Add `OnContextThreshold` hook for programmatic mid-session handoff

Summary

Expose a hook that fires when session context crosses a configurable percentage threshold (e.g., 65%, 80%), so users can automate state-saving and session handoff before auto-compaction kicks in around 95%.

Use case

I run multi-day autonomous build pipelines where a single CLI session can chew through context over the course of a long task. The current options are bad:

• Auto-compaction at ~95% is too late — by then the model has lost coherence on early decisions and compaction itself can drop important state.
• Manual monitoring of the statusline context % works but defeats the point of an autonomous pipeline.
• No hook today exposes context usage to programmatic logic. PreCompact fires too late and gets no token-count payload.

What I want is to fire a hook at 65%, save my memory files (auto-memory dir + Mem0 entries) with full fidelity, write a handoff note, then exit cleanly so a wrapper can spawn a fresh session that reads the handoff and continues.

Proposed shape

// settings.json
{
  "hooks": {
    "OnContextThreshold": [{
      "threshold": 0.65,           // fires once when crossed
      "command": "bash ~/.claude/handoff.sh"
    }]
  }
}

Hook input (stdin JSON):

{
  "tokens_used": 130000,
  "tokens_remaining": 70000,
  "percent": 0.65,
  "session_id": "...",
  "transcript_path": "..."
}

Multiple thresholds should be allowed (e.g., 65% warn, 85% hard-stop).

Why this is broadly useful

• Long-running agentic pipelines (autonomous builds, monitoring loops, multi-stage refactors)
• Context-budget enforcement in CI/CD-style Claude Code usage
• Graceful handoff between sessions without relying on auto-compaction quality
• Pairs naturally with the existing memory/Mem0 ecosystem — finally lets persistence be triggered by time-to-overflow instead of guesswork

What exists today doesn't cover this

• \PreCompact\ hook → fires too late, no token data in payload
• Statusline → rendered for the user, not piped anywhere programmable
• \/context\ slash command → display only, not a trigger
• Cowork / \schedule\ routines → external scheduled processes can't read the running session's context %
• Wrapper scripts that tail the transcript file → fragile (file size ≠ tokens, no graceful exit signal)

A first-class \OnContextThreshold\ hook would unlock a clean primitive that today requires brittle wrappers.

Workaround I'm using meanwhile

Wrapper script that estimates tokens from transcript file size + a CLAUDE.md rule telling the model to self-check \/context\ every ~20 turns and write a handoff if over threshold. Both are heuristic.