inspect-ai 0.3.76__py3-none-any.whl → 0.3.78__py3-none-any.whl

inspect_ai/model/_providers/openai_responses.py

@@ -0,0 +1,177 @@
+from logging import getLogger
+from typing import Any
+
+from openai import (
+    AsyncAzureOpenAI,
+    AsyncOpenAI,
+    BadRequestError,
+)
+from openai._types import NOT_GIVEN
+from openai.types.responses import Response, ResponseFormatTextJSONSchemaConfigParam
+
+from inspect_ai._util.logger import warn_once
+from inspect_ai.tool import ToolChoice, ToolInfo
+
+from .._chat_message import ChatMessage
+from .._generate_config import GenerateConfig
+from .._model_call import ModelCall
+from .._model_output import (
+    ModelOutput,
+    ModelUsage,
+)
+from .._openai import (
+    OpenAIResponseError,
+    is_gpt,
+    is_o1_mini,
+    is_o1_preview,
+    is_o_series,
+    openai_handle_bad_request,
+    openai_media_filter,
+)
+from .._openai_responses import (
+    openai_responses_chat_choices,
+    openai_responses_inputs,
+    openai_responses_tool_choice,
+    openai_responses_tools,
+)
+from .util.hooks import HttpxHooks
+
+logger = getLogger(__name__)
+
+
+async def generate_responses(
+    client: AsyncAzureOpenAI | AsyncOpenAI,
+    http_hooks: HttpxHooks,
+    model_name: str,
+    input: list[ChatMessage],
+    tools: list[ToolInfo],
+    tool_choice: ToolChoice,
+    config: GenerateConfig,
+) -> ModelOutput | tuple[ModelOutput | Exception, ModelCall]:
+    # allocate request_id (so we can see it from ModelCall)
+    request_id = http_hooks.start_request()
+
+    # setup request and response for ModelCall
+    request: dict[str, Any] = {}
+    response: dict[str, Any] = {}
+
+    def model_call() -> ModelCall:
+        return ModelCall.create(
+            request=request,
+            response=response,
+            # TODO: is this the right filter?
+            filter=openai_media_filter,
+            time=http_hooks.end_request(request_id),
+        )
+
+    # prepare request (we do this so we can log the ModelCall)
+    request = dict(
+        input=await openai_responses_inputs(input, model_name),
+        tools=openai_responses_tools(tools) if len(tools) > 0 else NOT_GIVEN,
+        tool_choice=openai_responses_tool_choice(tool_choice)
+        if len(tools) > 0
+        else NOT_GIVEN,
+        extra_headers={HttpxHooks.REQUEST_ID_HEADER: request_id},
+        **completion_params_responses(model_name, config, len(tools) > 0),
+    )
+
+    try:
+        # generate response
+        model_response: Response = await client.responses.create(**request)
+
+        # check for error
+        if model_response.error is not None:
+            raise OpenAIResponseError(
+                code=model_response.error.code, message=model_response.error.message
+            )
+
+        # save response for model_call
+        response = model_response.model_dump()
+
+        # parse out choices
+        choices = openai_responses_chat_choices(model_response, tools)
+
+        # return output and call
+        return ModelOutput(
+            model=model_response.model,
+            choices=choices,
+            usage=(
+                ModelUsage(
+                    input_tokens=model_response.usage.input_tokens,
+                    output_tokens=model_response.usage.output_tokens,
+                    input_tokens_cache_read=(
+                        model_response.usage.input_tokens_details.cached_tokens
+                    ),
+                    reasoning_tokens=model_response.usage.output_tokens_details.reasoning_tokens,
+                    total_tokens=model_response.usage.total_tokens,
+                )
+                if model_response.usage
+                else None
+            ),
+        ), model_call()
+    except BadRequestError as e:
+        return openai_handle_bad_request(model_name, e), model_call()
+
+
+def completion_params_responses(
+    model_name: str, config: GenerateConfig, tools: bool
+) -> dict[str, Any]:
+    # TODO: we'll need a computer_use_preview bool for the 'include'
+    # and 'reasoning' parameters
+    def unsupported_warning(param: str) -> None:
+        warn_once(
+            logger,
+            f"OpenAI Responses API does not support the '{param}' parameter.",
+        )
+
+    params: dict[str, Any] = dict(model=model_name, store=False)
+    if config.max_tokens is not None:
+        params["max_output_tokens"] = config.max_tokens
+    if config.frequency_penalty is not None:
+        unsupported_warning("frequency_penalty")
+    if config.stop_seqs is not None:
+        unsupported_warning("stop_seqs")
+    if config.presence_penalty is not None:
+        unsupported_warning("presence_penalty")
+    if config.logit_bias is not None:
+        unsupported_warning("logit_bias")
+    if config.seed is not None:
+        unsupported_warning("seed")
+    if config.temperature is not None:
+        if is_o_series(model_name):
+            warn_once(
+                logger,
+                "o series models do not support the 'temperature' parameter (temperature is always 1).",
+            )
+        else:
+            params["temperature"] = config.temperature
+    if config.top_p is not None:
+        params["top_p"] = config.top_p
+    if config.num_choices is not None:
+        unsupported_warning("num_choices")
+    if config.logprobs is not None:
+        unsupported_warning("logprobs")
+    if config.top_logprobs is not None:
+        unsupported_warning("top_logprobs")
+    if tools and config.parallel_tool_calls is not None and not is_o_series(model_name):
+        params["parallel_tool_calls"] = config.parallel_tool_calls
+    if (
+        config.reasoning_effort is not None
+        and not is_gpt(model_name)
+        and not is_o1_mini(model_name)
+        and not is_o1_preview(model_name)
+    ):
+        params["reasoning"] = dict(effort=config.reasoning_effort)
+    if config.response_schema is not None:
+        params["text"] = dict(
+            format=ResponseFormatTextJSONSchemaConfigParam(
+                type="json_schema",
+                name=config.response_schema.name,
+                schema=config.response_schema.json_schema.model_dump(exclude_none=True),
+                description=config.response_schema.description
+                or config.response_schema.name,
+                strict=config.response_schema.strict,
+            )
+        )
+
+    return params

inspect_ai/model/_providers/openrouter.py

@@ -1,9 +1,11 @@
+import json
 import os
-from typing import Any
+from typing import Any, TypedDict
 
-from typing_extensions import override
+from typing_extensions import NotRequired, override
 
 from inspect_ai._util.error import PrerequisiteError
+from inspect_ai.model._openai import OpenAIResponseError
 from inspect_ai.model._providers.util import model_base_url
 from inspect_ai.model._providers.util.util import environment_prerequisite_error
 
@@ -13,6 +15,28 @@ from .openai import OpenAIAPI
 OPENROUTER_API_KEY = "OPENROUTER_API_KEY"
 
 
+class ErrorResponse(TypedDict):
+    code: int
+    message: str
+    metadata: NotRequired[dict[str, Any]]
+
+
+class OpenRouterError(Exception):
+    def __init__(self, response: ErrorResponse) -> None:
+        self.response = response
+
+    @property
+    def message(self) -> str:
+        return f"Error {self.response['code']} - {self.response['message']}"
+
+    def __str__(self) -> str:
+        return (
+            self.message + ("\n" + json.dumps(self.response["metadata"], indent=2))
+            if "metadata" in self.response
+            else ""
+        )
+
+
 class OpenRouterAPI(OpenAIAPI):
     def __init__(
         self,
@@ -67,6 +91,32 @@ class OpenRouterAPI(OpenAIAPI):
             **model_args,
         )
 
+    @override
+    def on_response(self, response: dict[str, Any]) -> None:
+        """Handle documented OpenRouter error conditions.
+
+        https://openrouter.ai/docs/api-reference/errors
+        """
+        # check if open-router yielded an error (raise explicit
+        # OpenAIResponseError for cases where we should retry)
+        error: ErrorResponse | None = response.get("error", None)
+        if error is not None:
+            if error["code"] == 429:
+                raise OpenAIResponseError("rate_limit_exceeded", error["message"])
+            elif error["code"] in [408, 502]:
+                raise OpenAIResponseError("server_error", error["message"])
+            else:
+                raise OpenRouterError(error)
+
+        # check for an empty response (which they document can occur on
+        # startup). for this we'll return a "server_error" which will
+        # trigger a retry w/ exponential backoff
+        elif response.get("choices", None) is None:
+            raise OpenAIResponseError(
+                "server_error",
+                "Model is warming up, please retry again after waiting for warmup.",
+            )
+
     @override
     def completion_params(self, config: GenerateConfig, tools: bool) -> dict[str, Any]:
         # default params

inspect_ai/model/_providers/providers.py

@@ -282,7 +282,7 @@ def goodfire() -> type[ModelAPI]:
 def validate_openai_client(feature: str) -> None:
     FEATURE = feature
     PACKAGE = "openai"
-    MIN_VERSION = "1.58.1"
+    MIN_VERSION = "1.68.0"
 
     # verify we have the package
     try:

inspect_ai/tool/__init__.py

@@ -26,6 +26,7 @@ from ._tools._bash_session import bash_session
 from ._tools._computer import computer
 from ._tools._execute import bash, python
 from ._tools._text_editor import text_editor
+from ._tools._think import think
 from ._tools._web_browser import web_browser
 from ._tools._web_search import web_search
 
@@ -36,6 +37,7 @@ __all__ = [
     "python",
     "web_browser",
     "web_search",
+    "think",
     "text_editor",
     "tool",
     "tool_with",

inspect_ai/tool/_tool.py

@@ -20,6 +20,7 @@ from inspect_ai._util.content import (
 )
 from inspect_ai._util.registry import (
     RegistryInfo,
+    is_registry_object,
     registry_add,
     registry_name,
     registry_tag,
@@ -200,7 +201,25 @@ def tool(
         # wrap instantiations of scorer so they carry registry info and metrics
         @wraps(tool_type)
         def tool_wrapper(*args: P.args, **kwargs: P.kwargs) -> Tool:
+            # create the tool
             tool = tool_type(*args, **kwargs)
+
+            # this might already have registry info, in that case
+            # capture it and use it as defaults
+            from inspect_ai.tool._tool_def import tool_registry_info
+
+            tool_parallel = parallel
+            tool_viewer = viewer
+            tool_model_input = model_input
+            if is_registry_object(tool):
+                _, _, reg_parallel, reg_viewer, reg_model_input = tool_registry_info(
+                    tool
+                )
+                tool_parallel = parallel and reg_parallel
+                tool_viewer = viewer or reg_viewer
+                tool_model_input = model_input or reg_model_input
+
+            # tag the object
             registry_tag(
                 tool_type,
                 tool,
@@ -209,10 +228,11 @@ def tool(
                     name=tool_name,
                     metadata={
                         TOOL_PROMPT: prompt,
-                        TOOL_PARALLEL: parallel,
-                        TOOL_VIEWER: viewer,
+                        TOOL_PARALLEL: tool_parallel,
+                        TOOL_VIEWER: tool_viewer,
                         TOOL_MODEL_INPUT: (
-                            model_input or getattr(tool, TOOL_INIT_MODEL_INPUT, None)
+                            tool_model_input
+                            or getattr(tool, TOOL_INIT_MODEL_INPUT, None)
                         ),
                     },
                 ),

inspect_ai/tool/_tools/_think.py

@@ -0,0 +1,48 @@
+from .._tool import Tool, tool
+from .._tool_call import ToolCall, ToolCallContent, ToolCallView, ToolCallViewer
+from .._tool_def import ToolDef
+
+
+@tool
+def think(
+    description: str | None = None,
+    thought_description: str | None = None,
+) -> Tool:
+    """Think tool for extra thinking.
+
+    Tool that provides models with the ability to include an additional thinking step as part of getting to its final answer.
+
+    Note that the `think()` tool is not a substitute for reasoning and extended thinking, but rather an an alternate way of letting models express thinking that is better suited to some tool use scenarios. Please see the documentation on using the [think tool](https://inspect.aisi.org.uk/tools-standard.html#sec-think) before using it in your evaluations.
+
+    Args:
+        description: Override the default description of the think tool.
+        thought_description: Override the default description of the thought parameter.
+    """
+
+    async def execute(thought: str) -> str:
+        """Use the tool to think about something.
+
+        The will not obtain new information or change the environment, but just append the thought to the log. Use it when complex reasoning or some cache memory is needed."
+
+        Args:
+            thought: A thought to think about.
+        """
+        return ""
+
+    return ToolDef(
+        execute,
+        name="think",
+        description=description,
+        parameters=(dict(thought=thought_description) if thought_description else None),
+        viewer=think_tool_viewer(),
+    ).as_tool()
+
+
+def think_tool_viewer() -> ToolCallViewer:
+    def viewer(tool_call: ToolCall) -> ToolCallView:
+        call = ToolCallContent(
+            format="markdown", content=tool_call.arguments["thought"]
+        )
+        return ToolCallView(call=call)
+
+    return viewer

{inspect_ai-0.3.76.dist-info → inspect_ai-0.3.78.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: inspect_ai
-Version: 0.3.76
+Version: 0.3.78
 Summary: Framework for large language model evaluations
 Author: UK AI Security Institute
 License: MIT License

{inspect_ai-0.3.76.dist-info → inspect_ai-0.3.78.dist-info}/RECORD

@@ -3,7 +3,7 @@ inspect_ai/__main__.py,sha256=oWX4YwDZDg3GS3-IG0yPGoSEOfSzWihELg7QmrUlxjM,67
 inspect_ai/py.typed,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 inspect_ai/_cli/cache.py,sha256=RVGuBYwwk3c45JfyfcSFJ419etSsv3-Z7AyfQE-Hul0,3912
 inspect_ai/_cli/common.py,sha256=v4ZRvj4fJSbCq61IbZukb-iR_ZGawrmYH4n8JP_piVg,3676
-inspect_ai/_cli/eval.py,sha256=Jv506NwbcgsU8D-JbeNFsvYSZc779zq1ljJq4sB-9BA,34732
+inspect_ai/_cli/eval.py,sha256=76igFkS-c3uX8pPUcHWwSFlyF4ji_STZh6Ses8XHvlo,35233
 inspect_ai/_cli/info.py,sha256=QMxaTG9TmzW95EiLrOgkzubvavoR-VHxo3eV7ppmrzI,1789
 inspect_ai/_cli/list.py,sha256=M8mazI8Zuq8Hp99YWKnxQd9UWx1Qi87zfXRzZYAAakk,2459
 inspect_ai/_cli/log.py,sha256=O-w7GqsE7tLojPnEtpWXPSh0Vu2Hbrbxjneyc1BFfpk,5817
@@ -21,7 +21,7 @@ inspect_ai/_display/core/footer.py,sha256=_EI6yhv0TErl2Xku-mJBBSJXjiFnUJyvAp_TJR
 inspect_ai/_display/core/group.py,sha256=z8CIwQ-8Mm9adQ8JDuMjw94ih9GfymU5s-1qnbKoEPs,2871
 inspect_ai/_display/core/panel.py,sha256=gyGYnsqHurUkUC51MyVuh3oGAtUEaFtyRwewOB6pDts,3828
 inspect_ai/_display/core/progress.py,sha256=2dIRbpJGUx-Wz89ZABoACBGvJEGWJ3SDrFsuCrrpL7w,4198
-inspect_ai/_display/core/results.py,sha256=aFLmG1Ij0fxYk2848QgQlesfMeRdHVEg_W9esmeL_S0,7355
+inspect_ai/_display/core/results.py,sha256=zWzHZsZAdkcBTB0ZFRPvkFFjiHP9ekkRcKLJl8qh4Uw,7548
 inspect_ai/_display/core/rich.py,sha256=GPzc-0PWZVOPWxnjfQmNSK66uZXc3x8joz4ethgv_4M,2729
 inspect_ai/_display/core/textual.py,sha256=kzMTt8ijrodwhDB5V50pP2IBhnUCusVbP86TytU_rA8,870
 inspect_ai/_display/plain/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
@@ -45,12 +45,12 @@ inspect_ai/_display/textual/widgets/toggle.py,sha256=ToYs-S4n90yuxWcAW2OTg6AbRf0
 inspect_ai/_display/textual/widgets/transcript.py,sha256=zaxlDixT6Fie0acAWBM9Hltnk3Qf7G2awHGhK3vC4Nk,11752
 inspect_ai/_eval/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
 inspect_ai/_eval/context.py,sha256=gWTjEEMVTJMJpCCKLRs4joZDkG00rzE7-HXZFyzSC_I,1283
-inspect_ai/_eval/eval.py,sha256=lmDIfrfx9hmlIOSYEUkH6QzpX1dDV6JUgEB_fo4GANg,37824
-inspect_ai/_eval/evalset.py,sha256=1N4sJ-gvGe-viZtIUIjJ4DcUCjAHxUrB6ZN0SYGp0-4,23541
+inspect_ai/_eval/eval.py,sha256=yo5SFToYovoTxt7mvvuNEl_9b-ML125xHF8LvdIkgBg,38138
+inspect_ai/_eval/evalset.py,sha256=HGrz0LkTMsBbYDPZEMVnZCmFi_pYegZtSoqRVYbRDiE,23682
 inspect_ai/_eval/list.py,sha256=VbZ-2EI6MqrXvCN7VTz21TQSoU5K5_Q0hqhxmj5A_m0,3744
 inspect_ai/_eval/loader.py,sha256=yOj8HqYBFQntx0_GY4Wxqm6jivlT4N4WiQ1T8J9uRVA,23606
 inspect_ai/_eval/registry.py,sha256=9Q-Re9uZagQ2nw-W7hA6zhrmCQFmo3KcxncTcG24EG8,5315
-inspect_ai/_eval/run.py,sha256=1i-r9ha4afTyKKEwcA5PlMZ1SHZM80pt-N2MUKBzAk8,19093
+inspect_ai/_eval/run.py,sha256=9TtOp1vc18KclxstjvWhEL5VtcLBL3SRbRwsoRXaah8,19179
 inspect_ai/_eval/score.py,sha256=qf9T8XwUmfE6A8QK1-D-jUbNlLETVM-HXhucPWZ0Ro0,9591
 inspect_ai/_eval/task/__init__.py,sha256=6FvojMW3yo36L7xDacppCHDxt6A8_tzj_ftg5bQ6eNk,199
 inspect_ai/_eval/task/constants.py,sha256=quAKMw-4-3xKd1T_KwXCZvHYoKRXt1ZGuaHbBcWJwnA,72
@@ -73,7 +73,7 @@ inspect_ai/_util/ansi.py,sha256=fMxOAn72Nl5NG_kdlY408nC62bKhsW29L__A5FwwMJQ,983
 inspect_ai/_util/appdirs.py,sha256=lhURbDS9xT2YBzWOe0jjxsdK4ZdiVAv_WwXQC83V_jw,563
 inspect_ai/_util/config.py,sha256=nuWVZbShE8IPnotDfRJx0uBZJxwbV36M0qKVYsQDEEI,848
 inspect_ai/_util/constants.py,sha256=ZkpYBs6qnulFNBRhEBVUgQhnNw_phpuCdViA6FQhe3c,1031
-inspect_ai/_util/content.py,sha256=RU3RIMIBWNumMI-bSWtdco5-oSUUmIfHrGS0EmNKFkY,2128
+inspect_ai/_util/content.py,sha256=mYnN1m-VZ__leOyX8czbi1JRKukYCCxlQZgPwCk0aXE,2214
 inspect_ai/_util/datetime.py,sha256=WeQKSgT8VnmmJcHZbS-lWtVSDTPbQ4vO_V835wdTU7Y,270
 inspect_ai/_util/decorator.py,sha256=AEyOt-4VYMwXkEvMVyxMiB6xg_ZKQ89Q5gmJLf-dcpU,2595
 inspect_ai/_util/deprecation.py,sha256=Ng-_MXXf6zbsHa5FaFLTTvVeUMI5oQw7bl392hsiV1g,6670
@@ -144,7 +144,7 @@ inspect_ai/_view/www/.vscode/settings.json,sha256=g5hrVnMaYxM06JpiJD2EuE2xjcbF6x
 inspect_ai/_view/www/dist/index.html,sha256=gpdu6SR-SOH9EWx15cCWHzujMZujnZR5tRlEfROJg2A,997
 inspect_ai/_view/www/dist/assets/favicon.svg,sha256=b9AHYZaO2zBzeKH6G4PwXZMGGW_UxY0omKHam-c9MAs,1508
 inspect_ai/_view/www/dist/assets/index.css,sha256=V0HHB6ss9UdKAlLKTmyHoYAZrstD6nvb-UhnWInyhQQ,895546
-inspect_ai/_view/www/dist/assets/index.js,sha256=zcV5qvFRVPEBJkVclKSbtg91IX6L-T7dmO2M8yY9mZ4,2748105
+inspect_ai/_view/www/dist/assets/index.js,sha256=-dI_BWa2w7TMj1_inAgHQmG4qq44oG_X2JM70W7DygY,2748559
 inspect_ai/_view/www/src/App.tsx,sha256=rhiZKs-f1y9amyX_OLay7aL4OXaQ_o0gNd04M3pZVuk,28852
 inspect_ai/_view/www/src/AppErrorBoundary.tsx,sha256=RyhZWbIMZj1QeUOUUXh9hUFvq6LoDEoHuTY0giswmL0,1169
 inspect_ai/_view/www/src/constants.ts,sha256=aLncMT1XjKzyphLF_jkXPZicsCJXMsBXcDomAC4EaIY,1228
@@ -247,9 +247,9 @@ inspect_ai/_view/www/src/samples/chat/MessageContent.tsx,sha256=9T_MjXhUPKXQHMws
 inspect_ai/_view/www/src/samples/chat/MessageContents.module.css,sha256=t9sk1CbuwDuYhwuBjWzWip7JPdN0PnzwzqNwdgVJCWU,35
 inspect_ai/_view/www/src/samples/chat/MessageContents.tsx,sha256=FMKAg7e_8OOqoIvrO_Bc_XLbfrnh7IL55VQByPknqVs,3338
 inspect_ai/_view/www/src/samples/chat/messages.ts,sha256=qm5AE2yQWtk1KbxDDSwgg_uvtU8-MStfhMpkTWSQ1DM,3070
-inspect_ai/_view/www/src/samples/chat/tools/ToolCallView.tsx,sha256=DdDHbVjRDZ4LD9tHBptMykO9mU8_pVk4zf7MwTvuQeQ,3261
+inspect_ai/_view/www/src/samples/chat/tools/ToolCallView.tsx,sha256=FcEhHVa2LpJqH-QWHYPFVXgfTm8GfmxzePKEzHVSD6w,3651
 inspect_ai/_view/www/src/samples/chat/tools/ToolInput.module.css,sha256=apEfqHAe683GyFx803xDO6Wid3m3xn_3ftGSLa4uxVI,190
-inspect_ai/_view/www/src/samples/chat/tools/ToolInput.tsx,sha256=B7_R2_Y8k8POQ69mR807kQdq3izxyN3CuURLb_FpvcQ,2009
+inspect_ai/_view/www/src/samples/chat/tools/ToolInput.tsx,sha256=MWVvXSRPER394f273p8OIPvHXYzqvJkkjd-WK41Hong,2041
 inspect_ai/_view/www/src/samples/chat/tools/ToolOutput.module.css,sha256=p8jLonoAawbhuaqNzFPzToD-KRa3tBwX9w4yVOf3xqo,266
 inspect_ai/_view/www/src/samples/chat/tools/ToolOutput.tsx,sha256=kOHkL97fnw1xwfgP_ys0JGl-9gd7rZDO_FVQzX8oQ9s,1616
 inspect_ai/_view/www/src/samples/chat/tools/ToolTitle.module.css,sha256=RhrswSho4BZcJLi_YXtfDu7Ea70Z11rkwzeMvzsYsrM,51
@@ -456,16 +456,16 @@ inspect_ai/model/_call_tools.py,sha256=XvPO-RrbDSIk3A6dCXrqHG2IYAfBart1a-uMKUydA
 inspect_ai/model/_chat_message.py,sha256=I2EPWuVFRZVwS9mqA80mZCXm37brsEvnD633cq99wM0,7249
 inspect_ai/model/_conversation.py,sha256=7KPqnCx5ETNkdhR1g1gsjV8sCrJlrOOPwlBz_Yoc9kg,2463
 inspect_ai/model/_generate_config.py,sha256=JMX_48p-zv7WW0dGm_PQPyGpyJoa4AuFXBaBXajWWpo,10973
-inspect_ai/model/_image.py,sha256=kpO2Bn_-c-dK80HuPOPH1eSNmcoc39kofwf4yTTiTFE,477
 inspect_ai/model/_model.py,sha256=LaRXqKvM2fl9HGCQAdrOOkNryfYU970LZGsyOxk1o00,48304
 inspect_ai/model/_model_call.py,sha256=VJ8wnl9Y81JaiClBYM8eyt1jVb3n-yc6Dd88ofRiJDc,2234
 inspect_ai/model/_model_output.py,sha256=1picjX2Y0NSmEZ-vdQi42QAx3QvMcWVRn8pXP2wtmN8,7733
-inspect_ai/model/_openai.py,sha256=VvEUg0lQdXCq0vV3iR7GfqW52WjHXNRA2QDslIjg9RY,16975
+inspect_ai/model/_openai.py,sha256=3_lQ3u-WJ8BOY5bMLM49LvAePzCZAn57Vi2SNLrWPm4,19093
+inspect_ai/model/_openai_responses.py,sha256=qav1Fj2R-DOCkUfkrtGxV_PeT578mBZNwKt-yL4bmT0,10049
 inspect_ai/model/_reasoning.py,sha256=qmR8WT6t_cb7NIsJOQHPyFZh2eLV0HmYxKo2vtvteQ4,929
 inspect_ai/model/_registry.py,sha256=Cr2y32EqLnOqLbSWoXHVK4ivTTzCUhJuACxoTyPt8kY,2032
 inspect_ai/model/_render.py,sha256=rWypNUjgrH4NGp0r-ESAze9gZz7lYNjheEP438vRYZE,922
-inspect_ai/model/_providers/anthropic.py,sha256=SuXXs-VmG4XzSyHin7N1HPAHJoy43jExqhio2OPNV0o,33369
-inspect_ai/model/_providers/azureai.py,sha256=3phdDrrsKzinphLtGj39bet_jVLcKgtLcWr_uILO0cc,14225
+inspect_ai/model/_providers/anthropic.py,sha256=qA6GOl64axTjuL5ciEakoYsRbUljfxM9SDZr13Kbjns,33396
+inspect_ai/model/_providers/azureai.py,sha256=tDWuePLhnZBcpHLVzX7J3Wx8VRPhW8tmtjiwLPmJouU,14232
 inspect_ai/model/_providers/bedrock.py,sha256=mLeMW2JkG4lF0VQGEdku73ZL00EBy-hEvEcyCdjDUDo,24153
 inspect_ai/model/_providers/cloudflare.py,sha256=0e0HPas21cVC8N9mNJlZnSZyIt6FUB9lTIAqPXJDrtE,4586
 inspect_ai/model/_providers/goodfire.py,sha256=EzebC1woEjIXfHLP_ixpMR6G1hC-LxbSUxiilq1c-Is,8868
@@ -474,14 +474,15 @@ inspect_ai/model/_providers/grok.py,sha256=dS88ueXiD-kHAFr0jCoTpTGLGa2VsUlB_TFP8
 inspect_ai/model/_providers/groq.py,sha256=Fr4fy8NmqllmUW7jhnQ3W94zGlxyr276qaGFS_iDI3Q,11189
 inspect_ai/model/_providers/hf.py,sha256=EZRiiRSzIoRCdFYKj3Otn5ebsROdjzx5YSQ6CzqOJxk,17969
 inspect_ai/model/_providers/llama_cpp_python.py,sha256=i2I56Damgb8VDhMuPxPca24fVhlajDHzxCTYFnT41uI,702
-inspect_ai/model/_providers/mistral.py,sha256=pwBaO7VoW-ZsiZfNjF3-m7Zr44OkNasOlI5N3SEE1dc,17377
+inspect_ai/model/_providers/mistral.py,sha256=TPJIB0AytktnFwwAlvg6Mz2hXarJK8m0G6ggJw5Cbqg,17774
 inspect_ai/model/_providers/mockllm.py,sha256=gL9f-f5TOdE4a0GVENr3cOIIp2kv8zVXWPZ608rouGk,2440
 inspect_ai/model/_providers/none.py,sha256=6qLbZpHSoEZaaxFO7luieFjqig2Ju8Fu00DlRngAry8,935
 inspect_ai/model/_providers/ollama.py,sha256=mBPSxaEkiH_RnlHKqOyFBlXObQhc2dfjL-rCKrea5u8,675
-inspect_ai/model/_providers/openai.py,sha256=9ze3TOLQ5ycRGquJ67bwCPOgfMPYwRZZbN-h2lN4Gys,15058
+inspect_ai/model/_providers/openai.py,sha256=Qg0MM2OftmzzZ8aio-hvg3GZv77-PyuuY0uk3wtIM60,15189
 inspect_ai/model/_providers/openai_o1.py,sha256=wURSSI7aCBxbZONQBaEOYPrQleBExL2c2uSIdJXLi1U,12763
-inspect_ai/model/_providers/openrouter.py,sha256=5G8qS8xA7Gy4IGodEJd04xwjsN-O_as4oeU8DTsKB5s,2932
-inspect_ai/model/_providers/providers.py,sha256=GI68ptXsLoEkzuLMZseJQ1_jI2f1uqfZwKy9y8IX-Yg,6508
+inspect_ai/model/_providers/openai_responses.py,sha256=9mVdfcKgIULv6YpwCFKFUj63JVi7szvBzHczj12ZdOc,6096
+inspect_ai/model/_providers/openrouter.py,sha256=pDimDmm_4FzS4GZx0n9z8z717mQf3IQlgEy30huzpc4,4730
+inspect_ai/model/_providers/providers.py,sha256=TdRXKzNQmeWZyTaAWkalmxVylRHkft0O8LqGJ054EiI,6508
 inspect_ai/model/_providers/together.py,sha256=Hwg-5jlFR3K61Epp_NnmDR5lX-bHZ2QT77GjM9dQOKY,9461
 inspect_ai/model/_providers/vertex.py,sha256=3PX3pVvzydA-IczZaEAeqQeiTtimBa3a6t418GAtjhY,17235
 inspect_ai/model/_providers/vllm.py,sha256=KNEjdu-oeCXH-vhpMeF1Pmnf8-mkLpQ5pbyFJMlaT9c,14172
@@ -544,8 +545,8 @@ inspect_ai/solver/_human_agent/commands/note.py,sha256=yOHXPkm3yFrXzf-CK-IS6B3M0
 inspect_ai/solver/_human_agent/commands/score.py,sha256=Pkcd3g-8BAYlxQnvJsMFKPZCnqWsFjUYAM7T4Dl5rtM,2207
 inspect_ai/solver/_human_agent/commands/status.py,sha256=uUO5M4skWDp29OS8sqVKAqZw0OcM3MSesBYQNbRypJ0,1934
 inspect_ai/solver/_human_agent/commands/submit.py,sha256=D2p1M2ApvAcaVZhbP3fFofG9ZsPVvmxivSLIF5xQxtA,6524
-inspect_ai/tool/__init__.py,sha256=krnJakKZ2A8hC_Lz1SyMKx5nEeLUGwgZL-vmBqUTVZM,2420
-inspect_ai/tool/_tool.py,sha256=e5a2fHtCg5_AObaDfJliV6s-bJZNfa6HEbj587MjDuM,6285
+inspect_ai/tool/__init__.py,sha256=M1xR4GLN4LBCzWWbMbIGXE-XnagdvC9mK5_a4qK35Ew,2466
+inspect_ai/tool/_tool.py,sha256=VjnbMnsXPLpNqglc6Oss7OK7cVHHG7W2qzWtjcOc5Us,7055
 inspect_ai/tool/_tool_call.py,sha256=yCkcQBl5iSdESX2Mv6z54IOFytmo6spzIk6UWJo3s7g,2294
 inspect_ai/tool/_tool_choice.py,sha256=L8QNmcnSnilzKg2HU3G65W5aYaSngO09z4FQ0fQlATM,496
 inspect_ai/tool/_tool_def.py,sha256=wS6yX2_64-O6L4k1kMihUp52-OnGypMn11L3oXMNeQc,8191
@@ -560,6 +561,7 @@ inspect_ai/tool/_tools/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3
 inspect_ai/tool/_tools/_bash_session.py,sha256=_5Y8e5qGX_WTtLsg4bCmj0WH_xNY2locCz4rQCOlvM8,3480
 inspect_ai/tool/_tools/_execute.py,sha256=SZdtVxxGoQNErWFtYaJgfTtLsZbBP28lnY112mt-VWQ,3435
 inspect_ai/tool/_tools/_text_editor.py,sha256=9Z6vB4_sgwQw2OH22reA6FSw6LgFslbQ5X7ufvaVl5U,3939
+inspect_ai/tool/_tools/_think.py,sha256=FndjEwPbkkdqEmxpVHpecEwRJo6DYWFJzwUnNNLmy9Y,1821
 inspect_ai/tool/_tools/_web_search.py,sha256=GYkGasfmZWG3TY1XqBfa6yJnG7ysPtdGCVmPUJHQRSM,8012
 inspect_ai/tool/_tools/_computer/__init__.py,sha256=fq4BSM4aDhtEtE4279xm47NiO6vyiZHhhw7cq-azFzk,56
 inspect_ai/tool/_tools/_computer/_common.py,sha256=nX8Cf9jagsnjwZb9f9eWADUjzETf6PSNRcFrzp0_KLg,5949
@@ -625,9 +627,9 @@ inspect_ai/util/_sandbox/docker/internal.py,sha256=c8X8TLrBPOvsfnq5TkMlb_bzTALyc
 inspect_ai/util/_sandbox/docker/prereqs.py,sha256=0j6_OauBBnVlpBleADcZavIAAQZy4WewVjbRn9c0stg,3355
 inspect_ai/util/_sandbox/docker/service.py,sha256=hhHIWH1VDFLwehdGd19aUBD_VKfDO3GCPxpw1HSwVQk,2437
 inspect_ai/util/_sandbox/docker/util.py,sha256=EeInihCNXgUWxaqZ4dNOJd719kXL2_jr63QCoXn68vA,3154
-inspect_ai-0.3.76.dist-info/licenses/LICENSE,sha256=xZPCr8gTiFIerrA_DRpLAbw-UUftnLFsHxKeW-NTtq8,1081
-inspect_ai-0.3.76.dist-info/METADATA,sha256=qQk_DES_FDFGm7FFo85Tz9LI3Vrkj9MnUb-EHIi-ym0,4997
-inspect_ai-0.3.76.dist-info/WHEEL,sha256=1tXe9gY0PYatrMPMDd6jXqjfpz_B-Wqm32CPfRC58XU,91
-inspect_ai-0.3.76.dist-info/entry_points.txt,sha256=WGGLmzTzDWLzYfiyovSY6oEKuf-gqzSDNOb5V-hk3fM,54
-inspect_ai-0.3.76.dist-info/top_level.txt,sha256=Tp3za30CHXJEKLk8xLe9qGsW4pBzJpEIOMHOHNCXiVo,11
-inspect_ai-0.3.76.dist-info/RECORD,,
+inspect_ai-0.3.78.dist-info/licenses/LICENSE,sha256=xZPCr8gTiFIerrA_DRpLAbw-UUftnLFsHxKeW-NTtq8,1081
+inspect_ai-0.3.78.dist-info/METADATA,sha256=NNzcpeDyLChQBQ9aFqlE2iVcY-sod8sd5B72_pI2qq8,4997
+inspect_ai-0.3.78.dist-info/WHEEL,sha256=DK49LOLCYiurdXXOXwGJm6U4DkHkg4lcxjhqwRa0CP4,91
+inspect_ai-0.3.78.dist-info/entry_points.txt,sha256=WGGLmzTzDWLzYfiyovSY6oEKuf-gqzSDNOb5V-hk3fM,54
+inspect_ai-0.3.78.dist-info/top_level.txt,sha256=Tp3za30CHXJEKLk8xLe9qGsW4pBzJpEIOMHOHNCXiVo,11
+inspect_ai-0.3.78.dist-info/RECORD,,

{inspect_ai-0.3.76.dist-info → inspect_ai-0.3.78.dist-info}/WHEEL

@@ -1,5 +1,5 @@
 Wheel-Version: 1.0
-Generator: setuptools (77.0.3)
+Generator: setuptools (78.0.2)
 Root-Is-Purelib: true
 Tag: py3-none-any
 

inspect_ai/model/_image.py

@@ -1,15 +0,0 @@
-from copy import copy
-
-from pydantic import JsonValue
-
-from inspect_ai._util.constants import BASE_64_DATA_REMOVED
-
-
-def image_url_filter(key: JsonValue | None, value: JsonValue) -> JsonValue:
-    # remove images from raw api call
-    if key == "image_url" and isinstance(value, dict) and "url" in value:
-        url = str(value.get("url"))
-        if url.startswith("data:"):
-            value = copy(value)
-            value.update(url=BASE_64_DATA_REMOVED)
-    return value