Merge pull request #44 from AgentX-ai/feat/group-score-with-session

group scorer with session

Author: RobinXL

EVALUATIONS.md

@@ -432,7 +432,7 @@ Score strictly: any missing policy detail is a failing response.""",
 ```
 
 - `judge_prompt` is a raw template. `{input}`, `{output}`, and `{expected}` are substituted in; everything else (chain of thought, capabilities/references, criteria, per-question `judge_guideline`, delegation notes) is appended automatically after it, so a custom prompt can restructure the grading philosophy without ever losing that context. Omit it to keep the default rubric.
-- `judge_model` accepts any OpenAI or Anthropic model id (`client.evaluations.list_models(provider="Anthropic")` to discover valid ones). Omit it to keep the default (`gpt-5.5`).
+- `judge_model` accepts any OpenAI or Anthropic model id (`client.evaluations.list_models(provider="Anthropic")` to discover valid ones). Omit it to keep the engine default (`gpt-5.6-luna`).
 - `list_models()` is **hosted platform only**: it calls the hosted API's `/custom-agent-evaluations/models` registry, which the self-host engine does not serve (404, surfaced as `AgentXEvaluationsError`). On self-host, pass any model id your engine's judge keys can reach.
 
 ---
@@ -902,7 +902,7 @@ This run + gate flow is the **self-host CI path**. The separate CI-runs API in [
 report = client.evaluations.run(...).execute(my_agent).finalize().analyze(
     mode="auto",                                    # "auto" | "sync" | "batch"
     quality_mode="quality_first",                   # "quality_first" | "balanced"
-    judges=["gpt-5.5", "claude-opus-4-8"],           # 1-3 model ids; omit for a single gpt-5.5 judge
+    judges=["gpt-5.6-luna", "claude-opus-4-8"],      # 1-3 model ids; omit for the platform default judge
 )
 
 report.summary                 # str | None, overall narrative summary
@@ -933,7 +933,7 @@ This is separate from, and available even without, the numeric `average_rating`/
 |---|---|---|
 | `mode` | `"auto"` (default), `"sync"`, `"batch"` | How item scoring executes server-side; `"auto"` picks based on run size |
 | `quality_mode` | `"quality_first"`, `"balanced"` | `"quality_first"` runs a second judge on every item; `"balanced"` samples based on risk |
-| `judges` | 1-3 model ids | Which LLM(s) score each response. The first always runs; a second confirms, a third only breaks a tie between the first two. Defaults to a single judge, `["gpt-5.5"]`, if omitted. |
+| `judges` | 1-3 model ids | Which LLM(s) score each response. The first always runs; a second confirms, a third only breaks a tie between the first two. Omit to score with a single judge, the engine's platform default model. |
 | `poll_interval` | seconds, default `5.0` | How often to check job status while waiting |
 | `timeout` | seconds, default `1800.0` | Give up waiting after this long (the job keeps running server-side; call `get_report()` later to check on it) |
 

agentx/evaluations/datasets.py

@@ -48,7 +48,7 @@ def __init__(
             "questions": [],
         }
         # LLM-as-judge overrides for this dataset's own grading config. Omit either to keep the
-        # server default (raw prompt template / OpenAI gpt-5.5, see EVALUATIONS.md). judge_model
+        # server default (raw prompt template / gpt-5.6-luna, see EVALUATIONS.md). judge_model
         # must be one of client.evaluations.list_models() (OpenAI or Anthropic).
         if judge_prompt is not None:
             self._payload["judgePrompt"] = judge_prompt

agentx/evaluations/evaluation_settings.py

@@ -45,7 +45,7 @@ def __init__(
             "evaluationCriteria": evaluation_criteria,
         }
         # LLM-as-judge overrides. Omit either to keep the server default (raw prompt template /
-        # OpenAI gpt-5.5, see EVALUATIONS.md). judge_model must be one of
+        # gpt-5.6-luna, see EVALUATIONS.md). judge_model must be one of
         # client.evaluations.list_models() (OpenAI or Anthropic).
         if judge_prompt is not None:
             self._payload["judgePrompt"] = judge_prompt

agentx/evaluations/models.py

@@ -113,8 +113,8 @@ class EvaluationSettings(BaseModel):
     evaluation_criteria: Optional[str] = Field(default=None, alias="evaluationCriteria")
     # Custom code scorers on this scorer's offline profile - [{ id, name, code, enabled }].
     code_scorers: Optional[List[Dict[str, Any]]] = Field(default=None, alias="codeScorers")
-    # LLM-as-judge overrides. None means "use the server default" (raw prompt template / OpenAI
-    # gpt-5.5). See client.evaluations.settings.builder(judge_prompt=..., judge_model=...).
+    # LLM-as-judge overrides. None means "use the server default" (raw prompt template /
+    # gpt-5.6-luna). See client.evaluations.settings.builder(judge_prompt=..., judge_model=...).
     judge_prompt: Optional[str] = Field(default=None, alias="judgePrompt")
     judge_model: Optional[str] = Field(default=None, alias="judgeModel")
     status: str = "published"

agentx/evaluations/runner.py

@@ -63,9 +63,6 @@ def _say(*args, **kwargs) -> None:
     "l4_final_reduce": "writing final report",
 }
 
-_DEFAULT_JUDGE_MODEL = "gpt-5.5"
-
-
 class GateResult:
     """Wire result of the CI gate (GET /runs/:id/gate) with attribute access for the fields a
     CI script actually branches on."""
@@ -456,16 +453,16 @@ def analyze(
         Args:
             mode: "auto" (default), "sync", or "batch" - how item scoring executes server-side.
             quality_mode: "quality_first" or "balanced" - how many items get a second/third judge.
-            judges: 1-3 model ids, e.g. ``["gpt-5.5", "claude-opus-4-8"]``. Defaults to a single
-                judge, ``["gpt-5.5"]``, rather than the dashboard's 3-judge default - SDK runs are
-                typically lighter-weight, quick-start evaluations.
+            judges: 1-3 model ids from ``client.evaluations.list_models()``, e.g.
+                ``["gpt-5.6-luna", "claude-opus-4-8"]``. Omit to let the engine score with its
+                platform default model (a single judge, rather than the dashboard's 3-judge
+                default - SDK runs are typically lighter-weight, quick-start evaluations).
             poll_interval: seconds between status checks while waiting.
             timeout: give up waiting after this many seconds (the job keeps running server-side;
                 call ``get_report()`` later to check on it).
         """
         if judges is not None and not (1 <= len(judges) <= 3):
             raise ValueError("judges must contain 1-3 model ids")
-        resolved_judges = judges if judges is not None else [_DEFAULT_JUDGE_MODEL]
 
         _say()
         with Spinner("Analyzing - AI is reviewing your results") as spinner:
@@ -474,7 +471,7 @@ def analyze(
                     self._run.run_id,
                     mode=mode,
                     quality_mode=quality_mode,
-                    judges=resolved_judges,
+                    judges=judges,
                 )
                 deadline = time.monotonic() + timeout
                 status = self._client.get_analysis_status(self._run.run_id)

agentx/monitor/client.py

@@ -429,6 +429,24 @@ def list_session_spans(self, session_id: str) -> List[dict]:
         data = self._request("GET", f"/ingest/sessions/{session_id}/spans", base=self._api_root())
         return data.get("spans", []) if isinstance(data, dict) else data
 
+    def list_session_scores(self, session_id: str) -> List[dict]:
+        """Every session-level verdict on the session, newest first: session-scoped online
+        evaluators (kind ``online-eval:<id>``), session-scoped scorer groups
+        (``scorer-group:<id>``), and legacy coherence rows."""
+        data = self._request(
+            "GET", f"/agent-monitoring/sessions/{session_id}/scores", base=self._api_root()
+        )
+        return data.get("scores", []) if isinstance(data, dict) else data
+
+    def run_session_sweep(self) -> dict:
+        """Run the idle-session sweep once, now - the tick that scores quiet multi-turn
+        sessions with every enabled session-scoped evaluator and scorer group. Production
+        engines run this automatically every minute; the manual trigger exists for demos,
+        tests, and backfills. Returns ``{"judged": n}``."""
+        return self._request(
+            "POST", "/agent-monitoring/session-sweep/run", base=self._api_root(), timeout=300
+        )
+
     # ------------------------------------------------------------------
     # Model portability (self-host): replay a trace's input against other models
     # ------------------------------------------------------------------

agentx/monitor/scorer_groups.py

@@ -65,7 +65,9 @@ def create(
         online: Optional[Dict[str, Any]] = None,
     ) -> ScorerGroup:
         """``members``: [{"kind": "judge"|"pattern"|"custom", "refId": ..., "weight": 1, "gate": False}].
-        ``online``: {"enabled": True, "sampleRate": 0.1, "alertThreshold": 5, "severity": "medium"}
+        ``online``: {"enabled": True, "sampleRate": 0.1, "alertThreshold": 5, "severity": "medium"}.
+        Add ``"scope": "session", "idleSeconds": 120`` to score whole multi-turn sessions once
+        idle, instead of each sampled trace.
         or None for offline-only."""
         payload: Dict[str, Any] = {"name": name, "members": members}
         if description is not None:

agentx/monitor/sessions.py

@@ -19,3 +19,15 @@ def coherence_check(self, session_id: str) -> dict:
     def spans(self, session_id: str) -> List[dict]:
         """Every span in the session (roots and children), oldest first."""
         return self._client.list_session_spans(session_id)
+
+    def scores(self, session_id: str) -> List[dict]:
+        """Session-level verdicts, newest first. ``kind`` says who scored: a session-scoped
+        online evaluator (``online-eval:<id>``) or a session-scoped scorer group
+        (``scorer-group:<id>``)."""
+        return self._client.list_session_scores(session_id)
+
+    def run_sweep(self) -> dict:
+        """Trigger the idle-session sweep once (normally automatic, every minute) - scores
+        idle multi-turn sessions with every enabled session-scoped evaluator and scorer
+        group. Returns ``{"judged": n}``."""
+        return self._client.run_session_sweep()

tests/test_integrations.py

@@ -599,7 +599,9 @@ def test_crewai_captures_real_per_task_timing_via_event_bus():
     durations are provably unequal — the old "divide latency evenly across
     tasks" approximation would have reported them as identical.
     """
-    crewai = pytest.importorskip("crewai")
+    # exc_type=Exception: on older Pythons crewai can raise TypeError (PEP 604 syntax) at
+    # import time, and a broken optional integration should skip this test, not fail it.
+    crewai = pytest.importorskip("crewai", exc_type=Exception)
     from crewai.events.event_bus import crewai_event_bus
     from crewai.events.types.task_events import TaskCompletedEvent, TaskStartedEvent
     from crewai.tasks.task_output import TaskOutput

tests/test_selfhost_analysis_fallback.py

@@ -210,12 +210,12 @@ def test_the_fallback_request_gets_the_long_analysis_timeout():
         {("POST", f"{API_ROOT}/evaluate/analyze/{RUN}"): FakeResponse(200, {"status": "completed"})}
     )
 
-    client.analyze_run(RUN, judges=["gpt-5.5"])
+    client.analyze_run(RUN, judges=["gpt-5.6-luna"])
 
     method, url, kwargs = session.calls[-1]
     assert url == f"{API_ROOT}/evaluate/analyze/{RUN}"
     assert kwargs["timeout"] > 60, "a synchronous judge pass needs more than the 30s default"
-    assert kwargs["json"]["judges"] == [{"model": "gpt-5.5"}]
+    assert kwargs["json"]["judges"] == [{"model": "gpt-5.6-luna"}]
 
 
 # ---------------------------------------------------------------------------