feat: enhance ai quality effect review

2026-06-23 11:31:45 +08:00
parent f1fcf90797
commit 396259ebe9
7 changed files with 341 additions and 25 deletions
@@ -1759,6 +1759,20 @@ components:
            current_quality_score:
              type: number
              format: float
+            post_change_total_runs:
+              type: integer
+            post_change_affected_runs:
+              type: integer
+            post_change_pending_labels:
+              type: integer
+            post_change_quality_score:
+              type: number
+              format: float
+            post_change_affected_rate:
+              type: number
+              format: float
+            post_change_observation_days:
+              type: integer
            affected_rate_delta:
              type: number
              format: float
@@ -1767,6 +1781,16 @@ components:
            quality_score_delta:
              type: number
              format: float
+            post_change_affected_rate_delta:
+              type: number
+              format: float
+            post_change_pending_labels_delta:
+              type: integer
+            post_change_quality_score_delta:
+              type: number
+              format: float
+            related_change_actions_count:
+              type: integer
            effect_status:
              type: string
              enum: [pending_data, improved, unchanged, regressed]
@@ -9,7 +9,7 @@ from typing import Iterable, Protocol
 from uuid import uuid4

 from pydantic import BaseModel, Field
-from sqlalchemy import create_engine, delete, select, text
+from sqlalchemy import create_engine, delete, or_, select, text
 from sqlalchemy.engine import Engine
 from sqlalchemy.orm import Session, sessionmaker
 from sqlalchemy.pool import StaticPool
@@ -365,9 +365,19 @@ class AiQualityImprovementAction(AiQualityImprovementActionCreate):
    current_affected_runs: int = 0
    current_pending_labels: int = 0
    current_quality_score: float = 0.0
+    post_change_total_runs: int = 0
+    post_change_affected_runs: int = 0
+    post_change_pending_labels: int = 0
+    post_change_quality_score: float = 0.0
+    post_change_affected_rate: float = 0.0
+    post_change_observation_days: int = 0
    affected_rate_delta: float = 0.0
    pending_labels_delta: int = 0
    quality_score_delta: float = 0.0
+    post_change_affected_rate_delta: float = 0.0
+    post_change_pending_labels_delta: int = 0
+    post_change_quality_score_delta: float = 0.0
+    related_change_actions_count: int = 0
    effect_status: str = "pending_data"
    effect_summary: str | None = None
    resolution_note: str | None = None
@@ -896,7 +906,11 @@ class InMemoryAiPlatformRepository:
                **payload.model_dump(exclude={"title"}),
            )
            self.quality_improvement_actions[action.id] = action
-            return _materialize_improvement_action(action, self.ai_run_audits)
+            return _materialize_improvement_action(
+                action,
+                self.ai_run_audits,
+                all_actions=self.quality_improvement_actions.values(),
+            )

    def update_quality_improvement_action(
        self,
@@ -939,7 +953,11 @@ class InMemoryAiPlatformRepository:
                update["linked_import_batch_id"] = payload.linked_import_batch_id
            updated = action.model_copy(update=update)
            self.quality_improvement_actions[action_id] = updated
-            return _materialize_improvement_action(updated, self.ai_run_audits)
+            return _materialize_improvement_action(
+                updated,
+                self.ai_run_audits,
+                all_actions=self.quality_improvement_actions.values(),
+            )

    def update_ai_run_quality_label(
        self,
@@ -1557,9 +1575,21 @@ class SqlAlchemyAiPlatformRepository:
                .order_by(AiQualityImprovementActionModel.updated_at.desc())
                .limit(20)
            ).all()
+            all_action_rows = session.scalars(
+                select(AiQualityImprovementActionModel).order_by(
+                    AiQualityImprovementActionModel.updated_at.desc()
+                )
+            ).all()
            action_days = [row.evaluation_days for row in action_rows]
+            resolved_since = [
+                _ensure_utc(row.resolved_at)
+                for row in action_rows
+                if row.resolved_at is not None
+            ]
            stats_cutoff = datetime.now(UTC) - timedelta(days=max([days, *action_days], default=days) - 1)
            stats_cutoff = stats_cutoff.replace(hour=0, minute=0, second=0, microsecond=0)
+            if resolved_since:
+                stats_cutoff = min(stats_cutoff, min(resolved_since))
            rows = session.scalars(
                select(AiRunAuditModel)
                .where(AiRunAuditModel.created_at >= stats_cutoff)
@@ -1572,6 +1602,7 @@ class SqlAlchemyAiPlatformRepository:
            actions = _materialize_improvement_actions(
                [_quality_improvement_action_from_model(row) for row in action_rows],
                all_runs,
+                all_actions=[_quality_improvement_action_from_model(row) for row in all_action_rows],
            )
        return _build_quality_dashboard(
            dashboard_runs,
@@ -1608,15 +1639,57 @@ class SqlAlchemyAiPlatformRepository:
            max_days = max((row.evaluation_days for row in rows), default=14)
            cutoff = datetime.now(UTC) - timedelta(days=max(max_days - 1, 0))
            cutoff = cutoff.replace(hour=0, minute=0, second=0, microsecond=0)
+            resolved_since = [
+                _ensure_utc(row.resolved_at)
+                for row in rows
+                if row.resolved_at is not None
+            ]
+            if resolved_since:
+                cutoff = min(cutoff, min(resolved_since))
            audit_rows = session.scalars(
                select(AiRunAuditModel)
                .where(AiRunAuditModel.created_at >= cutoff)
                .order_by(AiRunAuditModel.created_at.desc())
            ).all()
            runs = [_ai_run_audit_from_model(row) for row in audit_rows]
+            related_statement = select(AiQualityImprovementActionModel)
+            if rows:
+                prompt_keys = [
+                    (row.linked_prompt_template_id, row.linked_prompt_template_version)
+                    for row in rows
+                    if row.linked_prompt_template_id
+                ]
+                import_batch_ids = [row.linked_import_batch_id for row in rows if row.linked_import_batch_id]
+                knowledge_base_ids = [
+                    row.linked_knowledge_base_id
+                    for row in rows
+                    if row.linked_knowledge_base_id and not row.linked_import_batch_id
+                ]
+                related_filters = []
+                for prompt_id, prompt_version in prompt_keys:
+                    related_filters.append(
+                        (AiQualityImprovementActionModel.linked_prompt_template_id == prompt_id)
+                        & (AiQualityImprovementActionModel.linked_prompt_template_version == prompt_version)
+                    )
+                if import_batch_ids:
+                    related_filters.append(
+                        AiQualityImprovementActionModel.linked_import_batch_id.in_(import_batch_ids)
+                    )
+                if knowledge_base_ids:
+                    related_filters.append(
+                        AiQualityImprovementActionModel.linked_knowledge_base_id.in_(knowledge_base_ids)
+                    )
+                if related_filters:
+                    related_statement = related_statement.where(or_(*related_filters))
+            related_rows = session.scalars(related_statement).all() if rows else []
+            related_actions = [_quality_improvement_action_from_model(row) for row in related_rows]
+            row_actions = [_quality_improvement_action_from_model(row) for row in rows]
+            related_by_id = {action.id: action for action in related_actions}
+            related_by_id.update({action.id: action for action in row_actions})
            actions = _materialize_improvement_actions(
-                [_quality_improvement_action_from_model(row) for row in rows],
+                row_actions,
                runs,
+                all_actions=related_by_id.values(),
            )
        return _filter_improvement_actions(
            actions,
@@ -1682,7 +1755,12 @@ class SqlAlchemyAiPlatformRepository:
            )
            session.add(row)
            session.flush()
-            return _materialize_improvement_action(_quality_improvement_action_from_model(row), runs)
+            action_rows = session.scalars(select(AiQualityImprovementActionModel)).all()
+            return _materialize_improvement_action(
+                _quality_improvement_action_from_model(row),
+                runs,
+                all_actions=[_quality_improvement_action_from_model(action_row) for action_row in action_rows],
+            )

    def update_quality_improvement_action(
        self,
@@ -1724,6 +1802,8 @@ class SqlAlchemyAiPlatformRepository:

            cutoff = now - timedelta(days=max(row.evaluation_days - 1, 0))
            cutoff = cutoff.replace(hour=0, minute=0, second=0, microsecond=0)
+            if row.resolved_at is not None:
+                cutoff = min(cutoff, _ensure_utc(row.resolved_at))
            audit_rows = session.scalars(
                select(AiRunAuditModel)
                .where(AiRunAuditModel.created_at >= cutoff)
@@ -1731,7 +1811,12 @@ class SqlAlchemyAiPlatformRepository:
            ).all()
            runs = [_ai_run_audit_from_model(row) for row in audit_rows]
            session.flush()
-            return _materialize_improvement_action(_quality_improvement_action_from_model(row), runs)
+            action_rows = session.scalars(select(AiQualityImprovementActionModel)).all()
+            return _materialize_improvement_action(
+                _quality_improvement_action_from_model(row),
+                runs,
+                all_actions=[_quality_improvement_action_from_model(action_row) for action_row in action_rows],
+            )

    def update_ai_run_quality_label(
        self,
@@ -2413,20 +2498,38 @@ def _finalize_attribution_groups(groups: dict[str, dict[str, object]]) -> list[A
 def _materialize_improvement_actions(
    actions: Iterable[AiQualityImprovementAction],
    runs: list[AiRunAudit],
+    *,
+    all_actions: Iterable[AiQualityImprovementAction] | None = None,
 ) -> list[AiQualityImprovementAction]:
-    return [_materialize_improvement_action(action, runs) for action in actions]
+    action_list = list(actions)
+    related_actions = list(all_actions) if all_actions is not None else action_list
+    return [
+        _materialize_improvement_action(action, runs, all_actions=related_actions)
+        for action in action_list
+    ]


 def _materialize_improvement_action(
    action: AiQualityImprovementAction,
    runs: list[AiRunAudit],
+    *,
+    all_actions: Iterable[AiQualityImprovementAction] | None = None,
 ) -> AiQualityImprovementAction:
    current = _calculate_improvement_action_stats(action, runs)
+    post_change = (
+        _calculate_improvement_action_stats(action, runs, since=action.resolved_at)
+        if action.resolved_at
+        else {"total_runs": 0, "affected_runs": 0, "pending_labels": 0, "quality_score": 0.0}
+    )
    baseline_rate = _safe_rate(action.baseline_affected_runs, action.baseline_total_runs)
    current_rate = _safe_rate(current["affected_runs"], current["total_runs"])
    affected_rate_delta = round(current_rate - baseline_rate, 4)
    pending_labels_delta = current["pending_labels"] - action.baseline_pending_labels
    quality_score_delta = round(current["quality_score"] - action.baseline_quality_score, 2)
+    post_change_rate = _safe_rate(post_change["affected_runs"], post_change["total_runs"])
+    post_change_affected_rate_delta = round(post_change_rate - baseline_rate, 4)
+    post_change_pending_labels_delta = post_change["pending_labels"] - action.baseline_pending_labels
+    post_change_quality_score_delta = round(post_change["quality_score"] - action.baseline_quality_score, 2)
    effect_status, effect_summary = _summarize_improvement_effect(
        action,
        current_total_runs=current["total_runs"],
@@ -2434,6 +2537,11 @@ def _materialize_improvement_action(
        affected_rate_delta=affected_rate_delta,
        pending_labels_delta=pending_labels_delta,
        quality_score_delta=quality_score_delta,
+        post_change_total_runs=post_change["total_runs"],
+        post_change_rate=post_change_rate,
+        post_change_affected_rate_delta=post_change_affected_rate_delta,
+        post_change_pending_labels_delta=post_change_pending_labels_delta,
+        post_change_quality_score_delta=post_change_quality_score_delta,
    )
    return action.model_copy(
        update={
@@ -2441,9 +2549,19 @@ def _materialize_improvement_action(
            "current_affected_runs": current["affected_runs"],
            "current_pending_labels": current["pending_labels"],
            "current_quality_score": current["quality_score"],
+            "post_change_total_runs": post_change["total_runs"],
+            "post_change_affected_runs": post_change["affected_runs"],
+            "post_change_pending_labels": post_change["pending_labels"],
+            "post_change_quality_score": post_change["quality_score"],
+            "post_change_affected_rate": post_change_rate,
+            "post_change_observation_days": _post_change_observation_days(action),
            "affected_rate_delta": affected_rate_delta,
            "pending_labels_delta": pending_labels_delta,
            "quality_score_delta": quality_score_delta,
+            "post_change_affected_rate_delta": post_change_affected_rate_delta,
+            "post_change_pending_labels_delta": post_change_pending_labels_delta,
+            "post_change_quality_score_delta": post_change_quality_score_delta,
+            "related_change_actions_count": _count_related_change_actions(action, all_actions or ()),
            "effect_status": effect_status,
            "effect_summary": effect_summary,
        }
@@ -2484,9 +2602,12 @@ def _filter_improvement_actions(
 def _calculate_improvement_action_stats(
    action: AiQualityImprovementActionCreate | AiQualityImprovementAction,
    runs: list[AiRunAudit],
+    *,
+    since: datetime | None = None,
 ) -> dict[str, int | float]:
-    cutoff = datetime.now(UTC) - timedelta(days=max(action.evaluation_days - 1, 0))
-    cutoff = cutoff.replace(hour=0, minute=0, second=0, microsecond=0)
+    cutoff = _ensure_utc(since) if since else datetime.now(UTC) - timedelta(days=max(action.evaluation_days - 1, 0))
+    if not since:
+        cutoff = cutoff.replace(hour=0, minute=0, second=0, microsecond=0)
    matched_runs: list[AiRunAudit] = []
    affected_runs: set[str] = set()
    pending_labels = 0
@@ -2550,6 +2671,37 @@ def _has_source_issue(run: AiRunAudit) -> bool:
    return any(label.code in {"source_insufficient", "source_limited"} for label in run.quality_labels)


+def _post_change_observation_days(action: AiQualityImprovementAction) -> int:
+    if not action.resolved_at:
+        return 0
+    elapsed = datetime.now(UTC).date() - _ensure_utc(action.resolved_at).date()
+    return max(elapsed.days + 1, 1)
+
+
+def _count_related_change_actions(
+    action: AiQualityImprovementAction,
+    actions: Iterable[AiQualityImprovementAction],
+) -> int:
+    change_key = _quality_action_change_key(action)
+    if change_key is None:
+        return 0
+    return sum(1 for candidate in actions if _quality_action_change_key(candidate) == change_key)
+
+
+def _quality_action_change_key(action: AiQualityImprovementAction) -> tuple[str, str, str | None] | None:
+    if action.linked_prompt_template_id:
+        return (
+            "prompt",
+            action.linked_prompt_template_id,
+            action.linked_prompt_template_version,
+        )
+    if action.linked_import_batch_id:
+        return ("knowledge_import", action.linked_import_batch_id, None)
+    if action.linked_knowledge_base_id:
+        return ("knowledge_base", action.linked_knowledge_base_id, None)
+    return None
+
+
 def _safe_rate(numerator: int | float, denominator: int | float) -> float:
    return round(float(numerator) / float(denominator), 4) if denominator else 0.0

@@ -2562,26 +2714,50 @@ def _summarize_improvement_effect(
    affected_rate_delta: float,
    pending_labels_delta: int | float,
    quality_score_delta: float,
+    post_change_total_runs: int | float,
+    post_change_rate: float,
+    post_change_affected_rate_delta: float,
+    post_change_pending_labels_delta: int | float,
+    post_change_quality_score_delta: float,
 ) -> tuple[str, str]:
    baseline_rate = _safe_rate(action.baseline_affected_runs, action.baseline_total_runs)
-    if action.baseline_total_runs <= 0 or current_total_runs <= 0:
+    resolved = action.status in {"resolved", "false_positive", "closed"}
+    if action.baseline_total_runs <= 0:
        return "pending_data", "Need more AI run data to evaluate this improvement action."
+    if resolved and post_change_total_runs <= 0:
+        return (
+            "pending_data",
+            "Need post-change AI run data after this action was resolved to evaluate effect.",
+        )
+    if not resolved and current_total_runs <= 0:
+        return "pending_data", "Need more AI run data to evaluate this improvement action."
+    effective_rate_delta = post_change_affected_rate_delta if resolved else affected_rate_delta
+    effective_pending_delta = post_change_pending_labels_delta if resolved else pending_labels_delta
+    effective_quality_delta = post_change_quality_score_delta if resolved else quality_score_delta
    improved = (
-        affected_rate_delta < -0.01
-        or pending_labels_delta < 0
-        or quality_score_delta > 1
+        effective_rate_delta < -0.01
+        or effective_pending_delta < 0
+        or effective_quality_delta > 1
    )
    regressed = (
-        affected_rate_delta > 0.01
-        or pending_labels_delta > 0
-        or quality_score_delta < -1
+        effective_rate_delta > 0.01
+        or effective_pending_delta > 0
+        or effective_quality_delta < -1
    )
    status = "improved" if improved and not regressed else "regressed" if regressed and not improved else "unchanged"
-    summary = (
-        f"Baseline affected rate {baseline_rate:.0%}, current {current_rate:.0%}; "
-        f"pending labels delta {pending_labels_delta:+.0f}; "
-        f"quality score delta {quality_score_delta:+.1f}."
-    )
+    if resolved:
+        summary = (
+            f"Baseline affected rate {baseline_rate:.0%}, post-change {post_change_rate:.0%} "
+            f"over {post_change_total_runs:.0f} run(s); "
+            f"pending labels delta {post_change_pending_labels_delta:+.0f}; "
+            f"quality score delta {post_change_quality_score_delta:+.1f}."
+        )
+    else:
+        summary = (
+            f"Baseline affected rate {baseline_rate:.0%}, current {current_rate:.0%}; "
+            f"pending labels delta {pending_labels_delta:+.0f}; "
+            f"quality score delta {quality_score_delta:+.1f}."
+        )
    return status, summary


@@ -1,3 +1,5 @@
+from datetime import UTC, datetime
+
 import pytest
 from fastapi.testclient import TestClient

@@ -450,6 +452,63 @@ def test_quality_improvement_action_resolve_api_returns_422_without_real_change(
    assert "linked_knowledge_base_id" in resolve_response.json()["detail"]


+def test_quality_improvement_action_effect_review_uses_post_resolution_window() -> None:
+    repository = InMemoryAiPlatformRepository()
+    repository.seed_defaults()
+    repository.add_ai_run_audit(
+        AiRunAudit(
+            id="run-baseline",
+            operation="legal_qa",
+            provider="deepseek",
+            model="deepseek-chat",
+            prompt_template_id="legal_qa.default",
+            retrieved_sources=[],
+            quality_score=88,
+            quality_status="needs_review",
+            quality_labels=[],
+            created_at=datetime.now(UTC),
+        )
+    )
+    action = repository.create_quality_improvement_action(
+        AiQualityImprovementActionCreate(
+            dimension="prompt_template",
+            key="legal_qa.default",
+            label="legal_qa.default",
+            action_type="tune_prompt",
+            evaluation_days=7,
+        )
+    )
+    assert action.current_total_runs >= 1
+    resolved = repository.update_quality_improvement_action(
+        action.id,
+        AiQualityImprovementActionUpdate(
+            status="resolved",
+            linked_prompt_template_id=action.linked_prompt_template_id,
+            linked_prompt_template_version=action.linked_prompt_template_version,
+            resolution_note="Prompt improved.",
+        ),
+    )
+    repository.add_ai_run_audit(
+        AiRunAudit(
+            id="run-post-change",
+            operation="legal_qa",
+            provider="deepseek",
+            model="deepseek-chat",
+            prompt_template_id="legal_qa.default",
+            retrieved_sources=[],
+            quality_score=96,
+            quality_status="passed",
+            quality_labels=[],
+            created_at=resolved.resolved_at or datetime.now(UTC),
+        )
+    )
+    refreshed = repository.list_quality_improvement_actions(limit=5)[0]
+    assert refreshed.post_change_observation_days >= 1
+    assert refreshed.post_change_total_runs >= 1
+    assert refreshed.related_change_actions_count >= 1
+    assert "post-change" in (refreshed.effect_summary or "")
+
+
 def test_knowledge_document_can_be_added_and_searched() -> None:
    client = make_client()

@@ -375,9 +375,19 @@ class AiQualityImprovementAction(AiQualityImprovementActionCreate):
    current_affected_runs: int = 0
    current_pending_labels: int = 0
    current_quality_score: float = 0.0
+    post_change_total_runs: int = 0
+    post_change_affected_runs: int = 0
+    post_change_pending_labels: int = 0
+    post_change_quality_score: float = 0.0
+    post_change_affected_rate: float = 0.0
+    post_change_observation_days: int = 0
    affected_rate_delta: float = 0.0
    pending_labels_delta: int = 0
    quality_score_delta: float = 0.0
+    post_change_affected_rate_delta: float = 0.0
+    post_change_pending_labels_delta: int = 0
+    post_change_quality_score_delta: float = 0.0
+    related_change_actions_count: int = 0
    effect_status: str = "pending_data"
    effect_summary: str | None = None
    resolution_note: str | None = None
@@ -128,11 +128,21 @@ def test_ai_platform_client_covers_admin_data_endpoints() -> None:
            "current_affected_runs": 1,
            "current_pending_labels": 1,
            "current_quality_score": 90.0,
+            "post_change_total_runs": 2,
+            "post_change_affected_runs": 0,
+            "post_change_pending_labels": 0,
+            "post_change_quality_score": 98.0,
+            "post_change_affected_rate": 0.0,
+            "post_change_observation_days": 3,
            "affected_rate_delta": 0.0,
            "pending_labels_delta": 0,
            "quality_score_delta": 0.0,
+            "post_change_affected_rate_delta": -1.0,
+            "post_change_pending_labels_delta": -1,
+            "post_change_quality_score_delta": 8.0,
+            "related_change_actions_count": 2,
            "effect_status": "unchanged",
-            "effect_summary": "Baseline affected rate 100%, current 100%; pending labels delta +0; quality score delta +0.0.",
+            "effect_summary": "Baseline affected rate 100%, post-change 0% over 2 run(s); pending labels delta -1; quality score delta +8.0.",
            "resolution_note": resolution_note,
            "created_at": "2026-06-22T00:00:00Z",
            "updated_at": "2026-06-22T00:02:00Z",
@@ -817,6 +827,8 @@ def test_ai_platform_client_covers_admin_data_endpoints() -> None:
    assert created_action.linked_import_batch_id == "kb-batch-1"
    assert created_action.baseline_affected_runs == 1
    assert created_action.effect_status == "unchanged"
+    assert created_action.post_change_total_runs == 2
+    assert created_action.related_change_actions_count == 2
    listed_actions = client.list_quality_improvement_actions(
        status="open",
        owner="admin",
@@ -837,6 +849,8 @@ def test_ai_platform_client_covers_admin_data_endpoints() -> None:
    )
    assert updated_action.status == "resolved"
    assert updated_action.resolved_at == "2026-06-22T00:05:00Z"
+    assert updated_action.post_change_total_runs == 2
+    assert updated_action.effect_summary and "post-change" in updated_action.effect_summary
    updated_audit = client.update_ai_run_quality_label(
        "run-1",
        "citation_present",
@@ -100,9 +100,19 @@ type QualityImprovementAction = {
  current_affected_runs: number;
  current_pending_labels: number;
  current_quality_score: number;
+  post_change_total_runs: number;
+  post_change_affected_runs: number;
+  post_change_pending_labels: number;
+  post_change_quality_score: number;
+  post_change_affected_rate: number;
+  post_change_observation_days: number;
  affected_rate_delta: number;
  pending_labels_delta: number;
  quality_score_delta: number;
+  post_change_affected_rate_delta: number;
+  post_change_pending_labels_delta: number;
+  post_change_quality_score_delta: number;
+  related_change_actions_count: number;
  effect_status?: string;
  effect_summary?: string | null;
  resolution_note?: string | null;
@@ -774,6 +784,7 @@ function QualityActionItem({
  const needsPromptDraft = action.action_type === "tune_prompt";
  const needsKnowledgeImport = action.action_type === "improve_knowledge";
  const completionBlockReason = getCompletionBlockReason(action);
+  const hasPostChange = action.post_change_total_runs > 0;

  return (
    <div className="quality-action-item">
@@ -796,13 +807,29 @@ function QualityActionItem({
        <span className="item-meta">{action.description || action.root_cause_hint}</span>
      )}
      <div className="quality-action-stats">
-        <span>异常率 {formatPercent(baselineRate)} → {formatPercent(currentRate)}</span>
-        <span>待处理 {action.baseline_pending_labels} → {action.current_pending_labels}</span>
-        <span>质量分 {action.baseline_quality_score} → {action.current_quality_score}</span>
+        <span>改前异常率 {formatPercent(baselineRate)}</span>
+        <span>当前异常率 {formatPercent(currentRate)}</span>
+        <span>完成后异常率 {hasPostChange ? formatPercent(action.post_change_affected_rate) : "待观察"}</span>
+        <span>观察 {action.post_change_observation_days || action.evaluation_days} 天</span>
+        <span>改前待处理 {action.baseline_pending_labels}</span>
+        <span>完成后待处理 {hasPostChange ? action.post_change_pending_labels : "待观察"}</span>
+        <span>改前质量分 {action.baseline_quality_score}</span>
+        <span>完成后质量分 {hasPostChange ? action.post_change_quality_score : "待观察"}</span>
        <span>变化 {formatSignedPercent(action.affected_rate_delta)}</span>
      </div>
      {action.resolution_note && <span className="item-meta">处理说明：{action.resolution_note}</span>}
      <LinkedChangeSummary action={action} />
+      <div className="quality-action-closure">
+        <span className="item-meta">关联改进项 {action.related_change_actions_count}</span>
+        {hasPostChange && (
+          <span className="item-meta">
+            完成后变化：异常率 {formatSignedPercent(action.post_change_affected_rate_delta)} / 待处理{" "}
+            {action.post_change_pending_labels_delta >= 0
+              ? `+${action.post_change_pending_labels_delta}`
+              : action.post_change_pending_labels_delta} / 质量分 {formatSignedPercent(action.post_change_quality_score_delta)}
+          </span>
+        )}
+      </div>
      {action.effect_summary && <span className="item-meta">效果复盘：{action.effect_summary}</span>}
      {["open", "in_progress"].includes(action.status) && (needsPromptDraft || needsKnowledgeImport) && (
        <div className="quality-action-execution">
@@ -812,6 +812,12 @@ select {
  background: #f1f5f9;
 }

+.quality-action-closure {
+  display: grid;
+  gap: 6px;
+  min-width: 0;
+}
+
 .quality-action-execution {
  display: grid;
  gap: 10px;