a3s_code_core/

prompts.rs

// Prompt Registry
//
// Central registry for all system prompts and prompt templates used in A3S Code.
// Every LLM-facing prompt is externalized here as a compile-time `include_str!`
// so the full agentic design is visible in one place.
//
// Directory layout:
//   prompts/
//   ├── subagent_explore.md         — Explore subagent system prompt
//   ├── subagent_plan.md            — Plan subagent system prompt
//   ├── subagent_title.md           — Title generation subagent prompt
//   ├── subagent_summary.md         — Summary generation subagent prompt
//   ├── context_compact.md          — Context compaction / summarization
//   ├── title_generate.md           — Session title generation
//   ├── llm_plan_system.md          — LLM planner: plan creation (JSON)
//   ├── llm_goal_extract_system.md  — LLM planner: goal extraction (JSON)
//   ├── llm_goal_check_system.md    — LLM planner: goal achievement (JSON)
//   ├── team_lead.md                — Team lead decomposition prompt
//   ├── team_reviewer.md            — Team reviewer validation prompt
//   └── skills_catalog_header.md    — Skill catalog system prompt header

// ============================================================================
// Default System Prompt
// ============================================================================

use crate::llm::LlmClient;
use anyhow::Context;

/// Default agentic system prompt — injected when no system prompt is configured.
///
/// Instructs the LLM to behave as an autonomous coding agent: use tools to act,
/// verify results, and keep working until the task is fully complete.
pub const SYSTEM_DEFAULT: &str = include_str!("../prompts/system_default.md");

/// Continuation message — injected as a user turn when the LLM stops without
/// completing the task (i.e. stops calling tools mid-task).
pub const CONTINUATION: &str = include_str!("../prompts/continuation.md");

// ============================================================================
// Subagent Prompts
// ============================================================================

/// Explore subagent — read-only codebase exploration
pub const SUBAGENT_EXPLORE: &str = include_str!("../prompts/subagent_explore.md");

/// Plan subagent — read-only planning and analysis
pub const SUBAGENT_PLAN: &str = include_str!("../prompts/subagent_plan.md");

/// Title subagent — generate concise conversation title
pub const SUBAGENT_TITLE: &str = include_str!("../prompts/subagent_title.md");

/// Summary subagent — summarize conversation key points
pub const SUBAGENT_SUMMARY: &str = include_str!("../prompts/subagent_summary.md");

// ============================================================================
// Session — Context Compaction
// ============================================================================

/// User template for context compaction. Placeholder: `{conversation}`
pub const CONTEXT_COMPACT: &str = include_str!("../prompts/context_compact.md");

/// Prefix for compacted summary messages
pub const CONTEXT_SUMMARY_PREFIX: &str = include_str!("../prompts/context_summary_prefix.md");

// ============================================================================
// Session — Title Generation
// ============================================================================

/// User template for session title generation. Placeholder: `{conversation}`
#[allow(dead_code)]
pub const TITLE_GENERATE: &str = include_str!("../prompts/title_generate.md");

// ============================================================================
// LLM Planner — JSON-structured prompts
// ============================================================================

/// System prompt for LLM planner: plan creation (JSON output)
pub const LLM_PLAN_SYSTEM: &str = include_str!("../prompts/llm_plan_system.md");

/// System prompt for LLM planner: goal extraction (JSON output)
pub const LLM_GOAL_EXTRACT_SYSTEM: &str = include_str!("../prompts/llm_goal_extract_system.md");

/// System prompt for LLM planner: goal achievement check (JSON output)
pub const LLM_GOAL_CHECK_SYSTEM: &str = include_str!("../prompts/llm_goal_check_system.md");

// ============================================================================
// Plan Execution (inline templates — no file needed)
// ============================================================================

/// Template for initial plan execution message
pub const PLAN_EXECUTE_GOAL: &str = include_str!("../prompts/plan_execute_goal.md");

/// Template for per-step execution prompt
pub const PLAN_EXECUTE_STEP: &str = include_str!("../prompts/plan_execute_step.md");

/// Template for fallback plan step description
pub const PLAN_FALLBACK_STEP: &str = include_str!("../prompts/plan_fallback_step.md");

/// Template for merging results from parallel step execution
pub const PLAN_PARALLEL_RESULTS: &str = include_str!("../prompts/plan_parallel_results.md");

/// Team lead prompt for decomposing a goal into worker tasks.
pub const TEAM_LEAD: &str = include_str!("../prompts/team_lead.md");

/// Team reviewer prompt for approving or rejecting completed tasks.
pub const TEAM_REVIEWER: &str = include_str!("../prompts/team_reviewer.md");

/// Skill catalog header injected before listing available skill names/descriptions.
pub const SKILLS_CATALOG_HEADER: &str = include_str!("../prompts/skills_catalog_header.md");

// ============================================================================
// Side Question (btw)
// ============================================================================

/// System prompt for `/btw` ephemeral side questions.
///
/// Used by [`crate::agent_api::AgentSession::btw()`] — the answer is never
/// added to conversation history.
pub const BTW_SYSTEM: &str = include_str!("../prompts/btw_system.md");

// ============================================================================
// Verification Agent
// ============================================================================

/// Verification agent — adversarial specialist that tries to break code
pub const AGENT_VERIFICATION: &str = include_str!("../prompts/agent_verification.md");

/// Tool restrictions for verification agent
pub const AGENT_VERIFICATION_RESTRICTIONS: &str =
    include_str!("../prompts/agent_verification_restrictions.md");

// ============================================================================
// Intent Classification
// ============================================================================

/// System prompt for LLM-based intent classification
pub const INTENT_CLASSIFY_SYSTEM: &str = include_str!("../prompts/intent_classify_system.md");

// ============================================================================
// Session Memory
// ============================================================================

/// Session memory template with structured sections
pub const SESSION_MEMORY_TEMPLATE: &str = include_str!("../prompts/system_session_memory.md");

// ============================================================================
// Prompt Suggestion Service
// ============================================================================

/// Prompt suggestion service with filtering rules
pub const PROMPT_SUGGESTION: &str = include_str!("../prompts/service_prompt_suggestion.md");

// ============================================================================
// Undercover Mode
// ============================================================================

/// Undercover mode instructions for commit/PR prompts
pub const UNDERCOVER_INSTRUCTIONS: &str = include_str!("../prompts/undercover_instructions.md");

// ============================================================================
// Planning Mode (Auto-Detection)
// ============================================================================

use serde::{Deserialize, Serialize};

/// Planning mode — controls when planning phase is used.
///
/// When set to `Auto` (the default), the system detects from the user's
/// message whether planning should be enabled. When explicitly `Enabled`,
/// planning runs on every execution. When `Disabled`, planning is skipped.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default, Serialize, Deserialize)]
pub enum PlanningMode {
    /// Automatically detect from message content — enables planning when the
    /// message contains planning-related keywords (plan, design, architecture).
    /// Also automatically uses `AgentStyle::Plan` detection.
    #[default]
    Auto,
    /// Explicitly disabled — never use planning phase.
    Disabled,
    /// Explicitly enabled — always use planning phase.
    Enabled,
}

impl PlanningMode {
    /// Returns true if planning should be used based on this mode and message.
    pub fn should_plan(&self, message: &str) -> bool {
        match self {
            PlanningMode::Auto => AgentStyle::detect_from_message(message).requires_planning(),
            PlanningMode::Enabled => true,
            PlanningMode::Disabled => false,
        }
    }
}

// ============================================================================
// Agent Style (Intent-Based Prompt Selection)
// ============================================================================

/// Agent style — determines which system prompt template is used.
///
/// Each style has a different focus and behavior, selected based on the user's
/// apparent intent from their message.
#[derive(Debug, Clone, Copy, PartialEq, Eq, Default)]
pub enum AgentStyle {
    /// Default — general purpose coding agent for research and multi-step tasks.
    #[default]
    GeneralPurpose,
    /// Read-only planning and architecture analysis.
    /// Prohibited from modifying files, focuses on design and planning.
    Plan,
    /// Adversarial verification specialist — tries to break code, not confirm it works.
    Verification,
    /// Fast file search and codebase exploration.
    /// Read-only, optimized for finding files and patterns quickly.
    Explore,
    /// Code review focused — analyzes code quality, best practices, potential issues.
    CodeReview,
}

/// Detection confidence level for style detection.
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
pub enum DetectionConfidence {
    /// High confidence — very specific keywords, skip LLM classification.
    High,
    /// Medium confidence — some indicators present, LLM classification helpful.
    Medium,
    /// Low confidence — no clear indicators, LLM classification recommended.
    Low,
}

impl AgentStyle {
    /// Returns the base system prompt for this style.
    pub fn base_prompt(&self) -> &'static str {
        match self {
            AgentStyle::GeneralPurpose => SYSTEM_DEFAULT,
            AgentStyle::Plan => SUBAGENT_PLAN,
            AgentStyle::Verification => AGENT_VERIFICATION,
            AgentStyle::Explore => SUBAGENT_EXPLORE,
            AgentStyle::CodeReview => SYSTEM_DEFAULT, // Uses general purpose with review guidelines
        }
    }

    /// Returns style-specific guidelines if any.
    pub fn guidelines(&self) -> Option<&'static str> {
        match self {
            AgentStyle::GeneralPurpose => None,
            AgentStyle::Plan => None, // Already embedded in subagent_plan.md
            AgentStyle::Verification => None, // Already embedded in agent_verification.md
            AgentStyle::Explore => None, // Already embedded in subagent_explore.md
            AgentStyle::CodeReview => Some(CODE_REVIEW_GUIDELINES),
        }
    }

    /// Returns a one-line description of this style.
    pub fn description(&self) -> &'static str {
        match self {
            AgentStyle::GeneralPurpose => {
                "General purpose coding agent for research and multi-step tasks"
            }
            AgentStyle::Plan => "Read-only planning and architecture analysis agent",
            AgentStyle::Verification => "Adversarial verification specialist — tries to break code",
            AgentStyle::Explore => "Fast read-only file search and codebase exploration agent",
            AgentStyle::CodeReview => "Code review focused — analyzes quality and best practices",
        }
    }

    /// Returns true if this style benefits from a planning phase.
    ///
    /// Planning is beneficial for styles that involve multi-step execution
    /// or where a structured approach improves outcomes.
    pub fn requires_planning(&self) -> bool {
        matches!(self, AgentStyle::Plan | AgentStyle::GeneralPurpose)
    }

    /// Detects the most appropriate agent style based on user message content,
    /// along with a confidence level.
    ///
    /// Use this for fast-path keyword matching. When confidence is [`Low`],
    /// consider using [`detect_with_llm`](AgentStyle::detect_with_llm) for
    /// more accurate classification.
    pub fn detect_with_confidence(message: &str) -> (Self, DetectionConfidence) {
        let lower = message.to_lowercase();

        // === HIGH CONFIDENCE: Very specific patterns ===

        // Strong verification indicators
        if lower.contains("try to break")
            || lower.contains("find vulnerabilities")
            || lower.contains("adversarial")
            || lower.contains("security audit")
        {
            return (AgentStyle::Verification, DetectionConfidence::High);
        }

        // Strong plan indicators
        if lower.contains("help me plan")
            || lower.contains("help me design")
            || lower.contains("create a plan")
            || lower.contains("implementation plan")
            || lower.contains("step-by-step plan")
        {
            return (AgentStyle::Plan, DetectionConfidence::High);
        }

        // Strong exploration indicators
        if lower.contains("find all files")
            || lower.contains("search for all")
            || lower.contains("locate all")
        {
            return (AgentStyle::Explore, DetectionConfidence::High);
        }

        // === MEDIUM CONFIDENCE: Specific but less definitive ===

        // Verification keywords
        if lower.contains("verify")
            || lower.contains("verification")
            || lower.contains("break")
            || lower.contains("debug")
            || lower.contains("test")
            || lower.contains("check if")
        {
            return (AgentStyle::Verification, DetectionConfidence::Medium);
        }

        // Plan keywords
        if lower.contains("plan")
            || lower.contains("design")
            || lower.contains("architecture")
            || lower.contains("approach")
        {
            return (AgentStyle::Plan, DetectionConfidence::Medium);
        }

        // Explore keywords
        if lower.contains("find")
            || lower.contains("search")
            || lower.contains("where is")
            || lower.contains("where's")
            || lower.contains("locate")
            || lower.contains("explore")
            || lower.contains("look for")
        {
            return (AgentStyle::Explore, DetectionConfidence::Medium);
        }

        // Code review keywords
        if lower.contains("review")
            || lower.contains("code review")
            || lower.contains("analyze")
            || lower.contains("assess")
            || lower.contains("quality")
            || lower.contains("best practice")
        {
            return (AgentStyle::CodeReview, DetectionConfidence::Medium);
        }

        // No clear indicators
        (AgentStyle::GeneralPurpose, DetectionConfidence::Low)
    }

    /// Detects the most appropriate agent style based on user message content.
    ///
    /// This is a simple keyword-based heuristic. For more precise control,
    /// users can explicitly set the style via `SystemPromptSlots::with_style()`.
    pub fn detect_from_message(message: &str) -> Self {
        Self::detect_with_confidence(message).0
    }

    /// Classifies user intent using LLM when keyword confidence is low.
    ///
    /// This is called when [`detect_with_confidence`] returns [`Low`] confidence,
    /// indicating the message doesn't have clear keyword indicators.
    ///
    /// Uses a lightweight classification prompt that returns a single word.
    pub async fn detect_with_llm(llm: &dyn LlmClient, message: &str) -> anyhow::Result<Self> {
        use crate::llm::Message;

        let system = INTENT_CLASSIFY_SYSTEM;
        let messages = vec![Message::user(message)];

        let response = llm
            .complete(&messages, Some(system), &[])
            .await
            .context("LLM intent classification failed")?;

        let text = response.text().trim().to_lowercase();

        let style = match text.as_str() {
            "plan" => AgentStyle::Plan,
            "explore" => AgentStyle::Explore,
            "verification" => AgentStyle::Verification,
            "codereview" | "code review" => AgentStyle::CodeReview,
            _ => AgentStyle::GeneralPurpose,
        };

        Ok(style)
    }
}

/// Code review guidelines — appended when CodeReview style is selected.
const CODE_REVIEW_GUIDELINES: &str = r#"## Code Review Focus

When reviewing code, pay attention to:

1. **Correctness** — Does the code do what it's supposed to do? Are edge cases handled?
2. **Security** — Are there potential vulnerabilities (injection, auth bypass, data exposure)?
3. **Performance** — Are there obvious inefficiencies (N+1 queries, unnecessary allocations)?
4. **Maintainability** — Is the code readable? Are names clear? Is there appropriate documentation?
5. **Best Practices** — Does it follow language/framework conventions? Are there anti-patterns?

Be specific in your review. Quote the actual code you're referring to. Suggest concrete improvements with examples where possible.

Remember: your job is to find issues, not to be nice. A code review that finds nothing is a missed opportunity."#;

// ============================================================================
// System Prompt Slots
// ============================================================================

/// Slot-based system prompt customization with intent-based style selection.
///
/// Users can customize specific parts of the system prompt without overriding
/// the core agentic capabilities (tool usage, autonomous behavior, completion
/// criteria). The default agentic core is ALWAYS included.
///
/// ## Assembly Order
///
/// ```text
/// [role]            ← Custom identity/role (e.g. "You are a Python expert")
/// [CORE]            ← Always present: Core Behaviour + Tool Usage Strategy + Completion Criteria
/// [guidelines]      ← Custom coding rules / constraints
/// [response_style]  ← Custom response format (replaces default Response Format section)
/// [extra]           ← Freeform additional instructions
/// ```
///
/// ## Intent-Based Selection
///
/// When `style` is left as `AgentStyle::GeneralPurpose` (the default), the
/// system will attempt to detect the user's intent from their first message and
/// automatically select an appropriate style. To override this behavior, explicitly
/// set the `style` field.
#[derive(Debug, Clone, Default)]
pub struct SystemPromptSlots {
    /// Agent style — determines which base prompt template is used.
    ///
    /// When `None` (default), the style is auto-detected from the user's message.
    /// Explicitly set this to force a particular style regardless of message content.
    pub style: Option<AgentStyle>,

    /// Custom role/identity prepended before the core prompt.
    ///
    /// Example: "You are a senior Python developer specializing in FastAPI."
    /// When set, replaces the default "You are A3S Code, an expert AI coding agent" line.
    pub role: Option<String>,

    /// Custom coding guidelines appended after the core prompt sections.
    ///
    /// Example: "Always use type hints. Follow PEP 8. Prefer dataclasses over dicts."
    pub guidelines: Option<String>,

    /// Custom response style that replaces the default "Response Format" section.
    ///
    /// When `None`, the default response format is used.
    pub response_style: Option<String>,

    /// Freeform extra instructions appended at the very end.
    ///
    /// This is the backward-compatible slot: setting `system_prompt` in the old API
    /// maps to this field.
    pub extra: Option<String>,
}

/// The default role line in SYSTEM_DEFAULT that gets replaced when `role` slot is set.
const DEFAULT_ROLE_LINE: &str = include_str!("../prompts/system_default_role_line.md");

/// The default response format section.
const DEFAULT_RESPONSE_FORMAT: &str = include_str!("../prompts/system_default_response_format.md");

impl SystemPromptSlots {
    /// Build the final system prompt by assembling slots around the core prompt.
    ///
    /// The core agentic behavior (Core Behaviour, Tool Usage Strategy, Completion
    /// Criteria) is always preserved. Users can only customize the edges.
    ///
    /// Note: This uses `AgentStyle::GeneralPurpose` as the base. Use
    /// `build_with_message()` to enable automatic intent-based style detection.
    pub fn build(&self) -> String {
        self.build_with_style(self.style.unwrap_or_default())
    }

    /// Build the final system prompt, auto-detecting style from the initial message.
    ///
    /// If `self.style` is explicitly set, that style is used regardless of message content.
    /// Otherwise, the style is detected from `initial_message` using keyword analysis.
    pub fn build_with_message(&self, initial_message: &str) -> String {
        let style = self
            .style
            .unwrap_or_else(|| AgentStyle::detect_from_message(initial_message));
        self.build_with_style(style)
    }

    /// Build the prompt with an explicitly specified style.
    fn build_with_style(&self, style: AgentStyle) -> String {
        let mut parts: Vec<String> = Vec::new();

        // Normalize line endings: strip \r so string matching works on Windows
        // where include_str! may produce \r\n if the file has CRLF endings.
        let base_prompt = style.base_prompt().replace('\r', "");
        let default_role_line = DEFAULT_ROLE_LINE.replace('\r', "");
        let default_response_format = DEFAULT_RESPONSE_FORMAT.replace('\r', "");

        // 1. Role: for GeneralPurpose, replace the default role line.
        // For other styles (Plan, Explore, Verification), prepend custom role since
        // those prompts have their own identity embedded.
        let core = if let Some(ref role) = self.role {
            if style == AgentStyle::GeneralPurpose {
                let custom_role = format!(
                    "{}. You operate in an agentic loop: you\nthink, use tools, observe results, and keep working until the task is fully complete.",
                    role.trim_end_matches('.')
                );
                base_prompt.replace(&default_role_line, &custom_role)
            } else {
                // Prepend custom role for other styles
                format!("{}\n\n{}", role, base_prompt)
            }
        } else {
            base_prompt
        };

        // 2. Core: strip the default response format section if custom one is provided
        let core = if self.response_style.is_some() {
            core.replace(&default_response_format, "")
                .trim_end()
                .to_string()
        } else {
            core.trim_end().to_string()
        };

        parts.push(core);

        // 3. Custom response style (replaces default Response Format)
        if let Some(ref style) = self.response_style {
            parts.push(format!("## Response Format\n\n{}", style));
        }

        // 4. Guidelines: style-specific + custom
        let style_guidelines = style.guidelines();
        if style_guidelines.is_some() || self.guidelines.is_some() {
            let mut guidelines_parts = Vec::new();
            if let Some(sg) = style_guidelines {
                guidelines_parts.push(sg.to_string());
            }
            if let Some(ref g) = self.guidelines {
                guidelines_parts.push(g.clone());
            }
            parts.push(format!(
                "## Guidelines\n\n{}",
                guidelines_parts.join("\n\n")
            ));
        }

        // 5. Extra (freeform, backward-compatible with old system_prompt)
        if let Some(ref extra) = self.extra {
            parts.push(extra.clone());
        }

        parts.join("\n\n")
    }

    /// Create slots from a legacy full system prompt string.
    ///
    /// For backward compatibility: the entire string is placed in the `extra` slot,
    /// and the default core prompt is still prepended.
    pub fn from_legacy(prompt: String) -> Self {
        Self {
            extra: Some(prompt),
            ..Default::default()
        }
    }

    /// Returns true if all slots are empty (use pure default prompt).
    pub fn is_empty(&self) -> bool {
        self.style.is_none()
            && self.role.is_none()
            && self.guidelines.is_none()
            && self.response_style.is_none()
            && self.extra.is_none()
    }

    /// Set the agent style explicitly.
    pub fn with_style(mut self, style: AgentStyle) -> Self {
        self.style = Some(style);
        self
    }

    /// Set the role/identity.
    pub fn with_role(mut self, role: impl Into<String>) -> Self {
        self.role = Some(role.into());
        self
    }

    /// Set custom guidelines.
    pub fn with_guidelines(mut self, guidelines: impl Into<String>) -> Self {
        self.guidelines = Some(guidelines.into());
        self
    }

    /// Set custom response style.
    pub fn with_response_style(mut self, style: impl Into<String>) -> Self {
        self.response_style = Some(style.into());
        self
    }

    /// Set extra instructions (backward-compatible with old system_prompt).
    pub fn with_extra(mut self, extra: impl Into<String>) -> Self {
        self.extra = Some(extra.into());
        self
    }
}

// ============================================================================
// Helper Functions
// ============================================================================

/// Render a template by replacing `{key}` placeholders with values
pub fn render(template: &str, vars: &[(&str, &str)]) -> String {
    let mut result = template.to_string();
    for (key, value) in vars {
        result = result.replace(&format!("{{{}}}", key), value);
    }
    result
}

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

    #[test]
    fn test_all_prompts_loaded() {
        // Verify all prompts are non-empty at compile time
        assert!(!SYSTEM_DEFAULT.is_empty());
        assert!(!CONTINUATION.is_empty());
        assert!(!SUBAGENT_EXPLORE.is_empty());
        assert!(!SUBAGENT_PLAN.is_empty());
        assert!(!SUBAGENT_TITLE.is_empty());
        assert!(!SUBAGENT_SUMMARY.is_empty());
        assert!(!CONTEXT_COMPACT.is_empty());
        assert!(!TITLE_GENERATE.is_empty());
        assert!(!LLM_PLAN_SYSTEM.is_empty());
        assert!(!LLM_GOAL_EXTRACT_SYSTEM.is_empty());
        assert!(!LLM_GOAL_CHECK_SYSTEM.is_empty());
        assert!(!TEAM_LEAD.is_empty());
        assert!(!TEAM_REVIEWER.is_empty());
        assert!(!SKILLS_CATALOG_HEADER.is_empty());
        assert!(!BTW_SYSTEM.is_empty());
        assert!(!PLAN_EXECUTE_GOAL.is_empty());
        assert!(!PLAN_EXECUTE_STEP.is_empty());
        assert!(!PLAN_FALLBACK_STEP.is_empty());
        assert!(!PLAN_PARALLEL_RESULTS.is_empty());
    }

    #[test]
    fn test_render_template() {
        let result = render(
            PLAN_EXECUTE_GOAL,
            &[("goal", "Build app"), ("steps", "1. Init")],
        );
        assert!(result.contains("Build app"));
        assert!(!result.contains("{goal}"));
    }

    #[test]
    fn test_render_multiple_placeholders() {
        let template = "Goal: {goal}\nCriteria: {criteria}\nState: {current_state}";
        let result = render(
            template,
            &[
                ("goal", "Build a REST API"),
                ("criteria", "- Endpoint works\n- Tests pass"),
                ("current_state", "API is deployed"),
            ],
        );
        assert!(result.contains("Build a REST API"));
        assert!(result.contains("Endpoint works"));
        assert!(result.contains("API is deployed"));
    }

    #[test]
    fn test_subagent_prompts_contain_guidelines() {
        assert!(SUBAGENT_EXPLORE.contains("Guidelines"));
        assert!(SUBAGENT_EXPLORE.contains("read-only"));
        assert!(SUBAGENT_PLAN.contains("Guidelines"));
        assert!(SUBAGENT_PLAN.contains("read-only"));
    }

    #[test]
    fn test_context_summary_prefix() {
        assert!(CONTEXT_SUMMARY_PREFIX.contains("Context Summary"));
    }

    // ── SystemPromptSlots tests ──

    #[test]
    fn test_slots_default_builds_system_default() {
        let slots = SystemPromptSlots::default();
        let built = slots.build();
        assert!(built.contains("Core Behaviour"));
        assert!(built.contains("Tool Usage Strategy"));
        assert!(built.contains("Completion Criteria"));
        assert!(built.contains("Response Format"));
        assert!(built.contains("A3S Code"));
    }

    #[test]
    fn test_slots_custom_role_replaces_default() {
        let slots = SystemPromptSlots {
            role: Some("You are a senior Python developer".to_string()),
            ..Default::default()
        };
        let built = slots.build();
        assert!(built.contains("You are a senior Python developer"));
        assert!(!built.contains("You are A3S Code"));
        // Core sections still present
        assert!(built.contains("Core Behaviour"));
        assert!(built.contains("Tool Usage Strategy"));
    }

    #[test]
    fn test_slots_custom_guidelines_appended() {
        let slots = SystemPromptSlots {
            guidelines: Some("Always use type hints. Follow PEP 8.".to_string()),
            ..Default::default()
        };
        let built = slots.build();
        assert!(built.contains("## Guidelines"));
        assert!(built.contains("Always use type hints"));
        assert!(built.contains("Core Behaviour"));
    }

    #[test]
    fn test_slots_custom_response_style_replaces_default() {
        let slots = SystemPromptSlots {
            response_style: Some("Be concise. Use bullet points.".to_string()),
            ..Default::default()
        };
        let built = slots.build();
        assert!(built.contains("Be concise. Use bullet points."));
        // Default response format content should be gone
        assert!(!built.contains("emit tool calls, no prose"));
        // But core is still there
        assert!(built.contains("Core Behaviour"));
    }

    #[test]
    fn test_slots_extra_appended() {
        let slots = SystemPromptSlots {
            extra: Some("Remember: always write tests first.".to_string()),
            ..Default::default()
        };
        let built = slots.build();
        assert!(built.contains("Remember: always write tests first."));
        assert!(built.contains("Core Behaviour"));
    }

    #[test]
    fn test_slots_from_legacy() {
        let slots = SystemPromptSlots::from_legacy("You are a helpful assistant.".to_string());
        let built = slots.build();
        // Legacy prompt goes into extra, core is still present
        assert!(built.contains("You are a helpful assistant."));
        assert!(built.contains("Core Behaviour"));
        assert!(built.contains("Tool Usage Strategy"));
    }

    #[test]
    fn test_slots_all_slots_combined() {
        let slots = SystemPromptSlots {
            style: None,
            role: Some("You are a Rust expert".to_string()),
            guidelines: Some("Use clippy. No unwrap.".to_string()),
            response_style: Some("Short answers only.".to_string()),
            extra: Some("Project uses tokio.".to_string()),
        };
        let built = slots.build();
        assert!(built.contains("You are a Rust expert"));
        assert!(built.contains("Core Behaviour"));
        assert!(built.contains("## Guidelines"));
        assert!(built.contains("Use clippy"));
        assert!(built.contains("Short answers only"));
        assert!(built.contains("Project uses tokio"));
        // Default response format replaced
        assert!(!built.contains("emit tool calls, no prose"));
    }

    #[test]
    fn test_slots_is_empty() {
        assert!(SystemPromptSlots::default().is_empty());
        assert!(!SystemPromptSlots {
            role: Some("test".to_string()),
            ..Default::default()
        }
        .is_empty());
        assert!(!SystemPromptSlots {
            style: Some(AgentStyle::Plan),
            ..Default::default()
        }
        .is_empty());
    }

    // ── AgentStyle tests ──

    #[test]
    fn test_agent_style_default_is_general_purpose() {
        assert_eq!(AgentStyle::default(), AgentStyle::GeneralPurpose);
    }

    #[test]
    fn test_agent_style_base_prompt() {
        assert_eq!(AgentStyle::GeneralPurpose.base_prompt(), SYSTEM_DEFAULT);
        assert_eq!(AgentStyle::Plan.base_prompt(), SUBAGENT_PLAN);
        assert_eq!(AgentStyle::Explore.base_prompt(), SUBAGENT_EXPLORE);
        assert_eq!(AgentStyle::Verification.base_prompt(), AGENT_VERIFICATION);
        // CodeReview uses GeneralPurpose base + review guidelines
        assert_eq!(AgentStyle::CodeReview.base_prompt(), SYSTEM_DEFAULT);
    }

    #[test]
    fn test_agent_style_guidelines() {
        assert!(AgentStyle::GeneralPurpose.guidelines().is_none());
        assert!(AgentStyle::Plan.guidelines().is_none()); // embedded in prompt
        assert!(AgentStyle::Explore.guidelines().is_none());
        assert!(AgentStyle::Verification.guidelines().is_none());
        assert!(AgentStyle::CodeReview.guidelines().is_some());
        assert!(AgentStyle::CodeReview
            .guidelines()
            .unwrap()
            .contains("Correctness"));
    }

    #[test]
    fn test_agent_style_detect_plan() {
        assert_eq!(
            AgentStyle::detect_from_message("Help me plan a new feature"),
            AgentStyle::Plan
        );
        assert_eq!(
            AgentStyle::detect_from_message("Design the architecture for this"),
            AgentStyle::Plan
        );
        assert_eq!(
            AgentStyle::detect_from_message("What's the implementation approach?"),
            AgentStyle::Plan
        );
    }

    #[test]
    fn test_agent_style_detect_verification() {
        assert_eq!(
            AgentStyle::detect_from_message("Verify that this works correctly"),
            AgentStyle::Verification
        );
        assert_eq!(
            AgentStyle::detect_from_message("Test the login flow"),
            AgentStyle::Verification
        );
        assert_eq!(
            AgentStyle::detect_from_message("Check if the API handles edge cases"),
            AgentStyle::Verification
        );
    }

    #[test]
    fn test_agent_style_detect_explore() {
        assert_eq!(
            AgentStyle::detect_from_message("Find all files related to auth"),
            AgentStyle::Explore
        );
        assert_eq!(
            AgentStyle::detect_from_message("Where is the user model defined?"),
            AgentStyle::Explore
        );
        assert_eq!(
            AgentStyle::detect_from_message("Search for password hashing code"),
            AgentStyle::Explore
        );
    }

    #[test]
    fn test_agent_style_detect_code_review() {
        assert_eq!(
            AgentStyle::detect_from_message("Review the PR changes"),
            AgentStyle::CodeReview
        );
        assert_eq!(
            AgentStyle::detect_from_message("Analyze this code for best practices"),
            AgentStyle::CodeReview
        );
        assert_eq!(
            AgentStyle::detect_from_message("Assess code quality"),
            AgentStyle::CodeReview
        );
    }

    #[test]
    fn test_agent_style_detect_default_is_general_purpose() {
        // "Implement" was removed from Plan keywords as too generic
        assert_eq!(
            AgentStyle::detect_from_message("Implement the new feature"),
            AgentStyle::GeneralPurpose
        );
        // "Write tests" contains "test" so it's detected as Verification
        assert_eq!(
            AgentStyle::detect_from_message("Write code for the API"),
            AgentStyle::GeneralPurpose
        );
    }

    #[test]
    fn test_build_with_message_auto_detects_style() {
        let slots = SystemPromptSlots::default();
        let built = slots.build_with_message("Help me plan a new feature");
        // Should use Plan style
        assert!(built.contains("planning agent") || built.contains("READ-ONLY"));
    }

    #[test]
    fn test_build_with_message_explicit_style_overrides() {
        let slots = SystemPromptSlots {
            style: Some(AgentStyle::Verification),
            ..Default::default()
        };
        let built = slots.build_with_message("Help me plan a new feature");
        // Should use Verification style, not Plan
        assert!(built.contains("verification specialist") || built.contains("try to break"));
    }

    #[test]
    fn test_build_with_message_plan_style() {
        let slots = SystemPromptSlots::default();
        let built = slots.build_with_message("Design the system architecture");
        assert!(built.contains("planning agent") || built.contains("READ-ONLY"));
    }

    #[test]
    fn test_build_with_message_explore_style() {
        let slots = SystemPromptSlots::default();
        let built = slots.build_with_message("Find all authentication files");
        assert!(built.contains("exploration agent") || built.contains("explore"));
    }

    #[test]
    fn test_build_with_message_code_review_style() {
        let slots = SystemPromptSlots::default();
        let built = slots.build_with_message("Review this code");
        // Should include code review guidelines
        assert!(built.contains("Correctness") || built.contains("Code Review"));
    }

    #[test]
    fn test_builder_methods() {
        let slots = SystemPromptSlots::default()
            .with_style(AgentStyle::Plan)
            .with_role("You are a Python expert")
            .with_guidelines("Use type hints")
            .with_response_style("Be brief")
            .with_extra("Additional instructions");

        assert_eq!(slots.style, Some(AgentStyle::Plan));
        assert_eq!(slots.role, Some("You are a Python expert".to_string()));
        assert_eq!(slots.guidelines, Some("Use type hints".to_string()));
        assert_eq!(slots.response_style, Some("Be brief".to_string()));
        assert_eq!(slots.extra, Some("Additional instructions".to_string()));

        let built = slots.build();
        assert!(built.contains("Python expert"));
        assert!(built.contains("Use type hints"));
        assert!(built.contains("Be brief"));
        assert!(built.contains("Additional instructions"));
    }

    #[test]
    fn test_code_review_guidelines_appended() {
        let slots = SystemPromptSlots {
            style: Some(AgentStyle::CodeReview),
            ..Default::default()
        };
        let built = slots.build();
        assert!(built.contains("Correctness"));
        assert!(built.contains("Security"));
        assert!(built.contains("Performance"));
        assert!(built.contains("Maintainability"));
    }
}