Warm Compacting in Power Claude — skip the context-wall wait
Neural-LLM Admin··3 min read
Long Claude Code sessions eventually hit a context wall. Claude Code then spends a long turn compacting history — you wait. Power Claude’s Warm Compacting pre-builds that summary in the background so when the wall arrives, the expensive wait is already done.
TL;DR
| Question | Answer |
|---|---|
| What is it? | Pre-build of the compaction summary before Claude Code autocompacts |
| Default | **On** (`powerClaude.warmCompacting.enabled`) |
| What you feel | Fewer multi-minute “stuck compacting” pauses on long jobs |
| Risk knobs | Token caps, fire %, apply-rewrite gate |
| Not | Bypassing model limits or Anthropic rate limits |
Why compacting hurts
Claude Code’s conversation grows with tool output, diffs, and plans. Near the model window, the product **compacts** older turns into a shorter summary. That compact turn is useful — and slow. Heavy agent fleets hit it mid-sprint.
Warm Compacting watches proximity to the compact threshold (configurable percent of the window), builds an abstractive summary of the earlier conversation **before** you’re blocked, and can inject it when the real compact would fire.
How to enable (VS Code)
1. Install Power Claude.
2. Open Settings → search `warmCompacting`.
3. Confirm **Warm Compacting: Enabled** is on (default).
4. Optional: set `tokenSaver.level` / Token Saver so less junk enters the window in the first place.
5. Run a long session; watch Debug / events for warm-compact activity if you need proof.
Settings that matter
| Setting | Role |
|---|---|
| `powerClaude.warmCompacting.enabled` | Master switch |
| `…autocompactPercentOfWindow` | Where “the wall” is estimated (default ~90%) |
| `…triggerPercentOfCompactThreshold` | When to start pre-build (default ~80% of that wall) |
| `…maxWarmSummaryTokens` | Cap size of the warm summary |
| `…applyRewrite` | Whether to actually splice the warm compact live |
| `…maxWarmCompactsPerSession` | Efficiency cap (default 1) |
Exact ranges: Power Claude **Settings** catalogue / `docs/system-index.md` in the extension repo.
When to leave it on
- Overnight or multi-hour agent loops
- Sub-agent fan-out that floods tool output
- Any workflow where a random 2–5 minute compact kills flow
When to tune or disable
- You need maximum fidelity of early-session detail (raise summary token cap carefully, or disable apply-rewrite and observe).
- You’re debugging compact behavior itself — use shadow-style caution and Debug tab.
- Emergency: global kill switch `~/.power-claude/state/emergency-off`.
Pair with Token Saver
Warm Compacting reduces **wait**. Token Saver reduces **how fast you race toward the wall**. Together they compound: less junk in context + less stall when compacting.
→ Token Saver guide
FAQ
Does this violate Anthropic terms?
No. It works with **local session context** and official Claude Code behavior. It does not invent extra Anthropic quota.
Is the summary lossy?
Yes — compaction is always lossy. Warm Compacting’s job is timing and continuity, not perfect archival. Critical facts still belong in files/tests, not only chat memory.
How do I know it fired?
Dashboard events / Debug diagnostics; long sessions that previously froze at the wall should continue with a short warm path instead of a full stall.
- Power Claude system index: Warm Compacting settings
- Video storyboard: `media/video/scenes/warm-compacting/` in the extension repo
- Product: neural-llm.com/power-claude
