Context Checkpoint

Measured usage for this session (may be unavailable):

!python "${CLAUDE_SKILL_DIR}/scripts/context_usage.py" "${CLAUDE_SESSION_ID}"

You are deciding whether the session should reset, and recommending — you cannot run /compact or /clear yourself; only the user can.

1. Get a real number

• If the measurement above shows a percentage, use it.
• Otherwise ask the user to run /context (or read their status line) and tell you the figure. Never guess a percentage, and never act on a guessed one.

2. Under 40%: report and continue

One line — "Context at ~N%, no reset needed" — then back to the task.

3. At or past 40%: recommend a reset at the next clean boundary

A clean boundary is a just-committed step, a just-delivered review or QA report — any moment the working state is written down. First finish reaching the boundary (complete the commit, deliver the artifact); never stop mid-edit. Then recommend exactly one of:

Clear and reseed — preferred when state is written down. If the plan document is committed and current: have the user fill the session-handoff template (bundled docs/templates/SESSION-HANDOFF.md if the repo has it; otherwise: current step, decisions not yet in the plan, next action, known landmines — you can draft it for them), then /clear, then seed the new session with the plan and the handoff. In Claude Code, /clear keeps the old conversation available under /resume, so nothing is lost.

Compact — when state is NOT yet written down. Mid-task with decisions that exist only in conversation: offer to write the handoff note now, and give the user a ready-to-paste command with explicit preservation instructions, e.g.:

/compact preserve: the plan at docs/plans/<file>, current step <N> and its
status, decisions made this session, unresolved findings, the exact next action

Past 60%, or if auto-compact has already fired this session: recommend clear-and-reseed, not compact — the deliberate reset is overdue.

4. Rot symptoms override the number

If the session shows two or more of — re-reading files already read, re-asking answered questions, contradicting a recorded decision, forgetting a context-file or plan constraint — recommend the reset now, at whatever the usage number is, and name the symptoms. Do not offer "just push on" as the default option.

Rules

• A measured or user-provided number only; never invented.
• Never claim to have compacted or cleared — recommend, with the exact command, and let the user run it.
• 40% is the default threshold (rationale: docs/07-context-hygiene.md); honor a different threshold if the user has set one.
• The percentage is of the actual window. On Fable 5 / Sonnet 5 that's a 1M window, so 40% there is far more tokens than 40% of 200K — the rule still holds, but the rot symptoms matter more than the raw number. The bundled script sizes the window from the model (Fable, Mythos, Sonnet 5, and [1m] names); for other 1M configs like opus[1m] it assumes 200K until usage passes that, so set CONTEXT_CHECKPOINT_WINDOW or confirm with /context. If Fable 5 itself offers to start a fresh session, that's the model being cautious, not this rule firing — the deliberate reset still triggers on the threshold and the symptoms.
• This skill checks only when invoked. The automatic nudge comes from the optional hook (scripts/context_hook.py next to this file — setup instructions in its header). If the user relies on the 40% rule, recommend once per project that they install the hook and a status line showing context_window.used_percentage.