GrowthBook experiment tengu_amber_redwood silently reduces autocompact window to ~80k on Sonnet 200k (v2.1.92+ regression)

Summary

GrowthBook experiment tengu_amber_redwood silently reduces the autocompact window on Sonnet 4.6 (200k context) to an effective ~80k tokens, causing compaction at ~23-25% context usage. This is the same class of bug as #43989 (Opus 1M → 400k) but affecting standard 200k context models — a much wider user base.

The built-in notification jz7() that should warn users about experiment-driven compaction ("compacted at {X} · override with CLAUDE_CODE_AUTO_COMPACT_WINDOW=...") does not appear to fire or render visibly.

Evidence

Measured compaction events from JSONL session transcripts (v2.1.100)

| Session | Model | preTokens | % of native window | Expected threshold |
|---------|-------|-------------|---------------------|--------------------|
| bb7608fc | sonnet-4-6 | 48,275 | 24.1% of 200k | 167k |
| bb7608fc | sonnet-4-6 | 45,281 | 22.6% of 200k | 167k |
| bb7608fc | sonnet-4-6 | 46,771 | 23.4% of 200k | 167k |
| 02395a86 | sonnet-4-6 | 50,206 | 25.1% of 200k | 167k |
| 904ab744 | opus-4-6 | 75,414 | 7.5% of 1M | 967k |
| 904ab744 | opus-4-6 | 47,353 | 4.7% of 1M | 967k |
| 904ab744 | opus-4-6 | 50,231 | 5.0% of 1M | 967k |

Cluster: 6/7 events trigger at 45,281–50,231 tokens (mean: 47,886). All well below the expected threshold.

Reverse-engineering the experiment value

From the source code (deobfuscated from v2.1.100 binary):

autoCompactThreshold = contextWindow - maxOutputTokens(20k) - buffer(13k)

If threshold ≈ 47k → effectiveWindow ≈ 60k → experiment is setting window ≈ 80k.

For Sonnet 200k: 80k is a 60% reduction of the native context window.
For Opus 1M: the 400k reported in #43989 is also a 60% reduction — consistent ratio.

Window resolution priority (from binary)

function vt(model, settingsWindow) {
    let nativeWindow = getModelWindow(model);
    // 1. env var CLAUDE_CODE_AUTO_COMPACT_WINDOW → highest priority
    // 2. user settings (autoCompactWindow)
    // 3. GrowthBook "tengu_amber_redwood" ← the problem
    // 4. model native window ← never reached when enrolled
}

Notification that should fire but doesn't

function jz7(app, model, window) {
    let {source, configured} = vt(model, window);
    if (source !== "experiment") return;
    app.addNotification?.({
        key: "autocompact-experiment-hint",
        text: `compacted at ${configured} · override with CLAUDE_CODE_AUTO_COMPACT_WINDOW=1000000`,
        priority: "medium"
    });
}

This notification exists specifically for experiment-enrolled users but either doesn't render or is dismissed before it can be seen.

Steps to Reproduce

1. Install Claude Code v2.1.100 on macOS arm64
2. Use model sonnet-4-6 (standard 200k context, NOT the 1M variant)
3. Have a conversation with 3-5 turns involving tool use (file reads, bash commands)
4. Observe compaction triggers at ~45-50k tokens (~23% of 200k)
5. Check JSONL transcript: compact_boundary events show preTokens ≈ 47-50k

Expected Behavior

Sonnet 4.6 with 200k context should compact at:

• 200k - 20k(output) - 13k(buffer) = 167k tokens (~83.5% usage)

Actual Behavior

Compaction triggers at ~47-50k tokens (23-25% usage), wasting 75% of the context window.

Sessions become unusable: after compaction reduces to ~20-30k tokens, only 2-4 tool-use turns fit before the next compaction — creating a thrashing loop. In one session, 3 compactions occurred within 6 minutes.

Additional issues

1. CLAUDE_AUTOCOMPACT_PCT_OVERRIDE cannot raise the threshold — the code uses Math.min(userPct, defaultThreshold), silently ignoring any value above default (#31806)
2. DISABLE_AUTO_COMPACT=1 does not affect microcompacts — three additional silent compaction mechanisms (time-based microcompact, cached microcompact, tool-result budget via tengu_hawthorn_window) run independently
3. No opt-out for experiments — users cannot inspect or disable GrowthBook flags without reverse-engineering the binary. The experiment silently overrides model capabilities that users are paying for.

Workaround

Add to ~/.claude/settings.json:

{
  "env": {
    "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "1000000"
  }
}

This takes priority position 1 in vt(), overriding the experiment. Note: reportedly does not work on all arm64 native builds (#43989 comment by @creatornader).

Environment

• Claude Code version: 2.1.100
• Model: claude-sonnet-4-6 (200k) and claude-opus-4-6 (opusplan mode)
• Platform: macOS Darwin 25.4.0 (Apple Silicon arm64)
• Terminal: iTerm2 / zsh
• Regression since: v2.1.92 (confirmed by #43989)

Related issues

• #43989 — same experiment, Opus 1M → 400k
• #42375 — same class of bug, Opus 1M at 6%
• #42817 — consolidated report: multiple disable methods silently ignored
• #31806 — CLAUDE_AUTOCOMPACT_PCT_OVERRIDE capped by Math.min
• #45977 — DISABLE_AUTO_COMPACT=1 bypassed on session resume
• #42542 — silent microcompacts not affected by any user override