moai-foundation-context

Track usage against the 200K token budget with warning at 75% and critical compression at 85%. Execute /clear at strategic checkpoints such as after SPEC creation, and reload only Tier 1 files like CLAUDE.md afterward.

Persist the session ID, current SPEC ID, phase, and completed steps to a state snapshot before clearing. After /clear, reload CLAUDE.md and the active SPEC file to resume work without losing progress.

Run /clear after /moai:1-plan completion, when context exceeds 150K tokens, when conversations pass 50 messages, before major phase transitions, and during model switches between Haiku and Sonnet.

Create a handoff package containing only the SPEC ID, current phase, next step, and key artifact references. Validate agent compatibility, checksum integrity, and a minimum 30K available token buffer before transfer.

Reasoning quality degrades in the final 20% of the context window, roughly above 160K tokens. Avoid this zone by clearing aggressively and keeping a 55K token emergency reserve for recovery operations.

Summarization compresses conversation to about 30% of original size but loses fine detail, so original content must be archived for recovery. It is unsuitable when exact code or decision wording must be preserved verbatim.