fix(agent): show static args in conversation tool call events [UV-17565] by andrewwan-uipath · Pull Request #1180 · UiPath/uipath-langchain-python

andrewwan-uipath · 2026-10-30T01:31:05Z

Problem

UV-17565: when a conversational agent's tool argument is bound with an argument expression, the tool runs with the bound value, but the chat and the Studio Web debug window show the value the model inferred.

Static args are applied to the tool calls after the model call (StaticArgsHandler.apply_to_response), but startToolCall / executingToolCall are built from the streamed chunks, which still carry the model's args.

Fix

StaticArgsHandler also keeps display values per tool, with sensitive values masked as <hidden>.
The LLM node passes them in the model run metadata. LangGraph forwards that metadata with each messages stream chunk.
The runtime passes the stream metadata to map_event, and the mapper applies the values to the tool call args when building the tool call events.

Testing

New mapper tests for the streamed and full-message paths.
A handler test for the display values and masking.
An end-to-end LLM node test that streams through a real LangGraph graph. It fails on main with LLM-GUESS-123.
ruff, mypy and the full test suite pass locally.

🤖 Generated with Claude Code

Argument expressions (static args) are applied to the tool calls after the model call, but startToolCall/executingToolCall are built from the streamed chunks, so the chat and debug window showed the model-inferred args instead of the ones the tool is called with. The LLM node now passes the resolved display values (sensitive values masked) in the model run metadata, and the message mapper applies them when building the tool call events. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Copilot

🟡 Changes recommended

Confirmation approval can overwrite sensitive server-side tool arguments with the literal <hidden> placeholder, causing tools to receive masked values instead of their bound secrets.

1 open issue

Preserve masked sensitive arguments during tool confirmation · New

What changed in this PR

This PR makes conversational tool-call events display resolved static arguments rather than model-inferred values, aligning chat and Studio Web debugging with tool execution.

Changes:

Pass static-argument display values through model-run metadata, masking sensitive values.
Apply those values when mapping streamed and full-message tool-call events.
Add regression coverage and bump the package version to 0.18.52.

File	Description
`uv.lock`	Update the locked package version.
`tests/runtime/test_runtime_unit.py`	Update the mapper mock to accept metadata.
`tests/runtime/test_chat_message_mapper.py`	Test display overrides and tool isolation.
`tests/agent/tools/test_static_args.py`	Test display values and sensitive-value masking.
`tests/agent/react/test_llm_node.py`	Add a LangGraph streaming integration test.
`src/uipath_langchain/runtime/runtime.py`	Forward stream metadata to the event mapper.
`src/uipath_langchain/runtime/messages.py`	Apply static display values to tool-call events.
`src/uipath_langchain/agent/tools/static_args.py`	Cache display values with sensitive arguments masked.
`src/uipath_langchain/agent/react/llm_node.py`	Attach display values to model-run metadata.
`src/uipath_langchain/agent/contracts/static_args.py`	Define the shared metadata key.
`pyproject.toml`	Bump the package version to `0.18.52`.

_{Review effort: Balanced}

The tool call events show sensitive static args masked, so the approval input may carry the placeholder or drop the arg. Restore sensitive static args from the server-side call args so the tool still gets the real value. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Restore all static args from the server-side call args on approval, so an edited arg (e.g. host) can't redirect a restored secret. Keep sensitive static args masked in the approved args and cancel metadata sent back to the model and client. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Copilot

🟡 Changes recommended

Edited confirmations can leak restored secrets into conversation output and subsequent model input, while empty wildcard arrays can leave bound tool inputs unrestored.

2 open issues

1 previously missed issue

In code that hasn't changed since last review

Restore bound values for empty wildcard arrays

src/uipath_langchain/chat/hitl.py:169

For a sensitive binding such as $['files'][*], approving {"files": []} leaves the list empty instead of restoring its bound value, so the tool executes without that input. apply_static_args() already handles this JSONPath limitation explicitly. Handle empty wildcard arrays in the restoration helper while preserving unrelated approved edits, and add a confirmation regression test.

_{Review effort: Balanced}

    # Mutate call args so the tool executes with the approved values
    call["args"] = approved_args
    return ConfirmationResult(
        cancelled=None,
        args_modified=approved_args != original_args,
        approved_args=approved_args,
    )


Restore a [*] static path as the whole array, and remove a static path the server-side call args don't have instead of keeping the client's value. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Copilot

🟡 Changes recommended

The restore logic can revert approved dynamic-child edits and overwrite distinct nested arrays with the first match, causing data loss and incorrect argument-modification tracking.

3 open issues

1 resolved since last review

Preserve masked sensitive arguments during tool confirmation

_{Review effort: Balanced}

+    argument_properties = getattr(tool, "argument_properties", None) or {}
+    restored = copy.deepcopy(approved_args)
+    for json_path in argument_properties:


+        expr = parse(restore_path)
+        matches = expr.find(original_args)
+        if matches:
+            expr.update_or_create(restored, copy.deepcopy(matches[0].value))


…call Restore static args on approval with the same dedup and apply_static_args semantics used on the model's call, so objectBuilder markers don't revert edits to dynamic children and nested [*] paths keep each array's length. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

sonarqubecloud · 2026-10-30T20:09:35Z

Quality Gate passed

Issues
3 New issues
0 Accepted issues

Measures
0 Security Hotspots
95.5% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud

andrewwan-uipath requested a balanced review from Copilot October 30, 2026 17:59

Copilot started reviewing on behalf of andrewwan-uipath October 30, 2026 18:00

View session

Copilot AI reviewed Oct 30, 2026

View reviewed changes

Comment thread src/uipath_langchain/runtime/messages.py Resolved

andrewwan-uipath requested a balanced review from Copilot October 30, 2026 19:02

Copilot started reviewing on behalf of andrewwan-uipath October 30, 2026 19:03

View session

Copilot AI reviewed Oct 30, 2026

View reviewed changes

Comment thread src/uipath_langchain/chat/hitl.py

Comment on lines 198 to 204

# Mutate call args so the tool executes with the approved values

call["args"] = approved_args

return ConfirmationResult(

cancelled=None,

args_modified=approved_args != original_args,

approved_args=approved_args,

)

fix(agent): restore [*] and unset static args on tool confirmation

d82ea3c

Restore a [*] static path as the whole array, and remove a static path the server-side call args don't have instead of keeping the client's value. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

andrewwan-uipath requested a balanced review from Copilot October 30, 2026 19:36

Copilot started reviewing on behalf of andrewwan-uipath October 30, 2026 19:37

View session

Copilot AI reviewed Oct 30, 2026

View reviewed changes

Navigation Menu

Repository navigation

fix(agent): show static args in conversation tool call events [UV-17565] - #1180

fix(agent): show static args in conversation tool call events [UV-17565]#1180
andrewwan-uipath wants to merge 5 commits into
mainfrom
fix/conversation-tool-call-static-args

andrewwan-uipath commented Oct 30, 2026 •

edited by atlassian Bot

Loading

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI left a comment

Uh oh!

sonarqubecloud Bot commented Oct 30, 2026

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

andrewwan-uipath commented Oct 30, 2026 • edited by atlassian Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Problem

Fix

Testing

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

🟡 Changes recommended

Uh oh!

sonarqubecloud Bot commented Oct 30, 2026

Quality Gate passed

Uh oh!

This branch has not been deployed

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

andrewwan-uipath commented Oct 30, 2026 •

edited by atlassian Bot

Loading