claude-smart 0.2.44 → 0.2.46

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

@@ -12,6 +12,7 @@ from typing import Literal, Self
 from pydantic import BaseModel, ConfigDict, Field, model_validator
 
 from reflexio.models.api_schema.validators import NonEmptyStr
+from reflexio.models.structured_output import StrictStructuredOutput
 
 HeroStateLiteral = Literal["full", "early", "shadow_off", "empty"]
 BucketLiteral = Literal["day", "week"]
@@ -98,6 +99,24 @@ class ScoreDistribution(BaseModel):
     labels: list[str]
 
 
+class EvaluationSourceSetRequest(BaseModel):
+    """One labeled request-source cohort for evaluation comparison.
+
+    ``sources`` match ``Request.source`` exactly. The empty string is valid and
+    represents requests published without a source.
+    """
+
+    label: NonEmptyStr
+    sources: list[str] = Field(min_length=1)
+
+    @model_validator(mode="after")
+    def validate_sources_are_unique(self) -> Self:
+        """Reject duplicate source values within a single set."""
+        if len(set(self.sources)) != len(self.sources):
+            raise ValueError("source set sources must be unique")
+        return self
+
+
 class GetEvaluationOverviewRequest(BaseModel):
     """Input for the overview endpoint.
 
@@ -107,18 +126,34 @@ class GetEvaluationOverviewRequest(BaseModel):
         bucket (BucketLiteral): Granularity of the hero trend buckets.
         include_shadow (bool): When False, skip the shadow-side aggregations
             (cheaper) — the hero will degrade to shadow_off state.
+        source_sets (list[EvaluationSourceSetRequest]): Optional labeled
+            request-source cohorts to compare. Sources are matched exactly
+            against the first request in each evaluated session.
     """
 
     from_ts: int = Field(ge=0)
     to_ts: int = Field(ge=0)
     bucket: BucketLiteral = "week"
     include_shadow: bool = True
+    source_sets: list[EvaluationSourceSetRequest] = Field(default_factory=list)
 
     @model_validator(mode="after")
-    def validate_time_window(self) -> Self:
-        """Ensure the requested time window is ordered."""
+    def validate_request(self) -> Self:
+        """Ensure the requested time window and source sets are valid."""
         if self.from_ts > self.to_ts:
             raise ValueError("from_ts must be <= to_ts")
+        labels = [s.label for s in self.source_sets]
+        if len(set(labels)) != len(labels):
+            raise ValueError("source set labels must be unique")
+        seen_sources: set[str] = set()
+        for source_set in self.source_sets:
+            overlap = seen_sources.intersection(source_set.sources)
+            if overlap:
+                overlap_list = ", ".join(sorted(repr(s) for s in overlap))
+                raise ValueError(
+                    f"source values cannot appear in multiple source sets: {overlap_list}"
+                )
+            seen_sources.update(source_set.sources)
         return self
 
 
@@ -139,41 +174,6 @@ class BraintrustTileRow(BaseModel):
     delta: float
 
 
-class TrendPoint(BaseModel):
-    """One point on a grouped success-rate trend curve (F2).
-
-    Args:
-        ts (int): Unix epoch seconds — bucket start.
-        rate (float): Success rate for sessions in this bucket, [0.0, 1.0].
-        n (int): Session count backing the rate. Must be non-negative.
-    """
-
-    ts: int
-    rate: float
-    n: int = Field(ge=0)
-
-
-class SuccessRateTrendByGroup(BaseModel):
-    """Group-split trend data for the dashboard's dual-curve chart (F2).
-
-    Grouping is by ``Request.metadata.reflexio_retrieval_enabled``, read from
-    the first request of each session in the window. Sessions whose first
-    request has the key absent OR a non-bool value land in ``untagged``.
-
-    Args:
-        treatment (list[TrendPoint]): Curve for sessions where the first
-            request had ``metadata.reflexio_retrieval_enabled = True``.
-        control (list[TrendPoint]): Curve for ``... = False``.
-        untagged (list[TrendPoint]): Curve for sessions where the key is
-            absent or non-bool — surfaced (not silently coerced) so
-            customers can see how many of their sessions are untagged.
-    """
-
-    treatment: list[TrendPoint] = Field(default_factory=list)
-    control: list[TrendPoint] = Field(default_factory=list)
-    untagged: list[TrendPoint] = Field(default_factory=list)
-
-
 class ShadowWinRateTrendPoint(BaseModel):
     """One daily bucket of per-turn shadow-comparison verdicts (F1).
 
@@ -243,18 +243,41 @@ class ShadowWinRateTrend(BaseModel):
     judge_prompt_version: str = Field(default="v1.0.0")
 
 
+class SourceSetEvaluationMetrics(BaseModel):
+    """Metrics for one labeled request-source set."""
+
+    label: str
+    sources: list[str]
+    session_count: int = Field(ge=0)
+    session_ids: list[str] = Field(default_factory=list)
+    success_rate_pp: float
+    buckets: list[HeroBucket]
+    context_tiles: ContextTile
+    score_distribution: ScoreDistribution
+    rule_attribution: list[RuleAttributionRow]
+    braintrust_tiles: list[BraintrustTileRow] = Field(default_factory=list)
+
+
+class SourceSetComparison(BaseModel):
+    """Comparison payload for request-source cohorts."""
+
+    available_sources: list[str] = Field(default_factory=list)
+    sets: list[SourceSetEvaluationMetrics] = Field(default_factory=list)
+    unmatched_session_count: int = Field(default=0, ge=0)
+
+
 class GetEvaluationOverviewResponse(BaseModel):
     hero: HeroBlock
     context_tiles: ContextTile
     rule_attribution: list[RuleAttributionRow]
     score_distribution: ScoreDistribution
     braintrust_tiles: list[BraintrustTileRow] = Field(default_factory=list)
-    success_rate_trend_by_group: SuccessRateTrendByGroup = Field(
-        default_factory=SuccessRateTrendByGroup
-    )
     shadow_win_rate_trend: ShadowWinRateTrend = Field(
         default_factory=ShadowWinRateTrend
     )
+    source_set_comparison: SourceSetComparison = Field(
+        default_factory=SourceSetComparison
+    )
 
 
 # ---------------------------------------------------------------------------
@@ -406,7 +429,7 @@ class GradeOnDemandResponse(BaseModel):
 # ---------------------------------------------------------------------------
 
 
-class ShadowComparisonOutput(BaseModel):
+class ShadowComparisonOutput(StrictStructuredOutput):
     """LLM judge verdict for a per-turn Reflexio-vs-Shadow comparison (F1).
 
     Args:

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

@@ -26,3 +26,22 @@ class SessionDescriptor(NamedTuple):
     session_id: str
     agent_version: str
     source: str
+
+
+class SessionFirstRequest(NamedTuple):
+    """Earliest request metadata for one session."""
+
+    session_id: str
+    user_id: str
+    source: str
+    created_at: int
+
+
+class SessionCitation(NamedTuple):
+    """One cited rule/profile occurrence, keyed by user/session."""
+
+    user_id: str
+    session_id: str
+    kind: str
+    real_id: str
+    title: str

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

@@ -3,14 +3,21 @@
 from __future__ import annotations
 
 from datetime import datetime
-from typing import Any
+from enum import StrEnum
+from typing import TYPE_CHECKING, Any
 
 from pydantic import BaseModel, ConfigDict, Field
 
-from reflexio.server.services.storage.storage_base import (
-    PendingToolCallRecord,
-    PendingToolCallStatus,
-)
+if TYPE_CHECKING:
+    from reflexio.server.services.storage.storage_base import PendingToolCallRecord
+
+
+class PendingToolCallStatus(StrEnum):
+    PENDING = "pending"
+    RESOLVED = "resolved"
+    EXPIRED = "expired"
+    SUPERSEDED = "superseded"
+    CANCELLED = "cancelled"
 
 
 class PendingToolCallResponse(BaseModel):

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
@@ -185,6 +191,7 @@ class GetUserProfilesRequest(BaseModel):
     end_time: datetime | None = None
     top_k: int | None = Field(default=30, gt=0)
     status_filter: list[Status | None] | None = None
+    tags: list[str] | None = None
 
     @model_validator(mode="after")
     def check_time_range(self) -> Self:
@@ -219,6 +226,7 @@ class GetUserPlaybooksRequest(BaseModel):
     playbook_name: str | None = None
     agent_version: str | None = None
     status_filter: list[Status | None] | None = None
+    tags: list[str] | None = None
 
 
 class GetUserPlaybooksResponse(BaseModel):
@@ -233,6 +241,11 @@ class GetAgentPlaybooksRequest(BaseModel):
     agent_version: str | 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
 
 
 class GetAgentPlaybooksResponse(BaseModel):
@@ -241,6 +254,30 @@ 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 +300,15 @@ 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 +358,15 @@ 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:
@@ -621,6 +668,11 @@ 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 +664,50 @@ 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)
+
+
 @dataclass(frozen=True)
 class EffectivePendingToolCallConfig:
     """Resolved pending-tool-call settings after applying tool overrides."""
@@ -733,6 +807,8 @@ 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 +825,9 @@ 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 +845,8 @@ 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)
     # Optional non-blocking async information tools for classic extraction.
     pending_tool_call_config: PendingToolCallConfig = Field(
         default_factory=PendingToolCallConfig
@@ -824,6 +904,7 @@ class Config(BaseModel):
                 "stride_size",
                 "reflection_config",
                 "playbook_optimizer_config",
+                "lineage_gc",
                 "pending_tool_call_config",
                 "retrieval_floor",
             ):