strategic-compact

Strategic compaction

Use when a long session is approaching its host's context limit or changing phases. Use actual context telemetry when available; tool-call counts and slower responses are not reliable measures of remaining tokens. Do not assume a particular window size, hook API, or /compact command exists in every runtime.

1. Prefer a coherent boundary: exploration to an accepted plan, completion of a milestone, or a change of subsystem. Avoid interrupting an unresolved edit or failure investigation if the host permits finishing that bounded step.
2. Write a compact working set using working-set-snapshot or handoff-doc: objective, accepted decisions, user constraints, revision/worktree, edited files, verification evidence, unresolved failures, and next concrete action.
3. Record evidence locations rather than dumping all tool output. Separate facts from hypotheses and preserve the original task when a recent message merely steers it. Do not promote unverified session guesses into permanent rules.
4. Check that the next agent can resume from those files. Commit status and artifact paths matter; a plan that points at vanished temporary files is not durable state.
5. Use the host's documented compaction mechanism when available. Re-read the working set and relevant project instructions afterward, then continue the same task without rerunning completed work solely because context changed.

Disk files and Git state persist independently of a conversation, but the host's summary retention guarantees vary. Do not claim all preferences or tool results survive automatically. Load skills and references on demand, avoid duplicate reads, and bound query results by the current question. Measure context savings in the actual harness rather than importing unsupported reduction percentages.