feat: expose review controls and feedback trends

Persist feedback-eval attention gaps into a reusable history file so calibration drift can be tracked over time alongside eval output. Expose the newer review pipeline controls in the web settings UI and align the model picker with the frontier defaults so the capabilities we added are actually configurable from the app.

Made-with: Cursor

Author: haasonsaas

src/commands/feedback_eval/command.rs

@@ -12,6 +12,7 @@ use report::emit_feedback_eval_report;
 pub async fn feedback_eval_command(
     input_path: PathBuf,
     output_path: Option<PathBuf>,
+    trend_path: Option<PathBuf>,
     confidence_threshold: f32,
     eval_report_path: Option<PathBuf>,
 ) -> Result<()> {
@@ -23,6 +24,7 @@ pub async fn feedback_eval_command(
     emit_feedback_eval_report(
         &loaded,
         output_path.as_deref(),
+        trend_path.as_deref(),
         confidence_threshold,
         eval_report.as_ref(),
     )

src/commands/feedback_eval/command/report.rs

@@ -4,13 +4,15 @@ use std::path::Path;
 use crate::commands::eval::EvalReport;
 
 use super::super::report::{
-    build_feedback_eval_report, print_feedback_eval_report, write_feedback_eval_report,
+    build_feedback_eval_report, print_feedback_eval_report, update_feedback_eval_trend,
+    write_feedback_eval_report,
 };
 use super::super::LoadedFeedbackEvalInput;
 
 pub(super) async fn emit_feedback_eval_report(
     loaded: &LoadedFeedbackEvalInput,
     output_path: Option<&Path>,
+    trend_path: Option<&Path>,
     confidence_threshold: f32,
     eval_report: Option<&EvalReport>,
 ) -> Result<()> {
@@ -21,6 +23,9 @@ pub(super) async fn emit_feedback_eval_report(
     if let Some(path) = output_path {
         write_feedback_eval_report(&report, path).await?;
     }
+    if let Some(path) = trend_path {
+        update_feedback_eval_trend(&report, eval_report, path).await?;
+    }
 
     Ok(())
 }

src/commands/feedback_eval/report.rs

@@ -4,6 +4,9 @@ mod build;
 mod examples;
 #[path = "report/output.rs"]
 mod output;
+#[path = "report/trend.rs"]
+mod trend;
 
 pub(super) use build::build_feedback_eval_report;
 pub(super) use output::{print_feedback_eval_report, write_feedback_eval_report};
+pub(super) use trend::update_feedback_eval_trend;

src/commands/feedback_eval/report/trend.rs

@@ -0,0 +1,245 @@
+use anyhow::{Context, Result};
+use chrono::Utc;
+use std::path::Path;
+
+use crate::commands::eval::EvalReport;
+
+use super::super::types::{FeedbackEvalTrend, FeedbackEvalTrendEntry, FeedbackEvalTrendGap};
+use super::super::{FeedbackEvalReport, FeedbackEvalRuleCorrelation};
+
+const MAX_ATTENTION_GAPS: usize = 5;
+
+pub(in super::super) async fn update_feedback_eval_trend(
+    report: &FeedbackEvalReport,
+    eval_report: Option<&EvalReport>,
+    path: &Path,
+) -> Result<()> {
+    let mut trend = if path.exists() {
+        let content = tokio::fs::read_to_string(path)
+            .await
+            .with_context(|| format!("failed to read feedback trend file {}", path.display()))?;
+        FeedbackEvalTrend::from_json(&content)
+            .with_context(|| format!("failed to parse feedback trend file {}", path.display()))?
+    } else {
+        FeedbackEvalTrend::new()
+    };
+    trend
+        .entries
+        .push(trend_entry_for_report(report, eval_report));
+
+    if let Some(parent) = path.parent() {
+        tokio::fs::create_dir_all(parent)
+            .await
+            .with_context(|| format!("failed to create {}", parent.display()))?;
+    }
+    tokio::fs::write(path, trend.to_json()?)
+        .await
+        .with_context(|| format!("failed to write feedback trend file {}", path.display()))?;
+    Ok(())
+}
+
+fn trend_entry_for_report(
+    report: &FeedbackEvalReport,
+    eval_report: Option<&EvalReport>,
+) -> FeedbackEvalTrendEntry {
+    let confidence_metrics = report.confidence_metrics;
+    let correlation = report.eval_correlation.as_ref();
+    FeedbackEvalTrendEntry {
+        timestamp: Utc::now().to_rfc3339(),
+        labeled_comments: report.labeled_comments,
+        accepted: report.accepted,
+        rejected: report.rejected,
+        acceptance_rate: report.acceptance_rate,
+        confidence_threshold: report.confidence_threshold,
+        confidence_agreement_rate: confidence_metrics.map(|metrics| metrics.agreement_rate),
+        confidence_precision: confidence_metrics.map(|metrics| metrics.precision),
+        confidence_recall: confidence_metrics.map(|metrics| metrics.recall),
+        confidence_f1: confidence_metrics.map(|metrics| metrics.f1),
+        eval_label: eval_report.and_then(|report| report.run.label.clone()),
+        eval_model: eval_report.map(|report| report.run.model.clone()),
+        eval_provider: eval_report.and_then(|report| report.run.provider.clone()),
+        attention_by_category: correlation
+            .map(|report| {
+                report
+                    .attention_by_category
+                    .iter()
+                    .take(MAX_ATTENTION_GAPS)
+                    .map(|category| FeedbackEvalTrendGap {
+                        name: category.name.clone(),
+                        feedback_total: category.feedback_total,
+                        high_confidence_total: category.high_confidence_total,
+                        high_confidence_acceptance_rate: category.high_confidence_acceptance_rate,
+                        eval_score: category.eval_micro_f1,
+                        gap: category.high_confidence_vs_eval_gap,
+                    })
+                    .collect()
+            })
+            .unwrap_or_default(),
+        attention_by_rule: correlation
+            .map(|report| {
+                report
+                    .attention_by_rule
+                    .iter()
+                    .take(MAX_ATTENTION_GAPS)
+                    .map(rule_gap)
+                    .collect()
+            })
+            .unwrap_or_default(),
+    }
+}
+
+fn rule_gap(rule: &FeedbackEvalRuleCorrelation) -> FeedbackEvalTrendGap {
+    FeedbackEvalTrendGap {
+        name: rule.rule_id.clone(),
+        feedback_total: rule.feedback_total,
+        high_confidence_total: rule.high_confidence_total,
+        high_confidence_acceptance_rate: rule.high_confidence_acceptance_rate,
+        eval_score: rule.eval_f1,
+        gap: rule.high_confidence_vs_eval_gap,
+    }
+}
+
+#[cfg(test)]
+mod tests {
+    use tempfile::tempdir;
+
+    use crate::commands::eval::EvalRunMetadata;
+    use crate::commands::feedback_eval::{
+        FeedbackEvalBucket, FeedbackEvalCategoryCorrelation, FeedbackEvalCorrelationReport,
+        FeedbackEvalReport, FeedbackEvalRuleCorrelation, FeedbackThresholdMetrics,
+    };
+
+    use super::*;
+
+    fn sample_feedback_report() -> FeedbackEvalReport {
+        FeedbackEvalReport {
+            total_comments_seen: 12,
+            total_reviews_seen: 3,
+            labeled_comments: 8,
+            labeled_reviews: 2,
+            accepted: 3,
+            rejected: 5,
+            acceptance_rate: 0.375,
+            confidence_threshold: 0.75,
+            vague_comments: FeedbackEvalBucket {
+                name: "vague".to_string(),
+                total: 1,
+                accepted: 0,
+                rejected: 1,
+                acceptance_rate: 0.0,
+            },
+            confidence_metrics: Some(FeedbackThresholdMetrics {
+                total_scored: 6,
+                true_positive: 2,
+                false_positive: 1,
+                true_negative: 2,
+                false_negative: 1,
+                precision: 0.67,
+                recall: 0.67,
+                f1: 0.67,
+                agreement_rate: 0.67,
+            }),
+            by_category: vec![],
+            by_rule: vec![],
+            high_confidence_by_category: vec![],
+            high_confidence_by_rule: vec![],
+            by_severity: vec![],
+            by_repo: vec![],
+            by_file_pattern: vec![],
+            eval_correlation: Some(FeedbackEvalCorrelationReport {
+                by_category: vec![],
+                by_rule: vec![],
+                attention_by_category: vec![FeedbackEvalCategoryCorrelation {
+                    name: "Security".to_string(),
+                    feedback_total: 4,
+                    feedback_acceptance_rate: 0.25,
+                    high_confidence_total: 3,
+                    high_confidence_acceptance_rate: 0.0,
+                    eval_fixture_count: Some(5),
+                    eval_micro_f1: Some(0.9),
+                    eval_weighted_score: Some(0.91),
+                    feedback_vs_eval_gap: Some(0.65),
+                    high_confidence_vs_eval_gap: Some(0.9),
+                }],
+                attention_by_rule: vec![FeedbackEvalRuleCorrelation {
+                    rule_id: "sec.sql.injection".to_string(),
+                    feedback_total: 3,
+                    feedback_acceptance_rate: 0.33,
+                    high_confidence_total: 2,
+                    high_confidence_acceptance_rate: 0.0,
+                    eval_precision: Some(1.0),
+                    eval_recall: Some(1.0),
+                    eval_f1: Some(1.0),
+                    feedback_vs_eval_gap: Some(0.67),
+                    high_confidence_vs_eval_gap: Some(1.0),
+                }],
+            }),
+            showcase_candidates: vec![],
+            vague_rejections: vec![],
+        }
+    }
+
+    fn sample_eval_report() -> EvalReport {
+        EvalReport {
+            run: EvalRunMetadata {
+                label: Some("frontier-e2e".to_string()),
+                model: "anthropic/claude-opus-4.5".to_string(),
+                provider: Some("openrouter".to_string()),
+                ..Default::default()
+            },
+            fixtures_total: 0,
+            fixtures_passed: 0,
+            fixtures_failed: 0,
+            rule_metrics: vec![],
+            rule_summary: None,
+            benchmark_summary: None,
+            suite_results: vec![],
+            benchmark_by_category: Default::default(),
+            benchmark_by_language: Default::default(),
+            benchmark_by_difficulty: Default::default(),
+            suite_comparisons: vec![],
+            category_comparisons: vec![],
+            language_comparisons: vec![],
+            verification_health: None,
+            warnings: vec![],
+            threshold_failures: vec![],
+            results: vec![],
+        }
+    }
+
+    #[tokio::test]
+    async fn update_feedback_eval_trend_appends_attention_entries() {
+        let dir = tempdir().unwrap();
+        let path = dir.path().join("feedback-trend.json");
+
+        update_feedback_eval_trend(
+            &sample_feedback_report(),
+            Some(&sample_eval_report()),
+            &path,
+        )
+        .await
+        .unwrap();
+        update_feedback_eval_trend(
+            &sample_feedback_report(),
+            Some(&sample_eval_report()),
+            &path,
+        )
+        .await
+        .unwrap();
+
+        let content = tokio::fs::read_to_string(&path).await.unwrap();
+        let trend = FeedbackEvalTrend::from_json(&content).unwrap();
+        assert_eq!(trend.entries.len(), 2);
+        assert_eq!(trend.entries[0].eval_label.as_deref(), Some("frontier-e2e"));
+        assert_eq!(
+            trend.entries[0].eval_model.as_deref(),
+            Some("anthropic/claude-opus-4.5")
+        );
+        assert_eq!(trend.entries[0].attention_by_category.len(), 1);
+        assert_eq!(trend.entries[0].attention_by_category[0].name, "Security");
+        assert_eq!(
+            trend.entries[0].attention_by_rule[0].name,
+            "sec.sql.injection"
+        );
+    }
+}

src/commands/feedback_eval/types.rs

@@ -6,5 +6,6 @@ mod report;
 pub(super) use input::{FeedbackEvalComment, LoadedFeedbackEvalInput};
 pub(super) use report::{
     FeedbackEvalBucket, FeedbackEvalCategoryCorrelation, FeedbackEvalCorrelationReport,
-    FeedbackEvalExample, FeedbackEvalReport, FeedbackEvalRuleCorrelation, FeedbackThresholdMetrics,
+    FeedbackEvalExample, FeedbackEvalReport, FeedbackEvalRuleCorrelation, FeedbackEvalTrend,
+    FeedbackEvalTrendEntry, FeedbackEvalTrendGap, FeedbackThresholdMetrics,
 };

src/commands/feedback_eval/types/report.rs

@@ -75,6 +75,76 @@ pub(in super::super) struct FeedbackEvalCorrelationReport {
     pub(in super::super) attention_by_rule: Vec<FeedbackEvalRuleCorrelation>,
 }
 
+#[derive(Debug, Clone, Serialize, Deserialize)]
+pub(in super::super) struct FeedbackEvalTrendGap {
+    #[serde(default)]
+    pub(in super::super) name: String,
+    #[serde(default)]
+    pub(in super::super) feedback_total: usize,
+    #[serde(default)]
+    pub(in super::super) high_confidence_total: usize,
+    #[serde(default)]
+    pub(in super::super) high_confidence_acceptance_rate: f32,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub(in super::super) eval_score: Option<f32>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub(in super::super) gap: Option<f32>,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, Default)]
+pub(in super::super) struct FeedbackEvalTrendEntry {
+    #[serde(default)]
+    pub(in super::super) timestamp: String,
+    #[serde(default)]
+    pub(in super::super) labeled_comments: usize,
+    #[serde(default)]
+    pub(in super::super) accepted: usize,
+    #[serde(default)]
+    pub(in super::super) rejected: usize,
+    #[serde(default)]
+    pub(in super::super) acceptance_rate: f32,
+    #[serde(default)]
+    pub(in super::super) confidence_threshold: f32,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub(in super::super) confidence_agreement_rate: Option<f32>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub(in super::super) confidence_precision: Option<f32>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub(in super::super) confidence_recall: Option<f32>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub(in super::super) confidence_f1: Option<f32>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub(in super::super) eval_label: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub(in super::super) eval_model: Option<String>,
+    #[serde(default, skip_serializing_if = "Option::is_none")]
+    pub(in super::super) eval_provider: Option<String>,
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub(in super::super) attention_by_category: Vec<FeedbackEvalTrendGap>,
+    #[serde(default, skip_serializing_if = "Vec::is_empty")]
+    pub(in super::super) attention_by_rule: Vec<FeedbackEvalTrendGap>,
+}
+
+#[derive(Debug, Clone, Serialize, Deserialize, Default)]
+pub(in super::super) struct FeedbackEvalTrend {
+    #[serde(default)]
+    pub(in super::super) entries: Vec<FeedbackEvalTrendEntry>,
+}
+
+impl FeedbackEvalTrend {
+    pub(in super::super) fn new() -> Self {
+        Self::default()
+    }
+
+    pub(in super::super) fn to_json(&self) -> Result<String, serde_json::Error> {
+        serde_json::to_string_pretty(self)
+    }
+
+    pub(in super::super) fn from_json(json: &str) -> Result<Self, serde_json::Error> {
+        serde_json::from_str(json)
+    }
+}
+
 #[derive(Debug, Clone, Copy, Serialize, Deserialize, Default)]
 pub(in super::super) struct FeedbackThresholdMetrics {
     #[serde(default)]