claude-smart 0.2.45 → 0.2.47

package/plugin/vendor/reflexio/reflexio/models/api_schema/retriever_schema.py

@@ -50,6 +50,11 @@ class SearchInteractionRequest(BaseModel):
 class SearchUserProfileRequest(BaseModel):
     user_id: NonEmptyStr
     generated_from_request_id: str | None = None
+    # Caller correlation IDs for billing attribution on the Application line.
+    # Optional; when populated they flow into the usage event request_id /
+    # session_id columns via _meter_applied_learnings in server/api.py.
+    request_id: str | None = None
+    session_id: str | None = None
     query: str | None = None
     start_time: datetime | None = None
     end_time: datetime | None = None
@@ -59,6 +64,7 @@ class SearchUserProfileRequest(BaseModel):
     extractor_name: str | None = (
         None  # Deprecated compatibility field; accepted but ignored.
     )
+    tags: list[str] | None = None
     threshold: float | None = Field(default=0.4, ge=0.0, le=1.0)
     enable_reformulation: bool | None = False
     search_mode: SearchMode = SearchMode.HYBRID
@@ -181,10 +187,15 @@ class GetInteractionsResponse(BaseModel):
 
 class GetUserProfilesRequest(BaseModel):
     user_id: NonEmptyStr
+    profile_id: str | None = None
+    query: str | None = None
     start_time: datetime | None = None
     end_time: datetime | None = None
     top_k: int | None = Field(default=30, gt=0)
+    source: str | None = None
+    profile_time_to_live: str | None = None
     status_filter: list[Status | None] | None = None
+    tags: list[str] | None = None
 
     @model_validator(mode="after")
     def check_time_range(self) -> Self:
@@ -215,10 +226,22 @@ class SetConfigResponse(BaseModel):
 
 class GetUserPlaybooksRequest(BaseModel):
     limit: int | None = Field(default=100, gt=0)
+    user_playbook_id: int | None = Field(default=None, gt=0)
     user_id: str | None = None
+    request_id: str | None = None
+    query: str | None = None
     playbook_name: str | None = None
     agent_version: str | None = None
+    start_time: datetime | None = None
+    end_time: datetime | None = None
     status_filter: list[Status | None] | None = None
+    tags: list[str] | None = None
+
+    @model_validator(mode="after")
+    def check_time_range(self) -> Self:
+        """Validate that end_time is after start_time."""
+        TimeRangeValidatorMixin.validate_time_range(self.start_time, self.end_time)
+        return self
 
 
 class GetUserPlaybooksResponse(BaseModel):
@@ -229,10 +252,25 @@ class GetUserPlaybooksResponse(BaseModel):
 
 class GetAgentPlaybooksRequest(BaseModel):
     limit: int | None = Field(default=100, gt=0)
+    agent_playbook_id: int | None = Field(default=None, gt=0)
+    query: str | None = None
     playbook_name: str | None = None
     agent_version: str | None = None
+    start_time: datetime | None = None
+    end_time: datetime | None = None
     status_filter: list[Status | None] | None = None
     playbook_status_filter: PlaybookStatus | None = None
+    tags: list[str] | None = None
+    # Caller correlation IDs for billing attribution on the Application line.
+    # Optional; consumed by _meter_applied_learnings in server/api.py.
+    request_id: str | None = None
+    session_id: str | None = None
+
+    @model_validator(mode="after")
+    def check_time_range(self) -> Self:
+        """Validate that end_time is after start_time."""
+        TimeRangeValidatorMixin.validate_time_range(self.start_time, self.end_time)
+        return self
 
 
 class GetAgentPlaybooksResponse(BaseModel):
@@ -241,6 +279,30 @@ class GetAgentPlaybooksResponse(BaseModel):
     msg: str | None = None
 
 
+class GetLearningProvenanceRequest(BaseModel):
+    kind: Literal["profile", "user_playbook", "agent_playbook"]
+    id: str
+
+
+class SourceUserPlaybookProvenanceView(BaseModel):
+    user_playbook: UserPlaybookView
+    interactions: list[InteractionView] = Field(default_factory=list)
+    source_interaction_ids: list[int] = Field(default_factory=list)
+
+
+class LearningProvenanceViewResponse(BaseModel):
+    success: bool
+    target_kind: Literal["profile", "user_playbook", "agent_playbook"]
+    target_id: str
+    provenance_status: Literal["exact", "best_effort", "unavailable"] = "unavailable"
+    trigger_request_id: str | None = None
+    interactions: list[InteractionView] = Field(default_factory=list)
+    source_user_playbooks: list[SourceUserPlaybookProvenanceView] = Field(
+        default_factory=list
+    )
+    msg: str | None = None
+
+
 class SearchUserPlaybookRequest(BaseModel):
     """Request for searching user playbooks with semantic/text search and filtering.
 
@@ -263,10 +325,15 @@ class SearchUserPlaybookRequest(BaseModel):
     start_time: datetime | None = None
     end_time: datetime | None = None
     status_filter: list[Status | None] | None = None
+    tags: list[str] | None = None
     top_k: int | None = Field(default=10, gt=0)
     threshold: float | None = Field(default=0.4, ge=0.0, le=1.0)
     enable_reformulation: bool | None = False
     search_mode: SearchMode = SearchMode.HYBRID
+    # Caller correlation IDs for billing attribution on the Application line.
+    # Optional; consumed by _meter_applied_learnings in server/api.py.
+    request_id: str | None = None
+    session_id: str | None = None
 
     @model_validator(mode="after")
     def check_time_range(self) -> Self:
@@ -316,10 +383,15 @@ class SearchAgentPlaybookRequest(BaseModel):
     end_time: datetime | None = None
     status_filter: list[Status | None] | None = None
     playbook_status_filter: PlaybookStatus | list[PlaybookStatus] | None = None
+    tags: list[str] | None = None
     top_k: int | None = Field(default=10, gt=0)
     threshold: float | None = Field(default=0.4, ge=0.0, le=1.0)
     enable_reformulation: bool | None = False
     search_mode: SearchMode = SearchMode.HYBRID
+    # Caller correlation IDs for billing attribution on the Application line.
+    # Optional; consumed by _meter_applied_learnings in server/api.py.
+    request_id: str | None = None
+    session_id: str | None = None
 
     @model_validator(mode="after")
     def check_time_range(self) -> Self:
@@ -345,6 +417,14 @@ class SearchAgentPlaybookResponse(BaseModel):
 class GetAgentSuccessEvaluationResultsRequest(BaseModel):
     limit: int | None = Field(default=100, gt=0)
     agent_version: str | None = None
+    start_time: datetime | None = None
+    end_time: datetime | None = None
+
+    @model_validator(mode="after")
+    def check_time_range(self) -> Self:
+        """Validate that end_time is after start_time."""
+        TimeRangeValidatorMixin.validate_time_range(self.start_time, self.end_time)
+        return self
 
 
 class GetAgentSuccessEvaluationResultsResponse(BaseModel):
@@ -357,10 +437,18 @@ class GetRequestsRequest(BaseModel):
     user_id: str | None = None
     request_id: str | None = None
     session_id: str | None = None
+    source: str | None = None
     start_time: datetime | None = None
     end_time: datetime | None = None
-    top_k: int | None = Field(default=30, gt=0)
-    offset: int | None = Field(default=0, ge=0)
+    top_k: int | None = Field(
+        default=30,
+        gt=0,
+        description="Maximum number of sessions to return. Pagination is "
+        "per-session: every returned session includes all of its requests.",
+    )
+    offset: int | None = Field(
+        default=0, ge=0, description="Number of sessions to skip for pagination."
+    )
 
     @model_validator(mode="after")
     def check_time_range(self) -> Self:
@@ -450,7 +538,8 @@ class TimeSeriesDataPoint(BaseModel):
     """A single data point in a time series."""
 
     timestamp: int = Field(gt=0)  # Unix timestamp
-    value: int = Field(ge=0)  # Count or metric value
+    value: float = Field(ge=0)  # Count or metric value
+    count: int | None = Field(default=None, ge=0)  # Optional weight for rate metrics
 
 
 class PeriodStats(BaseModel):
@@ -621,6 +710,11 @@ class UnifiedSearchRequest(BaseModel):
     enable_reformulation: bool | None = False
     enable_agent_answer: bool | None = False
     search_mode: SearchMode = SearchMode.HYBRID
+    # Caller correlation IDs for billing attribution on the Application line.
+    # Optional; consumed by _meter_applied_learnings in server/api.py.
+    request_id: str | None = None
+    session_id: str | None = None
+    interaction_id: int | None = Field(default=None, gt=0)
 
 
 class UnifiedSearchResponse(BaseModel):

package/plugin/vendor/reflexio/reflexio/models/api_schema/ui/converters.py

@@ -78,6 +78,8 @@ def to_profile_view(profile: UserProfile) -> ProfileView:
         status=profile.status,
         extractor_names=profile.extractor_names,
         source_span=profile.source_span,
+        source_interaction_ids=profile.source_interaction_ids,
+        tags=profile.tags or [],
     )
 
 
@@ -104,6 +106,7 @@ def to_user_playbook_view(rf: UserPlaybook) -> UserPlaybookView:
         source=rf.source,
         source_interaction_ids=rf.source_interaction_ids,
         source_span=rf.source_span,
+        tags=rf.tags or [],
     )
 
 
@@ -127,6 +130,7 @@ def to_agent_playbook_view(fb: AgentPlaybook) -> AgentPlaybookView:
         playbook_status=fb.playbook_status,
         playbook_metadata=fb.playbook_metadata,
         status=fb.status,
+        tags=fb.tags or [],
     )
 
 
@@ -143,6 +147,7 @@ def to_evaluation_result_view(
     """
     return EvaluationResultView(
         result_id=result.result_id,
+        user_id=result.user_id,
         agent_version=result.agent_version,
         session_id=result.session_id,
         is_success=result.is_success,
@@ -173,5 +178,4 @@ def to_profile_change_log_view(log: ProfileChangeLog) -> ProfileChangeLogView:
         created_at=log.created_at,
         added_profiles=[to_profile_view(p) for p in log.added_profiles],
         removed_profiles=[to_profile_view(p) for p in log.removed_profiles],
-        mentioned_profiles=[to_profile_view(p) for p in log.mentioned_profiles],
     )

package/plugin/vendor/reflexio/reflexio/models/api_schema/ui/entities.py

@@ -65,6 +65,8 @@ class ProfileView(BaseModel):
     status: Status | None = None
     extractor_names: list[str] | None = None
     source_span: str | None = None
+    source_interaction_ids: list[int] = Field(default_factory=list)
+    tags: list[str] = Field(default_factory=list)
 
 
 class UserPlaybookView(BaseModel):
@@ -83,6 +85,7 @@ class UserPlaybookView(BaseModel):
     source: str | None = None
     source_interaction_ids: list[int] = Field(default_factory=list)
     source_span: str | None = None
+    tags: list[str] = Field(default_factory=list)
 
 
 class AgentPlaybookView(BaseModel):
@@ -98,12 +101,14 @@ class AgentPlaybookView(BaseModel):
     playbook_status: PlaybookStatus = PlaybookStatus.PENDING
     playbook_metadata: str = ""
     status: Status | None = None
+    tags: list[str] = Field(default_factory=list)
 
 
 class EvaluationResultView(BaseModel):
     """User-facing AgentSuccessEvaluationResult — excludes embedding."""
 
     result_id: int = 0
+    user_id: str = ""
     agent_version: str
     session_id: str
     is_success: bool
@@ -126,4 +131,3 @@ class ProfileChangeLogView(BaseModel):
     created_at: int = Field(default_factory=lambda: int(datetime.now(UTC).timestamp()))
     added_profiles: list[ProfileView]
     removed_profiles: list[ProfileView]
-    mentioned_profiles: list[ProfileView]

package/plugin/vendor/reflexio/reflexio/models/config_schema.py

@@ -95,7 +95,13 @@ def _migrate_dict(data: Any, mapping: dict[str, str]) -> Any:
     if isinstance(data, dict):
         data = dict(data)
         for old, new in mapping.items():
-            if old in data and new not in data:
+            if old not in data:
+                continue
+            # New name wins when both are present; always drop the old key so it
+            # doesn't survive into validation and trip ``extra="forbid"``.
+            if new in data:
+                data.pop(old)
+            else:
                 data[new] = data.pop(old)
     return data
 
@@ -397,6 +403,16 @@ class DeduplicationConfig(BaseModel):
 SINGLETON_PROFILE_EXTRACTOR_NAME = "profile"
 SINGLETON_USER_PLAYBOOK_NAME = "playbook"
 SINGLETON_AGENT_SUCCESS_EVALUATION_NAME = "agent_success"
+DEFAULT_AGENT_SUCCESS_SAMPLING_RATE = 0.05
+DEFAULT_AGENT_SUCCESS_DEFINITION_PROMPT = (
+    "Evaluate whether the AI agent successfully handled the user's session.\n\n"
+    "Mark the session successful when, by the end of the conversation, the agent:\n"
+    "1. Identified and addressed the user's main goal or question.\n"
+    "2. Provided a correct, useful, and actionable response or completed the requested action.\n"
+    "Mark the session unsuccessful when the agent failed to understand the request,\n"
+    "gave incorrect or unhelpful guidance, did not complete an available action,\n"
+    "ignored important constraints, or left the user unsatisfied."
+)
 
 
 class ProfileExtractorConfig(_ExtractorWindowOverrideCompatMixin, BaseModel):
@@ -405,7 +421,7 @@ class ProfileExtractorConfig(_ExtractorWindowOverrideCompatMixin, BaseModel):
     extractor_name: NonEmptyStr | None = None
     extraction_definition_prompt: SanitizedNonEmptyStr
     context_prompt: str | None = None
-    metadata_definition_prompt: str | None = None
+    tagging_definition_prompt: str | None = None
     should_extract_profile_prompt_override: str | None = None
     request_sources_enabled: list[str] | None = (
         None  # default enabled for all sources, if set, only extract profiles from the enabled request sources
@@ -456,7 +472,7 @@ class UserPlaybookExtractorConfig(_ExtractorWindowOverrideCompatMixin, BaseModel
     extractor_name: NonEmptyStr | None = None
     extraction_definition_prompt: SanitizedNonEmptyStr
     context_prompt: str | None = None
-    metadata_definition_prompt: str | None = None
+    tagging_definition_prompt: str | None = None
     aggregation_config: PlaybookAggregatorConfig | None = None
     deduplication_config: DeduplicationConfig | None = None
     request_sources_enabled: list[str] | None = (
@@ -488,9 +504,15 @@ class AgentSuccessConfig(_ExtractorWindowOverrideCompatMixin, BaseModel):
     evaluation_name: NonEmptyStr | None = None
     success_definition_prompt: SanitizedNonEmptyStr
     metadata_definition_prompt: str | None = None
+    request_sources_enabled: list[str] | None = (
+        None  # default enabled for all sources, if set, only evaluate requests from the enabled request sources
+    )
     sampling_rate: float = Field(
-        default=1.0, ge=0.0, le=1.0
-    )  # fraction of window of interactions to be sampled for success evaluation
+        default=DEFAULT_AGENT_SUCCESS_SAMPLING_RATE,
+        ge=0.0,
+        le=1.0,
+        description="Fraction of sessions to evaluate automatically.",
+    )
     window_size_override: int | None = Field(default=None, gt=0)
     stride_size_override: int | None = Field(default=None, gt=0)
 
@@ -500,6 +522,14 @@ class AgentSuccessConfig(_ExtractorWindowOverrideCompatMixin, BaseModel):
         return _migrate_dict(data, _EXTRACTOR_OVERRIDE_MIGRATION)
 
 
+def _default_agent_success_config() -> AgentSuccessConfig:
+    return AgentSuccessConfig(
+        evaluation_name=SINGLETON_AGENT_SUCCESS_EVALUATION_NAME,
+        success_definition_prompt=DEFAULT_AGENT_SUCCESS_DEFINITION_PROMPT,
+        sampling_rate=DEFAULT_AGENT_SUCCESS_SAMPLING_RATE,
+    )
+
+
 class ReflectionConfig(BaseModel):
     """Configuration for the sliding-window reflection step.
 
@@ -559,7 +589,7 @@ class RetrievalFloorConfig(BaseModel):
     weak tail that drives false-positive citations. Calibrate per arm on real data.
     """
 
-    enabled: bool = True
+    enabled: bool = False
     pool_size: int = Field(
         default=30,
         gt=0,
@@ -634,6 +664,62 @@ class PlaybookOptimizerConfig(BaseModel):
         return self
 
 
+class LineageGCConfig(BaseModel):
+    """Configuration for the tombstone garbage-collection job (enabled by default).
+
+    Purpose
+    -------
+    Retains tombstone content for ``tombstone_grace_window_days`` days after the
+    row's *retirement instant* (``retired_at``) to support audit, replay of dedup
+    and aggregation runs, and rollback.  After the window expires the row is
+    hard-deleted and a ``hard_delete`` lineage event is recorded.
+
+    Age basis
+    ---------
+    The GC ages on ``retired_at`` — the INTEGER epoch set when a row is tombstoned
+    (merged, superseded, or archived).  Rows with ``retired_at = NULL`` (created
+    before the column was added) are never eligible; they have no retirement clock.
+
+    Grace window
+    ------------
+    90 days is the default grace window.  This matches common 90-day soft-delete
+    retention policies and satisfies GDPR Art. 5(1)(e) storage-limitation for
+    personal data in profiles.  The value is a per-deployment policy knob; ratify
+    with your DPO before shortening it in production.  The 90-day floor also
+    preserves tombstones long enough for B3 changelog replay and rollback
+    consumers — raise ``tombstone_grace_window_days`` further if your replay
+    horizon exceeds 90 days.
+
+    Enabled by default
+    ------------------
+    GC is ON by default so tombstones created by the soft-delete flags (also ON by
+    default) are reclaimed automatically.  Disabling GC while soft-delete is enabled
+    allows tombstone counts to grow without bound — only do this deliberately (e.g.
+    extended audit hold) and with a plan to re-enable.
+
+    Disabling
+    ---------
+    Set ``enabled = False`` in your deployment config to hold all tombstones
+    indefinitely (e.g. for an extended audit window or rollback standby period).
+    """
+
+    enabled: bool = True
+    tombstone_grace_window_days: int = Field(default=90, gt=0)
+    poll_interval_seconds: int = Field(default=86400, gt=0)
+
+
+class GovernanceRetentionConfig(BaseModel):
+    """Audit-event retention policy. **Enterprise-only:** reclamation is performed
+    by the reflexio_ext GovernanceRetentionCapability. In an OSS-only deployment
+    these knobs are accepted but inert (the OSS lineage scheduler does not reclaim
+    audit events) and the server logs a startup warning when retention is enabled.
+    """
+
+    audit_events_retention_enabled: bool = False
+    audit_events_retention_days: int = Field(default=365, gt=0)
+    audit_events_delete_batch_limit: int = Field(default=500, gt=0)
+
+
 @dataclass(frozen=True)
 class EffectivePendingToolCallConfig:
     """Resolved pending-tool-call settings after applying tool overrides."""
@@ -733,6 +819,8 @@ def _default_user_playbook_extractor_config() -> UserPlaybookExtractorConfig:
 
 
 class Config(BaseModel):
+    model_config = ConfigDict(extra="forbid")
+
     # define where user configuration is stored at
     storage_config: StorageConfig
     storage_config_test: StorageConfigTest | None = StorageConfigTest.UNKNOWN
@@ -749,7 +837,9 @@ class Config(BaseModel):
         default_factory=_default_user_playbook_extractor_config
     )
     # agent level success
-    agent_success_config: AgentSuccessConfig | None = None
+    agent_success_config: AgentSuccessConfig | None = Field(
+        default_factory=_default_agent_success_config
+    )
     # extraction preset — selects bundled window_size/stride_size values
     extraction_preset: ExtractionPreset | None = None
     # extraction parameters
@@ -767,6 +857,11 @@ class Config(BaseModel):
     playbook_optimizer_config: PlaybookOptimizerConfig = Field(
         default_factory=PlaybookOptimizerConfig
     )
+    # Tombstone GC job gate (opt-in, off by default — see LineageGCConfig)
+    lineage_gc: LineageGCConfig = Field(default_factory=LineageGCConfig)
+    governance_retention: GovernanceRetentionConfig = Field(
+        default_factory=GovernanceRetentionConfig
+    )
     # Optional non-blocking async information tools for classic extraction.
     pending_tool_call_config: PendingToolCallConfig = Field(
         default_factory=PendingToolCallConfig
@@ -824,6 +919,8 @@ class Config(BaseModel):
                 "stride_size",
                 "reflection_config",
                 "playbook_optimizer_config",
+                "lineage_gc",
+                "governance_retention",
                 "pending_tool_call_config",
                 "retrieval_floor",
             ):

package/plugin/vendor/reflexio/reflexio/models/structured_output.py

@@ -0,0 +1,148 @@
+"""Provider-safe base class for LLM structured-output Pydantic models.
+
+Every Pydantic model sent to an LLM as a structured-output schema (a
+``response_format`` or a tool's parameter schema) should inherit
+``StrictStructuredOutput``. It carries a ``__get_pydantic_json_schema__`` hook
+(via ``ProviderSafeUnionMixin``) that folds a discriminated union's ``oneOf`` into
+``anyOf`` and drops ``discriminator`` in the *emitted* JSON schema, so strict
+structured-output endpoints (OpenAI, minimax) accept it **by construction** — with
+no provider-detection gate and no post-processing. Only the wire schema is
+rewritten; the core (validation) schema keeps the discriminator, so keyed dispatch
+and precise per-variant errors are preserved.
+
+This module lives under ``models/`` (depending only on pydantic) so both
+``models/api_schema/`` schemas and ``server/services/*`` models can inherit the
+base without a layering inversion (a ``models/`` file importing from
+``server/llm/`` would invert the dependency direction).
+"""
+
+from typing import TYPE_CHECKING, Any
+
+from pydantic import BaseModel, GetJsonSchemaHandler
+from pydantic.json_schema import JsonSchemaValue
+from pydantic_core import CoreSchema
+
+# Keys whose *values* are name→subschema maps: their entries are user-chosen names
+# (field names, $def names), not JSON-Schema keywords. Traversals must recurse into
+# the subschemas but never treat the names themselves as keywords — otherwise a
+# model with a field literally named ``oneOf`` would be mishandled.
+_SCHEMA_NAME_MAP_KEYS = ("properties", "$defs", "definitions", "patternProperties")
+
+
+def _fold_oneof_to_anyof(node: Any) -> None:
+    """Rewrite ``oneOf`` to ``anyOf`` and drop ``discriminator`` in place (recursive).
+
+    Pydantic emits ``oneOf`` + ``discriminator`` for a discriminated union, which
+    strict structured-output endpoints (OpenAI, minimax) reject. Folding into
+    ``anyOf`` keeps the identical variant set so generation stays constrained;
+    Pydantic still enforces the discriminator after parse, so semantics are
+    preserved.
+
+    Structure-aware: under a name→subschema map (``properties``/``$defs``/…) it
+    recurses into the subschemas but does NOT treat a property/``$def`` *name* as a
+    keyword — so a model field literally named ``oneOf``/``discriminator`` is
+    preserved (only keyword-position occurrences are folded). Precondition:
+    ``node`` is a finite acyclic tree, as produced by ``model_json_schema()``
+    (which expresses recursion via ``$ref``/``$defs`` strings, never in-memory
+    cycles) — there is no cycle guard.
+
+    Args:
+        node (Any): A JSON-schema fragment (dict, list, or scalar); mutated in place.
+    """
+    if isinstance(node, dict):
+        one_of = node.pop("oneOf", None)
+        node.pop("discriminator", None)
+        if isinstance(one_of, list):
+            node["anyOf"] = node.get("anyOf", []) + one_of
+        for key, value in node.items():
+            if key in _SCHEMA_NAME_MAP_KEYS and isinstance(value, dict):
+                for sub in value.values():
+                    _fold_oneof_to_anyof(sub)
+            else:
+                _fold_oneof_to_anyof(value)
+    elif isinstance(node, list):
+        for item in node:
+            _fold_oneof_to_anyof(item)
+
+
+# Statically the mixin must appear to derive from ``BaseModel`` so the ``super()``
+# call below type-checks; at runtime it is a bare mixin (``object`` base), and
+# ``super()`` resolves to ``BaseModel.__get_pydantic_json_schema__`` via the MRO
+# of the concrete model (e.g. ``class X(ProviderSafeUnionMixin, BaseModel)``).
+_ProviderSafeUnionBase = BaseModel if TYPE_CHECKING else object
+
+
+class ProviderSafeUnionMixin(_ProviderSafeUnionBase):
+    """Make a model emit a provider-safe JSON schema by construction.
+
+    A model containing a Pydantic discriminated union serializes to JSON Schema
+    with ``oneOf`` + ``discriminator``, which strict structured-output endpoints
+    reject (the Sentry ``PYTHON-FASTAPI-9J`` incident). Mixing this in folds
+    ``oneOf`` into ``anyOf`` (and drops ``discriminator``) at the model boundary,
+    so every caller of ``model_json_schema()`` — litellm, instructor, our own
+    path — gets a provider-safe schema **unconditionally**, without depending on
+    a provider-detection gate (e.g. ``litellm.supports_response_schema``, which
+    under-reports some providers). Only the JSON (wire) schema is rewritten; the
+    core validation schema keeps the discriminator, so keyed dispatch and precise
+    per-variant errors are preserved.
+
+    Prefer inheriting ``StrictStructuredOutput`` (below) on output models rather
+    than mixing this in directly.
+    """
+
+    @classmethod
+    def __get_pydantic_json_schema__(
+        cls, core_schema: CoreSchema, handler: GetJsonSchemaHandler
+    ) -> JsonSchemaValue:
+        # Chain via super() (not handler() directly) so a cooperative base that
+        # also customizes __get_pydantic_json_schema__ is not silently shadowed;
+        # BaseModel's default impl just invokes handler(core_schema).
+        schema = super().__get_pydantic_json_schema__(core_schema, handler)
+        _fold_oneof_to_anyof(schema)
+        return schema
+
+
+class StrictStructuredOutput(ProviderSafeUnionMixin, BaseModel):
+    """Base for every Pydantic model sent to an LLM as a structured-output schema.
+
+    Inherits the provider-safe schema hook from ``ProviderSafeUnionMixin`` so the
+    emitted JSON schema folds any discriminated union (``oneOf``→``anyOf``, drop
+    ``discriminator``) by construction — accepted by strict structured-output
+    providers with no provider gate, including any *future* discriminated union.
+
+    Deliberately adds NO ``model_config``: each model keeps its own ``extra=`` and
+    ``json_schema_extra`` exactly as before. This base unifies the *provider-safety
+    guarantee*, not the per-model config (centralizing config is a separate change —
+    ``extra=`` is not uniform across models, so a blanket dedup would silently shift
+    validation/serialization behavior).
+    """
+
+
+def find_schema_keyword(node: Any, keyword: str) -> bool:
+    """Report whether ``keyword`` appears as a JSON-Schema *keyword* anywhere in ``node``.
+
+    Structure-aware: an occurrence of ``keyword`` as a property/``$defs`` *name*
+    (e.g. a model field literally named ``oneOf``) is NOT a match — only an
+    occurrence in JSON-Schema keyword position counts. Use this (not a blind
+    ``in``-walk) to check a schema for provider-unsafe keywords.
+
+    Args:
+        node (Any): A JSON-schema fragment (dict, list, or scalar).
+        keyword (str): The JSON-Schema keyword to search for (e.g. ``"oneOf"``).
+
+    Returns:
+        bool: True if ``keyword`` appears in keyword position, else False.
+    """
+    if isinstance(node, dict):
+        if keyword in node:
+            return True
+        for key, value in node.items():
+            if key in _SCHEMA_NAME_MAP_KEYS and isinstance(value, dict):
+                if any(find_schema_keyword(sub, keyword) for sub in value.values()):
+                    return True
+            elif find_schema_keyword(value, keyword):
+                return True
+        return False
+    if isinstance(node, list):
+        return any(find_schema_keyword(item, keyword) for item in node)
+    return False

package/plugin/vendor/reflexio/reflexio/server/OVERVIEW.md

@@ -23,7 +23,7 @@ Client (SDK/Web)
         -> GenerationService (server/services/)
           ├─> ProfileGenerationService -> ProfileExtractor(s) -> Storage
           ├─> PlaybookGenerationService -> PlaybookExtractor(s) -> Storage
-          └─> GroupEvaluationScheduler (deferred 10 min) -> Evaluator(s) -> Storage
+          └─> agent_success_evaluation/scheduler.py:GroupEvaluationScheduler (deferred 10 min) -> agent_success_evaluation/runner.py:run_group_evaluation -> agent_success_evaluation/service.py -> agent_success_evaluation/components/evaluator.py -> Storage
 ```
 
 ## Prerequisites
@@ -41,7 +41,7 @@ cp .env.example .env                         # Configure environment (set at lea
 uv sync                                      # Install Python dependencies (includes workspace packages)
 npm --prefix src/website install         # Install frontend dependencies
 npm --prefix src/public_docs install     # Install docs dependencies
-./run_services.sh                             # Starts API (8081), Website (8080), Docs (8082)
+./run_services.sh                             # Starts API (8061), Website (8080), Docs (8062)
 ./stop_services.sh                            # Stop all services
 ```
 
@@ -56,7 +56,7 @@ npm --prefix src/public_docs install     # Install docs dependencies
 **Testing:**
 ```python
 import reflexio
-client = reflexio.ReflexioClient(api_key="your-api-key", url_endpoint="http://127.0.0.1:8081/")
+client = reflexio.ReflexioClient(api_key="your-api-key", url_endpoint="http://127.0.0.1:8061/")
 ```
 See `notebooks/reflexio_cookbook.ipynb` and `src/tests/readme.md`