Repository navigation
Description
What happened, and what did you expect?
tool_with() silently drops viewer and/or model_input when they are passed together with parallel (or with each other). Only the first matching argument of an elif chain is applied; the dropped attributes keep their previous value (None) with no error or warning, so a custom tool-call view or model-input presenter never takes effect:
tool_with(tool, parallel=True, viewer=viewer)applies onlyparalleland silently discardsviewertool_with(tool, viewer=viewer, model_input=model_input)applies onlyviewerand discardsmodel_input- with all three kwargs, only
parallelis applied
I expected each of parallel, viewer and model_input to be applied independently. The docstring of tool_with() ("Tool with modifications to various attributes", src/inspect_ai/tool/_tool_with.py:23-42) documents each argument separately with no exclusivity language, and the sibling code paths treat the three attributes as independent: ToolDef.__init__ assigns self.parallel, self.viewer and self.model_input in three separate statements (src/inspect_ai/tool/_tool_def.py:94-96), and the @tool decorator writes TOOL_PARALLEL, TOOL_VIEWER and TOOL_MODEL_INPUT into one metadata dict unconditionally (src/inspect_ai/tool/_tool.py:264-273).
This is reachable from the public API: tool_with is exported from inspect_ai.tool (src/inspect_ai/tool/__init__.py:46,86). For example, the web browser tools are themselves built as tool_with(tool, viewer=web_at_viewer) (src/inspect_ai/tool/_tools/_web_browser/_web_browser.py:182), so adding parallel=True to such a call silently loses the custom view. The attributes are consumed at eval time by tool_def_fields (src/inspect_ai/tool/_tool_def.py:267-276) for transcript rendering and model-input playback.
Steps to reproduce
from inspect_ai.tool import tool
from inspect_ai.tool._tool import TOOL_MODEL_INPUT, TOOL_PARALLEL, TOOL_VIEWER
from inspect_ai.tool._tool_call import ToolCallView
from inspect_ai.tool._tool_def import tool_registry_info
from inspect_ai.tool._tool_with import tool_with
@tool
def demo() -> None:
async def execute(x: int) -> int:
"""Add one.
Args:
x: Input number.
"""
return x + 1
return execute
def viewer(call) -> ToolCallView:
return ToolCallView()
def model_input(index, total, content, hints):
return content
def meta(t):
reg = tool_registry_info(t)
return {
"parallel": reg.parallel,
"viewer": reg.viewer,
"model_input": reg.model_input,
}
t = demo()
tool_with(t, parallel=True, viewer=viewer)
print("parallel+viewer ->", meta(t))
t2 = demo()
tool_with(t2, viewer=viewer, model_input=model_input)
print("viewer+model_input ->", meta(t2))
t3 = demo()
tool_with(t3, parallel=True, viewer=viewer, model_input=model_input)
print("all three ->", meta(t3))
t4 = demo()
tool_with(t4, viewer=viewer)
print("viewer alone ->", meta(t4))
t5 = demo()
tool_with(t5, model_input=model_input)
print("model_input alone ->", meta(t5))
t6 = demo()
tool_with(t6, parallel=True)
print("parallel alone ->", meta(t6))Actual output on this commit (function addresses vary per run; reproduced identically in three separate runs):
parallel+viewer -> {'parallel': True, 'viewer': None, 'model_input': None}
viewer+model_input -> {'parallel': False, 'viewer': <function viewer at 0x108e9bd80>, 'model_input': None}
all three -> {'parallel': True, 'viewer': None, 'model_input': None}
viewer alone -> {'parallel': False, 'viewer': <function viewer at 0x108e9bd80>, 'model_input': None}
model_input alone -> {'parallel': False, 'viewer': None, 'model_input': <function model_input at 0x105c8e020>}
parallel alone -> {'parallel': True, 'viewer': None, 'model_input': None}
Each argument alone works; combining them applies only the first and silently drops the rest.
Root cause: src/inspect_ai/tool/_tool_with.py:62-67 resolves the three attributes with an elif chain (if parallel is not None: ... elif viewer is not None: ... elif model_input is not None: ...), so only the first non-None argument is written into the registry metadata; the others never reach the metadata the eval-time consumer reads.
I'd be happy to open a PR that resolves each attribute with an independent guard, mirroring ToolDef.__init__ and the @tool decorator, plus a regression test combining two of the kwargs.
Searched for existing reports and competing work: a full-text search for tool_with over issues finds only #1471 (basic_agent infinite tool init) and #678 (feature request to expose private elements), both closed and unrelated; no open PR touches src/inspect_ai/tool/_tool_with.py (open PR #5778 concerns tool approval and failing viewers but does not modify this file).
Version and environment
inspect_ai at commit 65447aa49aaae4aa36755ea4b9fe7fa798bed518 (main, after release 0.3.281), Python 3.12.14, macOS 26.6.2 (arm64).
Note: this report was prepared with AI assistance — an AI coding agent ran the reproduction and verification steps described above; happy to be assigned or to open the PR.
Activity
BlueX888 commentedon Oct 16, 2026
Reproduced this. Root cause: src/inspect_ai/tool/_tool_with.py:62-67 resolves parallel, viewer and model_input through an if/elif chain, so only the first non-None argument is written into the tool's registry metadata; the others keep their previous value (None) with no error or warning.
Minimal reproduction:
Observed output:
Fix is ready on
BlueX888:fix/prep-tool-with-elif-drops-attrs(commita6289d0a). Replaces the elif chain with independent if guards for each attribute, matching how the @tool decorator and ToolDef.init already set all three, plus regression tests for the combined cases.Verified with:
python -m pytest tests/tools/test_tool_with.py (6 passed: 4 new regression tests for the combined cases + the 2 pre-existing tests)Happy to open the PR once you confirm the approach or assign this to me.
fix: tool_with() keeps a tool's current name and descriptions unless …