inspect-ai 0.3.74__py3-none-any.whl → 0.3.76__py3-none-any.whl

inspect_ai/_view/www/src/types/log.d.ts

@@ -148,7 +148,7 @@ export type Input =
       | ChatMessageAssistant
       | ChatMessageTool
     )[];
-export type Id1 = string;
+export type Id1 = string | null;
 export type Content =
   | string
   | (
@@ -175,7 +175,7 @@ export type Video = string;
 export type Format1 = "mp4" | "mpeg" | "mov";
 export type Source = ("input" | "generate") | null;
 export type Role = "system";
-export type Id2 = string;
+export type Id2 = string | null;
 export type Content1 =
   | string
   | (
@@ -188,7 +188,7 @@ export type Content1 =
 export type Source1 = ("input" | "generate") | null;
 export type Role1 = "user";
 export type ToolCallId = string[] | null;
-export type Id3 = string;
+export type Id3 = string | null;
 export type Content2 =
   | string
   | (
@@ -203,12 +203,13 @@ export type Role2 = "assistant";
 export type ToolCalls = ToolCall[] | null;
 export type Id4 = string;
 export type Function = string;
-export type Type8 = "function";
+export type Type8 = string;
+export type InternalName = string | null;
 export type ParseError = string | null;
 export type Title = string | null;
 export type Format2 = "text" | "markdown";
 export type Content3 = string;
-export type Id5 = string;
+export type Id5 = string | null;
 export type Content4 =
   | string
   | (
@@ -222,6 +223,7 @@ export type Source3 = ("input" | "generate") | null;
 export type Role3 = "tool";
 export type ToolCallId1 = string | null;
 export type Function1 = string | null;
+export type InternalName1 = string | null;
 export type Type9 =
   | "parsing"
   | "timeout"
@@ -369,6 +371,7 @@ export type Event6 = "tool";
 export type Type12 = "function";
 export type Id7 = string;
 export type Function2 = string;
+export type InternalName2 = string | null;
 export type Result1 =
   | string
   | number
@@ -911,6 +914,7 @@ export interface ToolCall {
   function: Function;
   arguments: Arguments;
   type: Type8;
+  internal_name: InternalName;
   parse_error: ParseError;
   view: ToolCallContent | null;
 }
@@ -933,6 +937,7 @@ export interface ChatMessageTool {
   role: Role3;
   tool_call_id: ToolCallId1;
   function: Function1;
+  internal_name: InternalName1;
   error: ToolCallError | null;
 }
 export interface ToolCallError {
@@ -1201,6 +1206,7 @@ export interface ToolEvent {
   id: Id7;
   function: Function2;
   arguments: Arguments1;
+  internal_name: InternalName2;
   view: ToolCallContent | null;
   result: Result1;
   truncated: Truncated;

inspect_ai/_view/www/src/workspace/navbar/PrimaryBar.tsx

@@ -3,6 +3,7 @@ import { FC, useCallback } from "react";
 import { SampleSummary } from "../../api/types";
 import { ApplicationIcons } from "../../appearance/icons";
 import { CopyButton } from "../../components/CopyButton";
+import { kModelNone } from "../../constants";
 import { EvalResults, EvalSpec, Status } from "../../types/log";
 import { filename } from "../../utils/path";
 import styles from "./PrimaryBar.module.css";
@@ -71,18 +72,22 @@ export const PrimaryBar: FC<PrimaryBarProps> = ({
             >
               {evalSpec?.task}
             </div>
-            <div
-              id="task-model"
-              className={clsx(
-                "task-model",
-                "text-truncate",
-                styles.taskModel,
-                "text-size-base",
-              )}
-              title={evalSpec?.model}
-            >
-              {evalSpec?.model}
-            </div>
+            {evalSpec?.model && evalSpec.model !== kModelNone ? (
+              <div
+                id="task-model"
+                className={clsx(
+                  "task-model",
+                  "text-truncate",
+                  styles.taskModel,
+                  "text-size-base",
+                )}
+                title={evalSpec?.model}
+              >
+                {evalSpec?.model}
+              </div>
+            ) : (
+              ""
+            )}
           </div>
           <div className={clsx("text-size-small", styles.secondaryContainer)}>
             <div className={clsx("navbar-secondary-text", "text-truncate")}>

inspect_ai/_view/www/src/workspace/sidebar/SidebarLogEntry.tsx

@@ -1,6 +1,7 @@
 import clsx from "clsx";
 import { FC, Fragment } from "react";
 import { EvalLogHeader } from "../../api/types";
+import { kModelNone } from "../../constants";
 import { EvalStatus } from "./EvalStatus";
 import styles from "./SidebarLogEntry.module.css";
 
@@ -51,7 +52,7 @@ export const SidebarLogEntry: FC<SidebarLogEntryProps> = ({
           </div>
           <small className={clsx("mb-1", "text-size-small")}>{timeStr}</small>
 
-          {model ? (
+          {model && model !== kModelNone ? (
             <div>
               <small className={clsx("mb-1", "text-size-small")}>{model}</small>
             </div>

inspect_ai/_view/www/yarn.lock

@@ -220,13 +220,20 @@
     "@codemirror/view" "^6.0.0"
     crelt "^1.0.5"
 
-"@codemirror/state@^6.0.0", "@codemirror/state@^6.4.0", "@codemirror/state@^6.5.0", "@codemirror/state@^6.5.1":
+"@codemirror/state@^6.0.0", "@codemirror/state@^6.4.0", "@codemirror/state@^6.5.0":
   version "6.5.1"
   resolved "https://registry.yarnpkg.com/@codemirror/state/-/state-6.5.1.tgz#e5c0599f7b43cf03f19e05861317df5425c07904"
   integrity sha512-3rA9lcwciEB47ZevqvD8qgbzhM9qMb8vCcQCNmDfVRPQG4JT9mSb0Jg8H7YjKGGQcFnLN323fj9jdnG59Kx6bg==
   dependencies:
     "@marijn/find-cluster-break" "^1.0.0"
 
+"@codemirror/state@^6.5.2":
+  version "6.5.2"
+  resolved "https://registry.yarnpkg.com/@codemirror/state/-/state-6.5.2.tgz#8eca3a64212a83367dc85475b7d78d5c9b7076c6"
+  integrity sha512-FVqsPqtPWKVVL3dPSxy8wEF/ymIEuVzF1PK3VbUgrxXpJUSHQWWZz4JMToquRxnkw+36LTamCZG2iua2Ptq0fA==
+  dependencies:
+    "@marijn/find-cluster-break" "^1.0.0"
+
 "@codemirror/view@^6.0.0", "@codemirror/view@^6.17.0", "@codemirror/view@^6.23.0", "@codemirror/view@^6.27.0", "@codemirror/view@^6.35.0":
   version "6.36.2"
   resolved "https://registry.yarnpkg.com/@codemirror/view/-/view-6.36.2.tgz#aeb644e161440734ac5a153bf6e5b4a4355047be"
@@ -862,10 +869,10 @@ argparse@^2.0.1:
   resolved "https://registry.yarnpkg.com/argparse/-/argparse-2.0.1.tgz#246f50f3ca78a3240f6c997e8a9bd1eac49e4b38"
   integrity sha512-8+9WqebbFzpX9OR+Wa6O29asIogeRMzcGtAINdpMHHyAg10f05aSFVBbcEqGf/PXw1EjAZ+q2/bEBg3DvurK3Q==
 
-asciinema-player@^3.8.2:
-  version "3.8.2"
-  resolved "https://registry.yarnpkg.com/asciinema-player/-/asciinema-player-3.8.2.tgz#12fbf475ddaeee0051ace17532e5f003475f6dfa"
-  integrity sha512-Lgcnj9u/H6sRpGRX1my7Azcay6llLmB/GVkCGcDbPwdTVTisS1ir8SQ9jRWRvjlLUjpSJkN0euruvy3sLRM8tw==
+asciinema-player@^3.9.0:
+  version "3.9.0"
+  resolved "https://registry.yarnpkg.com/asciinema-player/-/asciinema-player-3.9.0.tgz#c60742f85978e861b878fc7eb6289a5622c298af"
+  integrity sha512-SXVFImVzeNr8ZUdNIHABGuzlbnGWTKy245AquAjODsAnv+Lp6vxjYGN0LfA8ns30tnx/ag/bMrTbLq13TpHE6w==
   dependencies:
     "@babel/runtime" "^7.21.0"
     solid-js "^1.3.0"

inspect_ai/log/_log.py

@@ -215,7 +215,16 @@ class EvalSample(BaseModel):
         Returns:
           StoreModel: Instance of model_cls bound to sample store data.
         """
-        return model_cls(store=Store(self.store))
+        # un-namespace names for creation
+        data = {
+            k.replace(f"{model_cls.__name__}:", "", 1): v for k, v in self.store.items()
+        }
+
+        # since we are reading from the log provide a fully detached store
+        data["store"] = Store()
+
+        # create the model
+        return model_cls.model_validate(data)
 
     events: list[Event] = Field(default_factory=list)
     """Events that occurred during sample execution."""

inspect_ai/log/_recorders/eval.py

@@ -10,7 +10,7 @@ from pydantic import BaseModel, Field
 from pydantic_core import to_json
 from typing_extensions import override
 
-from inspect_ai._util.constants import LOG_SCHEMA_VERSION
+from inspect_ai._util.constants import DESERIALIZING_CONTEXT, LOG_SCHEMA_VERSION
 from inspect_ai._util.content import (
     ContentAudio,
     ContentImage,
@@ -224,7 +224,9 @@ class EvalRecorder(FileRecorder):
             with ZipFile(z, mode="r") as zip:
                 try:
                     with zip.open(_sample_filename(id, epoch), "r") as f:
-                        return EvalSample(**json.load(f))
+                        return EvalSample.model_validate(
+                            json.load(f), context=DESERIALIZING_CONTEXT
+                        )
                 except KeyError:
                     raise IndexError(
                         f"Sample id {id} for epoch {epoch} not found in log {location}"
@@ -414,7 +416,10 @@ def _read_log(log: BinaryIO, location: str, header_only: bool = False) -> EvalLo
         if REDUCTIONS_JSON in zip.namelist():
             with zip.open(REDUCTIONS_JSON, "r") as f:
                 reductions = [
-                    EvalSampleReductions(**reduction) for reduction in json.load(f)
+                    EvalSampleReductions.model_validate(
+                        reduction, context=DESERIALIZING_CONTEXT
+                    )
+                    for reduction in json.load(f)
                 ]
                 if evalLog.results is not None:
                     evalLog.reductions = reductions
@@ -425,7 +430,11 @@ def _read_log(log: BinaryIO, location: str, header_only: bool = False) -> EvalLo
             for name in zip.namelist():
                 if name.startswith(f"{SAMPLES_DIR}/") and name.endswith(".json"):
                     with zip.open(name, "r") as f:
-                        samples.append(EvalSample(**json.load(f)))
+                        samples.append(
+                            EvalSample.model_validate(
+                                json.load(f), context=DESERIALIZING_CONTEXT
+                            ),
+                        )
             sort_samples(samples)
             evalLog.samples = samples
         return evalLog
@@ -452,7 +461,10 @@ def _read_all_summaries(zip: ZipFile, count: int) -> list[SampleSummary]:
     if SUMMARIES_JSON in zip.namelist():
         summaries_raw = _read_json(zip, SUMMARIES_JSON)
         if isinstance(summaries_raw, list):
-            return [SampleSummary(**value) for value in summaries_raw]
+            return [
+                SampleSummary.model_validate(value, context=DESERIALIZING_CONTEXT)
+                for value in summaries_raw
+            ]
         else:
             raise ValueError(
                 f"Expected a list of summaries when reading {SUMMARIES_JSON}"
@@ -464,7 +476,14 @@ def _read_all_summaries(zip: ZipFile, count: int) -> list[SampleSummary]:
             summary_path = _journal_summary_path(summary_file)
             summary = _read_json(zip, summary_path)
             if isinstance(summary, list):
-                summaries.extend([SampleSummary(**value) for value in summary])
+                summaries.extend(
+                    [
+                        SampleSummary.model_validate(
+                            value, context=DESERIALIZING_CONTEXT
+                        )
+                        for value in summary
+                    ]
+                )
             else:
                 raise ValueError(
                     f"Expected a list of summaries when reading {summary_file}"
@@ -476,12 +495,12 @@ def _read_header(zip: ZipFile, location: str) -> EvalLog:
     # first see if the header is here
     if HEADER_JSON in zip.namelist():
         with zip.open(HEADER_JSON, "r") as f:
-            log = EvalLog(**json.load(f))
+            log = EvalLog.model_validate(json.load(f), context=DESERIALIZING_CONTEXT)
             log.location = location
             return log
     else:
         with zip.open(_journal_path(START_JSON), "r") as f:
-            start = LogStart(**json.load(f))
+            start = LogStart.model_validate(json.load(f), context=DESERIALIZING_CONTEXT)
         return EvalLog(
             version=start.version, eval=start.eval, plan=start.plan, location=location
         )

inspect_ai/log/_recorders/json.py

@@ -7,7 +7,7 @@ from pydantic import BaseModel
 from pydantic_core import from_json
 from typing_extensions import override
 
-from inspect_ai._util.constants import LOG_SCHEMA_VERSION
+from inspect_ai._util.constants import DESERIALIZING_CONTEXT, LOG_SCHEMA_VERSION
 from inspect_ai._util.error import EvalError
 from inspect_ai._util.file import absolute_file_path, file
 from inspect_ai._util.trace import trace_action
@@ -143,7 +143,7 @@ class JSONRecorder(FileRecorder):
         with file(location, "r") as f:
             # parse w/ pydantic
             raw_data = from_json(f.read())
-            log = EvalLog(**raw_data)
+            log = EvalLog.model_validate(raw_data, context=DESERIALIZING_CONTEXT)
             log.location = location
 
             # fail for unknown version
@@ -217,6 +217,11 @@ def _read_header_streaming(log_file: str) -> EvalLog:
 
         # Parse the log file, stopping before parsing samples
         status: Literal["started", "success", "cancelled", "error"] | None = None
+        eval: EvalSpec | None = None
+        plan: EvalPlan | None = None
+        results: EvalResults | None = None
+        stats: EvalStats | None = None
+        error: EvalError | None = None
         for k, v in ijson.kvitems(f, ""):
             if k == "status":
                 assert v in get_args(
@@ -239,6 +244,9 @@ def _read_header_streaming(log_file: str) -> EvalLog:
                 break
 
     assert status, "Must encounter a 'status'"
+    assert eval, "Must encounter a 'eval'"
+    assert plan, "Must encounter a 'plan'"
+    assert stats, "Must encounter a 'stats'"
 
     return EvalLog(
         eval=eval,

inspect_ai/log/_transcript.py

@@ -146,7 +146,7 @@ class ModelEvent(BaseEvent):
     """working time for model call that succeeded (i.e. was not retried)."""
 
     @field_serializer("completed")
-    def serialize_completed(self, dt: datetime) -> str:
+    def serialize_completed(self, dt: datetime | None) -> str | None:
         if dt is None:
             return None
         return dt.astimezone().isoformat()
@@ -170,6 +170,9 @@ class ToolEvent(BaseEvent):
     arguments: dict[str, JsonValue]
     """Arguments to function."""
 
+    internal_name: str | None = Field(default=None)
+    """Internal name for tool (if any)."""
+
     view: ToolCallContent | None = Field(default=None)
     """Custom view of tool call input."""
 
@@ -235,7 +238,9 @@ class ToolEvent(BaseEvent):
     """Required so that we can include '_cancel_fn' as a member."""
 
     @field_serializer("completed")
-    def serialize_completed(self, dt: datetime) -> str:
+    def serialize_completed(self, dt: datetime | None) -> str | None:
+        if dt is None:
+            return None
         return dt.astimezone().isoformat()
 
 
@@ -270,7 +275,9 @@ class SandboxEvent(BaseEvent):
     """Time that sandbox action completed (see `timestamp` for started)"""
 
     @field_serializer("completed")
-    def serialize_completed(self, dt: datetime) -> str:
+    def serialize_completed(self, dt: datetime | None) -> str | None:
+        if dt is None:
+            return None
         return dt.astimezone().isoformat()
 
 
@@ -412,7 +419,9 @@ class SubtaskEvent(BaseEvent):
     """Working time for subtask (i.e. time not spent waiting on semaphores or model retries)."""
 
     @field_serializer("completed")
-    def serialize_completed(self, dt: datetime) -> str:
+    def serialize_completed(self, dt: datetime | None) -> str | None:
+        if dt is None:
+            return None
         return dt.astimezone().isoformat()
 
 

inspect_ai/model/_call_tools.py

@@ -25,7 +25,6 @@ from typing import (
 if sys.version_info < (3, 11):
     from exceptiongroup import ExceptionGroup
 
-
 import anyio
 import yaml
 from anyio.streams.memory import MemoryObjectSendStream
@@ -168,6 +167,7 @@ async def call_tools(
                 id=call.id,
                 function=call.function,
                 arguments=call.arguments,
+                internal_name=call.internal_name,
                 result=content,
                 truncated=truncated,
                 view=call.view,
@@ -183,6 +183,7 @@ async def call_tools(
                             content=content,
                             tool_call_id=call.id,
                             function=call.function,
+                            internal_name=call.internal_name,
                             error=tool_error,
                         ),
                         event,
@@ -201,6 +202,7 @@ async def call_tools(
                 id=call.id,
                 function=call.function,
                 arguments=call.arguments,
+                internal_name=call.internal_name,
                 view=call.view,
                 pending=True,
             )
@@ -216,9 +218,7 @@ async def call_tools(
                     tg.start_soon(call_tool_task, call, send_stream)
                     event._set_cancel_fn(tg.cancel_scope.cancel)
                     async with receive_stream:
-                        async for result in receive_stream:
-                            tool_message, result_event = result
-                            break
+                        tool_message, result_event = await receive_stream.receive()
             except ExceptionGroup as ex:
                 raise ex.exceptions[0]
 
@@ -226,6 +226,7 @@ async def call_tools(
                 tool_message = ChatMessageTool(
                     content="",
                     function=call.function,
+                    internal_name=call.internal_name,
                     tool_call_id=call.id,
                     error=ToolCallError(
                         "timeout", "Command timed out before completing."
@@ -235,6 +236,7 @@ async def call_tools(
                     id=call.id,
                     function=call.function,
                     arguments=call.arguments,
+                    internal_name=call.internal_name,
                     result=tool_message.content,
                     truncated=None,
                     view=call.view,
@@ -508,6 +510,13 @@ def tool_parse_error_message(arguments: str, ex: Exception) -> str:
 def parse_tool_call(
     id: str, function: str, arguments: str, tools: list[ToolInfo] | None = None
 ) -> ToolCall:
+    """Parse a tool call from a JSON payload.
+
+    Note that this function doesn't know about internal tool names so the caller
+    should ammend the returned `ToolCall` by mapping the parsed `function` field from
+    from an internal name to an inspect tool name and fixing up the `ToolCall` object
+    as required to reflect this change.
+    """
     error: str | None = None
     arguments_dict: dict[str, Any] = {}
 

inspect_ai/model/_chat_message.py

@@ -4,6 +4,7 @@ from typing import Any, Literal, Type, Union
 from pydantic import BaseModel, Field, model_validator
 from shortuuid import uuid
 
+from inspect_ai._util.constants import DESERIALIZING
 from inspect_ai._util.content import Content, ContentReasoning, ContentText
 from inspect_ai.tool import ToolCall
 from inspect_ai.tool._tool_call import ToolCallError
@@ -16,7 +17,7 @@ logger = getLogger(__name__)
 class ChatMessageBase(BaseModel):
     """Base class for chat messages."""
 
-    id: str = Field(default_factory=uuid)
+    id: str | None = Field(default=None)
     """Unique identifer for message."""
 
     content: str | list[Content]
@@ -25,6 +26,16 @@ class ChatMessageBase(BaseModel):
     source: Literal["input", "generate"] | None = Field(default=None)
     """Source of message."""
 
+    def model_post_init(self, __context: Any) -> None:
+        # check if deserializing
+        is_deserializing = isinstance(__context, dict) and __context.get(
+            DESERIALIZING, False
+        )
+
+        # Generate ID if needed and not deserializing
+        if self.id is None and not is_deserializing:
+            self.id = uuid()
+
     @property
     def text(self) -> str:
         """Get the text content of this message.
@@ -147,6 +158,9 @@ class ChatMessageTool(ChatMessageBase):
     function: str | None = Field(default=None)
     """Name of function called."""
 
+    internal_name: str | None = Field(default=None)
+    """Internal name for tool (if any)."""
+
     error: ToolCallError | None = Field(default=None)
     """Error which occurred during tool call."""
 

inspect_ai/model/_model.py

@@ -33,6 +33,7 @@ from inspect_ai._util.content import (
 from inspect_ai._util.hooks import init_hooks, override_api_key, send_telemetry
 from inspect_ai._util.interrupt import check_sample_interrupt
 from inspect_ai._util.logger import warn_once
+from inspect_ai._util.notgiven import NOT_GIVEN, NotGiven
 from inspect_ai._util.platform import platform_init
 from inspect_ai._util.registry import (
     RegistryInfo,
@@ -77,7 +78,7 @@ class ModelAPI(abc.ABC):
     by the user. You can then pass these on to the approriate place in
     your model initialisation code (for example, here is what many
     of the built-in providers do with the `model_args` passed to them:
-    https://inspect.ai-safety-institute.org.uk/models.html#model-args)
+    https://inspect.aisi.org.uk/models.html#model-args)
     """
 
     def __init__(
@@ -232,15 +233,19 @@ class Model:
     config: GenerateConfig
     """Generation config."""
 
-    def __init__(self, api: ModelAPI, config: GenerateConfig) -> None:
+    def __init__(
+        self, api: ModelAPI, config: GenerateConfig, model_args: dict[str, Any] = {}
+    ) -> None:
         """Create a model.
 
         Args:
            api: Model API provider.
            config: Model configuration.
+           model_args: Optional model args
         """
         self.api = api
         self.config = config
+        self.model_args = model_args
 
         # state indicating whether our lifetime is bound by a context manager
         self._context_bound = False
@@ -449,6 +454,7 @@ class Model:
         async def generate() -> ModelOutput:
             check_sample_interrupt()
 
+            cache_entry: CacheEntry | None
             if cache:
                 if isinstance(cache, CachePolicy):
                     policy = cache
@@ -476,6 +482,8 @@ class Model:
                         call=None,
                     )
                     return existing
+            else:
+                cache_entry = None
 
             # verify that model apis are allowed
             self.verify_model_apis()
@@ -545,7 +553,7 @@ class Model:
                     json.dumps(dict(model=str(self), usage=output.usage.model_dump())),
                 )
 
-            if cache:
+            if cache and cache_entry:
                 cache_store(entry=cache_entry, output=output)
 
             return output
@@ -773,6 +781,10 @@ def get_model(
     if isinstance(model, Model):
         return model
 
+    # next see if this is the special "none" model
+    if model == "none":
+        model = "none/none"
+
     # now try finding an 'ambient' model (active or env var)
     if model is None:
         # return active_model if there is one
@@ -835,7 +847,7 @@ def get_model(
             config=config,
             **model_args,
         )
-        m = Model(modelapi_instance, config)
+        m = Model(modelapi_instance, config, model_args)
         if memoize:
             _models[model_cache_key] = m
         return m
@@ -860,17 +872,25 @@ def cached_model(key: str) -> Model | None:
 
 
 def resolve_models(
-    model: str | Model | list[str] | list[Model] | None,
+    model: str | Model | list[str] | list[Model] | None | NotGiven = NOT_GIVEN,
     model_base_url: str | None = None,
     model_args: dict[str, Any] = dict(),
     config: GenerateConfig = GenerateConfig(),
 ) -> list[Model]:
+    # resolve NotGiven to current INSPECT_EVAL_MODEL
+    if isinstance(model, NotGiven):
+        model = os.getenv("INSPECT_EVAL_MODEL", None)
+
+    # resolve None to NoModel
+    if model is None:
+        return [get_model("none")]
+
     # reflect back a plain model
     if isinstance(model, Model):
         return [model]
 
     # helper to resolve model of various types
-    def resolve_model(m: str | Model | None) -> Model:
+    def resolve_model(m: str | Model) -> Model:
         return get_model(
             model=m,
             base_url=model_base_url,
@@ -878,11 +898,8 @@ def resolve_models(
             **model_args,
         )
 
-    # resolve None and str to list
-    if model is None or isinstance(model, str):
-        model = model or os.getenv("INSPECT_EVAL_MODEL", None)
-        if model is None:
-            raise ValueError("No model specified (and no INSPECT_EVAL_MODEL defined)")
+    # str to list
+    if isinstance(model, str):
         model = [m.strip() for m in model.split(",")]
 
     # resolve models
@@ -1098,6 +1115,7 @@ def tool_result_images_reducer(
                     content=edited_tool_message_content,
                     tool_call_id=message.tool_call_id,
                     function=message.function,
+                    internal_name=message.internal_name,
                 )
             ],
             pending_content + new_user_message_content,
@@ -1236,7 +1254,7 @@ def active_model() -> Model | None:
 
 
 # shared contexts for asyncio tasks
-active_model_context_var: ContextVar[Model] = ContextVar("active_model")
+active_model_context_var: ContextVar[Model | None] = ContextVar("active_model")
 
 
 def handle_sample_message_limit(input: str | list[ChatMessage]) -> None:

inspect_ai/model/_model_output.py

@@ -188,8 +188,10 @@ class ModelOutput(BaseModel):
         model: str,
         tool_name: str,
         tool_arguments: dict[str, Any],
+        internal_tool_name: str | None = None,
         tool_call_id: str | None = None,
         content: str | None = None,
+        type: str = "function",
     ) -> "ModelOutput":
         """
         Returns a ModelOutput for requesting a tool call.
@@ -197,6 +199,8 @@ class ModelOutput(BaseModel):
         Args:
             model: model name
             tool_name: The name of the tool.
+            internal_tool_name: The model's internal name for the tool (if any).
+            type: The model's type for the tool. e.g. "function", "computer_use_preview"
             tool_arguments: The arguments passed to the tool.
             tool_call_id: Optional ID for the tool call. Defaults to a random UUID.
             content: Optional content to include in the message. Defaults to "tool call for tool {tool_name}".
@@ -221,8 +225,9 @@ class ModelOutput(BaseModel):
                             ToolCall(
                                 id=tool_call_id,
                                 function=tool_name,
+                                internal_name=internal_tool_name,
                                 arguments=tool_arguments,
-                                type="function",
+                                type=type,
                             )
                         ],
                     ),