kolega-code 0.41.4__tar.gz → 0.43.0__tar.gz

{kolega_code-0.41.4 → kolega_code-0.43.0}/CHANGELOG.md

@@ -6,6 +6,44 @@ This project uses GitHub Releases for detailed generated release notes. This fil
 
 ## Unreleased
 
+## 0.43.0 - 2026-10-13
+
+### Added
+
+- **Strict tool schemas.** Code that drives Kolega's LLM client directly can
+  mark a tool definition `strict=True`. The Responses providers then ask the
+  provider to hold the tool call to the schema exactly — every required field
+  present and nothing extra. Tools that are not marked strict behave as before.
+
+### Changed
+
+- DeepSeek requests now use `https://api.deepseek.com/beta`. DeepSeek enforces
+  strict tool schemas only on that endpoint and silently ignores the flag on the
+  bare host, so models that required a field could previously return a tool call
+  without it.
+
+### Fixed
+
+- A tool call whose arguments contain a raw control character — an unescaped tab
+  or newline copied into a JSON string — is now kept instead of degrading to an
+  empty call. Observed live with `deepseek-flash`.
+
+## 0.42.0 - 2026-10-07
+
+### Added
+
+- **Native Mistral AI support.** Use Mistral's own API with `MISTRAL_API_KEY`
+  or a key saved in Settings. The model picker defaults to Mistral Medium 3.5
+  and also includes Mistral Small 4, Mistral Large 4 **preview**, Ministral,
+  and Codestral, with streaming, tool calls, usage reporting, and vision and
+  reasoning where supported.
+
+### Changed
+
+- Refreshed the OpenRouter, Ollama Cloud, and Perplexity Agent model catalogs
+  with newly available models, removal of retired entries, updated model
+  limits and capabilities, and a refreshed OpenRouter featured ordering.
+
 ## 0.41.4 - 2026-10-06
 
 ### Security

{kolega_code-0.41.4 → kolega_code-0.43.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: kolega-code
-Version: 0.41.4
+Version: 0.43.0
 Summary: Multi-agent coding in the terminal
 Project-URL: Homepage, https://kolega-ai.github.io/kolega-code/
 Project-URL: Documentation, https://kolega-ai.github.io/kolega-code/
@@ -263,6 +263,7 @@ Supported model providers:
 - OpenAI API
 - OpenAI via ChatGPT subscription sign-in
 - Google
+- Mistral AI (native API)
 - Groq
 - Together.ai
 - Fireworks.ai
@@ -304,6 +305,29 @@ directory unless `KOLEGA_CODE_STATE_DIR` is set. See the
 [Configuration docs](https://kolega-ai.github.io/kolega-code/configuration/settings-and-api-keys/)
 for the full story.
 
+For Mistral's own API, save a key under **Mistral AI** in Settings or export
+`MISTRAL_API_KEY`, then select a model explicitly:
+
+```bash
+export MISTRAL_API_KEY=your-mistral-api-key
+kolega-code . --provider mistral --model mistral-medium-3-5
+```
+
+The picker defaults to Mistral Medium 3.5 and includes Mistral Small 4
+(`mistral-small-2603`), Mistral Large 4 **preview** (`mistral-large-4`),
+Ministral 3 (3B/8B/14B), Codestral, and eligible native aliases. Large 4 access
+depends on the account/subscription; a catalog entry does not guarantee access.
+Aliases, especially `latest`, can change their target and capabilities.
+
+Medium 3.5, Small 4, Large 4, and Ministral 3 support image input; Codestral is
+text-only. Medium/Small/Large expose `none` (default) or `high` thinking, with
+same-provider reasoning replayed in Mistral's native format. Token counting is
+a conservative local estimate with a safety margin, not an exact Mistral
+tokenizer. The catalog's 32,768-token output reservation is Kolega's budgeting
+policy, not a verified native maximum-output ceiling. Kolega Code runs the local
+function-tool loop; Mistral-hosted server tools are not enabled. Native Mistral
+keys and model IDs are separate from gateway credentials and routes.
+
 ## Requirements
 
 - **Python 3.11+**

{kolega_code-0.41.4 → kolega_code-0.43.0}/README.md

@@ -172,6 +172,7 @@ Supported model providers:
 - OpenAI API
 - OpenAI via ChatGPT subscription sign-in
 - Google
+- Mistral AI (native API)
 - Groq
 - Together.ai
 - Fireworks.ai
@@ -213,6 +214,29 @@ directory unless `KOLEGA_CODE_STATE_DIR` is set. See the
 [Configuration docs](https://kolega-ai.github.io/kolega-code/configuration/settings-and-api-keys/)
 for the full story.
 
+For Mistral's own API, save a key under **Mistral AI** in Settings or export
+`MISTRAL_API_KEY`, then select a model explicitly:
+
+```bash
+export MISTRAL_API_KEY=your-mistral-api-key
+kolega-code . --provider mistral --model mistral-medium-3-5
+```
+
+The picker defaults to Mistral Medium 3.5 and includes Mistral Small 4
+(`mistral-small-2603`), Mistral Large 4 **preview** (`mistral-large-4`),
+Ministral 3 (3B/8B/14B), Codestral, and eligible native aliases. Large 4 access
+depends on the account/subscription; a catalog entry does not guarantee access.
+Aliases, especially `latest`, can change their target and capabilities.
+
+Medium 3.5, Small 4, Large 4, and Ministral 3 support image input; Codestral is
+text-only. Medium/Small/Large expose `none` (default) or `high` thinking, with
+same-provider reasoning replayed in Mistral's native format. Token counting is
+a conservative local estimate with a safety margin, not an exact Mistral
+tokenizer. The catalog's 32,768-token output reservation is Kolega's budgeting
+policy, not a verified native maximum-output ceiling. Kolega Code runs the local
+function-tool loop; Mistral-hosted server tools are not enabled. Native Mistral
+keys and model IDs are separate from gateway credentials and routes.
+
 ## Requirements
 
 - **Python 3.11+**

{kolega_code-0.41.4 → kolega_code-0.43.0}/kolega_code/__init__.py

@@ -174,4 +174,4 @@ __all__ = [
     "parse_git_status_output",
 ]
 
-__version__ = "0.41.4"
+__version__ = "0.43.0"

{kolega_code-0.41.4 → kolega_code-0.43.0}/kolega_code/agent/__init__.py

@@ -51,4 +51,4 @@ __all__ = [
     "ToolExtension",
 ]
 
-__version__ = "0.41.4"
+__version__ = "0.43.0"

{kolega_code-0.41.4 → kolega_code-0.43.0}/kolega_code/cli/config.py

@@ -60,6 +60,7 @@ API_KEY_ENV = {
     ModelProvider.ANTHROPIC: "ANTHROPIC_API_KEY",
     ModelProvider.OPENAI: "OPENAI_API_KEY",
     ModelProvider.GOOGLE: "GOOGLE_API_KEY",
+    ModelProvider.MISTRAL: "MISTRAL_API_KEY",
     ModelProvider.GROQ: "GROQ_API_KEY",
     ModelProvider.TOGETHER: "TOGETHER_API_KEY",
     ModelProvider.FIREWORKS: "FIREWORKS_API_KEY",
@@ -967,6 +968,7 @@ def build_agent_config(
             anthropic_api_key=_api_key_for_provider(ModelProvider.ANTHROPIC, loaded_env, settings),
             openai_api_key=_api_key_for_provider(ModelProvider.OPENAI, loaded_env, settings),
             google_api_key=_api_key_for_provider(ModelProvider.GOOGLE, loaded_env, settings),
+            mistral_api_key=_api_key_for_provider(ModelProvider.MISTRAL, loaded_env, settings),
             groq_api_key=_api_key_for_provider(ModelProvider.GROQ, loaded_env, settings),
             together_api_key=_api_key_for_provider(ModelProvider.TOGETHER, loaded_env, settings),
             fireworks_api_key=_api_key_for_provider(ModelProvider.FIREWORKS, loaded_env, settings),

{kolega_code-0.41.4 → kolega_code-0.43.0}/kolega_code/cli/provider_registry.py

@@ -37,6 +37,7 @@ PROVIDER_LABELS: dict[ModelProvider, str] = {
     ModelProvider.OPENAI: "OpenAI",
     ModelProvider.OPENAI_CHATGPT: "OpenAI (ChatGPT subscription)",
     ModelProvider.GOOGLE: "Google",
+    ModelProvider.MISTRAL: "Mistral AI",
     ModelProvider.XAI: "xAI",
     ModelProvider.GROQ: "Groq",
     ModelProvider.FIREWORKS: "Fireworks",
@@ -110,6 +111,25 @@ MODEL_LABELS: dict[str, str] = {
     "gemini-3.5-flash-lite": "Gemini 3.5 Flash-Lite",
     "gemini-3.1-pro-preview": "Gemini 3.1 Pro",
     "gemini-3.5-flash": "Gemini 3.5 Flash",
+    # Mistral (native API)
+    "mistral-medium-3-5": "Mistral Medium 3.5",
+    "mistral-medium-latest": "Mistral Medium (latest alias)",
+    "mistral-medium": "Mistral Medium (alias)",
+    "mistral-medium-3.5": "Mistral Medium 3.5 (alias)",
+    "mistral-medium-3": "Mistral Medium 3 (alias)",
+    "mistral-medium-2604": "Mistral Medium (2604 alias)",
+    "mistral-small-2603": "Mistral Small 4",
+    "mistral-small-latest": "Mistral Small (latest alias)",
+    "mistral-large-4": "Mistral Large 4 (Preview)",
+    "mistral-large-4-0": "Mistral Large 4 (Preview, alias)",
+    "ministral-3b-2512": "Ministral 3 3B",
+    "ministral-3b-latest": "Ministral 3 3B (latest alias)",
+    "ministral-8b-2512": "Ministral 3 8B",
+    "ministral-8b-latest": "Ministral 3 8B (latest alias)",
+    "ministral-14b-2512": "Ministral 3 14B",
+    "ministral-14b-latest": "Ministral 3 14B (latest alias)",
+    "codestral-2508": "Codestral (2508)",
+    "codestral-latest": "Codestral (latest alias)",
     # xAI
     "grok-4.6": "Grok 4.6",
     "grok-4.5": "Grok 4.5",
@@ -176,6 +196,7 @@ PROVIDER_DEFAULT_MODEL: dict[ModelProvider, str] = {
     ModelProvider.OPENAI: "gpt-5.6-sol",
     ModelProvider.OPENAI_CHATGPT: "gpt-5.6-sol",
     ModelProvider.GOOGLE: "gemini-3.8-flash",
+    ModelProvider.MISTRAL: "mistral-medium-3-5",
     ModelProvider.XAI: "grok-4.5",
     ModelProvider.FIREWORKS: "accounts/fireworks/models/glm-5p2",
     ModelProvider.TOGETHER: "moonshotai/Kimi-K2.7-Code",

{kolega_code-0.41.4 → kolega_code-0.43.0}/kolega_code/config.py

@@ -24,6 +24,7 @@ class ModelProvider(str, Enum):
     OPENAI = "openai"
     OPENAI_CHATGPT = "openai_chatgpt"  # OpenAI via ChatGPT-subscription OAuth (Responses API)
     GOOGLE = "google"
+    MISTRAL = "mistral"
     GROQ = "groq"
     TOGETHER = "together"
     FIREWORKS = "fireworks"
@@ -205,6 +206,7 @@ class AgentConfig(BaseModel):
     anthropic_api_key: Optional[str] = Field(default=None, description="API key for Anthropic")
     openai_api_key: Optional[str] = Field(default=None, description="API key for OpenAI")
     google_api_key: Optional[str] = Field(default=None, description="API key for Google")
+    mistral_api_key: Optional[str] = Field(default=None, description="API key for Mistral AI")
     groq_api_key: Optional[str] = Field(default=None, description="API key for Groq")
     together_api_key: Optional[str] = Field(default=None, description="API key for Together.ai")
     fireworks_api_key: Optional[str] = Field(default=None, description="API key for Fireworks.ai")
@@ -416,6 +418,7 @@ class AgentConfig(BaseModel):
                 self.openai_chatgpt_tokens.access_token if self.openai_chatgpt_tokens else None
             ),
             ModelProvider.GOOGLE: self.google_api_key,
+            ModelProvider.MISTRAL: self.mistral_api_key,
             ModelProvider.GROQ: self.groq_api_key,
             ModelProvider.TOGETHER: self.together_api_key,
             ModelProvider.FIREWORKS: self.fireworks_api_key,

{kolega_code-0.41.4 → kolega_code-0.43.0}/kolega_code/llm/client.py

@@ -48,6 +48,7 @@ if TYPE_CHECKING:
     # module imports its own SDK at module load).
     from .providers.anthropic import AnthropicProvider
     from .providers.google import GoogleProvider
+    from .providers.mistral import MistralProvider
     from .providers.openai import OpenAIProvider
     from .providers.tinker import TinkerProvider
 
@@ -158,6 +159,10 @@ class LLMClient:
             from .providers.google import GoogleProvider
 
             return GoogleProvider
+        if p == "mistral":
+            from .providers.mistral import MistralProvider
+
+            return MistralProvider
         if p == "tinker":
             from .providers.tinker import TinkerProvider
 
@@ -172,7 +177,7 @@ class LLMClient:
         tokens_per_minute: Optional[int] = None,
         base_url: Optional[str] = None,
         api_style: Optional[str] = None,
-    ) -> "Union[AnthropicProvider, OpenAIProvider, GoogleProvider, TinkerProvider]":
+    ) -> "Union[AnthropicProvider, OpenAIProvider, GoogleProvider, MistralProvider, TinkerProvider]":
         """Initialize the appropriate LLM provider based on the provider name.
 
         Args:
@@ -235,6 +240,7 @@ class LLMClient:
                 "fireworks": "https://api.fireworks.ai/inference/v1",
                 "llama": "http://localhost:8000/v1",
                 "google": "https://generativelanguage.googleapis.com",
+                "mistral": "https://api.mistral.ai/v1/",
                 "xai": "https://api.x.ai/v1",
                 "dashscope": "https://dashscope-intl.aliyuncs.com/compatible-mode/v1",
                 "moonshot": "https://api.moonshot.ai/anthropic",

{kolega_code-0.41.4 → kolega_code-0.43.0}/kolega_code/llm/instrumented_client.py

@@ -303,7 +303,7 @@ class InstrumentedLLMClient(LLMClient):
                         "cache_read_input_tokens": usage_details.get("cache_read_input_tokens", 0),
                         "cache_creation_input_tokens": usage_details.get("cache_write_input_tokens", 0),
                     }
-                elif provider in ["openai", "openai_chatgpt", "fireworks", "deepseek"]:
+                elif provider in ["openai", "openai_chatgpt", "fireworks", "deepseek", "mistral"]:
                     normalized_usage = {
                         "input": usage_details.get("prompt_tokens", 0),
                         "output": usage_details.get("completion_tokens", 0),
@@ -576,7 +576,7 @@ class MinimalLangfuseStreamWrapper:
                 "cache_read_input_tokens": usage_metadata.get("cache_read_input_tokens", 0),
                 "cache_creation_input_tokens": usage_metadata.get("cache_write_input_tokens", 0),
             }
-        elif provider in ["openai", "openai_chatgpt", "fireworks", "deepseek"]:
+        elif provider in ["openai", "openai_chatgpt", "fireworks", "deepseek", "mistral"]:
             return {
                 "input": usage_metadata.get("prompt_tokens", 0),
                 "output": usage_metadata.get("completion_tokens", 0),

{kolega_code-0.41.4 → kolega_code-0.43.0}/kolega_code/llm/models.py

@@ -100,6 +100,7 @@ def safe_parse_tool_arguments(raw: str) -> Dict[str, Any]:
     """Parse OpenAI tool function.arguments into a dict safely.
 
     - Tries strict json.loads first
+    - Accepts raw control characters inside strings (a literal tab or newline)
     - Applies minimal, conservative repairs (trim, remove trailing commas, balance one missing brace)
     - Returns a fallback dict with _raw_arguments and _parse_error on failure
     """
@@ -108,11 +109,16 @@ def safe_parse_tool_arguments(raw: str) -> Dict[str, Any]:
             return {}
         return json.loads(raw)
     except Exception as first_err:
+        # Models sometimes leave a raw tab or newline inside a string.
+        try:
+            return json.loads(raw, strict=False)
+        except Exception:
+            pass
         repaired = raw.strip()
         repaired = _remove_trailing_commas(repaired)
         repaired = _balance_brackets(repaired)
         try:
-            return json.loads(repaired)
+            return json.loads(repaired, strict=False)
         except Exception as second_err:
             # Last resort: do not crash; surface raw args for downstream handling
             snippet = raw if len(raw) <= 200 else raw[:200] + "..."
@@ -359,18 +365,29 @@ class ThinkingBlock(ContentBlock):
 
     TYPE_NAME = "thinking"
 
-    def __init__(self, thinking: str, cache_checkpoint: bool = False, signature: Optional[str] = None):
+    def __init__(
+        self,
+        thinking: str,
+        cache_checkpoint: bool = False,
+        signature: Optional[str] = None,
+        provider_metadata: Optional[Dict[str, Any]] = None,
+    ):
         super().__init__(type=self.TYPE_NAME, cache_checkpoint=cache_checkpoint)
         self.thinking = thinking
         self.signature = signature
+        # Preserve replay-critical native structure without changing how other
+        # providers serialize this block (e.g. Mistral nested chunks/closed).
+        self.provider_metadata = provider_metadata or {}
 
     def to_dict(self) -> Dict[str, Any]:
-        result = {
+        result: Dict[str, Any] = {
             "type": self.type,
             "thinking": self.thinking,
         }
-        if self.signature:
+        if self.signature or "signature" in self.provider_metadata:
             result["signature"] = self.signature
+        if self.provider_metadata:
+            result["provider_metadata"] = self.provider_metadata
         return result
 
     @classmethod
@@ -378,6 +395,7 @@ class ThinkingBlock(ContentBlock):
         return cls(
             thinking=data["thinking"],
             signature=data.get("signature"),
+            provider_metadata=data.get("provider_metadata"),
         )
 
     def to_anthropic(self) -> Dict[str, Any]:
@@ -764,6 +782,7 @@ class ToolDefinition(ContentBlock):
         input_kind: ToolInputKind = "json",
         freeform_format: Optional[Dict[str, str]] = None,
         cache_ttl: Optional[str] = None,
+        strict: bool = False,
     ):
         super().__init__(type="tool_definition", cache_checkpoint=cache_checkpoint, cache_ttl=cache_ttl)
         self.name = name
@@ -778,6 +797,8 @@ class ToolDefinition(ContentBlock):
         # Provider-native constraint for a freeform tool. OpenAI Responses accepts
         # {type: "grammar", syntax: "lark"|"regex", definition: "..."}.
         self.freeform_format = freeform_format
+        # Sent by the Responses providers; the schema must meet the provider's strict rules.
+        self.strict = strict
 
     def _object_schema(self) -> Dict[str, Any]:
         """The JSON-schema object describing this tool's input."""
@@ -818,6 +839,8 @@ class ToolDefinition(ContentBlock):
             public["input_kind"] = self.input_kind
         if self.freeform_format:
             public["freeform_format"] = dict(self.freeform_format)
+        if self.strict:
+            public["strict"] = True
         return public
 
     @classmethod
@@ -1544,7 +1567,32 @@ class Message:
                 if thinking_blocks and (rest_blocks or has_tool_calls):
                     reasoning_text = "\n\n".join(b.thinking for b in thinking_blocks)
                     non_tool_blocks = rest_blocks
-            content = [item.to_openai() for item in non_tool_blocks]
+            if provider and _provider_value(provider) == "mistral":
+                content = []
+                for item in non_tool_blocks:
+                    if isinstance(item, ThinkingBlock):
+                        if (
+                            self.role != "assistant"
+                            or source_provider is None
+                            or _provider_value(source_provider) != "mistral"
+                        ):
+                            continue
+                        native = item.provider_metadata
+                        thinking_chunk: Dict[str, Any] = {
+                            "type": "thinking",
+                            "thinking": native.get("thinking", [{"type": "text", "text": item.thinking}]),
+                        }
+                        if "closed" in native:
+                            thinking_chunk["closed"] = native["closed"]
+                        if "signature" in native:
+                            thinking_chunk["signature"] = native["signature"]
+                        if item.signature is not None:
+                            thinking_chunk["signature"] = item.signature
+                        content.append(thinking_chunk)
+                    elif not isinstance(item, RedactedThinkingBlock):
+                        content.append(item.to_openai())
+            else:
+                content = [item.to_openai() for item in non_tool_blocks]
             # Strict OpenAI-compatible servers (e.g. DeepSeek) prefer an empty
             # string over an empty list when an assistant message carries only
             # tool_calls and/or native reasoning.

{kolega_code-0.41.4 → kolega_code-0.43.0}/kolega_code/llm/providers/deepseek_responses.py

@@ -9,7 +9,7 @@ other providers (Fireworks/OpenRouter/Ollama-Cloud) still use Chat Completions v
 Request building, streaming, and token counting are shared with the OpenAI
 Responses providers via
 :class:`~kolega_code.llm.providers.responses_common.ResponsesProviderBase`; only the
-transport differs — a plain api key against ``api.deepseek.com``.
+transport differs — a plain api key against ``api.deepseek.com/beta``.
 
 Reasoning continuity: DeepSeek returns no ``reasoning.encrypted_content`` (the
 ``include`` param the shared builder sends is silently ignored) but exposes the raw
@@ -31,8 +31,9 @@ back to the API" (findings/flash-output-token-efficiency-levers.md), while a
 minimal single-round history with a foreign call_id — fabricated or a genuine
 flash id from another conversation — is simply tolerated (probed 2026-08-04).
 
-The base URL is the bare host — DeepSeek's Responses guide uses
-``https://api.deepseek.com`` (no ``/v1``).
+The base URL is DeepSeek's beta endpoint, ``https://api.deepseek.com/beta`` (no
+``/v1``): strict tool schemas are enforced only there, and the bare host silently
+ignores ``strict: true`` (probed live 2026-10-12).
 """
 
 from __future__ import annotations
@@ -48,7 +49,7 @@ from .base import BaseLLMProvider
 from .models import GenerationParams
 from .responses_common import ResponsesProviderBase
 
-DEFAULT_BASE_URL = "https://api.deepseek.com"
+DEFAULT_BASE_URL = "https://api.deepseek.com/beta"
 DEFAULT_MODEL = "deepseek-flash"