Supervision board¶

When you have several agents running at once, you can't watch them all. The supervision board does it for you: vupai board (or just say "board") splits a dedicated pane (right, ~40%) that shows, per named agent pane, a one-line summary of its main conclusion or pending action, so a glance tells you who's done, who's stuck, and who needs you.

• Tool-agnostic. Works with any agentic CLI, not just Claude Code: pane activity is detected from terminal-output churn, and the summarizer is a swappable command (board_summarizer_cmd), not a fixed model. vupai appends the pane's scrollback tail as the command's last argument and takes its last stdout line as the summary, so any command that follows that contract works:

  To summarize with…	Set board_summarizer_cmd to
  Claude Haiku (default, streaming)	python -m vupai.claude_summarize --model claude-haiku-4-5
  Claude (plain, buffered)	claude -p --model claude-haiku-4-5
  Codex	codex exec
  Gemini	gemini -p
  Ollama (local/remote)	python scripts/ollama_summarize.py --host http://BOX:11434 --model qwen2.5:7b

  The model is whatever that command uses (e.g. Codex's own config/profile). If the command is missing or fails, the board falls back to a non-LLM last-line summary.

• Cheap by design. A pane is summarized only when it settles (finishes a burst of work), skipped when nothing changed, and throttled per pane (board_min_summary_interval).

• Speak it too. "read board" reads the digest aloud; "read atlas" reads a single pane.

One board per session. Close the pane to stop it.