See what fills your context window

Why is this agent slow, expensive, or drifting? Logs can now answer with a context-window breakdown on every execution.

• Composition bar. The log details panel shows what filled the context: system prompt, tool definitions, record context, conversation history, tool results, and the current message, measured on the actual model calls (including everything the tool loop added).
• Per-iteration ledger. The Trace tab charts context per agent-loop iteration, with compact bars on the trace tree, so you can watch history and tool results accumulate turn by turn.
• A meaningful headline. The summary is absolute estimated input tokens with a band label ("Typical range", "Elevated", "Very high") instead of a percent of the model's window, so the signal doesn't jump when you switch models and still works on 1M-context models.
• One hint at a time. When a run leaves the typical band, you get at most one targeted suggestion, like truncating large tool outputs or trimming unused tool definitions.
• Live tail pills. In live tail, active executions surface a mini context bar you can click into.

The breakdown also rides the execution stream as an optional estimatedContextBreakdown field on turn_complete, available in the API and both SDKs.