car-multi 0.26.0

//! Advisor — executor stays in control, stronger model is consulted on demand.
//!
//! Unlike delegation patterns, the advisor does not take over execution.
//! It only returns a bounded verdict that the caller may apply or ignore.

use crate::error::MultiError;
use crate::mailbox::Mailbox;
use crate::runner::AgentRunner;
use crate::shared::SharedInfra;
use crate::types::{AgentOutput, AgentSpec};
use serde::{Deserialize, Serialize};
use std::collections::HashMap;
use std::sync::atomic::{AtomicU32, Ordering};
use std::sync::Arc;
use tracing::instrument;

/// Structured guidance returned by an advisor consultation.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "snake_case")]
pub enum AdvisorVerdict {
    Continue { rationale: String },
    Plan { guidance: String },
    Correction { guidance: String },
    Stop { reason: String },
    Uncertain { reason: String },
}

/// Result of one advisor consultation.
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct AdvisorResult {
    pub verdict: AdvisorVerdict,
    pub advisor_output: AgentOutput,
    pub used: u32,
    pub max_uses: u32,
}

/// Bounded consultation primitive for stronger-model judgment.
pub struct Advisor {
    pub advisor: AgentSpec,
    pub max_uses: u32,
    used: AtomicU32,
}

impl Advisor {
    pub fn new(advisor: AgentSpec, max_uses: u32) -> Self {
        Self {
            advisor,
            max_uses,
            used: AtomicU32::new(0),
        }
    }

    pub fn used(&self) -> u32 {
        self.used.load(Ordering::Relaxed)
    }

    pub fn remaining(&self) -> u32 {
        self.max_uses.saturating_sub(self.used())
    }

    #[instrument(name = "multi.advisor", skip_all)]
    pub async fn consult(
        &self,
        task: &str,
        runner: &Arc<dyn AgentRunner>,
        infra: &SharedInfra,
    ) -> Result<AdvisorResult, MultiError> {
        let prior = self.used.fetch_add(1, Ordering::Relaxed);
        if prior >= self.max_uses {
            return Err(MultiError::AdvisorExhausted {
                used: prior,
                max_uses: self.max_uses,
            });
        }

        let advisory_task = format!(
            r#"You are an advisor. The executor remains in control of the loop.

Return exactly one verdict as JSON:
```json
{{
  "verdict": "continue|plan|correction|stop|uncertain",
  "rationale": "text when verdict is continue",
  "guidance": "text when verdict is plan or correction",
  "reason": "text when verdict is stop or uncertain"
}}
```

Task or draft to review:
{task}"#
        );

        let mailbox = Mailbox::default();
        let rt = infra.make_runtime();
        let output = runner
            .run(&self.advisor, &advisory_task, &rt, &mailbox)
            .await
            .map_err(|e| {
                MultiError::AgentFailed(
                    self.advisor.name.clone(),
                    format!("advisor consultation failed: {}", e),
                )
            })?;

        let verdict = parse_verdict(&output.answer);
        Ok(AdvisorResult {
            verdict,
            advisor_output: output,
            used: prior + 1,
            max_uses: self.max_uses,
        })
    }
}

/// Risk level assigned by the caller or fixture harness.
#[derive(Debug, Clone, Copy, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "snake_case")]
pub enum TaskRisk {
    Low,
    Medium,
    High,
}

/// Inputs for deciding whether the student should consult the advisor.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
pub struct AdvisorTriggerContext {
    pub repeated_failures: u32,
    pub explicit_uncertainty: bool,
    pub missing_guidance: bool,
    pub prior_advisor_calls: u32,
    #[serde(default = "default_task_risk")]
    pub task_risk: TaskRisk,
    #[serde(default)]
    pub labels: Vec<String>,
}

const fn default_task_risk() -> TaskRisk {
    TaskRisk::Low
}

impl Default for AdvisorTriggerContext {
    fn default() -> Self {
        Self {
            repeated_failures: 0,
            explicit_uncertainty: false,
            missing_guidance: false,
            prior_advisor_calls: 0,
            task_risk: TaskRisk::Low,
            labels: Vec::new(),
        }
    }
}

/// Explicit outcome of trigger evaluation.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
#[serde(rename_all = "snake_case")]
pub enum AdvisorTriggerDecision {
    NoConsult { reason: String },
    OptionalConsult { reason: String },
    MustConsult { reason: String },
}

/// Bounded policy for when the student should consult the advisor.
#[derive(Debug, Clone, Serialize, Deserialize, PartialEq, Eq)]
pub struct AdvisorTriggerPolicy {
    pub max_calls_per_run: u32,
    pub repeated_failure_threshold: u32,
    pub consult_on_missing_guidance: bool,
    pub require_on_high_risk: bool,
}

impl Default for AdvisorTriggerPolicy {
    fn default() -> Self {
        Self {
            max_calls_per_run: 3,
            repeated_failure_threshold: 2,
            consult_on_missing_guidance: true,
            require_on_high_risk: true,
        }
    }
}

impl AdvisorTriggerPolicy {
    pub fn evaluate(&self, ctx: &AdvisorTriggerContext) -> AdvisorTriggerDecision {
        if ctx.prior_advisor_calls >= self.max_calls_per_run {
            return AdvisorTriggerDecision::NoConsult {
                reason: "advisor budget exhausted".to_string(),
            };
        }

        if self.require_on_high_risk && matches!(ctx.task_risk, TaskRisk::High) {
            return AdvisorTriggerDecision::MustConsult {
                reason: "high-risk task".to_string(),
            };
        }

        if ctx.repeated_failures >= self.repeated_failure_threshold {
            return AdvisorTriggerDecision::MustConsult {
                reason: format!(
                    "repeated failures reached threshold ({}/{})",
                    ctx.repeated_failures, self.repeated_failure_threshold
                ),
            };
        }

        if ctx.explicit_uncertainty {
            return AdvisorTriggerDecision::OptionalConsult {
                reason: "executor signaled uncertainty".to_string(),
            };
        }

        if self.consult_on_missing_guidance && ctx.missing_guidance {
            return AdvisorTriggerDecision::OptionalConsult {
                reason: "no approved guidance matched".to_string(),
            };
        }

        AdvisorTriggerDecision::NoConsult {
            reason: "no trigger matched".to_string(),
        }
    }
}

fn parse_verdict(response: &str) -> AdvisorVerdict {
    if let Some(json_str) = car_ir::json_extract::extract_json_object(response) {
        if let Ok(parsed) = serde_json::from_str::<HashMap<String, serde_json::Value>>(&json_str) {
            let verdict = parsed.get("verdict").and_then(|v| v.as_str()).unwrap_or("");
            let rationale = parsed
                .get("rationale")
                .and_then(|v| v.as_str())
                .unwrap_or("")
                .trim();
            let guidance = parsed
                .get("guidance")
                .and_then(|v| v.as_str())
                .unwrap_or("")
                .trim();
            let reason = parsed
                .get("reason")
                .and_then(|v| v.as_str())
                .unwrap_or("")
                .trim();
            return match verdict {
                "continue" if !rationale.is_empty() => AdvisorVerdict::Continue {
                    rationale: rationale.to_string(),
                },
                "plan" if !guidance.is_empty() => AdvisorVerdict::Plan {
                    guidance: guidance.to_string(),
                },
                "correction" if !guidance.is_empty() => AdvisorVerdict::Correction {
                    guidance: guidance.to_string(),
                },
                "stop" if !reason.is_empty() => AdvisorVerdict::Stop {
                    reason: reason.to_string(),
                },
                "uncertain" if !reason.is_empty() => AdvisorVerdict::Uncertain {
                    reason: reason.to_string(),
                },
                _ => AdvisorVerdict::Uncertain {
                    reason: "advisor response could not be parsed cleanly".to_string(),
                },
            };
        }
    }

    let lower = response.to_lowercase();
    if lower.contains("verdict: continue") {
        AdvisorVerdict::Continue {
            rationale: response.trim().to_string(),
        }
    } else if lower.contains("verdict: stop") {
        AdvisorVerdict::Stop {
            reason: response.trim().to_string(),
        }
    } else if lower.contains("verdict: correction") {
        AdvisorVerdict::Correction {
            guidance: response.trim().to_string(),
        }
    } else if lower.contains("verdict: plan") {
        AdvisorVerdict::Plan {
            guidance: response.trim().to_string(),
        }
    } else {
        AdvisorVerdict::Uncertain {
            reason: response.trim().to_string(),
        }
    }
}

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

    struct StaticRunner {
        response: String,
    }

    #[async_trait::async_trait]
    impl crate::runner::AgentRunner for StaticRunner {
        async fn run(
            &self,
            spec: &AgentSpec,
            _task: &str,
            _runtime: &Runtime,
            _mailbox: &Mailbox,
        ) -> Result<AgentOutput, MultiError> {
            Ok(AgentOutput {
                name: spec.name.clone(),
                answer: self.response.clone(),
                turns: 1,
                tool_calls: 0,
                duration_ms: 3.0,
                error: None,
                outcome: None,
                tokens: None,
                tools_used: Vec::new(),
            })
        }
    }

    #[tokio::test]
    async fn consult_returns_parsed_verdict() {
        let advisor = Advisor::new(AgentSpec::new("advisor", "review"), 2);
        let runner: Arc<dyn crate::runner::AgentRunner> = Arc::new(StaticRunner {
            response: r#"{"verdict":"plan","guidance":"check state before editing"}"#.to_string(),
        });
        let infra = SharedInfra::new();

        let result = advisor.consult("help", &runner, &infra).await.unwrap();
        assert_eq!(
            result.verdict,
            AdvisorVerdict::Plan {
                guidance: "check state before editing".to_string(),
            }
        );
        assert_eq!(result.used, 1);
        assert_eq!(advisor.remaining(), 1);
    }

    #[tokio::test]
    async fn consult_enforces_budget() {
        let advisor = Advisor::new(AgentSpec::new("advisor", "review"), 1);
        let runner: Arc<dyn crate::runner::AgentRunner> = Arc::new(StaticRunner {
            response: r#"{"verdict":"uncertain","reason":"need more evidence"}"#.to_string(),
        });
        let infra = SharedInfra::new();

        advisor.consult("first", &runner, &infra).await.unwrap();
        let err = advisor
            .consult("second", &runner, &infra)
            .await
            .unwrap_err();
        match err {
            MultiError::AdvisorExhausted { used, max_uses } => {
                assert_eq!(used, 1);
                assert_eq!(max_uses, 1);
            }
            other => panic!("unexpected error: {other:?}"),
        }
    }

    #[test]
    fn trigger_policy_requires_high_risk() {
        let policy = AdvisorTriggerPolicy::default();
        let ctx = AdvisorTriggerContext {
            task_risk: TaskRisk::High,
            ..Default::default()
        };
        assert_eq!(
            policy.evaluate(&ctx),
            AdvisorTriggerDecision::MustConsult {
                reason: "high-risk task".to_string(),
            }
        );
    }

    #[test]
    fn trigger_policy_honors_budget() {
        let policy = AdvisorTriggerPolicy::default();
        let ctx = AdvisorTriggerContext {
            prior_advisor_calls: 3,
            explicit_uncertainty: true,
            ..Default::default()
        };
        assert_eq!(
            policy.evaluate(&ctx),
            AdvisorTriggerDecision::NoConsult {
                reason: "advisor budget exhausted".to_string(),
            }
        );
    }

    #[test]
    fn parse_verdict_falls_back_to_uncertain() {
        let verdict = parse_verdict("I am not sure.");
        assert_eq!(
            verdict,
            AdvisorVerdict::Uncertain {
                reason: "I am not sure.".to_string(),
            }
        );
    }

    #[test]
    fn parse_verdict_reads_continue_json() {
        let verdict = parse_verdict(
            r#"{"verdict":"continue","rationale":"current completion claim is supported"}"#,
        );
        assert_eq!(
            verdict,
            AdvisorVerdict::Continue {
                rationale: "current completion claim is supported".to_string(),
            }
        );
    }
}