feat(skills): add skills support to advanced agents with sandboxed cli tool

[mariadhakalUipath]

Title: Uipath Skills with Cli tool for advanced agents.

Summary

Adds skills feature for advanced agents. Agent can act on it — run uip commands, review what it ran, and verify what it scaffolded.

[Copilot]

Pull request overview

Adds initial “skills” plumbing to advanced agents and introduces a sandbox-oriented internal tool for running uip commands inside an agent workspace.

Changes:

• Introduces a new create_uipath_cli_tool() (uipath_cli) StructuredTool that runs one uip command per invocation with command validation.
• Adds unit tests covering command parsing, validation, subprocess result mapping, and timeout behavior for the new tool.
• Extends advanced-agent builders to accept and forward a skills parameter into the underlying deepagents agent construction.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 4 comments.

File	Description
tests/agent/tools/internal_tools/test_uipath_cli_tool.py	Adds test coverage for the new sandboxed CLI tool.
src/uipath_langchain/agent/tools/internal_tools/uipath_cli_tool.py	Implements the sandboxed uipath_cli tool and its input/output schemas.
src/uipath_langchain/agent/tools/internal_tools/__init__.py	Exports create_uipath_cli_tool from the internal tools package.
src/uipath_langchain/agent/advanced/agent.py	Adds skills parameter propagation to deepagents agent creation/wrappers.

[Copilot]

Pull request overview

Copilot reviewed 4 out of 4 changed files in this pull request and generated 4 comments.

Comments suppressed due to low confidence (1)

src/uipath_langchain/agent/advanced/agent.py:45

• The create_advanced_agent docstring documents memory but not the newly added skills parameter, which is part of the public API. Update the docstring to describe what skills does and how an empty value is handled.

    """Create a deepagents agent with planning, filesystem, and sub-agent tools.

    ``memory`` is a list of file paths loaded via deepagents' ``MemoryMiddleware``:
    each is read from ``backend`` and injected into the system prompt every turn,
    and the model maintains them with ``edit_file``. Empty disables the middleware.
    """

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 3 comments.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 2 comments.

Comments suppressed due to low confidence (1)

src/uipath_langchain/agent/advanced/agent.py:42

• create_advanced_agent now accepts skills, but the docstring only documents memory. Since empty skills collapses to None (disabling skills support), documenting this behavior would help callers understand how to enable/disable the feature.

    skills: Sequence[str] = (),
) -> CompiledStateGraph[Any, Any, Any, Any]:
    """Create a deepagents agent with planning, filesystem, and sub-agent tools.

    ``memory`` is a list of file paths loaded via deepagents' ``MemoryMiddleware``:

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated no new comments.

Suppressed comments (3)

src/uipath_langchain/agent/tools/internal_tools/uipath_cli_tool.py:37

• The tool description says “A negative exit_code means the command was refused”. That’s inaccurate because real subprocess exit codes can be negative on Unix when terminated by a signal. The contract in UiPathCliOutput/test cases is specifically “exit_code == -1000 indicates refusal”, so the description should match to avoid confusing the agent.

    "Run a single UiPath `uip` command in the agent workspace; returns "
    "exit_code/stdout/stderr. Only the uip/uipath binaries, one command per call; "
    "shell chaining is refused. A negative exit_code means the command was refused "
    "before it ran and stderr explains why. Use `subdir` to target a scaffolded "
    "project folder."

src/uipath_langchain/agent/tools/internal_tools/uipath_cli_tool.py:167

• On timeout, proc.kill() can raise ProcessLookupError if the process exits between the timeout firing and the kill call. That would crash the tool instead of returning the recoverable _REJECTED_EXIT_CODE payload.

    except asyncio.TimeoutError:
        proc.kill()
        await proc.wait()
        return _rejected(
            echoed, f"Command timed out after {_COMMAND_TIMEOUT_SECONDS}s."

src/uipath_langchain/agent/advanced/agent.py:110

• This PR introduces a sandboxed uipath_cli tool, but create_advanced_agent only forwards the skills list into deepagents and never registers the new tool. The new tool module’s docstring says it is “injected programmatically by the agent graph builder when UiPath skills are active”, but there is currently no call site for create_uipath_cli_tool in src/ (verified by repo search). If skills are expected to prescribe uip commands, the agent needs the tool added to its tool list when skills are enabled (and a FilesystemBackend workspace exists).

        memory=list(memory) or None,
        middleware=list(middleware),
        skills=list(skills) if skills else None,

[radu-mocanu]

codeagent init is not a valid command; the cli returns unknown command 'codeagent'. use codedagent init here, and add a schema assertion so the example stays aligned with the actual cli.

[mariadhakalUipath]

Thank you for correcting me. I have changed the example command and also added the schema assertion for the example commands.

[radu-mocanu]

this is a logic-changing public change, but the pr does not bump the package version. main is already at 0.16.2 while this branch still has 0.16.1. rebase and bump to the next version in pyproject.toml before merge.

[mariadhakalUipath]

Okay will do this before I merge.

[radu-mocanu]

this still executes the cli arguments directly, without the cli owning and enforcing a safe mode. the parser currently accepts commands such as solution delete and admin tenants delete, so resolving the previous thread did not address the version-skew problem. let s add the safety flag in the cli and always append it here before shipping this tool.

[mariadhakalUipath]

This is really good point we were thinking about this, the reason I don't have it implemented is:

1. We are using the user's token to run these commands so it is already tenant and folder scoped, agent can only run the commands that user is allowed.
2. This tool is used by both conversational and autonomous low-code agent, conversational agent uses the interactive mode to ask user to run the command and only the autonomous low-code agent is running the command by itself so I added prompt-based guardrail to end the execution if there is any destructive command.
3. It is only attached to agent if the skills are enabled so it will only follow the instructions there and if the user is using this tool in coded agent then it becomes their responsibility to filter the commands.
   Please guide me if I am missing anything that I should have considered for this, and I can work through it.

[radu-mocanu]

proc.kill() can raise ProcessLookupError when the child exits in the timeout and kill race. i reproduced that path and the tool crashes instead of returning the recoverable timeout payload. catch that race and still await proc.wait().

[mariadhakalUipath]

Fix:

except asyncio.TimeoutError:
    try:
        proc.kill()
    except ProcessLookupError:
        pass
    await proc.wait()

wait_for timing out only means the child hadn't finished when the timer fired it can
exit before kill() runs, and the transport teardown that nulls _proc makes kill()
raise. await proc.wait() stays outside the try so it runs on both paths (safe after
teardown: _wait() returns the cached returncode immediately). Catching only
ProcessLookupError so a real PermissionError still surfaces.

Added Regression test: test_run_times_out_when_child_exits_during_kill to verify it fails
with an uncaught ProcessLookupError against the unpatched code.

[radu-mocanu]

this tool is still not wired into the advanced agent. with skills enabled, we only forward the skill paths and leave tools unchanged, so the agent has no way to call uipath_cli. let s add it here for the filesystem-backed skills path and cover the final tool list in a test.

[mariadhakalUipath]

We have this tool behind feature flag and we wired it when creating advanced agent graph for low-code agents if skills are enabled instead of wiring it to advanced agent. The scope for now is to attach the tool for low-code agents.

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
97.8% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud