limbo-code 0.2__tar.gz → 0.2.1__tar.gz

{limbo_code-0.2 → limbo_code-0.2.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: limbo-code
-Version: 0.2
+Version: 0.2.1
 Summary: A minimal terminal AI coding agent
 Requires-Python: >=3.11
 Requires-Dist: httpx>=0.27

{limbo_code-0.2 → limbo_code-0.2.1}/src/limbo/llm/catalog.py

@@ -175,16 +175,19 @@ def _kimi_coding(
     *,
     context_window: int = 262_144,
     max_tokens: int = 32_768,
+    vision: bool = True,
 ) -> ModelSpec:
     # All Kimi For Coding models reason with Anthropic adaptive thinking
     # (thinking: {type: adaptive} + output_config.effort); thinking cannot
     # be disabled. Replayed thinking blocks may carry an empty signature.
+    # Every Kimi model is multimodal (accepts image input).
     return ModelSpec(
         id=model_id,
         provider=KIMI_CODING,
         context_window=context_window,
         max_tokens=max_tokens,
         reasoning=True,
+        vision=vision,
         thinking_format="anthropic-adaptive",
         thinking_levels={"low": "low", "high": "high", "max": "max"},
         thinking_can_disable=False,
@@ -197,7 +200,9 @@ def _moonshot(
     context_window: int = 262_144,
     max_tokens: int = 262_144,
     reasoning: bool = False,
-    vision: bool = False,
+    # Every Kimi (Moonshot) model is multimodal: image input is supported
+    # across the family, so the helper defaults to vision-capable.
+    vision: bool = True,
     thinking_format: str | None = None,
     thinking_levels: dict[str, str] | None = None,
     thinking_can_disable: bool = True,
@@ -264,6 +269,16 @@ CATALOG: dict[str, ModelSpec] = {
         reasoning=True,
         thinking_format="deepseek",
     ),
+    # Vision-enabled experimental flash variant (accepts image input).
+    "deepseek-v4-flash-vision-exp": ModelSpec(
+        id="deepseek-v4-flash-vision-exp",
+        provider=DEEPSEEK,
+        context_window=1_000_000,
+        max_tokens=65_536,
+        reasoning=True,
+        vision=True,
+        thinking_format="deepseek",
+    ),
     # -- Kimi (Moonshot AI) -------------------------------------------------
     "kimi-k2-0711-preview": _moonshot(
         "kimi-k2-0711-preview", context_window=131_072, max_tokens=16_384
@@ -278,7 +293,7 @@ CATALOG: dict[str, ModelSpec] = {
     ),
     # Kimi K2.5 is Moonshot's multimodal generation (accepts image input).
     "kimi-k2.5": _moonshot(
-        "kimi-k2.5", reasoning=True, thinking_format="deepseek", vision=True
+        "kimi-k2.5", reasoning=True, thinking_format="deepseek"
     ),
     "kimi-k2.6": _moonshot(
         "kimi-k2.6", reasoning=True, thinking_format="deepseek"

{limbo_code-0.2 → limbo_code-0.2.1}/tests/test_catalog.py

@@ -5,6 +5,7 @@ import pytest
 from limbo.config import Config, ProviderOverride
 from limbo.llm.anthropic_client import AnthropicMessagesClient
 from limbo.llm.catalog import (
+    CATALOG,
     DEEPSEEK,
     DEFAULT_BASE_URL,
     GENERIC_OPENAI,
@@ -298,13 +299,41 @@ def test_register_client_adds_dialect():
 def test_vision_defaults_to_false():
     # Text-only catalog entries stay non-vision (conservative default).
     assert resolve_model("deepseek-v4-flash").vision is False
-    assert resolve_model("kimi-k3").vision is False
+    assert resolve_model("glm-5.2").vision is False
+
+
+def test_deepseek_flash_vision_exp_is_vision_capable():
+    # Experimental flash variant: accepts image input, same spec as
+    # deepseek-v4-flash otherwise.
+    spec = resolve_model("deepseek-v4-flash-vision-exp")
+    assert spec.provider is DEEPSEEK
+    assert spec.vision is True
+    assert spec.reasoning is True
+    assert spec.thinking_format == "deepseek"
+
+
+def test_all_kimi_models_are_vision_capable():
+    # Every Kimi model (Moonshot API + Kimi For Coding) is multimodal.
+    kimi_ids = {
+        m for m in CATALOG
+        if m.startswith("kimi-") or m in {"k3", "kimi-for-coding", "kimi-for-coding-highspeed"}
+    }
+    assert kimi_ids, "expected at least one Kimi model in the catalog"
+    for model_id in kimi_ids:
+        assert resolve_model(model_id).vision is True, model_id
 
 
 def test_kimi_k2_5_is_vision_capable():
     assert resolve_model("kimi-k2.5").vision is True
 
 
+def test_kimi_k3_is_vision_capable():
+    # k3 (Kimi For Coding) and kimi-k3 (Moonshot) are the same multimodal
+    # model: both accept image input and must pass the vision gate.
+    assert resolve_model("kimi-k3").vision is True
+    assert resolve_model("k3").vision is True
+
+
 def test_unknown_model_vision_defaults_to_false():
     # GENERIC_OPENAI fallback: image attachments degrade to path references
     # rather than breaking an unrecognized endpoint with base64 payloads.