claudectl 0.35.0

#![allow(dead_code)]

use std::collections::HashMap;

use super::decisions::{DecisionRecord, DistilledPreferences};
use super::insights::{Insight, InsightCategory, InsightSeverity, epoch_now};

// ────────────────────────────────────────────────────────────────────────────
// Detection algorithms
// ────────────────────────────────────────────────────────────────────────────

/// Extract a command keyword for grouping (first two tokens).
/// Duplicated from decisions.rs because that function is private.
pub(crate) fn extract_command_keyword(command: Option<&str>) -> Option<String> {
    let cmd = command?.trim();
    if cmd.is_empty() {
        return None;
    }
    let tokens: Vec<&str> = cmd.split_whitespace().take(2).collect();
    Some(tokens.join(" "))
}

/// Detect tools/commands that are repeatedly rejected by the user.
pub(crate) fn detect_friction_patterns(decisions: &[DecisionRecord]) -> Vec<Insight> {
    let mut groups: HashMap<(String, Option<String>), (u32, u32)> = HashMap::new();

    for d in decisions {
        let tool = d.tool.clone().unwrap_or_else(|| "*".to_string());
        let cmd = extract_command_keyword(d.command.as_deref());
        let key = (tool, cmd);
        let entry = groups.entry(key).or_insert((0, 0));
        entry.0 += 1; // total
        if d.is_negative() {
            entry.1 += 1; // rejected
        }
    }

    let now = epoch_now();
    let mut insights = Vec::new();

    for ((tool, cmd), (total, rejected)) in &groups {
        if *rejected < 3 || *total < 3 {
            continue;
        }
        let rejection_rate = *rejected as f64 / *total as f64;
        if rejection_rate < 0.6 {
            continue;
        }

        let cmd_part = cmd
            .as_ref()
            .map(|c| format!(" \"{c}\""))
            .unwrap_or_default();

        let severity = if rejection_rate >= 0.9 {
            InsightSeverity::Warning
        } else {
            InsightSeverity::Suggestion
        };

        insights.push(Insight {
            fingerprint: format!("friction:{}:{}", tool, cmd.as_deref().unwrap_or("*")),
            generated_at: now,
            category: InsightCategory::FrictionPattern,
            severity,
            summary: format!(
                "[{tool}]{cmd_part} rejected {rejected}/{total} times ({:.0}%)",
                rejection_rate * 100.0
            ),
            suggestion: Some(format!("consider adding deny rule for [{tool}]{cmd_part}")),
            evidence_count: *total,
        });
    }

    insights
}

/// Detect repeated errors from the same tool across sessions.
pub(crate) fn detect_error_loops(decisions: &[DecisionRecord]) -> Vec<Insight> {
    // Group by PID, then find consecutive errors for same tool
    let mut pid_groups: HashMap<u32, Vec<&DecisionRecord>> = HashMap::new();
    for d in decisions {
        pid_groups.entry(d.pid).or_default().push(d);
    }

    // Count how many sessions had error loops for each (tool, cmd) combo
    let mut loop_counts: HashMap<(String, Option<String>), u32> = HashMap::new();

    for session_decisions in pid_groups.values() {
        let mut streak_tool: Option<String> = None;
        let mut streak_cmd: Option<String> = None;
        let mut streak_count: u32 = 0;

        for d in session_decisions {
            let has_error = d
                .context
                .as_ref()
                .map(|c| c.last_tool_error)
                .unwrap_or(false);
            let tool = d.tool.clone().unwrap_or_default();
            let cmd = extract_command_keyword(d.command.as_deref());

            if has_error && Some(&tool) == streak_tool.as_ref() {
                streak_count += 1;
            } else if has_error {
                // New error streak
                streak_tool = Some(tool.clone());
                streak_cmd = cmd.clone();
                streak_count = 1;
            } else {
                // No error — check if previous streak was long enough
                if streak_count >= 3 {
                    if let Some(ref t) = streak_tool {
                        *loop_counts
                            .entry((t.clone(), streak_cmd.clone()))
                            .or_insert(0) += 1;
                    }
                }
                streak_tool = None;
                streak_cmd = None;
                streak_count = 0;
            }
        }
        // Check trailing streak
        if streak_count >= 3 {
            if let Some(ref t) = streak_tool {
                *loop_counts
                    .entry((t.clone(), streak_cmd.clone()))
                    .or_insert(0) += 1;
            }
        }
    }

    let now = epoch_now();
    loop_counts
        .into_iter()
        .filter(|(_, count)| *count >= 1)
        .map(|((tool, cmd), count)| {
            let cmd_part = cmd
                .as_ref()
                .map(|c| format!(" \"{c}\""))
                .unwrap_or_default();
            Insight {
                fingerprint: format!("error_loop:{}:{}", tool, cmd.as_deref().unwrap_or("*")),
                generated_at: now,
                category: InsightCategory::ErrorLoop,
                severity: if count >= 3 {
                    InsightSeverity::Warning
                } else {
                    InsightSeverity::Suggestion
                },
                summary: format!(
                    "[{tool}]{cmd_part} hit 3+ consecutive errors in {count} session(s)"
                ),
                suggestion: Some(format!("investigate why [{tool}]{cmd_part} keeps failing")),
                evidence_count: count,
            }
        })
        .collect()
}

/// Detect sessions frequently hitting high context usage.
pub(crate) fn detect_context_blowouts(decisions: &[DecisionRecord]) -> Vec<Insight> {
    // Group by PID, check if any decision in session had context > 80%
    let mut pid_max_context: HashMap<u32, u8> = HashMap::new();
    for d in decisions {
        if let Some(ref ctx) = d.context {
            let entry = pid_max_context.entry(d.pid).or_insert(0);
            if ctx.context_pct > *entry {
                *entry = ctx.context_pct;
            }
        }
    }

    if pid_max_context.is_empty() {
        return Vec::new();
    }

    // Only look at recent sessions (last 20 PIDs by insertion order)
    let recent: Vec<u8> = pid_max_context
        .values()
        .copied()
        .collect::<Vec<_>>()
        .into_iter()
        .rev()
        .take(20)
        .collect();

    let blowout_count = recent.iter().filter(|&&pct| pct > 80).count();
    let total = recent.len();
    let blowout_rate = blowout_count as f64 / total as f64;

    if blowout_rate < 0.4 || blowout_count < 2 {
        return Vec::new();
    }

    vec![Insight {
        fingerprint: "context_blowout:global".to_string(),
        generated_at: epoch_now(),
        category: InsightCategory::ContextBlowout,
        severity: if blowout_rate >= 0.7 {
            InsightSeverity::Warning
        } else {
            InsightSeverity::Suggestion
        },
        summary: format!(
            "Context >80% in {blowout_count}/{total} recent sessions ({:.0}%)",
            blowout_rate * 100.0
        ),
        suggestion: Some("consider earlier /compact when context approaches 70%".to_string()),
        evidence_count: blowout_count as u32,
    }]
}

/// Detect high-confidence patterns that could become AutoRules.
pub(crate) fn detect_missing_rules(
    _decisions: &[DecisionRecord],
    prefs: &DistilledPreferences,
) -> Vec<Insight> {
    let now = epoch_now();
    let mut insights = Vec::new();

    for p in &prefs.patterns {
        if p.sample_count < 5 || p.confidence < 0.8 {
            continue;
        }

        // High-confidence approve patterns
        if p.accept_rate >= 0.9 {
            let cmd_part = p
                .command_pattern
                .as_ref()
                .map(|c| format!(" \"{c}\""))
                .unwrap_or_default();

            insights.push(Insight {
                fingerprint: format!(
                    "missing_rule:approve:{}:{}",
                    p.tool,
                    p.command_pattern.as_deref().unwrap_or("*")
                ),
                generated_at: now,
                category: InsightCategory::MissingRule,
                severity: InsightSeverity::Suggestion,
                summary: format!(
                    "approve [{}]{cmd_part} (accepted {:.0}%, n={})",
                    p.tool,
                    p.accept_rate * 100.0,
                    p.sample_count,
                ),
                suggestion: Some(format!(
                    "add to .claudectl.toml: [[rules]] match_tool=\"{}\" match_command=\"{}\" action=\"approve\"",
                    p.tool,
                    p.command_pattern.as_deref().unwrap_or("*"),
                )),
                evidence_count: p.sample_count,
            });
        }

        // High-confidence deny patterns
        if p.accept_rate <= 0.1 {
            let cmd_part = p
                .command_pattern
                .as_ref()
                .map(|c| format!(" \"{c}\""))
                .unwrap_or_default();

            insights.push(Insight {
                fingerprint: format!(
                    "missing_rule:deny:{}:{}",
                    p.tool,
                    p.command_pattern.as_deref().unwrap_or("*")
                ),
                generated_at: now,
                category: InsightCategory::MissingRule,
                severity: InsightSeverity::Suggestion,
                summary: format!(
                    "deny [{}]{cmd_part} (rejected {:.0}%, n={})",
                    p.tool,
                    (1.0 - p.accept_rate) * 100.0,
                    p.sample_count,
                ),
                suggestion: Some(format!(
                    "add to .claudectl.toml: [[rules]] match_tool=\"{}\" match_command=\"{}\" action=\"deny\"",
                    p.tool,
                    p.command_pattern.as_deref().unwrap_or("*"),
                )),
                evidence_count: p.sample_count,
            });
        }
    }

    insights
}

/// Detect tools where brain accuracy is low.
pub(crate) fn detect_accuracy_gaps(prefs: &DistilledPreferences) -> Vec<Insight> {
    let now = epoch_now();
    prefs
        .tool_accuracy
        .iter()
        .filter(|ta| ta.total >= 5 && ta.confidence_threshold > 0.7)
        .map(|ta| {
            let accuracy = if ta.total > 0 {
                (ta.correct as f64 / ta.total as f64) * 100.0
            } else {
                0.0
            };
            Insight {
                fingerprint: format!("accuracy_gap:{}", ta.tool),
                generated_at: now,
                category: InsightCategory::AccuracyGap,
                severity: if accuracy < 50.0 {
                    InsightSeverity::Warning
                } else {
                    InsightSeverity::Suggestion
                },
                summary: format!(
                    "Brain accuracy for [{}] is {:.0}% (threshold raised to {:.2})",
                    ta.tool, accuracy, ta.confidence_threshold,
                ),
                suggestion: Some(format!(
                    "more training data needed for [{}] — brain defers these to manual review",
                    ta.tool,
                )),
                evidence_count: ta.total,
            }
        })
        .collect()
}

/// Convert temporal patterns from distillation into insights.
pub(crate) fn detect_temporal_friction(prefs: &DistilledPreferences) -> Vec<Insight> {
    let now = epoch_now();
    prefs
        .temporal
        .iter()
        .filter(|tp| tp.strength > 0.3 && tp.sample_count >= 3)
        .map(|tp| {
            // Use first 40 chars of description as fingerprint suffix
            let fp_suffix: String = tp
                .description
                .chars()
                .filter(|c| c.is_alphanumeric() || *c == ' ')
                .take(40)
                .collect::<String>()
                .replace(' ', "_");
            Insight {
                fingerprint: format!("temporal:{fp_suffix}"),
                generated_at: now,
                category: InsightCategory::TemporalFriction,
                severity: InsightSeverity::Info,
                summary: tp.description.clone(),
                suggestion: None,
                evidence_count: tp.sample_count,
            }
        })
        .collect()
}

/// Detect increasing burn rate trends and cost outliers.
pub(crate) fn detect_cost_patterns(decisions: &[DecisionRecord]) -> Vec<Insight> {
    let burn_rates: Vec<f64> = decisions
        .iter()
        .filter_map(|d| d.context.as_ref().map(|c| c.burn_rate_per_hr))
        .filter(|r| *r > 0.0)
        .collect();

    if burn_rates.len() < 10 {
        return Vec::new();
    }

    let mut insights = Vec::new();
    let now = epoch_now();

    // Compare first half vs second half burn rates
    let mid = burn_rates.len() / 2;
    let first_avg: f64 = burn_rates[..mid].iter().sum::<f64>() / mid as f64;
    let second_avg: f64 = burn_rates[mid..].iter().sum::<f64>() / (burn_rates.len() - mid) as f64;

    if first_avg > 0.0 {
        let increase = (second_avg - first_avg) / first_avg;
        if increase > 0.5 {
            insights.push(Insight {
                fingerprint: "cost_trend:increasing".to_string(),
                generated_at: now,
                category: InsightCategory::CostPattern,
                severity: if increase > 1.0 {
                    InsightSeverity::Warning
                } else {
                    InsightSeverity::Suggestion
                },
                summary: format!(
                    "Burn rate trending up: ${:.2}/hr -> ${:.2}/hr ({:+.0}%)",
                    first_avg,
                    second_avg,
                    increase * 100.0,
                ),
                suggestion: Some(
                    "consider setting a budget with --budget or reviewing costly operations"
                        .to_string(),
                ),
                evidence_count: burn_rates.len() as u32,
            });
        }
    }

    // Detect cost outlier sessions
    let mut per_session_cost: HashMap<u32, f64> = HashMap::new();
    for d in decisions {
        if let Some(ref ctx) = d.context {
            let entry = per_session_cost.entry(d.pid).or_insert(0.0);
            if ctx.cost_usd > *entry {
                *entry = ctx.cost_usd;
            }
        }
    }

    if per_session_cost.len() >= 3 {
        let costs: Vec<f64> = per_session_cost.values().copied().collect();
        let avg: f64 = costs.iter().sum::<f64>() / costs.len() as f64;
        let outlier_count = costs.iter().filter(|&&c| c > avg * 2.0 && c > 1.0).count();

        if outlier_count >= 2 {
            insights.push(Insight {
                fingerprint: "cost_trend:outliers".to_string(),
                generated_at: now,
                category: InsightCategory::CostPattern,
                severity: InsightSeverity::Info,
                summary: format!(
                    "{outlier_count} sessions cost >2x average (avg ${avg:.2})"
                ),
                suggestion: Some(
                    "review high-cost sessions — consider budget limits or earlier session restarts"
                        .to_string(),
                ),
                evidence_count: outlier_count as u32,
            });
        }
    }

    insights
}

// ────────────────────────────────────────────────────────────────────────────
// Tests
// ────────────────────────────────────────────────────────────────────────────

#[cfg(test)]
mod tests {
    use super::*;
    use crate::brain::decisions::{
        DecisionContext, DecisionType, DistilledPreferences, PreferencePattern, TemporalPattern,
        ToolAccuracy,
    };

    fn make_decision(tool: &str, command: &str, user_action: &str, pid: u32) -> DecisionRecord {
        DecisionRecord {
            timestamp: "0".to_string(),
            pid,
            project: "test".to_string(),
            tool: Some(tool.to_string()),
            command: Some(command.to_string()),
            brain_action: "approve".to_string(),
            brain_confidence: 0.8,
            brain_reasoning: String::new(),
            user_action: user_action.to_string(),
            context: None,
            outcome: None,
            decision_type: DecisionType::Session,
            suggested_at: None,
        }
    }

    #[allow(clippy::too_many_arguments)]
    fn make_decision_with_context(
        tool: &str,
        command: &str,
        user_action: &str,
        pid: u32,
        context_pct: u8,
        last_error: bool,
        burn_rate: f64,
        cost: f64,
    ) -> DecisionRecord {
        let mut d = make_decision(tool, command, user_action, pid);
        d.context = Some(DecisionContext {
            cost_usd: cost,
            context_pct,
            last_tool_error: last_error,
            error_message: None,
            model: "test".to_string(),
            elapsed_secs: 100,
            files_modified_count: 0,
            total_tool_calls: 10,
            has_file_conflict: false,
            status: "Processing".to_string(),
            burn_rate_per_hr: burn_rate,
            recent_error_count: 0,
            subagent_count: 0,
            hour: Some(10),
        });
        d
    }

    #[test]
    fn test_friction_patterns_detected() {
        let decisions: Vec<DecisionRecord> = (0..10)
            .map(|i| make_decision("Bash", "npm install", "reject", i))
            .collect();

        let insights = detect_friction_patterns(&decisions);
        assert_eq!(insights.len(), 1);
        assert_eq!(insights[0].category, InsightCategory::FrictionPattern);
        assert!(insights[0].summary.contains("npm install"));
        assert!(insights[0].summary.contains("10/10"));
    }

    #[test]
    fn test_friction_below_threshold_not_detected() {
        // 2 rejections out of 5 = 40%, below 60% threshold
        let mut decisions = Vec::new();
        for i in 0..3 {
            decisions.push(make_decision("Bash", "cargo test", "accept", i));
        }
        for i in 3..5 {
            decisions.push(make_decision("Bash", "cargo test", "reject", i));
        }

        let insights = detect_friction_patterns(&decisions);
        assert!(insights.is_empty());
    }

    #[test]
    fn test_error_loops_detected() {
        // 4 consecutive errors for same tool in one session
        let decisions: Vec<DecisionRecord> = (0..4)
            .map(|_| {
                make_decision_with_context(
                    "Write",
                    "src/main.rs",
                    "accept",
                    100,
                    50,
                    true,
                    1.0,
                    0.5,
                )
            })
            .collect();

        let insights = detect_error_loops(&decisions);
        assert_eq!(insights.len(), 1);
        assert_eq!(insights[0].category, InsightCategory::ErrorLoop);
    }

    #[test]
    fn test_context_blowouts_detected() {
        // 5 sessions all hitting >80% context
        let decisions: Vec<DecisionRecord> = (0..5)
            .map(|pid| {
                make_decision_with_context("Read", "file.rs", "accept", pid, 85, false, 1.0, 0.5)
            })
            .collect();

        let insights = detect_context_blowouts(&decisions);
        assert_eq!(insights.len(), 1);
        assert_eq!(insights[0].category, InsightCategory::ContextBlowout);
    }

    #[test]
    fn test_missing_rules_detected() {
        let prefs = DistilledPreferences {
            patterns: vec![
                PreferencePattern {
                    tool: "Bash".to_string(),
                    command_pattern: Some("cargo test".to_string()),
                    preferred_action: "approve".to_string(),
                    sample_count: 15,
                    accept_rate: 1.0,
                    conditions: Vec::new(),
                    confidence: 1.0,
                },
                PreferencePattern {
                    tool: "Bash".to_string(),
                    command_pattern: Some("rm -rf".to_string()),
                    preferred_action: "deny".to_string(),
                    sample_count: 6,
                    accept_rate: 0.0,
                    conditions: Vec::new(),
                    confidence: 1.0,
                },
            ],
            tool_accuracy: Vec::new(),
            total_decisions: 21,
            overall_accuracy: 0.8,
            temporal: Vec::new(),
        };

        let insights = detect_missing_rules(&[], &prefs);
        assert_eq!(insights.len(), 2);
        assert!(insights.iter().any(|i| i.summary.contains("cargo test")));
        assert!(insights.iter().any(|i| i.summary.contains("rm -rf")));
    }

    #[test]
    fn test_accuracy_gaps_detected() {
        let prefs = DistilledPreferences {
            patterns: Vec::new(),
            tool_accuracy: vec![ToolAccuracy {
                tool: "Edit".to_string(),
                total: 10,
                correct: 4,
                confidence_threshold: 0.85,
            }],
            total_decisions: 10,
            overall_accuracy: 0.4,
            temporal: Vec::new(),
        };

        let insights = detect_accuracy_gaps(&prefs);
        assert_eq!(insights.len(), 1);
        assert!(insights[0].summary.contains("Edit"));
        assert!(insights[0].summary.contains("40%"));
    }

    #[test]
    fn test_cost_trend_detected() {
        let mut decisions = Vec::new();
        // First 10: low burn rate
        for i in 0..10 {
            decisions.push(make_decision_with_context(
                "Bash", "cmd", "accept", i, 50, false, 1.0, 0.5,
            ));
        }
        // Next 10: high burn rate (3x increase)
        for i in 10..20 {
            decisions.push(make_decision_with_context(
                "Bash", "cmd", "accept", i, 50, false, 3.0, 1.5,
            ));
        }

        let insights = detect_cost_patterns(&decisions);
        assert!(
            insights
                .iter()
                .any(|i| i.fingerprint == "cost_trend:increasing")
        );
    }

    #[test]
    fn test_temporal_friction_detected() {
        let prefs = DistilledPreferences {
            patterns: Vec::new(),
            tool_accuracy: Vec::new(),
            total_decisions: 50,
            overall_accuracy: 0.8,
            temporal: vec![TemporalPattern {
                description: "After 3+ errors: user usually denies (n=5)".to_string(),
                sample_count: 5,
                strength: 0.6,
            }],
        };

        let insights = detect_temporal_friction(&prefs);
        assert_eq!(insights.len(), 1);
        assert_eq!(insights[0].category, InsightCategory::TemporalFriction);
    }
}