lm-deluge 0.0.79__py3-none-any.whl → 0.0.81__py3-none-any.whl

lm_deluge/__init__.py

@@ -1,7 +1,7 @@
 from .client import APIResponse, LLMClient, SamplingParams
 from .file import File
 from .prompt import Conversation, Message
-from .tool import Tool, ToolParams
+from .tool import Tool
 
 try:
     from .mock_openai import (  # noqa
@@ -25,7 +25,6 @@ __all__ = [
     "Conversation",
     "Message",
     "Tool",
-    "ToolParams",
     "File",
 ]
 

lm_deluge/api_requests/anthropic.py

@@ -16,6 +16,7 @@ from lm_deluge.util.schema import (
     prepare_output_schema,
     transform_schema_for_anthropic,
 )
+from lm_deluge.warnings import maybe_warn
 
 from ..models import APIModel
 from .base import APIRequestBase, APIResponse
@@ -62,20 +63,45 @@ def _build_anthropic_request(
         "max_tokens": sampling_params.max_new_tokens,
     }
 
+    if model.id == "claude-4.5-opus" and sampling_params.global_effort:
+        request_json["output_config"] = {"effort": sampling_params.global_effort}
+        _add_beta(base_headers, "effort-2025-11-24")
+
     # handle thinking
-    if model.reasoning_model and sampling_params.reasoning_effort:
-        # translate reasoning effort of low, medium, high to budget tokens
-        budget = {"minimal": 256, "low": 1024, "medium": 4096, "high": 16384}.get(
-            sampling_params.reasoning_effort
-        )
-        request_json["thinking"] = {
-            "type": "enabled",
-            "budget_tokens": budget,
-        }
-        if "top_p" in request_json:
-            request_json["top_p"] = max(request_json["top_p"], 0.95)
-        request_json["temperature"] = 1.0
-        request_json["max_tokens"] += budget
+    if model.reasoning_model:
+        if (
+            sampling_params.thinking_budget is not None
+            and sampling_params.reasoning_effort is not None
+        ):
+            maybe_warn("WARN_THINKING_BUDGET_AND_REASONING_EFFORT")
+
+        if sampling_params.thinking_budget is not None:
+            budget = sampling_params.thinking_budget
+        elif sampling_params.reasoning_effort is not None:
+            # translate reasoning effort of low, medium, high to budget tokens
+            budget = {
+                "none": 0,
+                "minimal": 256,
+                "low": 1024,
+                "medium": 4096,
+                "high": 16384,
+            }.get(sampling_params.reasoning_effort)
+            assert isinstance(budget, int)
+        else:
+            budget = 0
+
+        if budget > 0:
+            request_json["thinking"] = {
+                "type": "enabled",
+                "budget_tokens": budget,
+            }
+            if "top_p" in request_json:
+                request_json["top_p"] = max(request_json["top_p"], 0.95)
+            request_json["temperature"] = 1.0
+            request_json["max_tokens"] += budget
+        else:
+            request_json["thinking"] = {"type": "disabled"}
+
     else:
         request_json["thinking"] = {"type": "disabled"}
         if sampling_params.reasoning_effort:
@@ -83,11 +109,13 @@ def _build_anthropic_request(
     if system_message is not None:
         request_json["system"] = system_message
 
-    # handle temp + top_p for opus 4.1/sonnet 4.5
+    # handle temp + top_p for opus 4.1/sonnet 4.5.
+    # TODO: make clearer / more user-friendly so there can be NotGiven
+    # and user can control which one they want to use
     if "4-1" in model.name or "4-5" in model.name:
-        if "temperature" in request_json and "top_p" in request_json:
-            request_json.pop("top_p")
+        request_json.pop("top_p")
 
+    # print(request_json)
     # Handle structured outputs (output_format)
     if context.output_schema:
         if model.supports_json:

lm_deluge/api_requests/base.py

@@ -222,6 +222,19 @@ class APIRequestBase(ABC):
                 usage=None,
             )
 
+        except aiohttp.ServerDisconnectedError:
+            return APIResponse(
+                id=self.context.task_id,
+                model_internal=self.context.model_name,
+                prompt=self.context.prompt,
+                sampling_params=self.context.sampling_params,
+                status_code=None,
+                is_error=True,
+                error_message="Server disconnected.",
+                content=None,
+                usage=None,
+            )
+
         except Exception as e:
             raise_if_modal_exception(e)
             tb = traceback.format_exc()

lm_deluge/api_requests/gemini.py

@@ -1,6 +1,5 @@
 import json
 import os
-from typing import Any
 
 from aiohttp import ClientResponse
 
@@ -52,47 +51,61 @@ async def _build_gemini_request(
         request_json["systemInstruction"] = {"parts": [{"text": system_message}]}
 
     # Handle reasoning models (thinking)
-    if model.reasoning_model:
-        thinking_config: dict[str, Any] | None = None
-        effort = sampling_params.reasoning_effort
-        is_gemini_3 = "gemini-3" in model.name.lower()
+    is_gemini_3 = "gemini-3" in model.name.lower()
+    if is_gemini_3:
+        # gemini3 MUST think
+        if not sampling_params.reasoning_effort:
+            maybe_warn("WARN_GEMINI3_NO_REASONING")
+            effort = "low"
+        else:
+            level_map = {
+                "none": "low",
+                "minimal": "low",
+                "low": "low",
+                "medium": "high",  # change when supported
+                "high": "high",
+            }
+            effort = level_map[sampling_params.reasoning_effort]
+        thinking_config = {"thinkingLevel": effort}
+        request_json["generationConfig"]["thinkingConfig"] = thinking_config
 
-        if is_gemini_3:
-            # Gemini 3 uses thinkingLevel instead of thinkingBudget
-            if effort in {"none", "minimal"}:
-                thinking_config = {"thinkingLevel": "low"}
-            elif effort is None:
-                # Default to high when reasoning is enabled but no preference was provided
-                thinking_config = {"thinkingLevel": "high"}
-            else:
-                # Map reasoning_effort to thinkingLevel
-                level_map = {
-                    "minimal": "low",
-                    "low": "low",
-                    "medium": "medium",  # Will work when supported
-                    "high": "high",
-                }
-                thinking_level = level_map.get(effort, "high")
-                thinking_config = {"thinkingLevel": thinking_level}
+    elif model.reasoning_model:
+        if (
+            sampling_params.thinking_budget is not None
+            and sampling_params.reasoning_effort is not None
+        ):
+            maybe_warn("WARN_THINKING_BUDGET_AND_REASONING_EFFORT")
+
+        if (
+            sampling_params.thinking_budget is not None
+            and sampling_params.thinking_budget > 0
+        ):
+            thinking_config = {
+                "includeThoughts": True,
+                "thinkingBudget": sampling_params.thinking_budget,
+            }
+        elif sampling_params.thinking_budget == -1:
+            # dynamic thinking
+            thinking_config = {"includeThoughts": True, "thinkingBudget": -1}
+        elif sampling_params.reasoning_effort not in [None, "none"]:
+            level_map = {
+                "minimal": 256,
+                "low": 1024,
+                "medium": 4096,
+                "high": 16384,
+            }
+            assert sampling_params.reasoning_effort in level_map
+            budget = level_map[sampling_params.reasoning_effort]
+            if "flash-lite" in model.id:
+                budget = max(budget, 512)
+            thinking_config = {"includeThoughts": True, "thinkingBudget": budget}
+        elif "2.5-pro" in model.id:
+            # 2.5 pro must think.
+            thinking_config = {"includeThoughts": True, "thinkingBudget": 128}
         else:
-            # Gemini 2.5 uses thinkingBudget (legacy)
-            if effort is None or effort == "none":
-                budget = 128 if "2.5-pro" in model.id else 0
-                # Explicitly disable thoughts when no effort is requested
-                thinking_config = {"includeThoughts": False, "thinkingBudget": budget}
-            else:
-                thinking_config = {"includeThoughts": True}
-                if (
-                    effort in {"minimal", "low", "medium", "high"}
-                    and "flash" in model.id
-                ):
-                    budget = {
-                        "minimal": 256,
-                        "low": 1024,
-                        "medium": 4096,
-                        "high": 16384,
-                    }[effort]
-                    thinking_config["thinkingBudget"] = budget
+            # no thoughts head empty
+            thinking_config = {"includeThoughts": False, "thinkingBudget": 0}
+
         request_json["generationConfig"]["thinkingConfig"] = thinking_config
 
     else:
@@ -160,7 +173,7 @@ class GeminiRequest(APIRequestBase):
         self.request_json = await _build_gemini_request(
             self.model,
             self.context.prompt,
-            self.context.tools,
+            self.context.tools,  # type: ignore
             self.context.sampling_params,
         )
 

lm_deluge/api_requests/openai.py

@@ -2,17 +2,18 @@ import json
 import os
 import traceback as tb
 from types import SimpleNamespace
+from typing import Sequence
 
 import aiohttp
 from aiohttp import ClientResponse
 
 from lm_deluge.request_context import RequestContext
 from lm_deluge.tool import MCPServer, Tool
-from lm_deluge.warnings import maybe_warn
 from lm_deluge.util.schema import (
     prepare_output_schema,
     transform_schema_for_openai,
 )
+from lm_deluge.warnings import maybe_warn
 
 from ..config import SamplingParams
 from ..models import APIModel
@@ -610,7 +611,7 @@ async def stream_chat(
     model_name: str,  # must correspond to registry
     prompt: Conversation,
     sampling_params: SamplingParams = SamplingParams(),
-    tools: list | None = None,
+    tools: Sequence[Tool | dict | MCPServer] | None = None,
     cache: CachePattern | None = None,
     extra_headers: dict[str, str] | None = None,
 ):

lm_deluge/client.py

@@ -79,7 +79,7 @@ class _LLMClient(BaseModel):
     background: bool = False
     # sampling params - if provided, and sampling_params is not,
     # these override the defaults
-    temperature: float = 0.75
+    temperature: float = 1.0
     top_p: float = 1.0
     json_mode: bool = False
     max_new_tokens: int = 512
@@ -337,7 +337,7 @@ class _LLMClient(BaseModel):
         if "sampling_params" not in data or len(data.get("sampling_params", [])) == 0:
             data["sampling_params"] = [
                 SamplingParams(
-                    temperature=data.get("temperature", 0.75),
+                    temperature=data.get("temperature", 1.0),
                     top_p=data.get("top_p", 1.0),
                     json_mode=data.get("json_mode", False),
                     max_new_tokens=data.get("max_new_tokens", 512),
@@ -572,7 +572,7 @@ class _LLMClient(BaseModel):
         *,
         return_completions_only: Literal[True],
         show_progress: bool = ...,
-        tools: list[Tool | dict | MCPServer] | None = ...,
+        tools: Sequence[Tool | dict | MCPServer] | None = ...,
         output_schema: type[BaseModel] | dict | None = ...,
         cache: CachePattern | None = ...,
         service_tier: Literal["auto", "default", "flex", "priority"] | None = ...,
@@ -585,7 +585,7 @@ class _LLMClient(BaseModel):
         *,
         return_completions_only: Literal[False] = ...,
         show_progress: bool = ...,
-        tools: list[Tool | dict | MCPServer] | None = ...,
+        tools: Sequence[Tool | dict | MCPServer] | None = ...,
         output_schema: type[BaseModel] | dict | None = ...,
         cache: CachePattern | None = ...,
         service_tier: Literal["auto", "default", "flex", "priority"] | None = ...,
@@ -597,7 +597,7 @@ class _LLMClient(BaseModel):
         *,
         return_completions_only: bool = False,
         show_progress: bool = True,
-        tools: list[Tool | dict | MCPServer] | None = None,
+        tools: Sequence[Tool | dict | MCPServer] | None = None,
         output_schema: type[BaseModel] | dict | None = None,
         cache: CachePattern | None = None,
         service_tier: Literal["auto", "default", "flex", "priority"] | None = None,
@@ -672,7 +672,7 @@ class _LLMClient(BaseModel):
         *,
         return_completions_only: bool = False,
         show_progress=True,
-        tools: list[Tool | dict | MCPServer] | None = None,
+        tools: Sequence[Tool | dict | MCPServer] | None = None,
         output_schema: type[BaseModel] | dict | None = None,
         cache: CachePattern | None = None,
     ):
@@ -705,7 +705,7 @@ class _LLMClient(BaseModel):
         self,
         prompt: Prompt,
         *,
-        tools: list[Tool | dict | MCPServer] | None = None,
+        tools: Sequence[Tool | dict | MCPServer] | None = None,
         output_schema: type[BaseModel] | dict | None = None,
         cache: CachePattern | None = None,
         service_tier: Literal["auto", "default", "flex", "priority"] | None = None,
@@ -742,7 +742,7 @@ class _LLMClient(BaseModel):
         self,
         prompt: Prompt,
         *,
-        tools: list[Tool | dict | MCPServer] | None = None,
+        tools: Sequence[Tool | dict | MCPServer] | None = None,
         output_schema: type[BaseModel] | dict | None = None,
         cache: CachePattern | None = None,
         service_tier: Literal["auto", "default", "flex", "priority"] | None = None,
@@ -835,7 +835,7 @@ class _LLMClient(BaseModel):
     async def stream(
         self,
         prompt: Prompt,
-        tools: list[Tool | dict | MCPServer] | None = None,
+        tools: Sequence[Tool | dict | MCPServer] | None = None,
     ):
         model, sampling_params = self._select_model()
         prompt = prompts_to_conversations([prompt])[0]
@@ -856,7 +856,7 @@ class _LLMClient(BaseModel):
         task_id: int,
         conversation: Conversation,
         *,
-        tools: list[Tool | dict | MCPServer] | None = None,
+        tools: Sequence[Tool | dict | MCPServer] | None = None,
         max_rounds: int = 5,
     ) -> AgentLoopResponse:
         """Internal method to run agent loop and return wrapped result."""
@@ -920,7 +920,7 @@ class _LLMClient(BaseModel):
         self,
         conversation: Prompt,
         *,
-        tools: list[Tool | dict | MCPServer] | None = None,
+        tools: Sequence[Tool | dict | MCPServer] | None = None,
         max_rounds: int = 5,
     ) -> int:
         """Start an agent loop without waiting for it to complete.
@@ -967,7 +967,7 @@ class _LLMClient(BaseModel):
         self,
         conversation: Prompt,
         *,
-        tools: list[Tool | dict | MCPServer] | None = None,
+        tools: Sequence[Tool | dict | MCPServer] | None = None,
         max_rounds: int = 5,
         show_progress: bool = False,
     ) -> tuple[Conversation, APIResponse]:
@@ -986,7 +986,7 @@ class _LLMClient(BaseModel):
         self,
         conversation: Prompt,
         *,
-        tools: list[Tool | dict | MCPServer] | None = None,
+        tools: Sequence[Tool | dict | MCPServer] | None = None,
         max_rounds: int = 5,
         show_progress: bool = False,
     ) -> tuple[Conversation, APIResponse]:
@@ -1067,7 +1067,7 @@ def LLMClient(
     extra_headers: dict[str, str] | None = None,
     use_responses_api: bool = False,
     background: bool = False,
-    temperature: float = 0.75,
+    temperature: float = 1.0,
     top_p: float = 1.0,
     json_mode: bool = False,
     max_new_tokens: int = 512,
@@ -1096,7 +1096,7 @@ def LLMClient(
     extra_headers: dict[str, str] | None = None,
     use_responses_api: bool = False,
     background: bool = False,
-    temperature: float = 0.75,
+    temperature: float = 1.0,
     top_p: float = 1.0,
     json_mode: bool = False,
     max_new_tokens: int = 512,
@@ -1124,7 +1124,7 @@ def LLMClient(
     extra_headers: dict[str, str] | None = None,
     use_responses_api: bool = False,
     background: bool = False,
-    temperature: float = 0.75,
+    temperature: float = 1.0,
     top_p: float = 1.0,
     json_mode: bool = False,
     max_new_tokens: int = 512,

lm_deluge/config.py

@@ -4,11 +4,13 @@ from pydantic import BaseModel
 
 
 class SamplingParams(BaseModel):
-    temperature: float = 0.0
+    temperature: float = 1.0  # more typical for new models
     top_p: float = 1.0
     json_mode: bool = False
     max_new_tokens: int = 2_048
+    global_effort: Literal["low", "medium", "high"] = "high"  # for opus-4.5
     reasoning_effort: Literal["low", "medium", "high", "minimal", "none", None] = None
+    thinking_budget: int | None = None
     logprobs: bool = False
     top_logprobs: int | None = None
     strict_tools: bool = True

lm_deluge/llm_tools/__init__.py

@@ -1,8 +1,15 @@
-from .extract import extract, extract_async
-from .score import score_llm
-from .subagents import SubAgentManager
-from .todos import TodoItem, TodoManager, TodoPriority, TodoStatus
-from .translate import translate, translate_async
+# Backward compatibility - re-export from new locations
+# Pipelines (workflow functions)
+from ..pipelines import extract, extract_async, score_llm, translate, translate_async
+
+# Prefab tools (Tool managers)
+from ..tool.prefab import (
+    SubAgentManager,
+    TodoItem,
+    TodoManager,
+    TodoPriority,
+    TodoStatus,
+)
 
 __all__ = [
     "extract",

lm_deluge/models/anthropic.py

@@ -10,6 +10,19 @@ ANTHROPIC_MODELS = {
     #                                                                   ░███
     #                                                                   █████
     #
+    "claude-4.5-opus": {
+        "id": "claude-4.5-opus",
+        "name": "claude-opus-4-5-20251101",
+        "api_base": "https://api.anthropic.com/v1",
+        "api_key_env_var": "ANTHROPIC_API_KEY",
+        "supports_json": False,
+        "api_spec": "anthropic",
+        "input_cost": 5.0,
+        "cached_input_cost": 0.50,
+        "cache_write_cost": 6.25,
+        "output_cost": 25.0,
+        "reasoning_model": True,
+    },
     "claude-4.5-haiku": {
         "id": "claude-4.5-haiku",
         "name": "claude-haiku-4-5-20251001",
@@ -21,6 +34,7 @@ ANTHROPIC_MODELS = {
         "cached_input_cost": 0.10,
         "cache_write_cost": 1.25,
         "output_cost": 3.0,
+        "reasoning_model": True,
     },
     "claude-4.5-sonnet": {
         "id": "claude-4.5-sonnet",
@@ -33,6 +47,7 @@ ANTHROPIC_MODELS = {
         "cached_input_cost": 0.30,
         "cache_write_cost": 3.75,
         "output_cost": 15.0,
+        "reasoning_model": True,
     },
     "claude-4.1-opus": {
         "id": "claude-4.1-opus",

lm_deluge/pipelines/__init__.py

@@ -0,0 +1,11 @@
+from .extract import extract, extract_async
+from .score import score_llm
+from .translate import translate, translate_async
+
+__all__ = [
+    "extract",
+    "extract_async",
+    "translate",
+    "translate_async",
+    "score_llm",
+]

lm_deluge/{llm_tools → pipelines}/score.py

@@ -1,4 +1,4 @@
-from ..client import LLMClient, APIResponse
+from ..client import _LLMClient, APIResponse
 from ..util.logprobs import extract_prob
 
 # def extract_prob_yes(logprobs: list[dict]):
@@ -24,7 +24,7 @@ from ..util.logprobs import extract_prob
 def score_llm(
     scoring_prompt_template: str,
     inputs: list[tuple | list | dict],  # to format the template
-    scoring_model: LLMClient,
+    scoring_model: _LLMClient,
     return_probabilities: bool,
     yes_token: str = "yes",
 ) -> list[bool | None] | list[float | None]:

lm_deluge/{llm_tools → pipelines}/translate.py

@@ -1,5 +1,5 @@
 import asyncio
-from ..client import LLMClient
+from ..client import _LLMClient
 
 translation_prompt = (
     "Translate the following text (enclosed in ```) into English. "
@@ -20,7 +20,9 @@ def is_english(text: str, low_memory: bool = True):
     return True
 
 
-async def translate_async(texts: list[str], client: LLMClient, low_memory: bool = True):
+async def translate_async(
+    texts: list[str], client: _LLMClient, low_memory: bool = True
+):
     to_translate_idxs = [
         i for i, text in enumerate(texts) if not is_english(text, low_memory=low_memory)
     ]
@@ -40,5 +42,5 @@ async def translate_async(texts: list[str], client: LLMClient, low_memory: bool
     return texts
 
 
-def translate(texts: list[str], client: LLMClient, low_memory: bool = True):
+def translate(texts: list[str], client: _LLMClient, low_memory: bool = True):
     return asyncio.run(translate_async(texts, client, low_memory))

lm_deluge/prompt.py

@@ -1598,6 +1598,111 @@ class Conversation:
 
         return {"messages": serialized}
 
+    def print(self, max_text_length: int = 500, indent: int = 2) -> None:
+        """Pretty-print the conversation to stdout.
+
+        Args:
+            max_text_length: Truncate text content longer than this (default 500 chars)
+            indent: JSON indentation for tool calls/results (default 2)
+        """
+        ROLE_COLORS = {
+            "system": "\033[95m",  # magenta
+            "user": "\033[94m",  # blue
+            "assistant": "\033[92m",  # green
+            "tool": "\033[93m",  # yellow
+        }
+        RESET = "\033[0m"
+        DIM = "\033[2m"
+        BOLD = "\033[1m"
+
+        def truncate(text: str, max_len: int) -> str:
+            if len(text) <= max_len:
+                return text
+            return (
+                text[:max_len] + f"{DIM}... [{len(text) - max_len} more chars]{RESET}"
+            )
+
+        def format_json(obj: dict | list, ind: int) -> str:
+            return json.dumps(obj, indent=ind, ensure_ascii=False)
+
+        print(f"\n{BOLD}{'=' * 60}{RESET}")
+        print(f"{BOLD}Conversation ({len(self.messages)} messages){RESET}")
+        print(f"{BOLD}{'=' * 60}{RESET}\n")
+
+        for i, msg in enumerate(self.messages):
+            role_color = ROLE_COLORS.get(msg.role, "")
+            print(f"{role_color}{BOLD}[{msg.role.upper()}]{RESET}")
+
+            for part in msg.parts:
+                if isinstance(part, Text):
+                    text = truncate(part.text, max_text_length)
+                    # Indent multiline text
+                    lines = text.split("\n")
+                    if len(lines) > 1:
+                        print("  " + "\n  ".join(lines))
+                    else:
+                        print(f"  {text}")
+
+                elif isinstance(part, Image):
+                    w, h = part.size
+                    print(f"  {DIM}<Image ({w}x{h})>{RESET}")
+
+                elif isinstance(part, File):
+                    size = part.size
+                    filename = getattr(part, "filename", None)
+                    if filename:
+                        print(f"  {DIM}<File: {filename} ({size} bytes)>{RESET}")
+                    else:
+                        print(f"  {DIM}<File ({size} bytes)>{RESET}")
+
+                elif isinstance(part, ToolCall):
+                    print(
+                        f"  {DIM}Tool Call:{RESET} {BOLD}{part.name}{RESET} (id: {part.id})"
+                    )
+                    if part.arguments:
+                        args_json = format_json(part.arguments, indent)
+                        # Indent the JSON
+                        indented = "\n".join(
+                            "    " + line for line in args_json.split("\n")
+                        )
+                        print(indented)
+
+                elif isinstance(part, ToolResult):
+                    print(f"  {DIM}Tool Result:{RESET} (call_id: {part.tool_call_id})")
+                    if isinstance(part.result, str):
+                        result_text = truncate(part.result, max_text_length)
+                        lines = result_text.split("\n")
+                        for line in lines:
+                            print(f"    {line}")
+                    elif isinstance(part.result, dict):
+                        result_json = format_json(part.result, indent)
+                        indented = "\n".join(
+                            "    " + line for line in result_json.split("\n")
+                        )
+                        print(indented)
+                    elif isinstance(part.result, list):
+                        print(f"    {DIM}<{len(part.result)} content blocks>{RESET}")
+                        for block in part.result:
+                            if isinstance(block, Text):
+                                block_text = truncate(block.text, max_text_length // 2)
+                                print(f"      [text] {block_text}")
+                            elif isinstance(block, Image):
+                                bw, bh = block.size
+                                print(f"      {DIM}<Image ({bw}x{bh})>{RESET}")
+
+                elif isinstance(part, Thinking):
+                    print(f"  {DIM}Thinking:{RESET}")
+                    thought = truncate(part.content, max_text_length)
+                    lines = thought.split("\n")
+                    for line in lines:
+                        print(f"    {DIM}{line}{RESET}")
+
+            # Separator between messages
+            if i < len(self.messages) - 1:
+                print(f"\n{'-' * 40}\n")
+
+        print(f"\n{BOLD}{'=' * 60}{RESET}\n")
+
     @classmethod
     def from_log(cls, payload: dict) -> "Conversation":
         """Re-hydrate a Conversation previously produced by `to_log()`."""

lm_deluge/request_context.py

@@ -1,6 +1,6 @@
 from dataclasses import dataclass, field
 from functools import cached_property
-from typing import Any, Callable, TYPE_CHECKING
+from typing import Any, Callable, Sequence, TYPE_CHECKING
 
 from .config import SamplingParams
 from .prompt import CachePattern, Conversation
@@ -34,7 +34,7 @@ class RequestContext:
     callback: Callable | None = None
 
     # Optional features
-    tools: list | None = None
+    tools: Sequence[Any] | None = None
     output_schema: "type[BaseModel] | dict | None" = None
     cache: CachePattern | None = None
     use_responses_api: bool = False