/auxiliary_client.py

@@ -83,7 +83,10 @@ _AUTH_JSON_PATH = get_hermes_home() / "auth.json"
 
 # Codex fallback: uses the Responses API (the only endpoint the Codex
 # OAuth token can access) with a fast model for auxiliary tasks.
-_CODEX_AUX_MODEL = "gpt-5.3-codex"
+# ChatGPT-backed Codex accounts currently reject gpt-5.3-codex for these
+# auxiliary flows, while gpt-5.2-codex remains broadly available and supports
+# vision via Responses.
+_CODEX_AUX_MODEL = "gpt-5.2-codex"
 _CODEX_AUX_BASE_URL = "https://chatgpt.com/backend-api/codex"
 
 

run_agent.py

@@ -21,6 +21,8 @@ Usage:
 """
 
 import atexit
+import asyncio
+import base64
 import concurrent.futures
 import copy
 import hashlib
@@ -31,6 +33,7 @@ import os
 import random
 import re
 import sys
+import tempfile
 import time
 import threading
 import weakref
@@ -504,6 +507,11 @@ class AIAgent:
         self._persist_user_message_idx = None
         self._persist_user_message_override = None
 
+        # Cache anthropic image-to-text fallbacks per image payload/URL so a
+        # single tool loop does not repeatedly re-run auxiliary vision on the
+        # same image history.
+        self._anthropic_image_fallback_cache: Dict[str, str] = {}
+
         # Initialize LLM client via centralized provider router.
         # The router handles auth resolution, base URL, headers, and
         # Codex/Anthropic wrapping for all known providers.
@@ -3034,13 +3042,156 @@ class AIAgent:
 
     # ── End provider fallback ──────────────────────────────────────────────
 
+    @staticmethod
+    def _content_has_image_parts(content: Any) -> bool:
+        if not isinstance(content, list):
+            return False
+        for part in content:
+            if isinstance(part, dict) and part.get("type") in {"image_url", "input_image"}:
+                return True
+        return False
+
+    @staticmethod
+    def _materialize_data_url_for_vision(image_url: str) -> tuple[str, Optional[Path]]:
+        header, _, data = str(image_url or "").partition(",")
+        mime = "image/jpeg"
+        if header.startswith("data:"):
+            mime_part = header[len("data:"):].split(";", 1)[0].strip()
+            if mime_part.startswith("image/"):
+                mime = mime_part
+        suffix = {
+            "image/png": ".png",
+            "image/gif": ".gif",
+            "image/webp": ".webp",
+            "image/jpeg": ".jpg",
+            "image/jpg": ".jpg",
+        }.get(mime, ".jpg")
+        tmp = tempfile.NamedTemporaryFile(prefix="anthropic_image_", suffix=suffix, delete=False)
+        with tmp:
+            tmp.write(base64.b64decode(data))
+        path = Path(tmp.name)
+        return str(path), path
+
+    def _describe_image_for_anthropic_fallback(self, image_url: str, role: str) -> str:
+        cache_key = hashlib.sha256(str(image_url or "").encode("utf-8")).hexdigest()
+        cached = self._anthropic_image_fallback_cache.get(cache_key)
+        if cached:
+            return cached
+
+        role_label = {
+            "assistant": "assistant",
+            "tool": "tool result",
+        }.get(role, "user")
+        analysis_prompt = (
+            "Describe everything visible in this image in thorough detail. "
+            "Include any text, code, UI, data, objects, people, layout, colors, "
+            "and any other notable visual information."
+        )
+
+        vision_source = str(image_url or "")
+        cleanup_path: Optional[Path] = None
+        if vision_source.startswith("data:"):
+            vision_source, cleanup_path = self._materialize_data_url_for_vision(vision_source)
+
+        description = ""
+        try:
+            from tools.vision_tools import vision_analyze_tool
+
+            result_json = asyncio.run(
+                vision_analyze_tool(image_url=vision_source, user_prompt=analysis_prompt)
+            )
+            result = json.loads(result_json) if isinstance(result_json, str) else {}
+            description = (result.get("analysis") or "").strip()
+        except Exception as e:
+            description = f"Image analysis failed: {e}"
+        finally:
+            if cleanup_path and cleanup_path.exists():
+                try:
+                    cleanup_path.unlink()
+                except OSError:
+                    pass
+
+        if not description:
+            description = "Image analysis failed."
+
+        note = f"[The {role_label} attached an image. Here's what it contains:\n{description}]"
+        if vision_source and not str(image_url or "").startswith("data:"):
+            note += (
+                f"\n[If you need a closer look, use vision_analyze with image_url: {vision_source}]"
+            )
+
+        self._anthropic_image_fallback_cache[cache_key] = note
+        return note
+
+    def _preprocess_anthropic_content(self, content: Any, role: str) -> Any:
+        if not self._content_has_image_parts(content):
+            return content
+
+        text_parts: List[str] = []
+        image_notes: List[str] = []
+        for part in content:
+            if isinstance(part, str):
+                if part.strip():
+                    text_parts.append(part.strip())
+                continue
+            if not isinstance(part, dict):
+                continue
+
+            ptype = part.get("type")
+            if ptype in {"text", "input_text"}:
+                text = str(part.get("text", "") or "").strip()
+                if text:
+                    text_parts.append(text)
+                continue
+
+            if ptype in {"image_url", "input_image"}:
+                image_data = part.get("image_url", {})
+                image_url = image_data.get("url", "") if isinstance(image_data, dict) else str(image_data or "")
+                if image_url:
+                    image_notes.append(self._describe_image_for_anthropic_fallback(image_url, role))
+                else:
+                    image_notes.append("[An image was attached but no image source was available.]")
+                continue
+
+            text = str(part.get("text", "") or "").strip()
+            if text:
+                text_parts.append(text)
+
+        prefix = "\n\n".join(note for note in image_notes if note).strip()
+        suffix = "\n".join(text for text in text_parts if text).strip()
+        if prefix and suffix:
+            return f"{prefix}\n\n{suffix}"
+        if prefix:
+            return prefix
+        if suffix:
+            return suffix
+        return "[A multimodal message was converted to text for Anthropic compatibility.]"
+
+    def _prepare_anthropic_messages_for_api(self, api_messages: list) -> list:
+        if not any(
+            isinstance(msg, dict) and self._content_has_image_parts(msg.get("content"))
+            for msg in api_messages
+        ):
+            return api_messages
+
+        transformed = copy.deepcopy(api_messages)
+        for msg in transformed:
+            if not isinstance(msg, dict):
+                continue
+            msg["content"] = self._preprocess_anthropic_content(
+                msg.get("content"),
+                str(msg.get("role", "user") or "user"),
+            )
+        return transformed
+
     def _build_api_kwargs(self, api_messages: list) -> dict:
         """Build the keyword arguments dict for the active API mode."""
         if self.api_mode == "anthropic_messages":
             from agent.anthropic_adapter import build_anthropic_kwargs
+            anthropic_messages = self._prepare_anthropic_messages_for_api(api_messages)
             return build_anthropic_kwargs(
                 model=self.model,
-                messages=api_messages,
+                messages=anthropic_messages,
                 tools=self.tools,
                 max_tokens=self.max_tokens,
                 reasoning_config=self.reasoning_config,

tests/agent/test_auxiliary_client.py

@@ -195,7 +195,7 @@ class TestGetTextAuxiliaryClient:
         with patch("agent.auxiliary_client._read_nous_auth", return_value=None), \
              patch("agent.auxiliary_client.OpenAI") as mock_openai:
             client, model = get_text_auxiliary_client()
-        assert model == "gpt-5.3-codex"
+        assert model == "gpt-5.2-codex"
         # Returns a CodexAuxiliaryClient wrapper, not a raw OpenAI client
         from agent.auxiliary_client import CodexAuxiliaryClient
         assert isinstance(client, CodexAuxiliaryClient)
@@ -288,7 +288,7 @@ class TestVisionClientFallback:
             client, model = get_vision_auxiliary_client()
         from agent.auxiliary_client import CodexAuxiliaryClient
         assert isinstance(client, CodexAuxiliaryClient)
-        assert model == "gpt-5.3-codex"
+        assert model == "gpt-5.2-codex"
 
     def test_vision_auto_falls_back_to_custom_endpoint(self, monkeypatch):
         """Custom endpoint is used as fallback in vision auto mode.
@@ -371,7 +371,7 @@ class TestVisionClientFallback:
             client, model = get_vision_auxiliary_client()
         from agent.auxiliary_client import CodexAuxiliaryClient
         assert isinstance(client, CodexAuxiliaryClient)
-        assert model == "gpt-5.3-codex"
+        assert model == "gpt-5.2-codex"
 
 
 class TestGetAuxiliaryProvider:
@@ -489,7 +489,7 @@ class TestResolveForcedProvider:
             client, model = _resolve_forced_provider("main")
         from agent.auxiliary_client import CodexAuxiliaryClient
         assert isinstance(client, CodexAuxiliaryClient)
-        assert model == "gpt-5.3-codex"
+        assert model == "gpt-5.2-codex"
 
     def test_forced_codex(self, codex_auth_dir, monkeypatch):
         with patch("agent.auxiliary_client._read_nous_auth", return_value=None), \
@@ -497,7 +497,7 @@ class TestResolveForcedProvider:
             client, model = _resolve_forced_provider("codex")
         from agent.auxiliary_client import CodexAuxiliaryClient
         assert isinstance(client, CodexAuxiliaryClient)
-        assert model == "gpt-5.3-codex"
+        assert model == "gpt-5.2-codex"
 
     def test_forced_codex_no_token(self, monkeypatch):
         with patch("agent.auxiliary_client._read_codex_access_token", return_value=None):

tests/test_anthropic_adapter.py

@@ -495,6 +495,59 @@ class TestConvertMessages:
         assert len(result) == 1
         assert result[0]["role"] == "user"
 
+    def test_converts_user_image_url_blocks_to_anthropic_image_blocks(self):
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "Can you see this?"},
+                    {"type": "image_url", "image_url": {"url": "https://example.com/cat.png"}},
+                ],
+            }
+        ]
+
+        _, result = convert_messages_to_anthropic(messages)
+
+        assert result == [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "Can you see this?"},
+                    {"type": "image", "source": {"type": "url", "url": "https://example.com/cat.png"}},
+                ],
+            }
+        ]
+
+    def test_converts_data_url_image_blocks_to_base64_anthropic_image_blocks(self):
+        messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "input_text", "text": "What is in this screenshot?"},
+                    {"type": "input_image", "image_url": "data:image/png;base64,AAAA"},
+                ],
+            }
+        ]
+
+        _, result = convert_messages_to_anthropic(messages)
+
+        assert result == [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "What is in this screenshot?"},
+                    {
+                        "type": "image",
+                        "source": {
+                            "type": "base64",
+                            "media_type": "image/png",
+                            "data": "AAAA",
+                        },
+                    },
+                ],
+            }
+        ]
+
     def test_converts_tool_calls(self):
         messages = [
             {

tests/test_provider_parity.py

@@ -543,7 +543,7 @@ class TestAuxiliaryClientProviderPriority:
              patch("agent.auxiliary_client._read_codex_access_token", return_value="codex-tok"), \
              patch("agent.auxiliary_client.OpenAI"):
             client, model = get_text_auxiliary_client()
-        assert model == "gpt-5.3-codex"
+        assert model == "gpt-5.2-codex"
         assert isinstance(client, CodexAuxiliaryClient)
 
 

tests/test_run_agent.py

@@ -12,7 +12,7 @@ import uuid
 from logging.handlers import RotatingFileHandler
 from pathlib import Path
 from types import SimpleNamespace
-from unittest.mock import MagicMock, patch
+from unittest.mock import AsyncMock, MagicMock, patch
 
 import pytest
 
@@ -1986,6 +1986,69 @@ class TestBuildApiKwargsAnthropicMaxTokens:
                 assert call_args[0][3] is None
 
 
+class TestAnthropicImageFallback:
+    def test_build_api_kwargs_converts_multimodal_user_image_to_text(self, agent):
+        agent.api_mode = "anthropic_messages"
+        agent.reasoning_config = None
+
+        api_messages = [{
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "Can you see this now?"},
+                {"type": "image_url", "image_url": {"url": "https://example.com/cat.png"}},
+            ],
+        }]
+
+        with (
+            patch("tools.vision_tools.vision_analyze_tool", new=AsyncMock(return_value=json.dumps({"success": True, "analysis": "A cat sitting on a chair."}))),
+            patch("agent.anthropic_adapter.build_anthropic_kwargs") as mock_build,
+        ):
+            mock_build.return_value = {"model": "claude-sonnet-4-20250514", "messages": [], "max_tokens": 4096}
+            agent._build_api_kwargs(api_messages)
+
+        kwargs = mock_build.call_args.kwargs or dict(zip(
+            ["model", "messages", "tools", "max_tokens", "reasoning_config"],
+            mock_build.call_args.args,
+        ))
+        transformed = kwargs["messages"]
+        assert isinstance(transformed[0]["content"], str)
+        assert "A cat sitting on a chair." in transformed[0]["content"]
+        assert "Can you see this now?" in transformed[0]["content"]
+        assert "vision_analyze with image_url: https://example.com/cat.png" in transformed[0]["content"]
+
+    def test_build_api_kwargs_reuses_cached_image_analysis_for_duplicate_images(self, agent):
+        agent.api_mode = "anthropic_messages"
+        agent.reasoning_config = None
+        data_url = "data:image/png;base64,QUFBQQ=="
+
+        api_messages = [
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "first"},
+                    {"type": "input_image", "image_url": data_url},
+                ],
+            },
+            {
+                "role": "user",
+                "content": [
+                    {"type": "text", "text": "second"},
+                    {"type": "input_image", "image_url": data_url},
+                ],
+            },
+        ]
+
+        mock_vision = AsyncMock(return_value=json.dumps({"success": True, "analysis": "A small test image."}))
+        with (
+            patch("tools.vision_tools.vision_analyze_tool", new=mock_vision),
+            patch("agent.anthropic_adapter.build_anthropic_kwargs") as mock_build,
+        ):
+            mock_build.return_value = {"model": "claude-sonnet-4-20250514", "messages": [], "max_tokens": 4096}
+            agent._build_api_kwargs(api_messages)
+
+        assert mock_vision.await_count == 1
+
+
 class TestFallbackAnthropicProvider:
     """Bug fix: _try_activate_fallback had no case for anthropic provider."""