Skip to main content

run_replay

shadow_core::replay::engine

Function run_replay 

Source 
pub async fn run_replay<B: LlmBackend + ?Sized>(
    baseline: &[Record],
    backend: &B,
    clock: &dyn Clock,
) -> Result<Vec<Record>, ReplayError>
Expand description

Run a replay: walk baseline, dispatch every chat_request to backend, and produce a fresh trace with the same structure.

Algorithm (SPEC §10.1):

  1. Emit a new metadata record with parent = None and an envelope meta.baseline_of = baseline_root_id.
  2. For each baseline record in file order:
    • chat_request: re-emit with a fresh ts and parent = previous output record id; then call backend.complete(request.payload) and emit a chat_response whose parent = the re-emitted request.
    • tool_call, tool_result, error: copy-through with fresh ts and relinked parent.
    • chat_response, metadata, replay_summary: skipped (the backend produces responses; replay_summary is added at the end; a baseline can only have one metadata record and it’s the root).
  3. Emit a replay_summary at the end.