runtime-communication

# Runtime Communication Use this skill only for the local `research_mvp` runtime. ## Files To Read First - `AGENTS.md` - `research_mvp/runtime.toml` ## Commands Use the configured runtime CLI: ```bash python -m research_mvp.runtime_cli --config research_mvp/runtime.toml status python -m research_mvp.runtime_cli --config research_mvp/runtime.toml thread tail -n 50 python -m research_mvp.runtime_cli --config research_mvp/runtime.toml inbox list leader ``` ## Messaging Rules - Shared thread is for summaries, planning context, progress reporting, and visibility. - Direct agent work must go through targeted inbox delegation. - If you need one specific agent to receive a message, use `delegate` or `thread send` through `runtime_cli`. Do not hand-edit `thread.jsonl`, because that does not create an inbox delivery. - Formal long-running training should be submitted to an external `train_service`, not left running inside the `trainer` tmux pane. - If the human request is to start a task under `recipe/<name>/`, first read `recipe/<name>/data.md`, `recipe/<name>/overview.md`, and `recipe/<name>/start_prompt.md`. - Treat `recipe/<name>/` tasks as Kaggle-style competition tasks by default unless the recipe explicitly says otherwise. - For a new `recipe/<name>/` task, do EDA before baseline iteration. Put EDA scripts, notes, and charts under `<workdir>/eda/`. - This runtime is a machine-learning baseline team. Default iteration should follow the repository's existing baseline namin...