Compact context automatically during long research sessions

Goal

Build a long-running research workflow that preserves the active question and key findings after compaction.

Steps

1. Run an uncompacted baseline for the same research task; note when the context window approaches its limit and whether later turns lose early evidence. [source07]
2. Configure the documented context_management compaction strategy, threshold, and summary instructions for the Managed Agent. [source07]
3. Repeat the task and inspect compaction events and summary text; confirm new evidence is incorporated, key sources remain, and the summary is not treated as a full transcript. [source07]

How to verify

Tests show automatic compaction occurs near the configured threshold and later answers retain key evidence while recognizing summary loss.

Troubleshooting

If compaction does not trigger or drops critical details, check model and beta API support, lower the threshold, or reduce input; a summary is not a verbatim audit log.