Struct PlanVerifier 

pub struct PlanVerifier<P>
where
    P: LlmProvider,
{ /* private fields */ }

LLM-backed post-task completeness verifier.

Uses a cheap provider for verification (configured via verify_provider). All failures are fail-open — verification never blocks task graph execution.

Implementations

impl<P> PlanVerifier<P>

where P: LlmProvider,

pub fn new( provider: P, max_tokens: u32, sanitizer: ContentSanitizer, ) -> PlanVerifier<P>

Create a new PlanVerifier.

pub async fn verify( &mut self, task: &TaskNode, output: &str, ) -> VerificationResult

Verify that a completed task’s output satisfies its description.

Returns VerificationResult { complete: true, gaps: [], confidence: 0.0 } on any LLM failure (fail-open). Logs ERROR after 3+ consecutive failures to surface systematic misconfiguration (critic S4).

The task stays Completed regardless of verification outcome. Downstream tasks are unblocked immediately on completion — verification does not gate dispatch.

pub async fn replan( &mut self, task: &TaskNode, gaps: &[Gap], graph: &TaskGraph, max_tasks: u32, ) -> Result<Vec<TaskNode>, OrchestrationError>

Generate new TaskNodes for critical and important gaps only.

Minor gaps are logged and skipped. New tasks depend on verified_task_id and are assigned IDs starting from next_id. Returns empty Vec on any LLM failure (fail-open).

Errors

Returns OrchestrationError::VerificationFailed only for hard invariant violations (e.g. too many tasks would exceed the graph limit). LLM errors are fail-open and never returned.