Transcript replay repeats a recorded generation exchange against the current compiler without contacting a model. The bundled adapter supports it through llmcpp-agent --replay.

$ llmc++     -fllm-regenerate     -fllm-agent='llmcpp-agent --replay trace.jsonl'     main.cpp

Regeneration bypasses ordinary body-cache reuse so that the recorded exchange can run. The replay adapter selects a successful matching task, checks its context, prompt, instructions, and effective policy, then performs the recorded compiler-tool calls.

Replay therefore differs from caching. A cache hit supplies a stored body directly; replay exercises the recorded tool interaction and submission against the compiler. Changes to the task can prevent a match rather than silently adapting an old conversation.

Keep the original source context and generation options when reproducing a transcript. The feature is useful for investigating compiler behavior without spending another model request or relying on a model to choose identical actions.

Replaying an accepted exchange does not make its algorithm correct. It reproduces the generation steps, while behavioral tests check whether their result fulfills the request.