car-reason 0.24.1

//! Action executor — runs individual reasoning actions.
//!
//! Key difference from v1: the `locate` action does REAL codebase search
//! (grep, glob, file reading) instead of asking an LLM to guess. This feeds
//! actual source code into subsequent actions so diagnose/generate_fix have
//! real context to work with.

use std::process::Command;

use car_inference::{GenerateParams, GenerateRequest};

use crate::handle::ReasoningInferenceHandle;
use crate::types::*;
use crate::ReasonError;

/// Maximum bytes of source code to feed as context.
const MAX_CODE_CONTEXT_BYTES: usize = 30_000;

/// Maximum symbols to include in AST-extracted context.
const MAX_CONTEXT_SYMBOLS: usize = 50;

/// Accumulated state threaded between actions.
#[derive(Debug, Clone)]
pub struct AccumulatedContext {
    pub problem: String,
    pub problem_class: ProblemClass,
    pub memory_context: String,
    pub locations: String,
    pub patterns: String,
    pub diagnosis: String,
    pub code: String,
    pub explanation: String,
    /// Actual source code from located files (fed to diagnose/generate_fix).
    pub source_code: String,
}

impl AccumulatedContext {
    pub fn new(problem: &str, memory_context: &str, problem_class: ProblemClass) -> Self {
        Self {
            problem: problem.to_string(),
            problem_class,
            memory_context: memory_context.to_string(),
            locations: String::new(),
            patterns: String::new(),
            diagnosis: String::new(),
            code: String::new(),
            explanation: String::new(),
            source_code: String::new(),
        }
    }

    pub fn integrate(&mut self, outcome: &ActionOutcome) {
        match outcome.action {
            ActionKind::Locate => self.locations = outcome.output.clone(),
            ActionKind::RetrievePatterns => self.patterns = outcome.output.clone(),
            ActionKind::Diagnose => self.diagnosis = outcome.output.clone(),
            ActionKind::GenerateFix => self.code = outcome.output.clone(),
            ActionKind::Explain => self.explanation = outcome.output.clone(),
            ActionKind::Classify | ActionKind::VerifyFix => {}
        }
    }
}

/// Execute just the Locate action (grep + AST). No model calls needed.
/// Useful for testing the codebase search pipeline in isolation.
pub fn execute_locate_only(problem: &str) -> Result<ActionOutcome, ReasonError> {
    let ctx = AccumulatedContext::new(problem, "", crate::types::ProblemClass::BugFix);
    execute_locate(&ctx)
}

/// Execute a single reasoning action.
pub async fn execute_action(
    engine: &dyn ReasoningInferenceHandle,
    action: ActionKind,
    config: &ActionConfig,
    ctx: &AccumulatedContext,
) -> Result<ActionOutcome, ReasonError> {
    match action {
        // RetrievePatterns: uses the graph, not a model
        ActionKind::RetrievePatterns => {
            return Ok(ActionOutcome {
                action,
                model_used: "memgine".into(),
                trace_id: String::new(),
                latency_ms: 0,
                output: ctx.patterns.clone(),
                confidence: 1.0,
                success: true,
            });
        }
        // Locate: REAL codebase search, no LLM needed
        ActionKind::Locate => {
            return execute_locate(ctx);
        }
        // Everything else: LLM-powered
        _ => {}
    }

    // Build the prompt with source code context for diagnose/generate_fix
    let prompt = render_prompt(action, config, ctx);

    // Build context: memory + source code for grounding
    let full_context = if ctx.source_code.is_empty() {
        if ctx.memory_context.is_empty() {
            None
        } else {
            Some(ctx.memory_context.clone())
        }
    } else {
        let mut c = String::new();
        if !ctx.memory_context.is_empty() {
            c.push_str(&ctx.memory_context);
            c.push_str("\n\n");
        }
        c.push_str("## Source Code\n\n");
        c.push_str(&ctx.source_code);
        Some(c)
    };

    // Route each action to the right model tier.
    // Architecture/explanation questions need frontier quality, not fast/cheap.
    let model = match action {
        ActionKind::Classify => {
            // Fast & cheap — local small model
            Some("Qwen3-0.6B".to_string())
        }
        ActionKind::Diagnose | ActionKind::GenerateFix => {
            // Complex reasoning — frontier model (best quality)
            pick_model_for_tier(engine, ModelTier::Frontier).await
        }
        ActionKind::Explain => {
            // Explain quality depends on problem class — architecture/explanation
            // questions need frontier; bug fix explanations can use fast tier
            match ctx.problem_class {
                ProblemClass::Architecture | ProblemClass::Explanation => {
                    pick_model_for_tier(engine, ModelTier::Frontier).await
                }
                _ => pick_model_for_tier(engine, ModelTier::Fast).await,
            }
        }
        ActionKind::VerifyFix => {
            // Verification — fast tier is fine
            pick_model_for_tier(engine, ModelTier::Fast).await
        }
        _ => None,
    };

    let req = GenerateRequest {
        prompt,
        model,
        params: GenerateParams {
            temperature: action.temperature(),
            max_tokens: action.max_tokens(),
            ..Default::default()
        },
        context: full_context,
        tools: None,
        images: None,
        messages: None,
        cache_control: false,
        response_format: None,
        intent: None,
    };

    let result = engine
        .generate_tracked(req)
        .await
        .map_err(|e| ReasonError::InferenceFailed(e.to_string()))?;

    let confidence = assess_confidence(&result.text, action);

    Ok(ActionOutcome {
        action,
        model_used: result.model_used,
        trace_id: result.trace_id,
        latency_ms: result.latency_ms,
        output: result.text,
        confidence,
        success: confidence > 0.2,
    })
}

#[derive(Debug, Clone, Copy)]
enum ModelTier {
    /// Best quality: Opus, GPT-5.3-codex
    Frontier,
    /// Good quality + fast: Sonnet, GPT-4o-mini
    Fast,
}

/// Pick the best available model for a given tier.
async fn pick_model_for_tier(
    engine: &dyn ReasoningInferenceHandle,
    tier: ModelTier,
) -> Option<String> {
    let candidates: &[(&str, &str)] = match tier {
        ModelTier::Frontier => &[
            ("claude-opus-4-7", "ANTHROPIC_API_KEY"),
            ("gpt-5.4", "OPENAI_API_KEY"),
            ("o3", "OPENAI_API_KEY"),
            ("claude-sonnet-4-6", "ANTHROPIC_API_KEY"),
            ("gpt-5.3-codex", "OPENAI_API_KEY"),
            ("gemini-2.5-pro", "GOOGLE_API_KEY"),
        ],
        ModelTier::Fast => &[
            ("gpt-5.4-mini", "OPENAI_API_KEY"),
            ("claude-haiku-4-5", "ANTHROPIC_API_KEY"),
            ("gpt-4.1-mini", "OPENAI_API_KEY"),
            ("gemini-2.5-flash", "GOOGLE_API_KEY"),
            ("claude-sonnet-4-6", "ANTHROPIC_API_KEY"),
            ("o4-mini", "OPENAI_API_KEY"),
        ],
    };

    for (model_name, env_var) in candidates {
        if std::env::var(env_var).is_ok() {
            if engine.find_model_by_name(model_name).await.is_some() {
                return Some(model_name.to_string());
            }
        }
    }

    None // Fall back to adaptive router (local models)
}

/// Execute the locate action by searching the codebase and extracting
/// structured symbols via tree-sitter AST parsing.
///
/// Flow: grep for files -> parse each with tree-sitter -> extract only
/// relevant symbols (matching search terms) -> produce compact context
/// with signatures and focused source instead of 30KB text dumps.
fn execute_locate(ctx: &AccumulatedContext) -> Result<ActionOutcome, ReasonError> {
    let start = std::time::Instant::now();
    let cwd = std::env::current_dir().unwrap_or_default();

    // Extract search terms from the problem
    let search_terms = extract_search_terms(&ctx.problem);

    let mut found_files: Vec<String> = Vec::new();

    // 1. Grep for relevant terms in source files
    for term in &search_terms {
        if let Ok(output) = Command::new("grep")
            .args([
                "-rl",
                "--include=*.rs",
                "--include=*.py",
                "--include=*.ts",
                "--include=*.js",
                "--include=*.go",
                "-i",
                term,
                ".",
            ])
            .current_dir(&cwd)
            .output()
        {
            if output.status.success() {
                let files = String::from_utf8_lossy(&output.stdout);
                for f in files.lines() {
                    let f = f.trim().to_string();
                    if !f.is_empty() && !found_files.contains(&f) {
                        found_files.push(f);
                    }
                }
            }
        }
    }

    // 2. Also check for files mentioned in the problem text
    let problem_lower = ctx.problem.to_lowercase();
    if let Ok(output) = Command::new("find")
        .args([
            ".", "-name", "*.rs", "-o", "-name", "*.py", "-o", "-name", "*.ts", "-o", "-name",
            "*.go", "-o", "-name", "*.js",
        ])
        .current_dir(&cwd)
        .output()
    {
        if output.status.success() {
            let all_files = String::from_utf8_lossy(&output.stdout);
            for f in all_files.lines() {
                let f = f.trim();
                if f.is_empty() {
                    continue;
                }
                let basename = f.rsplit('/').next().unwrap_or(f);
                if problem_lower.contains(&basename.to_lowercase())
                    && !found_files.contains(&f.to_string())
                {
                    found_files.insert(0, f.to_string());
                }
            }
        }
    }

    found_files.truncate(30);

    // 3. Parse files with tree-sitter and extract relevant symbols
    let mut source_content = String::new();
    let mut total_bytes = 0;
    let mut files_read = 0;
    let mut total_symbols = 0;

    for file_path in &found_files {
        if total_bytes >= MAX_CODE_CONTEXT_BYTES || total_symbols >= MAX_CONTEXT_SYMBOLS {
            break;
        }

        let full_path = cwd.join(file_path);
        let content = match std::fs::read_to_string(&full_path) {
            Ok(c) => c,
            Err(_) => continue,
        };

        // Try AST-based extraction first
        if let Some(parsed) = car_ast::parse_file(&content, file_path) {
            // Find symbols matching search terms
            let mut relevant: Vec<&car_ast::Symbol> = Vec::new();
            for term in &search_terms {
                relevant.extend(parsed.find_symbol_fuzzy(term));
            }

            // Deduplicate by name+kind
            relevant.sort_by(|a, b| a.span.start_byte.cmp(&b.span.start_byte));
            relevant.dedup_by(|a, b| a.name == b.name && a.kind == b.kind);

            if relevant.is_empty() {
                // No matching symbols but file was found by grep — include
                // all top-level signatures as a summary
                for sym in &parsed.symbols {
                    relevant.push(sym);
                }
            }

            if !relevant.is_empty() {
                let remaining = MAX_CONTEXT_SYMBOLS - total_symbols;
                if relevant.len() > remaining {
                    relevant.truncate(remaining);
                }

                source_content.push_str(&format!(
                    "### {} (AST: {} symbols)\n",
                    file_path,
                    relevant.len()
                ));

                // Include imports summary
                if !parsed.imports.is_empty() {
                    source_content.push_str("Imports: ");
                    let import_summary: Vec<_> = parsed
                        .imports
                        .iter()
                        .take(5)
                        .map(|i| i.path.as_str())
                        .collect();
                    source_content.push_str(&import_summary.join(", "));
                    if parsed.imports.len() > 5 {
                        source_content.push_str(&format!(" (+{} more)", parsed.imports.len() - 5));
                    }
                    source_content.push('\n');
                }

                // Include each relevant symbol's signature + source
                for sym in &relevant {
                    // Signature line
                    source_content.push_str(&format!("\n[{:?}] {}\n", sym.kind, sym.signature));

                    if let Some(doc) = &sym.doc_comment {
                        source_content
                            .push_str(&format!("  doc: {}\n", doc.lines().next().unwrap_or("")));
                    }

                    // Include children (methods) signatures
                    for child in &sym.children {
                        source_content
                            .push_str(&format!("  [{:?}] {}\n", child.kind, child.signature));
                    }

                    // Include the actual source for the symbol
                    let sym_source = car_ast::extract_source(sym, &content);
                    let remaining_bytes = MAX_CODE_CONTEXT_BYTES - total_bytes;
                    if sym_source.len() <= remaining_bytes {
                        source_content.push_str("```\n");
                        source_content.push_str(&sym_source);
                        source_content.push_str("\n```\n");
                        total_bytes += sym_source.len();
                    } else if remaining_bytes > 200 {
                        // Truncate but include what we can
                        let trunc = &sym_source[..sym_source.floor_char_boundary(remaining_bytes)];
                        source_content.push_str("```\n");
                        source_content.push_str(trunc);
                        source_content.push_str("\n// ... (truncated)\n```\n");
                        total_bytes += trunc.len();
                    }

                    total_symbols += 1;
                }

                source_content.push('\n');
                files_read += 1;
            }
        } else {
            // Fallback: language not supported by tree-sitter, dump raw text
            let remaining = MAX_CODE_CONTEXT_BYTES - total_bytes;
            if remaining < 200 {
                continue;
            }
            let truncated = if content.len() > remaining {
                &content[..content.floor_char_boundary(remaining)]
            } else {
                &content
            };
            source_content.push_str(&format!(
                "### {} (raw)\n```\n{}\n```\n\n",
                file_path, truncated
            ));
            total_bytes += truncated.len();
            files_read += 1;
        }
    }

    let latency_ms = start.elapsed().as_millis() as u64;

    let locations_summary = format!(
        "Found {} relevant files ({} read, {} symbols, {} bytes):\n{}",
        found_files.len(),
        files_read,
        total_symbols,
        total_bytes,
        found_files
            .iter()
            .enumerate()
            .map(|(i, f)| format!("  {}. {}", i + 1, f))
            .collect::<Vec<_>>()
            .join("\n"),
    );

    let output = format!(
        "{}\n\n---SOURCE_CODE_START---\n{}",
        locations_summary, source_content
    );

    Ok(ActionOutcome {
        action: ActionKind::Locate,
        model_used: "codebase-search+ast".into(),
        trace_id: String::new(),
        latency_ms,
        output,
        confidence: if files_read > 0 { 0.9 } else { 0.3 },
        success: files_read > 0,
    })
}

/// Extract meaningful search terms from a problem description.
fn extract_search_terms(problem: &str) -> Vec<String> {
    let mut terms = Vec::new();

    // Extract identifiers: words that look like code (camelCase, snake_case, PascalCase)
    for word in problem.split_whitespace() {
        let clean: String = word
            .chars()
            .filter(|c| c.is_alphanumeric() || *c == '_')
            .collect();
        if clean.len() < 3 {
            continue;
        }

        // Skip common English words
        let lower = clean.to_lowercase();
        let noise = [
            "the", "and", "for", "with", "this", "that", "from", "what", "how", "would", "should",
            "could", "does", "not", "all", "are", "but", "its", "has", "have", "was", "were",
            "will", "can", "use", "using", "used", "add", "fix", "bug", "error",
        ];
        if noise.contains(&lower.as_str()) {
            continue;
        }

        // Prefer code-like identifiers
        let has_underscore = clean.contains('_');
        let has_camel =
            clean.chars().any(|c| c.is_uppercase()) && clean.chars().any(|c| c.is_lowercase());
        let is_code_term = has_underscore || has_camel || clean.len() > 6;

        if is_code_term {
            terms.insert(0, clean); // Code terms first
        } else {
            terms.push(clean);
        }
    }

    terms.truncate(10); // Top 10 search terms
    terms
}

/// Render the prompt template with accumulated context.
fn render_prompt(action: ActionKind, config: &ActionConfig, ctx: &AccumulatedContext) -> String {
    let template = if config.prompt_template.is_empty() {
        default_prompt(action)
    } else {
        config.prompt_template.clone()
    };

    let diagnosis_section = if ctx.diagnosis.is_empty() {
        String::new()
    } else {
        format!("Diagnosis: {}\n\n", ctx.diagnosis)
    };

    let fix_section = if ctx.code.is_empty() {
        String::new()
    } else {
        format!("Proposed fix:\n{}\n\n", ctx.code)
    };

    template
        .replace("{problem}", &ctx.problem)
        .replace("{problem_class}", &ctx.problem_class.to_string())
        .replace("{context}", &ctx.memory_context)
        .replace("{locations}", &ctx.locations)
        .replace("{patterns}", &ctx.patterns)
        .replace("{diagnosis}", &ctx.diagnosis)
        .replace("{code}", &ctx.code)
        .replace("{diagnosis_section}", &diagnosis_section)
        .replace("{fix_section}", &fix_section)
}

/// Default prompts — these are grounded in real code context now.
fn default_prompt(action: ActionKind) -> String {
    match action {
        ActionKind::Classify => "Classify this code problem into one category: bug_fix, refactor, architecture, new_feature, performance, test_writing, or explanation.\n\nProblem: {problem}\n\nCategory:".into(),
        ActionKind::Locate => String::new(), // Not used — locate is code-driven
        ActionKind::RetrievePatterns => String::new(),
        ActionKind::Diagnose => "You are an expert code analyst. Analyze the root cause of this problem using the source code provided in the context.\n\nProblem: {problem}\n\nLocated files:\n{locations}\n\nProvide a precise root cause analysis referencing specific functions, types, and line numbers from the source code.".into(),
        ActionKind::GenerateFix => "You are an expert programmer. Generate a precise code fix for this problem. Reference the actual source code from the context.\n\nProblem: {problem}\n\nDiagnosis: {diagnosis}\n\nGenerate the fix as a code block with the file path. Only change what's necessary.".into(),
        ActionKind::VerifyFix => "Review this fix for correctness. Check for: off-by-one errors, null/None handling, type mismatches, missing edge cases.\n\nProposed fix:\n{code}\n\nVerification (PASS or FAIL with reason):".into(),
        ActionKind::Explain => "Explain clearly and concisely what was wrong and how the fix works.\n\nProblem: {problem}\n{diagnosis_section}{fix_section}Explanation:".into(),
    }
}

/// Heuristic confidence assessment.
fn assess_confidence(output: &str, action: ActionKind) -> f64 {
    if output.trim().is_empty() {
        return 0.0;
    }
    let length = output.len();

    match action {
        ActionKind::Classify => {
            if length < 50 {
                0.9
            } else {
                0.5
            }
        }
        ActionKind::Locate => {
            let has_paths = output.contains('/') || output.contains('.');
            if has_paths {
                0.8
            } else {
                0.4
            }
        }
        ActionKind::Diagnose => {
            if length > 200 {
                0.8
            } else if length > 50 {
                0.6
            } else {
                0.3
            }
        }
        ActionKind::GenerateFix => {
            let has_code =
                output.contains("```") || output.contains("fn ") || output.contains("def ");
            if has_code {
                0.7
            } else {
                0.3
            }
        }
        ActionKind::VerifyFix => {
            let upper = output.to_uppercase();
            if upper.contains("PASS") {
                0.9
            } else if upper.contains("FAIL") {
                0.3
            } else {
                0.5
            }
        }
        ActionKind::Explain => {
            if length > 100 {
                0.8
            } else {
                0.5
            }
        }
        ActionKind::RetrievePatterns => 1.0,
    }
}