car-agents 0.14.0

//! Planner agent — given a goal, produce an ordered list of action steps.
//!
//! Bridges natural language goals to CAR's ActionProposal format.
//! Uses the inference engine to generate plans, then optionally scores
//! them via car-planner for validation.

use crate::{AgentContext, AgentResult};
use car_inference::{GenerateParams, GenerateRequest};

/// Planner agent configuration.
#[derive(Debug, Clone)]
pub struct PlanConfig {
    pub max_tokens: usize,
    pub temperature: f64,
    pub model: Option<String>,
    /// If true, include available tools in the prompt.
    pub include_tools: bool,
}

impl Default for PlanConfig {
    fn default() -> Self {
        Self {
            max_tokens: 2048,
            temperature: 0.2,
            model: None,
            include_tools: true,
        }
    }
}

/// Planner: goal → ordered action steps.
pub struct PlannerAgent {
    ctx: AgentContext,
    config: PlanConfig,
}

impl PlannerAgent {
    pub fn new(ctx: AgentContext) -> Self {
        Self {
            ctx,
            config: PlanConfig::default(),
        }
    }

    pub fn with_config(ctx: AgentContext, config: PlanConfig) -> Self {
        Self { ctx, config }
    }

    /// Generate a plan for achieving a goal.
    pub async fn plan(&self, goal: &str, context: Option<&str>) -> AgentResult {
        let prompt = format!(
            "You are a planning agent. Break down the following goal into concrete, ordered steps.\n\n\
            Goal: {goal}\n\n\
            For each step, specify:\n\
            1. What to do (action)\n\
            2. What it depends on (which previous steps must complete)\n\
            3. What it produces (output/state change)\n\
            4. How to verify it worked\n\n\
            Format as a numbered list. Be specific and actionable — no vague steps like 'analyze the situation.'"
        );

        let start = std::time::Instant::now();
        let req = GenerateRequest {
            prompt,
            model: self.config.model.clone(),
            params: GenerateParams {
                temperature: self.config.temperature,
                max_tokens: self.config.max_tokens,
                ..Default::default()
            },
            context: context.map(String::from),
            tools: None,
            images: None,
            messages: None,
            cache_control: false,
            response_format: None,
            intent: None,
        };

        match self.ctx.inference.generate_tracked(req).await {
            Ok(result) => {
                // Count steps as a rough quality signal
                let step_count = result
                    .text
                    .lines()
                    .filter(|l| l.trim_start().starts_with(|c: char| c.is_ascii_digit()))
                    .count();
                let confidence = if step_count >= 3 { 0.8 } else { 0.5 };
                AgentResult {
                    agent: "planner".into(),
                    output: result.text,
                    confidence,
                    model_used: result.model_used,
                    latency_ms: start.elapsed().as_millis() as u64,
                }
            }
            Err(e) => AgentResult {
                agent: "planner".into(),
                output: format!("Planning failed: {}", e),
                confidence: 0.0,
                model_used: String::new(),
                latency_ms: start.elapsed().as_millis() as u64,
            },
        }
    }
}