strategic-compact

To preserve context across task phases in long AI sessions, use manual context compaction suggestions that trigger at configurable token thresholds. This retains essential phase-specific information while discarding bulky exploration history before auto-compaction arbitrarily removes it.

Auto-compaction discards important information during multi-phase workflows because it triggers arbitrarily at token limits without recognizing phase boundaries. Manual compaction suggestions detect tool call counts and prompt you to compact strategically between phases like research and implementation.

To set up PreToolUse hooks for context management in Claude, add a local suggester script that monitors Edit and Write events, counts tool calls against a configurable threshold, and emits periodic compaction prompts without requiring external network access.

Yes, you can configure token thresholds for compaction reminders. The system uses an environment-configurable threshold for the first suggestion and supports periodic reminder intervals afterward to fit different multi-phase workflow needs.

You should manually compact context instead of relying on automatic compaction when transitioning between workflow phases like research, planning, implementation, and testing. This ensures distilled plans are kept for the next phase while noisy exploration history is safely discarded.

No, this context management approach does not require external network access. The PreToolUse hook runs a local suggester that counts tool calls and emits compaction prompts entirely locally, making it suitable for offline or restricted environments.