Reliable context checkpoints

Create a small task-local checkpoint when a substantial phase finishes, the runtime signals context pressure, or another session must resume the work. Use actual runtime signals when available; tool-call counts and response latency do not measure remaining context. Do not install global hooks or modify personal memory merely to make a checkpoint.

Persist what changes continuation

Record the current objective, accepted corrections, authorization boundaries, hard constraints, and next actionable step. Preserve exact deadlines and not-before times with a timezone; elapsed conversation time is not permission. Distinguish required missing input from optional preferences.

For large collections, maintain a ledger of stable item IDs with explicit states, such as pending, reviewed, revised, and evaluated. Record the total item universe and evidence per item. Enumerating or opening a file is not completing its review. Keep archived records if sources disappear, and invalidate affected conclusions when their source commit or content hash changes.

Link the files, commits, commands, test logs, and artifacts needed to justify completed work. Record failed checks, unresolved symptoms, active process IDs, pending approvals, and meaningful decisions. Prefer pointers to raw evidence over repeating large tool outputs or preserving every abandoned thought.

Write and verify the checkpoint before requesting compaction or a handoff. Use the runtime's supported mechanism if one exists; the checkpoint remains useful without a /compact command or a hook script.

Resume without losing scope

Read the latest user instruction, repository guidance, checkpoint, and affected worktree state. User corrections take precedence over older checkpoint text. Verify volatile facts that affect the next action, such as remote heads, active processes, access, and time gates. Recheck saved evidence if its content or input signature changed; do not repeat already justified work solely because a new session began.

Continue from the next eligible incomplete item. Keep deferred work visible until its actual condition is satisfied. Update the checkpoint after each bounded batch, and distinguish achieved goals from work that merely stopped.