[BUG] Autocompact triggers at 17% on 1M context model (effectiveWindow likely capped at 200k)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

Auto-compaction triggers at ~17% context usage on the Opus 4.6 (1M context) model, rendering ~830k tokens effectively unusable.

Evidence from statusline debug logs

I instrumented the statusline to capture context_window data from the Claude Code API. Here are real sessions on claude-opus-4-6[1m]:

| Session | used_percentage | current_usage total | context_window_size |
|---------|----------------|--------------------|--------------------|
| ce85b6 | 15% | ~148k | 1,000,000 |
| 186445 | 17% | ~171k | 1,000,000 |
| 1376de | 21% | ~207k | 1,000,000 |
| 54f9ba | 34% | ~340k | 1,000,000 |

Session 186445 was observed compacting at 17% (screenshot available). The API correctly reports context_window_size: 1000000 and used_percentage: 17, yet compaction triggered.

Raw debug log (session that compacted at 17%)

{
  "model": { "id": "claude-opus-4-6[1m]", "display_name": "Opus 4.6 (1M context)" },
  "context_window": {
    "total_input_tokens": 8556,
    "total_output_tokens": 62054,
    "context_window_size": 1000000,
    "current_usage": {
      "input_tokens": 3,
      "cache_creation_input_tokens": 2458,
      "cache_read_input_tokens": 169011
    },
    "used_percentage": 17,
    "remaining_percentage": 83
  }
}

Root cause hypothesis

From deobfuscated source (referenced in #31806):

function getAutoCompactThreshold(model) {
  let effectiveWindow = getEffectiveWindow(model);
  let defaultThreshold = effectiveWindow - 13000;
  // ...
}

If getEffectiveWindow() returns 200k (not 1M) for the 1M model, then:

• defaultThreshold = 200,000 - 13,000 = 187,000
• 187k / 1,000,000 = 18.7% — matches observed behavior exactly

Expected Behavior

Compaction should trigger relative to the actual context_window_size (1M), not a hardcoded or capped effective window of 200k. At minimum, getEffectiveWindow() should return the real context window size reported by the API.

Actual Behavior

Compaction triggers at ~170-190k tokens regardless of the 1M context window, wasting ~80% of available context.

Is this a regression?

Unknown — behavior has been consistent since I started using the 1M model.

Environment

• Claude Code version: latest (March 2026)
• Model: claude-opus-4-6[1m] (Opus 4.6 with 1M context)
• OS: Windows 11 Pro
• Plan: Max (1M context enabled)

Related issues

• #31806 — CLAUDE_AUTOCOMPACT_PCT_OVERRIDE cannot raise threshold above default (~83%) due to Math.min clamp
• #35296 — 1M context window advertised capability concerns
• #28728 — Auto-compact threshold should be configurable