Struct Planner 

pub struct Planner { /* private fields */ }

Proposal scorer and ranker.

Implementations

impl Planner

pub fn new(config: PlannerConfig) -> Planner

pub fn score( &self, proposal: &ActionProposal, initial_state: Option<&HashMap<String, Value>>, registered_tools: Option<&HashSet<String>>, ) -> ScoredProposal

Score a single proposal using static verification.

pub fn rank( &self, candidates: &[ActionProposal], initial_state: Option<&HashMap<String, Value>>, registered_tools: Option<&HashSet<String>>, ) -> Vec<ScoredProposal>

Rank multiple candidate proposals. Returns scored proposals sorted by score descending (best first).

pub fn rank_with_feedback( &self, candidates: &[ActionProposal], initial_state: Option<&HashMap<String, Value>>, registered_tools: Option<&HashSet<String>>, feedback: Option<&ToolFeedback>, ) -> Vec<ScoredProposal>

Rank with historical tool feedback from trajectory store.

pub fn pick_best( &self, candidates: &[ActionProposal], initial_state: Option<&HashMap<String, Value>>, registered_tools: Option<&HashSet<String>>, ) -> Option<(usize, ScoredProposal)>

Pick the best valid proposal from candidates. Returns None if all candidates have errors.

pub fn rank_with_outcome( &self, candidates: &[ActionProposal], initial_state: Option<&HashMap<String, Value>>, registered_tools: Option<&HashSet<String>>, feedback: Option<&ToolFeedback>, model: &dyn EffectModel, value: &dyn ValueFunction, ) -> Vec<ScoredProposal>

Rank candidates by predicted outcome as well as validity/cost (Code World Models Slice 3). For each candidate, the predicted final state is computed with car_verify::simulate_with_model against the injected model — a verified Code World Model (e.g. a car_verify::cwm::GatedEffectModel) — and scored by value. The outcome value is blended into the static score with PlannerConfig::outcome_weight: score = static_score * (1 - outcome_weight) + outcome_value * outcome_weight for valid proposals (invalid ones stay 0.0).

This is the step that makes planning outcome-aware: a plan that reaches the goal can outrank a cheaper plan that doesn’t. It is library-only by construction — like car_verify::synthesize_cwm, it takes injected trait objects (the model and the value function), which don’t cross the FFI boundary; the stateless predictive primitive is simulate_with_predictions.

With outcome_weight == 0.0 (the default) the blend is a no-op and the ordering matches Planner::rank_with_feedback exactly — opting in an EffectModel never silently changes behavior.