lean_ctx/core/

intent_engine.rs

use super::intent_lang;

#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
pub enum TaskType {
    Generate,
    FixBug,
    Refactor,
    Explore,
    Test,
    Debug,
    Config,
    Deploy,
    Review,
}

impl TaskType {
    pub fn as_str(&self) -> &'static str {
        match self {
            Self::Generate => "generate",
            Self::FixBug => "fix_bug",
            Self::Refactor => "refactor",
            Self::Explore => "explore",
            Self::Test => "test",
            Self::Debug => "debug",
            Self::Config => "config",
            Self::Deploy => "deploy",
            Self::Review => "review",
        }
    }

    /// All task types in declaration order. The default (coding) intent
    /// taxonomy; personas can override it (12.16).
    pub fn all() -> &'static [TaskType] {
        &[
            Self::Generate,
            Self::FixBug,
            Self::Refactor,
            Self::Explore,
            Self::Test,
            Self::Debug,
            Self::Config,
            Self::Deploy,
            Self::Review,
        ]
    }

    pub fn thinking_budget(&self) -> ThinkingBudget {
        match self {
            Self::Generate | Self::FixBug | Self::Test | Self::Config | Self::Deploy => {
                ThinkingBudget::Minimal
            }
            Self::Refactor | Self::Explore | Self::Debug | Self::Review => ThinkingBudget::Medium,
        }
    }

    pub fn output_format(&self) -> OutputFormat {
        match self {
            Self::Generate | Self::Test | Self::Config => OutputFormat::CodeOnly,
            Self::FixBug | Self::Refactor => OutputFormat::DiffOnly,
            Self::Explore | Self::Review => OutputFormat::ExplainConcise,
            Self::Debug => OutputFormat::Trace,
            Self::Deploy => OutputFormat::StepList,
        }
    }
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum ThinkingBudget {
    Minimal,
    Medium,
    Trace,
    Deep,
}

impl ThinkingBudget {
    pub fn instruction(&self) -> &'static str {
        match self {
            Self::Minimal => "THINKING: Skip analysis. The task is clear — generate code directly.",
            Self::Medium => {
                "THINKING: 2-3 step analysis max. Identify what to change, then act. Do not over-analyze."
            }
            Self::Trace => {
                "THINKING: Short trace only. Identify root cause in 3 steps max, then generate fix."
            }
            Self::Deep => {
                "THINKING: Analyze structure and dependencies. Summarize findings concisely."
            }
        }
    }

    pub fn suppresses_thinking(&self) -> bool {
        matches!(self, Self::Minimal)
    }
}

#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum OutputFormat {
    CodeOnly,
    DiffOnly,
    ExplainConcise,
    Trace,
    StepList,
}

impl OutputFormat {
    pub fn instruction(&self) -> &'static str {
        match self {
            Self::CodeOnly => {
                "OUTPUT-HINT: Prefer code blocks. Minimize prose unless user asks for explanation."
            }
            Self::DiffOnly => "OUTPUT-HINT: Prefer showing only changed lines as +/- diffs.",
            Self::ExplainConcise => "OUTPUT-HINT: Brief summary, then code/data if relevant.",
            Self::Trace => "OUTPUT-HINT: Show cause→effect chain with code references.",
            Self::StepList => "OUTPUT-HINT: Numbered action list, one step at a time.",
        }
    }
}

#[derive(Debug)]
pub struct TaskClassification {
    pub task_type: TaskType,
    pub confidence: f64,
    pub targets: Vec<String>,
    pub keywords: Vec<String>,
}

const PHRASE_RULES: &[(&[&str], TaskType, f64)] = &[
    (
        &[
            "add",
            "create",
            "implement",
            "build",
            "write",
            "generate",
            "make",
            "new feature",
            "new",
        ],
        TaskType::Generate,
        0.9,
    ),
    (
        &[
            "fix",
            "bug",
            "broken",
            "crash",
            "error in",
            "not working",
            "fails",
            "wrong output",
        ],
        TaskType::FixBug,
        0.95,
    ),
    (
        &[
            "refactor",
            "clean up",
            "restructure",
            "rename",
            "move",
            "extract",
            "simplify",
            "split",
        ],
        TaskType::Refactor,
        0.9,
    ),
    (
        &[
            "how",
            "what",
            "where",
            "explain",
            "understand",
            "show me",
            "describe",
            "why does",
        ],
        TaskType::Explore,
        0.85,
    ),
    (
        &[
            "test",
            "spec",
            "coverage",
            "assert",
            "unit test",
            "integration test",
            "mock",
        ],
        TaskType::Test,
        0.9,
    ),
    (
        &[
            "debug",
            "trace",
            "inspect",
            "log",
            "breakpoint",
            "step through",
            "stack trace",
        ],
        TaskType::Debug,
        0.9,
    ),
    (
        &[
            "config",
            "setup",
            "install",
            "env",
            "configure",
            "settings",
            "dotenv",
        ],
        TaskType::Config,
        0.85,
    ),
    (
        &[
            "deploy", "release", "publish", "ship", "ci/cd", "pipeline", "docker",
        ],
        TaskType::Deploy,
        0.85,
    ),
    (
        &[
            "review",
            "check",
            "audit",
            "look at",
            "evaluate",
            "assess",
            "pr review",
        ],
        TaskType::Review,
        0.8,
    ),
];

pub fn classify(query: &str) -> TaskClassification {
    let q = query.to_lowercase();
    let words: Vec<&str> = q.split_whitespace().collect();
    let lang = intent_lang::detect_query_lang(&words);

    let mut best_type = TaskType::Explore;
    let mut best_score = 0.0_f64;

    apply_rules(
        &q,
        &words,
        PHRASE_RULES,
        |word, phrase| word == phrase,
        &mut best_type,
        &mut best_score,
    );
    // Multilingual stems (de/fr/es) match morphology-tolerant via prefix.
    apply_rules(
        &q,
        &words,
        intent_lang::STEM_RULES,
        |word, stem| word.starts_with(stem),
        &mut best_type,
        &mut best_score,
    );

    let targets = extract_targets(query, lang);
    let keywords = extract_keywords(&q, lang);

    if best_score < 0.1 {
        best_type = TaskType::Explore;
        best_score = 0.3;
    }

    TaskClassification {
        task_type: best_type,
        confidence: best_score,
        targets,
        keywords,
    }
}

/// Shared scoring loop for PHRASE_RULES (exact word match) and
/// STEM_RULES (prefix match). Multi-word entries match the whole query.
fn apply_rules(
    q: &str,
    words: &[&str],
    rules: &[(&[&str], TaskType, f64)],
    token_match: fn(&str, &str) -> bool,
    best_type: &mut TaskType,
    best_score: &mut f64,
) {
    for &(phrases, task_type, base_confidence) in rules {
        let mut match_count = 0usize;
        for phrase in phrases {
            if phrase.contains(' ') {
                if q.contains(phrase) {
                    match_count += 2;
                }
            } else if words.iter().any(|w| token_match(w, phrase)) {
                match_count += 1;
            }
        }
        if match_count > 0 {
            let score = base_confidence * (match_count as f64).min(2.0) / 2.0;
            if score > *best_score {
                *best_score = score;
                *best_type = task_type;
            }
        }
    }
}

fn extract_targets(query: &str, lang: intent_lang::QueryLang) -> Vec<String> {
    let mut targets = Vec::new();

    for word in query.split_whitespace() {
        if word.contains('.') && !word.starts_with('.') {
            let clean = word.trim_matches(|c: char| {
                !c.is_alphanumeric() && c != '.' && c != '/' && c != '_' && c != '-'
            });
            if looks_like_path(clean) {
                targets.push(clean.to_string());
            }
        }
        if word.contains('/') && !word.starts_with("//") && !word.starts_with("http") {
            let clean = word.trim_matches(|c: char| {
                !c.is_alphanumeric() && c != '.' && c != '/' && c != '_' && c != '-'
            });
            if clean.len() > 2 {
                targets.push(clean.to_string());
            }
        }
    }

    for word in query.split_whitespace() {
        let w = word.trim_matches(|c: char| !c.is_alphanumeric() && c != '_');
        if w.contains('_') && w.len() > 3 && !targets.contains(&w.to_string()) {
            targets.push(w.to_string());
        }
        if w.chars().any(char::is_uppercase)
            && w.len() > 2
            && !is_stop_word(w)
            && !intent_lang::is_stop_word_for(lang, &w.to_lowercase())
            && !targets.contains(&w.to_string())
        {
            targets.push(w.to_string());
        }
    }

    targets.truncate(5);
    targets
}

fn looks_like_path(s: &str) -> bool {
    let exts = [
        ".rs", ".ts", ".tsx", ".js", ".jsx", ".py", ".go", ".toml", ".yaml", ".yml", ".json", ".md",
    ];
    exts.iter().any(|ext| s.ends_with(ext)) || s.contains('/')
}

fn is_stop_word(w: &str) -> bool {
    matches!(
        w.to_lowercase().as_str(),
        "the"
            | "this"
            | "that"
            | "with"
            | "from"
            | "into"
            | "have"
            | "please"
            | "could"
            | "would"
            | "should"
            | "also"
            | "just"
            | "then"
            | "when"
            | "what"
            | "where"
            | "which"
            | "there"
            | "here"
            | "these"
            | "those"
            | "does"
            | "will"
            | "shall"
            | "can"
            | "may"
            | "must"
            | "need"
            | "want"
            | "like"
            | "make"
            | "take"
    )
}

fn extract_keywords(query: &str, lang: intent_lang::QueryLang) -> Vec<String> {
    query
        .split_whitespace()
        .filter(|w| w.len() > 3)
        .filter(|w| !is_stop_word(w))
        .map(|w| {
            w.trim_matches(|c: char| !c.is_alphanumeric() && c != '_')
                .to_lowercase()
        })
        .filter(|w| !w.is_empty())
        .filter(|w| !intent_lang::is_stop_word_for(lang, w))
        .take(8)
        .collect()
}

pub fn classify_complexity(
    query: &str,
    classification: &TaskClassification,
) -> super::adaptive::TaskComplexity {
    use super::adaptive::TaskComplexity;

    let q = query.to_lowercase();
    let word_count = q.split_whitespace().count();
    let target_count = classification.targets.len();

    let has_multi_file = target_count >= 3;
    let has_cross_cutting = q.contains("all files")
        || q.contains("across")
        || q.contains("everywhere")
        || q.contains("every")
        || q.contains("migration")
        || q.contains("architecture");

    let is_simple = word_count < 8
        && target_count <= 1
        && matches!(
            classification.task_type,
            TaskType::Generate | TaskType::Config
        );

    if is_simple {
        TaskComplexity::Mechanical
    } else if has_multi_file || has_cross_cutting {
        TaskComplexity::Architectural
    } else {
        TaskComplexity::Standard
    }
}

pub fn detect_multi_intent(query: &str) -> Vec<TaskClassification> {
    let delimiters = [" and then ", " then ", " also ", " + ", ". "];

    let mut parts: Vec<&str> = vec![query];
    for delim in &delimiters {
        let mut new_parts = Vec::new();
        for part in &parts {
            for sub in part.split(delim) {
                let trimmed = sub.trim();
                if !trimmed.is_empty() {
                    new_parts.push(trimmed);
                }
            }
        }
        parts = new_parts;
    }

    if parts.len() <= 1 {
        return vec![classify(query)];
    }

    parts.iter().map(|part| classify(part)).collect()
}

pub fn format_briefing_header(classification: &TaskClassification) -> String {
    format!(
        "[TASK:{} CONF:{:.0}% TARGETS:{} KW:{}]",
        classification.task_type.as_str(),
        classification.confidence * 100.0,
        if classification.targets.is_empty() {
            "-".to_string()
        } else {
            classification.targets.join(",")
        },
        classification.keywords.join(","),
    )
}

#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
pub enum IntentScope {
    SingleFile,
    MultiFile,
    CrossModule,
    ProjectWide,
}

#[derive(Debug, Clone, serde::Serialize, serde::Deserialize)]
pub struct StructuredIntent {
    pub task_type: TaskType,
    pub confidence: f64,
    pub targets: Vec<String>,
    pub keywords: Vec<String>,
    pub scope: IntentScope,
    pub language_hint: Option<String>,
    pub urgency: f64,
    pub action_verb: Option<String>,
}

impl StructuredIntent {
    pub fn from_query(query: &str) -> Self {
        let classification = classify(query);
        let complexity = classify_complexity(query, &classification);
        let file_targets = classification
            .targets
            .iter()
            .filter(|t| t.contains('.') || t.contains('/'))
            .count();
        let scope = match complexity {
            super::adaptive::TaskComplexity::Mechanical => IntentScope::SingleFile,
            super::adaptive::TaskComplexity::Standard => {
                if file_targets > 1 {
                    IntentScope::MultiFile
                } else {
                    IntentScope::SingleFile
                }
            }
            super::adaptive::TaskComplexity::Architectural => {
                let q = query.to_lowercase();
                if q.contains("all files") || q.contains("everywhere") || q.contains("migration") {
                    IntentScope::ProjectWide
                } else {
                    IntentScope::CrossModule
                }
            }
        };

        let language_hint = detect_language_hint(query, &classification.targets);
        let urgency = detect_urgency(query);
        let action_verb = extract_action_verb(query);

        StructuredIntent {
            task_type: classification.task_type,
            confidence: classification.confidence,
            targets: classification.targets,
            keywords: classification.keywords,
            scope,
            language_hint,
            urgency,
            action_verb,
        }
    }

    pub fn from_file_patterns(touched_files: &[String]) -> Self {
        if touched_files.is_empty() {
            return Self {
                task_type: TaskType::Explore,
                confidence: 0.3,
                targets: Vec::new(),
                keywords: Vec::new(),
                scope: IntentScope::SingleFile,
                language_hint: None,
                urgency: 0.0,
                action_verb: None,
            };
        }

        let has_tests = touched_files
            .iter()
            .any(|f| f.contains("test") || f.contains("spec"));
        let has_config = touched_files.iter().any(|f| {
            let p = std::path::Path::new(f.as_str());
            let is_config_ext = p.extension().is_some_and(|e| {
                e.eq_ignore_ascii_case("toml")
                    || e.eq_ignore_ascii_case("yaml")
                    || e.eq_ignore_ascii_case("yml")
                    || e.eq_ignore_ascii_case("json")
            });
            is_config_ext || f.contains("config") || f.contains(".env")
        });

        let dirs: std::collections::HashSet<&str> = touched_files
            .iter()
            .filter_map(|f| std::path::Path::new(f).parent()?.to_str())
            .collect();

        let task_type = if has_tests && touched_files.len() <= 3 {
            TaskType::Test
        } else if has_config && touched_files.len() <= 2 {
            TaskType::Config
        } else if dirs.len() > 3 {
            TaskType::Refactor
        } else {
            TaskType::Explore
        };

        let scope = match touched_files.len() {
            1 => IntentScope::SingleFile,
            2..=4 => IntentScope::MultiFile,
            _ => IntentScope::CrossModule,
        };

        let language_hint = detect_language_from_files(touched_files);

        Self {
            task_type,
            confidence: 0.5,
            targets: touched_files.to_vec(),
            keywords: Vec::new(),
            scope,
            language_hint,
            urgency: 0.0,
            action_verb: None,
        }
    }

    pub fn from_query_with_session(query: &str, touched_files: &[String]) -> Self {
        let mut intent = Self::from_query(query);

        // Text signals too weak (unknown wording or language)? Behavioral
        // signals from the session outrank a blind Explore fallback (#591).
        if intent.confidence < 0.5 && !touched_files.is_empty() {
            let behavioral = Self::from_file_patterns(touched_files);
            if behavioral.confidence > intent.confidence {
                intent.task_type = behavioral.task_type;
                intent.confidence = behavioral.confidence;
                intent.scope = behavioral.scope;
            }
        }

        if intent.language_hint.is_none() && !touched_files.is_empty() {
            intent.language_hint = detect_language_from_files(touched_files);
        }

        if intent.scope == IntentScope::SingleFile && touched_files.len() > 3 {
            let dirs: std::collections::HashSet<&str> = touched_files
                .iter()
                .filter_map(|f| std::path::Path::new(f).parent()?.to_str())
                .collect();
            if dirs.len() > 2 {
                intent.scope = IntentScope::MultiFile;
            }
        }

        intent
    }

    pub fn format_header(&self) -> String {
        format!(
            "[TASK:{} SCOPE:{} CONF:{:.0}%{}{}]",
            self.task_type.as_str(),
            match self.scope {
                IntentScope::SingleFile => "single",
                IntentScope::MultiFile => "multi",
                IntentScope::CrossModule => "cross",
                IntentScope::ProjectWide => "project",
            },
            self.confidence * 100.0,
            self.language_hint
                .as_ref()
                .map(|l| format!(" LANG:{l}"))
                .unwrap_or_default(),
            if self.urgency > 0.5 { " URGENT" } else { "" },
        )
    }
}

fn detect_language_hint(query: &str, targets: &[String]) -> Option<String> {
    for t in targets {
        let ext = std::path::Path::new(t).extension().and_then(|e| e.to_str());
        match ext {
            Some("rs") => return Some("rust".into()),
            Some("ts" | "tsx") => return Some("typescript".into()),
            Some("js" | "jsx") => return Some("javascript".into()),
            Some("py") => return Some("python".into()),
            Some("go") => return Some("go".into()),
            Some("rb") => return Some("ruby".into()),
            Some("java") => return Some("java".into()),
            Some("swift") => return Some("swift".into()),
            Some("zig") => return Some("zig".into()),
            _ => {}
        }
    }

    let q = query.to_lowercase();
    let lang_keywords: &[(&str, &str)] = &[
        ("rust", "rust"),
        ("python", "python"),
        ("typescript", "typescript"),
        ("javascript", "javascript"),
        ("golang", "go"),
        (" go ", "go"),
        ("ruby", "ruby"),
        ("java ", "java"),
        ("swift", "swift"),
    ];
    for &(kw, lang) in lang_keywords {
        if q.contains(kw) {
            return Some(lang.into());
        }
    }

    None
}

fn detect_language_from_files(files: &[String]) -> Option<String> {
    let mut counts: std::collections::HashMap<&str, usize> = std::collections::HashMap::new();
    for f in files {
        let ext = std::path::Path::new(f)
            .extension()
            .and_then(|e| e.to_str())
            .unwrap_or("");
        let lang = match ext {
            "rs" => "rust",
            "ts" | "tsx" => "typescript",
            "js" | "jsx" => "javascript",
            "py" => "python",
            "go" => "go",
            "rb" => "ruby",
            "java" => "java",
            _ => continue,
        };
        *counts.entry(lang).or_insert(0) += 1;
    }
    counts
        .into_iter()
        .max_by_key(|(_, c)| *c)
        .map(|(l, _)| l.to_string())
}

fn detect_urgency(query: &str) -> f64 {
    let q = query.to_lowercase();
    let urgent_words = [
        "urgent",
        "asap",
        "immediately",
        "critical",
        "hotfix",
        "emergency",
        "blocker",
        "breaking",
    ];
    let hits = urgent_words.iter().filter(|w| q.contains(*w)).count()
        + intent_lang::URGENT_WORDS_I18N
            .iter()
            .filter(|w| q.contains(*w))
            .count();
    (hits as f64 * 0.4).min(1.0)
}

fn extract_action_verb(query: &str) -> Option<String> {
    let verbs = [
        "fix",
        "add",
        "create",
        "implement",
        "refactor",
        "debug",
        "test",
        "write",
        "update",
        "remove",
        "delete",
        "rename",
        "move",
        "extract",
        "split",
        "merge",
        "deploy",
        "review",
        "check",
        "build",
        "generate",
        "optimize",
        "clean",
    ];
    let q = query.to_lowercase();
    let words: Vec<&str> = q.split_whitespace().collect();
    for v in &verbs {
        if words.first() == Some(v) || words.get(1) == Some(v) {
            return Some(v.to_string());
        }
    }
    for v in &verbs {
        if words.contains(v) {
            return Some(v.to_string());
        }
    }
    None
}

#[derive(Debug, Clone, Copy, PartialEq, Eq, serde::Serialize, serde::Deserialize)]
pub enum IntentDimension {
    What,
    How,
    Do,
}

impl IntentDimension {
    pub fn as_str(&self) -> &'static str {
        match self {
            Self::What => "what",
            Self::How => "how",
            Self::Do => "do",
        }
    }
}

#[derive(Debug, Clone, serde::Serialize)]
pub struct IntentRoute {
    pub dimension: IntentDimension,
    pub confidence: f64,
    pub reasoning: String,
}

pub fn route_intent(query: &str, classification: &TaskClassification) -> IntentRoute {
    let dimension = match classification.task_type {
        TaskType::Explore | TaskType::Debug => IntentDimension::What,
        TaskType::Review | TaskType::FixBug | TaskType::Test => IntentDimension::How,
        TaskType::Generate | TaskType::Refactor | TaskType::Deploy | TaskType::Config => {
            IntentDimension::Do
        }
    };
    let complexity = classify_complexity(query, classification);
    let reasoning = format!(
        "{}({}) + {}complexity",
        classification.task_type.as_str(),
        dimension.as_str(),
        match complexity {
            super::adaptive::TaskComplexity::Mechanical => "low ",
            super::adaptive::TaskComplexity::Standard => "",
            super::adaptive::TaskComplexity::Architectural => "high ",
        }
    );
    IntentRoute {
        dimension,
        confidence: classification.confidence,
        reasoning,
    }
}

#[cfg(test)]
pub mod tests {
    use super::*;

    #[test]
    fn classify_fix_bug() {
        let r = classify("fix the bug in entropy.rs where token_entropy returns NaN");
        assert_eq!(r.task_type, TaskType::FixBug);
        assert!(r.confidence > 0.5);
        assert!(r.targets.iter().any(|t| t.contains("entropy.rs")));
    }

    #[test]
    fn classify_generate() {
        let r = classify("add a new function normalized_token_entropy to entropy.rs");
        assert_eq!(r.task_type, TaskType::Generate);
        assert!(r.confidence > 0.5);
    }

    #[test]
    fn classify_refactor() {
        let r = classify("refactor the compression pipeline to split into smaller modules");
        assert_eq!(r.task_type, TaskType::Refactor);
    }

    #[test]
    fn classify_explore() {
        let r = classify("how does the session cache work?");
        assert_eq!(r.task_type, TaskType::Explore);
    }

    #[test]
    fn classify_debug() {
        let r = classify("debug why the compression ratio drops for large files");
        assert_eq!(r.task_type, TaskType::Debug);
    }

    #[test]
    fn classify_test() {
        let r = classify("write unit tests for the token_optimizer module");
        assert_eq!(r.task_type, TaskType::Test);
    }

    #[test]
    fn targets_extract_paths() {
        let r = classify("fix entropy.rs and update core/mod.rs");
        assert!(r.targets.iter().any(|t| t.contains("entropy.rs")));
        assert!(r.targets.iter().any(|t| t.contains("core/mod.rs")));
    }

    #[test]
    fn targets_extract_identifiers() {
        let r = classify("refactor SessionCache to use LRU eviction");
        assert!(r.targets.iter().any(|t| t == "SessionCache"));
    }

    #[test]
    fn fallback_to_explore() {
        let r = classify("xyz qqq bbb");
        assert_eq!(r.task_type, TaskType::Explore);
        assert!(r.confidence < 0.5);
    }

    #[test]
    fn multi_intent_detection() {
        let results = detect_multi_intent("fix the bug in auth.rs and then write unit tests");
        assert!(results.len() >= 2);
        assert_eq!(results[0].task_type, TaskType::FixBug);
        assert_eq!(results[1].task_type, TaskType::Test);
    }

    #[test]
    fn single_intent_no_split() {
        let results = detect_multi_intent("fix the bug in auth.rs");
        assert_eq!(results.len(), 1);
        assert_eq!(results[0].task_type, TaskType::FixBug);
    }

    #[test]
    fn complexity_mechanical() {
        let r = classify("add a comment");
        let c = classify_complexity("add a comment", &r);
        assert_eq!(c, super::super::adaptive::TaskComplexity::Mechanical);
    }

    #[test]
    fn complexity_architectural() {
        let r = classify("refactor auth across all files and update the migration");
        let c = classify_complexity(
            "refactor auth across all files and update the migration",
            &r,
        );
        assert_eq!(c, super::super::adaptive::TaskComplexity::Architectural);
    }

    #[test]
    fn route_explore_is_what() {
        let c = TaskClassification {
            task_type: TaskType::Explore,
            confidence: 0.8,
            targets: vec![],
            keywords: vec!["explore".into()],
        };
        let route = route_intent("explore the codebase", &c);
        assert_eq!(route.dimension, IntentDimension::What);
    }

    #[test]
    fn route_fixbug_is_how() {
        let c = TaskClassification {
            task_type: TaskType::FixBug,
            confidence: 0.9,
            targets: vec!["auth.rs".into()],
            keywords: vec!["fix".into(), "bug".into()],
        };
        let route = route_intent("fix the null pointer bug in auth.rs", &c);
        assert_eq!(route.dimension, IntentDimension::How);
    }

    #[test]
    fn route_generate_is_do() {
        let c = TaskClassification {
            task_type: TaskType::Generate,
            confidence: 0.85,
            targets: vec![],
            keywords: vec!["generate".into()],
        };
        let route = route_intent("generate a new module", &c);
        assert_eq!(route.dimension, IntentDimension::Do);
    }

    #[test]
    fn route_complexity_describes_context_without_model_advice() {
        let c = TaskClassification {
            task_type: TaskType::FixBug,
            confidence: 0.8,
            targets: vec!["auth.rs".into(), "middleware.rs".into()],
            keywords: vec!["fix".into()],
        };
        let route = route_intent("fix auth across all files and update the migration", &c);
        assert_eq!(route.dimension, IntentDimension::How);
        assert!(route.reasoning.contains("high complexity"));
        assert!(
            serde_json::to_value(&route)
                .unwrap()
                .get("model_tier")
                .is_none()
        );
    }

    #[test]
    fn route_low_confidence_preserves_classification() {
        let c = TaskClassification {
            task_type: TaskType::Explore,
            confidence: 0.3,
            targets: vec![],
            keywords: vec![],
        };
        let route = route_intent("something vague", &c);
        assert_eq!(route.dimension, IntentDimension::What);
        assert_eq!(route.confidence, c.confidence);
    }

    /// #591 acceptance: de/fr/es sentences per TaskType, ≥90% hit rate.
    #[test]
    fn classify_multilingual_table() {
        let table: &[(&str, TaskType)] = &[
            // German
            ("behebe den fehler in auth.rs", TaskType::FixBug),
            (
                "erstelle eine neue funktion für das datums-parsing",
                TaskType::Generate,
            ),
            ("räum die funktion auf", TaskType::Refactor),
            (
                "refaktorisiere das modul in kleinere teile",
                TaskType::Refactor,
            ),
            (
                "erkläre wie der session cache funktioniert",
                TaskType::Explore,
            ),
            ("schreibe tests für den parser", TaskType::Test),
            (
                "prüfe ob die validierung korrekt funktioniert",
                TaskType::Test,
            ),
            ("debugge warum der server abstürzt", TaskType::Debug),
            (
                "konfiguriere die umgebungsvariablen für den daemon",
                TaskType::Config,
            ),
            ("veröffentliche die neue version", TaskType::Deploy),
            ("überprüfe die änderungen vor dem merge", TaskType::Review),
            // French
            ("corrige le bug dans le parseur", TaskType::FixBug),
            ("ajoute une fonction de validation", TaskType::Generate),
            ("explique comment fonctionne le cache", TaskType::Explore),
            ("nettoie ce module pour le simplifier", TaskType::Refactor),
            ("vérifie que les tests passent", TaskType::Test),
            ("déploie la nouvelle version", TaskType::Deploy),
            ("pourquoi le serveur plante-t-il", TaskType::Explore),
            // Spanish
            ("corrige el error en el módulo de auth", TaskType::FixBug),
            ("crea una función para parsear fechas", TaskType::Generate),
            ("explica cómo funciona la caché", TaskType::Explore),
            ("agrega soporte para webhooks", TaskType::Generate),
            (
                "muestra dónde se define la configuración",
                TaskType::Explore,
            ),
            ("revisa este pull request", TaskType::Review),
            ("despliega la nueva versión", TaskType::Deploy),
            ("escribe pruebas para el parser", TaskType::Test),
        ];

        let misses: Vec<String> = table
            .iter()
            .filter_map(|(query, expected)| {
                let got = classify(query).task_type;
                (got != *expected).then(|| format!("'{query}': want {expected:?}, got {got:?}"))
            })
            .collect();

        let hit_rate = (table.len() - misses.len()) as f64 / table.len() as f64;
        assert!(
            hit_rate >= 0.9,
            "multilingual hit rate {:.0}% < 90%:\n{}",
            hit_rate * 100.0,
            misses.join("\n")
        );
    }

    /// #591 acceptance: German filler words must not pollute keywords.
    #[test]
    fn keywords_filter_german_fillers() {
        let r = classify("bitte kannst du diese funktion aufräumen, ich möchte sauberen code");
        for filler in ["bitte", "kannst", "möchte", "diese"] {
            assert!(
                !r.keywords.iter().any(|k| k == filler),
                "filler '{filler}' leaked into keywords: {:?}",
                r.keywords
            );
        }
        assert!(r.keywords.iter().any(|k| k == "funktion"));
        assert_eq!(r.task_type, TaskType::Refactor);
    }

    #[test]
    fn targets_filter_capitalized_german_fillers() {
        let r = classify("Bitte behebe den Fehler in der Konfiguration");
        assert!(!r.targets.iter().any(|t| t == "Bitte"), "{:?}", r.targets);
        assert_eq!(r.task_type, TaskType::FixBug);
    }

    #[test]
    fn urgency_detects_german_markers() {
        let intent = StructuredIntent::from_query("dringend: behebe den absturz sofort");
        assert!(intent.urgency > 0.5);
        assert_eq!(intent.task_type, TaskType::FixBug);
        assert!(intent.format_header().contains("URGENT"));
    }

    /// #591: with weak text confidence, session file patterns outrank the
    /// blind Explore fallback.
    #[test]
    fn behavioral_signals_beat_low_confidence_text() {
        let touched = vec![
            "tests/foo_test.rs".to_string(),
            "tests/bar_test.rs".to_string(),
        ];
        let intent = StructuredIntent::from_query_with_session("xyz qqq bbb", &touched);
        assert_eq!(intent.task_type, TaskType::Test);
        assert!((intent.confidence - 0.5).abs() < f64::EPSILON);
    }

    #[test]
    fn german_refactor_gets_diff_output() {
        let intent = StructuredIntent::from_query("räum die funktion auf");
        assert_eq!(intent.task_type, TaskType::Refactor);
        assert_eq!(intent.task_type.output_format(), OutputFormat::DiffOnly);
    }
}