feat: add jev agent tool

[ionut-mihalache-uipath]

Summary

This PR lets an agent definition declare the experimental Jev Classifier built-in tool (properties.toolType: "jev-classifier"). The tool sends a text, an object or a list to TypeSafe's Jev model. Jev answers typed questions about it and returns a probability for each answer.

This PR only adds models to uipath.agent.models.agent, so agent.json files that contain the tool parse. Executing the tool is out of scope here.

Changes

• AgentInternalToolType.JEV_CLASSIFIER and AgentInternalJevClassifierToolProperties, added to the internal tool properties union.

  • settings holds only model (default jev-latest).
  • The tool's inputs, state (what Jev evaluates) and questions (what it answers), are ordinary tool arguments. They are declared in the resource's inputSchema and supplied through argumentProperties, so they are not in settings.
  • Settings written by earlier builds of the editor (with state / questions inside settings) still parse. Those keys are kept as ignored extras.

• Question models (JevQuestion, a union keyed on type with case-insensitive values), for validating the questions argument. The limits are TypeSafe's:

  Type	Fields	Rules
  choice	options: [{name, description?}]	2–255 options, unique names
  score	levels: [str]	2–10 non-empty levels, ordered
  noul	criteria?: {true?, false?}	yes/no; the JSON keys true / false match TypeSafe's API

  • Every question has a name matching ^[A-Za-z_][A-Za-z0-9_]{0,63}$ and non-empty instructions.
  • A question is rejected if it carries a non-null field that belongs to another question type.

• objectBuilder argument properties variant (AgentToolObjectBuilderArgumentProperties). It marks an object argument whose properties each have their own value source, stored under <field>['<property>']. The Jev editor uses it for the state and the questions builder.

Testing

• New tests/agent/models/test_jev_classifier_tool.py. It covers settings and defaults, the case-insensitive tool type, legacy settings extras, all three question types, noul criteria serialization, invalid questions, and objectBuilder parsing.
• uv run pytest tests/agent/models: 177 passed.

🤖 Generated with Claude Code

https://claude.ai/code/session_014RnRusgxSnX8MbkqXBx4WG

[Copilot]

Copilot review overview

🟢 Approval recommended

The new models integrate with existing discriminated unions and are covered by focused validation tests.

Review effort: Balanced
Findings: None

What changed in this PR

Adds model support for the internal Jev classifier tool and its question formats.

Changes:

• Adds Jev tool, question, criteria, and object-builder models.
• Adds validation and parsing tests for Jev configurations.

File	Description
packages/​uipath/​src/​uipath/​agent/​models/​agent.py	Defines Jev classifier and object-builder models.
packages/​uipath/​tests/​agent/​models/​test_jev_classifier_tool.py	Tests parsing, validation, defaults, and argument properties.

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[IRONICBo]

Moving questions out of settings also removes the only aggregate uniqueness check (and the duplicate-name regression). _QUESTION now validates one item at a time, while inputSchema and argumentProperties remain generic, so two runtime questions with the same name pass every model added here. System One’s wire contract keys questions by name; converting that list to the request object must then either overwrite one question or fail later in the execution layer.

Could this retain a typed question collection/root model with an after-validator, or explicitly require the future execution adapter to reject duplicate names before serialization, with the two-question regression restored? This is the same keyed boundary used by Jev Social’s request builder: the stable question key is also how the matching answer is recovered.

[ionut-mihalache-uipath]

The check for uniqueness is at runtime. If question names were keys in an object, the input schema would have "additionalProperties: true" for the case when the questions are dynamically asked by the LLM. The langchain clients for gemini drop the "additionProperties" field and this is why we can't to that.

[github-actions]

🚨 Heads up: uipath-langchain cross-tests are FAILING 🚨

Your changes may break the uipath-langchain-python integration.

⚠️ These checks are NOT enforced by branch protection rules. Please review the failures before merging.

🔍 Inspect the failed run →

[sonarqubecloud]

Quality Gate passed

Issues
0 New issues
0 Accepted issues

Measures
0 Security Hotspots
100.0% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarQube Cloud