Struct ResumePrePass 

pub struct ResumePrePass<'a> {
    pub tool_specs: &'a [ToolSpec],
    pub approved_overrides: &'a HashMap<(String, String, String), ApprovalOverride>,
    pub denied_call_ids: &'a HashSet<(String, String, String)>,
}

The approval resume pre-pass: resolve the tool calls a human already decided.

Before the turn drives the model, execute the calls the human approved that are dangling in the resumed transcript, resolve signed or denied calls to synthetic results, splice those results in after the paused batch, and append any approver-injected context notes.

On an approval resume the control plane replays the paused turn’s assistant tool_use (which has NO paired tool_result — the call was paused, never executed) and forwards the signed decisions via RunTurnOptions::approved_call_ids / RunTurnOptions::denied_call_ids. The function-calling loop only executes tool calls the model emits this turn, so without this step an approval takes effect only if the model happens to RE-EMIT the same call. Resolving the dangling calls deterministically here makes an approval ALWAYS take effect, independent of whether the model re-emits.

Classification and the #141 approval binding mirror the in-loop batch so the two can’t drift. It runs only on a resume: a fresh turn carries an empty decision set, so this step is a no-op and the hot path is unchanged.

A forwarded signed decision must never resolve silently to nothing: whenever approved_call_ids is non-empty, Self::run logs a resolution summary and tracing::warn!s individually for every approved tuple that matches no unanswered call in the resumed transcript — distinguishing an already- answered (harmless) re-forward from a call that is missing outright (a projection loss upstream, e.g. one folded into a compaction summary). This is purely observational: an approval that resolves to nothing still resolves to nothing (re-pausing here would loop), but the loss is now loud instead of surfacing only as an unexplained model refusal downstream.

Borrows the turn’s read-only resume inputs — the pinned tool-spec set (for pause-card titles), the approver edits, and the signed denials — while the mutable working state (transcript, outputs, the sticky denial set, and the remaining approvals) rides the TurnCtx.

Fields

tool_specs: &'a [ToolSpec]

The turn’s pinned tool-spec set, read once for pause-card titles.

approved_overrides: &'a HashMap<(String, String, String), ApprovalOverride>

Approver edits (#67), keyed by the canonicalized approval identity.

denied_call_ids: &'a HashSet<(String, String, String)>

Verified signed denials as canonicalized (id, name, args) tuples.

Trait Implementations

impl<P, T> TurnStep<P, T> for ResumePrePass<'_>

where P: LlmProvider + ?Sized, T: ToolExecutor + ?Sized,

fn run<'life0, 'life1, 'life2, 'async_trait>( &'life0 self, ctx: &'life1 mut TurnCtx<'life2, P, T>, ) -> Pin<Box<dyn Future<Output = Result<StepOutcome, P::Error>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait,

Run this step against the turn context.