fix(deepseek): honor Flash 1M window leftovers and native vision

Users on native DeepSeek were told to pin model.context_length and model.supports_vision in config.yaml. That is the wrong layer: the 1M window is already in DEFAULT_CONTEXT_LENGTHS, and a global supports_vision pin would also mark text-only deepseek-v4-pro as multimodal. Two catalog gaps still produced the reported symptoms: - A leftover context_length_cache.yaml entry of 128K (the old ``deepseek`` catch-all) outlived the 1M catalog keys because deepseek-flash was missing from _PRE_CATALOG_STALE_KEYS. - When models.dev is empty/cold, Flash has no capability record, so image routing falls through to lossy text. Vendor docs (2026-09-10) mark deepseek-flash as vision-capable and deepseek-v4-pro as not. Discard those 128K leftovers, fill Flash (and retired Flash aliases) via _BUILTIN_MODEL_METADATA, and leave Pro catalog-only.
2026-09-11 07:37:05 -07:00
parent 7469c0f2a5
commit 759024bdff
6 changed files with 46 additions and 4 deletions
@@ -1303,6 +1303,9 @@ _PRE_CATALOG_STALE_KEYS = frozenset({
    "grok-4.3", "grok-4.6",  # 1M / 500K; "grok-4" catch-all persisted 256,000
    "grok-4-fast", "grok-4.20",  # 2M; fell through to the 256K fallback
    "qwen3.6-plus",  # 1M; "qwen" catch-all persisted 131,072
+    # V4 / V4.1 Flash: 1M. Pre-entry builds matched the family catch-all and persisted 128K.
+    "deepseek-flash", "deepseek-v4.1-flash", "deepseek-v4-flash", "deepseek-v4-pro",
+    "deepseek-chat", "deepseek-reasoner",
 })


@@ -525,6 +525,14 @@ _UNKNOWN_MODEL_BASE: Dict[str, Any] = {"limit": {"context": 200000, "output": 81
 # Account-gated models may be usable before models.dev has indexed them.  Keep
 # their capabilities available for an explicitly selected/discovered model
 # without adding them to any picker catalog.
+_DEEPSEEK_FLASH_VISION: Dict[str, Any] = {
+    "limit": {"context": 1_000_000, "output": 384_000},
+    "modalities": {"input": ["text", "image"], "output": ["text"]},
+    "tool_call": True,
+    "reasoning": True,
+    "family": "deepseek-flash",
+}
+
 _BUILTIN_MODEL_METADATA: Dict[Tuple[str, str], Dict[str, Any]] = {
    ("openai", "gpt-6-astra"): {
        "limit": {"context": 1_050_000, "output": 128_000},
@@ -533,6 +541,14 @@ _BUILTIN_MODEL_METADATA: Dict[Tuple[str, str], Dict[str, Any]] = {
        "reasoning": True,
        "family": "gpt-6",
    },
+    # Native DeepSeek V4.1-Flash is multimodal (https://api-docs.deepseek.com/guides/vision).
+    # models.dev lagged the 2026-09-10 rename; without this, a cold/empty cache treats
+    # ``deepseek-flash`` as unknown → image_input_mode falls through to lossy text.
+    # ``deepseek-v4-pro`` stays catalog-only: vendor docs still mark it text-only.
+    ("deepseek", "deepseek-flash"): _DEEPSEEK_FLASH_VISION,
+    ("deepseek", "deepseek-v4-flash"): _DEEPSEEK_FLASH_VISION,
+    ("deepseek", "deepseek-v4.1-flash"): _DEEPSEEK_FLASH_VISION,
+    ("deepseek", "deepseek-v4-flash-vision-exp"): _DEEPSEEK_FLASH_VISION,
 }


@@ -1720,6 +1720,13 @@ class TestGenericPreCatalogStaleGuard:
        assert not _stale_pre_catalog_cache_entry("grok-4.20", 2_000_000)
        # Sibling qwen slugs with legitimately small windows are untouched.
        assert not _stale_pre_catalog_cache_entry("qwen3-coder", 131_072)
+        # DeepSeek V4 / V4.1 Flash: 1M. Pre-entry builds persisted the 128K
+        # ``deepseek`` catch-all; a leftover must drop, a 1M value must not.
+        assert _stale_pre_catalog_cache_entry("deepseek-flash", 128_000)
+        assert _stale_pre_catalog_cache_entry("deepseek/deepseek-flash", 128_000)
+        assert _stale_pre_catalog_cache_entry("deepseek-v4-pro", 128_000)
+        assert not _stale_pre_catalog_cache_entry("deepseek-flash", 1_000_000)
+        assert not _stale_pre_catalog_cache_entry("deepseek", 128_000)

    def test_unknown_models_never_dropped(self):
        from agent.model_metadata import _stale_pre_catalog_cache_entry
@@ -836,6 +836,22 @@ class TestGetModelCapabilities:
        api_caps = get_model_capabilities("openai-api", "gpt-6-astra")
        assert api_caps == caps

+    def test_deepseek_flash_builtin_vision_fills_catalog_lag(self):
+        """Native Flash stays multimodal when models.dev is empty; Pro does not.
+
+        Vendor docs: deepseek-flash accepts images, deepseek-v4-pro does not.
+        A global model.supports_vision pin would lie about Pro.
+        """
+        with patch("agent.models_dev.fetch_models_dev", return_value={}):
+            flash = get_model_capabilities("deepseek", "deepseek-flash")
+            alias = get_model_capabilities("deepseek", "deepseek-v4-flash")
+            pro = get_model_capabilities("deepseek", "deepseek-v4-pro")
+
+        assert flash is not None and flash.supports_vision is True
+        assert flash.context_window == 1_000_000
+        assert alias is not None and alias.supports_vision is True
+        assert pro is None
+

 # ---------------------------------------------------------------------------
 # Per-model metadata overrides (model_overrides config)
@@ -200,8 +200,8 @@ When a user attaches an image — from the CLI clipboard, the gateway (Telegram/

 | Your model | What happens to the image |
 |---|---|
-| **Vision-capable** (GPT-4V, Claude with vision, Gemini, Qwen-VL, MiMo-VL, etc.) | Sent as **real pixels** using the provider's native image content format above. No text summary layer. |
-| **Text-only** (DeepSeek V3, smaller open-source models, older chat-only endpoints) | Routed through the `vision_analyze` auxiliary tool — an auxiliary vision model describes the image, and the text description is injected into the conversation. |
+| **Vision-capable** (GPT-4V, Claude with vision, Gemini, Qwen-VL, MiMo-VL, DeepSeek Flash / V4.1-Flash, etc.) | Sent as **real pixels** using the provider's native image content format above. No text summary layer. |
+| **Text-only** (DeepSeek V4 Pro, DeepSeek V3, smaller open-source models, older chat-only endpoints) | Routed through the `vision_analyze` auxiliary tool — an auxiliary vision model describes the image, and the text description is injected into the conversation. |

 You don't configure this — Hermes looks up your current model's capability in the provider metadata and picks the right path automatically. The practical effect: you can switch between vision and non-vision models mid-session and image handling "just works" without changing your workflow. Text-only models get coherent context about the image rather than a broken multimodal payload they'd have to reject.

@@ -196,8 +196,8 @@ powershell.exe -NoProfile -Command "Add-Type -AssemblyName System.Windows.Forms;

 | 你的模型 | 图像处理方式 |
 |---|---|
-| **支持视觉的模型**（GPT-4V、Claude with vision、Gemini、Qwen-VL、MiMo-VL 等） | 使用上述提供商原生图像内容格式，以**真实像素**发送。无文本摘要层。 |
-| **纯文本模型**（DeepSeek V3、较小的开源模型、旧版纯对话端点） | 通过 `vision_analyze` 辅助工具路由——辅助视觉模型描述图像，文本描述注入对话。 |
+| **支持视觉的模型**（GPT-4V、Claude with vision、Gemini、Qwen-VL、MiMo-VL、DeepSeek Flash / V4.1-Flash 等） | 使用上述提供商原生图像内容格式，以**真实像素**发送。无文本摘要层。 |
+| **纯文本模型**（DeepSeek V4 Pro、DeepSeek V3、较小的开源模型、旧版纯对话端点） | 通过 `vision_analyze` 辅助工具路由——辅助视觉模型描述图像，文本描述注入对话。 |

 无需手动配置——Hermes 在提供商元数据中查找当前模型的能力并自动选择正确路径。实际效果：你可以在会话中途切换视觉模型与非视觉模型，图像处理"开箱即用"，无需更改工作流。纯文本模型会获得关于图像的连贯上下文，而不是一个会被拒绝的损坏多模态载荷。