The context-usage chip
The chat header always shows a quiet usage chip — for example 9% · 12k. The percentage is the share of the model’s context window that the current conversation occupies; the number is the total tokens used so far in this session. Click the chip to open a full Usage breakdown showing input tokens, output tokens, cached tokens, and reasoning tokens. The chip’s aria label reads “Context usage{percent}% — view details”, so screen readers describe it fully even when the numbers are compact.
Approaching the context limit
When the conversation fills 80% or more of the context window, a warning banner appears above the composer: Approaching the context limit “This conversation uses{percent}% of the model’s context window. Compact it to keep the agent fast and accurate.”
You can dismiss this banner; it reappears if the conversation grows by another five percentage points.
Blocked at the context limit
When the conversation reaches 95% of the context window, the banner becomes non-dismissible and the composer is disabled: Context limit reached “This conversation no longer fits the model’s context window. Compact it to continue.” The message input shows “Compact the conversation to continue…”. You must compact before you can send another message.If compaction itself cannot reduce the conversation far enough, IMP shows: “This conversation can’t be compacted further — please start a new chat.” In that case, start a fresh conversation and paste in any key context you need to carry over.
Compact the conversation
1
Click Compact conversation
Press the Compact conversation button in the banner. A small optional input appears: “Anything to preserve in detail?” Type any specific figures, names, or conclusions you want the summary to include verbatim.
2
Wait for the summary
The button changes to “Summarizing conversation…” while IMP processes the history. Do not navigate away.
3
Continue chatting
Once complete, the banner disappears and the composer re-enables. The conversation shows a checkpoint marker in the history where compaction happened.
What compaction keeps and drops
Compaction replaces the earlier part of the conversation with a compact summary. The summary preserves the key conclusions, facts, and decisions from that part of the exchange. Fine-grained detail — long tool outputs, intermediate reasoning steps, raw document text — is condensed or omitted to free up space. The original messages are never deleted. The full history is still visible in the conversation and remains in the database. Only the model’s working context is summarized; you can always scroll back to read earlier turns.Next steps
Model override and usage
Inspect per-turn token counts in the Usage panel.
Budgets in chat
Understand how spending budgets relate to usage.