refactor(eval): move all agentic evaluation logic into gooddata_eval SDK

- Add agentic runners for metric_skill, alert_skill, search_tool,
  general_question, guardrail, and conversation test kinds
- agentic_search pass_at_k requires only tool_selected (matches original
  Tavern behavior; tool_correctness is a Langfuse quality metric only)
- Expose evaluate_agentic_* functions for use by Tavern thin shims
- Update uv.lock

JIRA: GDAI-1830
risk: nonprod

Author: zdenekmusil-gd

packages/gooddata-eval/src/gooddata_eval/cli/agentic_runner.py

@@ -0,0 +1,209 @@
+# (C) 2026 GoodData Corporation. All rights reserved.
+# SPDX-License-Identifier: LicenseRef-GoodData-Enterprise
+"""Agentic evaluation runner for gd-eval CLI — handles multi-turn agentic test kinds."""
+
+from __future__ import annotations
+
+import time
+from typing import Any
+
+from gooddata_eval.core.agentic._langfuse import HttpxLangfuseClient, make_langfuse_client
+from gooddata_eval.core.models import CreatedVisualization, DatasetItem
+from gooddata_eval.core.runner import EvalReport, ItemReport
+
+AGENTIC_TEST_KINDS = frozenset({
+    "vis_agentic",            # production: expected_output.visualization (single/multi CreatedVisualization)
+    "agentic_visualization",  # experimental: expected_output.expected_outputs (multi-candidate)
+    "agentic_metric_skill",
+    "agentic_alert_skill",
+    "agentic_search",
+    "agentic_general_question",
+    "agentic_guardrail",
+    "agentic_conversation",
+})
+
+
+def _parse_visualization_expected(expected_output: Any) -> list[CreatedVisualization]:
+    """Parse expected_output into a list of CreatedVisualization candidates.
+
+    Accepts:
+      {"expected_outputs": [{"visualization": {...}}, ...]}  <- agentic fixture format
+      {"visualization": {...}} or {"visualization": [{...}]}  <- single/multi candidate
+      [{"visualization": {...}}, ...]                          <- bare list
+    """
+    if isinstance(expected_output, dict):
+        raw_list = expected_output.get("expected_outputs")
+        if raw_list is not None:
+            return [
+                CreatedVisualization.model_validate(v.get("visualization", v) if isinstance(v, dict) else v)
+                for v in raw_list
+            ]
+        raw_viz = expected_output.get("visualization")
+        if raw_viz is not None:
+            if isinstance(raw_viz, list):
+                return [CreatedVisualization.model_validate(v) for v in raw_viz]
+            return [CreatedVisualization.model_validate(raw_viz)]
+    if isinstance(expected_output, list):
+        return [
+            CreatedVisualization.model_validate(v.get("visualization", v) if isinstance(v, dict) else v)
+            for v in expected_output
+        ]
+    raise ValueError(
+        f"Cannot parse agentic_visualization expected_output: {type(expected_output).__name__}. "
+        'Expected {"expected_outputs": [...]} or {"visualization": {...}}.'
+    )
+
+
+def _dispatch_agentic(
+    item: DatasetItem,
+    host: str,
+    token: str,
+    workspace_id: str,
+    k: int,
+    langfuse: Any,
+    run_ts: str,
+    model_version_override: str | None,
+) -> None:
+    """Call the appropriate evaluate_agentic_* function for the item's test_kind."""
+    kind = item.test_kind
+    eo = item.expected_output
+    lf_kw = dict(
+        langfuse=langfuse,
+        dataset_item_id=item.id,
+        dataset_name=item.dataset_name,
+        run_timestamp=run_ts,
+        model_version_override=model_version_override,
+    )
+
+    if kind in ("vis_agentic", "agentic_visualization"):
+        from gooddata_eval.core.agentic.visualization import evaluate_agentic_visualization  # noqa: PLC0415
+        evaluate_agentic_visualization(
+            host=host, token=token, workspace_id=workspace_id,
+            question=item.question,
+            expected_outputs=_parse_visualization_expected(eo),
+            k=k,
+            **lf_kw,
+        )
+    elif kind == "agentic_metric_skill":
+        from gooddata_eval.core.agentic.metric_skill import evaluate_agentic_metric_skill  # noqa: PLC0415
+        evaluate_agentic_metric_skill(
+            host=host, token=token, workspace_id=workspace_id,
+            question=item.question,
+            expected_output=eo if isinstance(eo, dict) else {},
+            k=k,
+            **lf_kw,
+        )
+    elif kind == "agentic_alert_skill":
+        from gooddata_eval.core.agentic.alert_skill import evaluate_agentic_alert_skill  # noqa: PLC0415
+        evaluate_agentic_alert_skill(
+            host=host, token=token, workspace_id=workspace_id,
+            question=item.question,
+            expected_output=eo if isinstance(eo, dict) else {},
+            k=k,
+            **lf_kw,
+        )
+    elif kind == "agentic_search":
+        from gooddata_eval.core.agentic.search_tool import evaluate_agentic_search_tool  # noqa: PLC0415
+        eo_dict = eo if isinstance(eo, dict) else {}
+        tool_call = eo_dict.get("tool_call", {})
+        expected_args = tool_call.get("function_arguments", eo_dict)
+        evaluate_agentic_search_tool(
+            host=host, token=token, workspace_id=workspace_id,
+            question=item.question,
+            expected_tool_call=expected_args,
+            k=k,
+            **lf_kw,
+        )
+    elif kind == "agentic_general_question":
+        from gooddata_eval.core.agentic.general_question import evaluate_agentic_general_question  # noqa: PLC0415
+        evaluate_agentic_general_question(
+            host=host, token=token, workspace_id=workspace_id,
+            question=item.question,
+            expected_output=eo if isinstance(eo, str) else str(eo),
+            k=k,
+            **lf_kw,
+        )
+    elif kind == "agentic_guardrail":
+        from gooddata_eval.core.agentic.guardrail import evaluate_agentic_guardrail  # noqa: PLC0415
+        evaluate_agentic_guardrail(
+            host=host, token=token, workspace_id=workspace_id,
+            question=item.question,
+            expected_output=eo if isinstance(eo, str) else str(eo),
+            k=k,
+            **lf_kw,
+        )
+    elif kind == "agentic_conversation":
+        from gooddata_eval.core.agentic.conversation import ConversationFixture, evaluate_agentic_conversation  # noqa: PLC0415
+        fixture_data = eo.get("fixture") or eo if isinstance(eo, dict) else {}
+        evaluate_agentic_conversation(
+            host=host, token=token, workspace_id=workspace_id,
+            fixture=ConversationFixture.model_validate(fixture_data),
+            **lf_kw,
+        )
+    else:
+        raise ValueError(f"Unknown agentic test kind: {kind!r}")
+
+
+def run_agentic_items(
+    items: list[DatasetItem],
+    host: str,
+    token: str,
+    workspace_id: str,
+    *,
+    k: int = 2,
+    model_version: str | None = None,
+    use_langfuse: bool = False,
+    run_ts: str,
+    on_item_start: Any = None,
+    on_item_done: Any = None,
+) -> EvalReport:
+    """Run agentic items through evaluate_agentic_* and return an EvalReport."""
+    langfuse = make_langfuse_client() if use_langfuse else None
+
+    report = EvalReport(model=model_version)
+    total = len(items)
+
+    for index, item in enumerate(items, start=1):
+        if on_item_start is not None:
+            try:
+                on_item_start(index, total, item)
+            except Exception:
+                pass
+
+        item_report = ItemReport(
+            id=item.id,
+            dataset_name=item.dataset_name,
+            test_kind=item.test_kind,
+            question=item.question,
+        )
+        t0 = time.perf_counter()
+        try:
+            _dispatch_agentic(item, host, token, workspace_id, k, langfuse, run_ts, model_version)
+            item_report.pass_at_k = True
+            item_report.runs = k
+        except AssertionError as exc:
+            item_report.pass_at_k = False
+            item_report.runs = k
+            print(f"[agentic] {item.id} FAIL: {exc}", flush=True)
+        except Exception as exc:
+            item_report.error = f"{type(exc).__name__}: {exc}"
+            item_report.runs = 0
+        finally:
+            item_report.latency_s = time.perf_counter() - t0
+
+        if on_item_done is not None:
+            try:
+                on_item_done(index, total, item_report)
+            except Exception:
+                pass
+
+        report.items.append(item_report)
+
+    if langfuse is not None:
+        try:
+            langfuse.flush()
+            langfuse.close()
+        except Exception:
+            pass
+
+    return report

packages/gooddata-eval/src/gooddata_eval/cli/main.py

@@ -19,6 +19,7 @@
 from gooddata_eval.core.models import ChatResult, DatasetItem
 from gooddata_eval.core.reporting.console import render_comparison, render_console
 from gooddata_eval.core.reporting.json_report import write_multi_model_report
+from gooddata_eval.cli.agentic_runner import AGENTIC_TEST_KINDS, run_agentic_items
 from gooddata_eval.core.runner import ItemReport, run_items
 from gooddata_eval.core.summary.http_client import SummaryClient
 from gooddata_eval.core.workspace import ModelResolutionError, WorkspaceModelController
@@ -62,6 +63,17 @@ def _build_parser() -> argparse.ArgumentParser:
     source = run.add_mutually_exclusive_group(required=True)
     source.add_argument("--dataset", help="Path to a folder of dataset JSON files.")
     source.add_argument("--langfuse-dataset", dest="langfuse_dataset", help="Langfuse dataset name.")
+    run.add_argument(
+        "--kind",
+        dest="kind",
+        default="visualization",
+        metavar="TEST_KIND",
+        help=(
+            "Default test kind for dataset items that don't embed one. "
+            "Use 'agentic_visualization', 'agentic_metric_skill', etc. for multi-turn agentic eval. "
+            "(default: visualization)"
+        ),
+    )
     run.add_argument(
         "--model",
         action="append",
@@ -165,7 +177,7 @@ def _load_dataset(config: RunConfig):
 
     if config.langfuse_dataset is None:  # pragma: no cover - argparse mutually-exclusive group guarantees one is set
         raise ValueError("Either --dataset or --langfuse-dataset is required.")
-    return load_langfuse_dataset(config.langfuse_dataset)
+    return load_langfuse_dataset(config.langfuse_dataset, default_test_kind=config.kind)
 
 
 def _list_models(host: str, token: str, workspace_id: str | None) -> int:
@@ -228,6 +240,8 @@ def _run(config: RunConfig) -> int:
         return _EXIT_OPERATIONAL_ERROR
 
     items = _load_dataset(config)
+    agentic_items = [i for i in items if i.test_kind in AGENTIC_TEST_KINDS]
+    non_agentic_items = [i for i in items if i.test_kind not in AGENTIC_TEST_KINDS]
     models = config.models or []
     run_ts = datetime.now(timezone.utc).strftime("%Y-%m-%d-%H-%M")
     n_models = len(models) if models else 1
@@ -287,13 +301,30 @@ def on_langfuse_item_done(
                 ) -> None:
                     _sink.log_item(report, dataset_item_id=report.id)
 
+            # --- agentic items (multi-turn, use evaluate_agentic_*) ---
+            agentic_report = None
+            if agentic_items:
+                agentic_report = run_agentic_items(
+                    agentic_items,
+                    host=config.host,
+                    token=config.token,
+                    workspace_id=config.workspace_id,
+                    k=config.runs,
+                    model_version=resolved.model_id,
+                    use_langfuse=config.log_to_langfuse,
+                    run_ts=run_ts,
+                    on_item_start=on_item_start,
+                    on_item_done=on_item_done,
+                )
+
+            # --- non-agentic items (single-turn, use Evaluator) ---
             backend = _RoutingBackend(
                 ChatClient(host=config.host, token=config.token, workspace_id=config.workspace_id),
                 SummaryClient(host=config.host, token=config.token, workspace_id=config.workspace_id),
             )
             try:
-                report = run_items(
-                    items,
+                single_report = run_items(
+                    non_agentic_items,
                     backend,
                     runs=config.runs,
                     model=resolved.model_id,
@@ -310,6 +341,19 @@ def on_langfuse_item_done(
                 if hasattr(backend, "close"):
                     backend.close()
 
+            # merge into a single report for display/export
+            from gooddata_eval.core.runner import EvalReport  # noqa: PLC0415
+            report = EvalReport(
+                model=resolved.model_id,
+                provider_name=resolved.provider_name or resolved.provider_id,
+                provider_type=resolved.provider_type,
+                workspace_id=config.workspace_id,
+            )
+            if agentic_report is not None:
+                report.items.extend(agentic_report.items)
+            report.items.extend(single_report.items)
+            report.wall_clock_s = (agentic_report.wall_clock_s if agentic_report else 0.0) + single_report.wall_clock_s
+
             skipped_kinds = sorted({i.test_kind for i in report.items if i.skipped})
             if skipped_kinds:
                 print(
@@ -363,6 +407,7 @@ def main(argv: list[str] | None = None) -> int:
             json_path=Path(args.json_path) if args.json_path else None,
             log_to_langfuse=args.langfuse,
             quiet=args.quiet,
+            kind=args.kind,
         )
         return _run(config)
     except (