Function log_llm_request 

pub fn log_llm_request(
    datalog_dir: &Path,
    messages: &[Message],
    tool_defs: &[ToolDef],
    model: &str,
    context_window: usize,
    step: usize,
    enabled: bool,
) -> Option<PathBuf>

Per-round LLM log files live under <datalog_dir>/llm/ where <datalog_dir> is the per-project directory resolved by DatalogWriter::resolve_log_dir (typically ~/.atomcode/datalog/<project-slug>/). The caller is responsible for passing in the resolved dir — see runner.rs. This keeps the JSONL dump in lockstep with the markdown writer’s [datalog].dir config; the prior hard-coded <working_dir>/datalog/llm/ would silently ignore the user’s dir override.

One file per LLM round-trip, containing both request and response sections. Filename = timestamp. calls.log is a one-line-per-round index.

Split-file layout (prior design) produced two JSONs per round plus a CSV entry per half — hard to read and review. One-file-per-round is both AI-friendly (single JSON to grep/diff/feed back) and human-friendly.

Request / response pairing: log_llm_request returns the path of the JSON file it wrote; the caller holds that PathBuf locally and passes it to log_llm_response so the response section merges into the SAME file. No process-wide shared state — safe for concurrent chat_stream handlers in the daemon (prior design used a static Mutex<Option<PathBuf>> that bled across sessions). Log the LLM request. Writes a JSON file containing the request section under <datalog_dir>/llm/<timestamp>.json.

Returns the path so the caller can pass it to log_llm_response for in-place merge. Returns None if enabled is false or the write failed.