Context and communication

Inspect agent evidence where the work happens.

Agent context

Open an agent's Visual progress view to inspect reported context usage and captured instructions, messages, workflow assignments, tool input/results, and replies. Categories use labels and colors; open entries to read the source evidence.

Captured character counts and characters-divided-by-four token estimates are distinct from CLI-reported window usage. Hidden instructions, compaction, unmanaged helpers, and uncaptured material remain unknown. A path or URL appearing in text does not prove it was read.

Team communication

Open Communications on Live to inspect directed messages, delivery/read state, traffic over time, payload sizes, and reported token usage. Filter routes and inspect message content before drawing conclusions about overhead.

The dashboard uses bounded loaded history, not necessarily lifetime totals. Queue-to-delivery latency is not model completion time. Prompt overhead estimates are not measured network or compute costs.

Evidence and outcomes

Preserve the distinction between planned work, work in progress, agent-reported completion, independently verified checks, and blockers. Tool success and file existence alone do not establish correctness.