Caching reuses accepted implementations across builds. The default cache sits in .llmcache/ beside the source; these commands instead name a reviewed cache directory:
# Generate missing implementations and save accepted bodies. $ llmc++ -fllm-cache-dir=reviewed-cache main.cpp -o main # Require reusable entries without contacting an agent. $ llmc++ \ -fllm-cache-dir=reviewed-cache \ -fllm-offline main.cpp -o main
The first build may contact a configured backend. The second fails on a missing or unusable entry instead of silently generating another body.
A cache hit still supplies C++ for the current compilation. Metadata validates the identity and schema, not behavioral correctness. Review the accepted source before preserving it as a trusted implementation. Identity decides what matches; lifetime and read/write policies decide whether an entry can be used or replaced.