Generation transcripts (llmcpp)

Generation transcripts record tasks, compiler-tool calls, and outcomes as JSON Lines. The repository's triangular-number example shows recording and replay:

# Record an exchange with a configured backend.
$ llmc++ \
    -fllm-transcript=trace.jsonl \
    examples/19_transcript.cpp
 
# Repeat a matching recorded exchange without a model.
$ llmc++ \
    -fllm-regenerate \
    -fllm-agent='llmcpp-agent --replay trace.jsonl' \
    examples/19_transcript.cpp

The first command records events that actually occur. If generation is skipped by a cache hit, it does not create a new model exchange. Replay needs a successful matching recorded task.

Transcripts help explain inspected declarations, failed candidates, and accepted submissions. Agent configuration is omitted and credential-bearing fields are redacted, but prompts and code can remain. A recorded history explains provenance; it does not prove behavioral correctness.

Source: 19_transcript.cpp.