Generation limits bound the effort spent implementing one body. The numbered examples give a score-counting function two rejected submissions and a 120-second deadline:
__llm__(max_attempts(2), timeout(120)) int passing(const std::vector<int> &scores) { Count scores greater than or equal to 60. }
These function settings override the corresponding driver defaults. A failed submit consumes an attempt; a failed try_compile uses a tool call but not a submission attempt.
Other limits cap compiler-tool calls and tokens per model response. Supported settings are passed with the task, and the compiler enforces its budgets. The bundled CLI adapters reject explicit output-token limits they cannot enforce. Limits control build cost and duration; they do not establish correctness.
Source: 17_generation_limits.cpp.