add-compact

To prevent context rot in long-running LLM agent sessions, use the /compact command to trigger Claude Agent SDK built-in compaction. This summarizes conversation history while preserving session continuity by forwarding the SDK-issued new session ID to the orchestrator.

LLM agents lose coherence during long conversations because the model context becomes cluttered, causing context rot. Manual session compaction clears this clutter by summarizing previous history while keeping the ongoing logical session intact.

To archive transcripts before compacting an agent session, trigger the /compact command. The system automatically saves conversation transcripts before the Claude Agent SDK compacts the context, ensuring you avoid losing auditability.

Untrusted users cannot trigger context compaction in operational chat sessions. Strict access controls enforce that only main trusted group members or admin and device-owner senders can execute the /compact command to reduce destructive memory impact.

Session compaction does not use synthetic reset prompts to clear context. It triggers the Claude Agent SDK built-in compaction mechanism directly, summarizing the conversation history while forwarding a new session ID to preserve continuity.

You should manually compact chat history in Claude Agent SDK workflows when long-running conversations lose coherence from context clutter. Execute /compact from the main trusted group to summarize history while maintaining the ongoing logical session.