Summary

scripts/hooks/suggest-compact.js currently triggers purely on cumulative tool-call count (first at COMPACT_THRESHOLD/50, then every 25). Tool count is a weak proxy for actual window pressure, and adding a context-size signal would make the suggestion fire when it actually matters.

Why tool count alone is weak

• A few large operations (big file reads, large MCP tool responses) can fill a large fraction of the window in very few tool calls — the count signal stays silent.
• Conversely, many tiny calls can cross 50 while the window is barely used — it nags early.
• In MCP/Bash-heavy sessions the count climbs without reflecting real context growth.

Proposal

Add a second (primary) signal based on the real context size, read from the latest assistant turn's usage in the session transcript. transcript_path is already provided in the PreToolUse hook stdin payload.

• Sum input_tokens + cache_read_input_tokens + cache_creation_input_tokens from the most recent usage record (these three partition the prompt, so the sum is the true context size of the turn).
• Compare to a window-scaled threshold (e.g. 250k on a 1M-context model, 160k on a 200k window; env-overridable), and re-remind every +N tokens (e.g. 60k).
• Detect the window: treat as 1M if the model id carries a [1m] marker, or if observed tokens already exceed 200k (covers logs that drop the suffix); else 200k.
• Surface via the same output({ hookSpecificOutput: { hookEventName: 'PreToolUse', additionalContext } }) path the count signal already uses (post-PreToolUse warn-only hooks: stderr output not visible to users #2082), so it stays non-blocking.

Reference sketch

function readContextTokens(transcriptPath) {
  const lines = fs.readFileSync(transcriptPath, 'utf8').split('\n');
  for (let i = lines.length - 1; i >= 0; i--) {
    const l = lines[i].trim(); if (!l) continue;
    let rec; try { rec = JSON.parse(l); } catch { continue; }
    const u = rec?.message?.usage; if (!u) continue;
    const t = (u.input_tokens||0)+(u.cache_read_input_tokens||0)+(u.cache_creation_input_tokens||0);
    if (t > 0) return { tokens: t, model: rec.message.model || '' };
  }
  return null;
}

Threshold crossing is tracked with a small per-session "bucket" so it only re-fires when context grows by another step.

I'm happy to open a PR if this direction is welcome.

_{Filed after auditing a local copy of the skill; surfaced via Claude Code.}