[FEATURE] Per-model configurable autocompact threshold

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

With the 1M context window now available for Opus, a single global CLAUDE_AUTOCOMPACT_PCT_OVERRIDE is no longer sufficient. Different models benefit from different compaction strategies, and users need per-model control over when autocompact triggers.

Motivation

When using Opus (1M context) as the primary model that spawns Sonnet subagents, the optimal autocompact threshold differs significantly between them:

Use Case Example

• Opus (1M): I may want to compact at ~30% (300K tokens) to keep costs manageable and responses fast, since the full 1M window is rarely needed
• Sonnet subagents: I want them to use a larger portion of their context (e.g., 80-90%) since they handle focused, bounded tasks where more context is useful

Currently, CLAUDE_AUTOCOMPACT_PCT_OVERRIDE applies globally — forcing a single trade-off that can't serve both use cases.

Proposed Solution

Allow per-model autocompact thresholds in settings.json:

{
  "autoCompactThreshold": {
    "default": 80,
    "opus": 30,
    "sonnet": 90,
    "haiku": 95
  }
}

Or alternatively, extend the environment variable approach:

CLAUDE_AUTOCOMPACT_PCT_OPUS=30
CLAUDE_AUTOCOMPACT_PCT_SONNET=90
CLAUDE_AUTOCOMPACT_PCT_HAIKU=95
CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=80  # fallback default

Why Not Use the Existing Global Override

| Scenario | Global at 30% | Global at 90% |
|----------|---------------|---------------|
| Opus (1M) primary session | Good — keeps cost/latency down | Bad — consumes up to 900K tokens before compacting |
| Sonnet subagents | Bad — compacts too early, loses useful context for focused tasks | Good — maximizes available context |

There is no single value that works well for both.

Priority

High - Significant impact on productivity

Feature Category

Configuration and settings

Additional Context

Related Issues

• #28728 — Auto-compact should trigger earlier / threshold should be configurable
• #15719 — Configurable Context Window Compaction Threshold
• #15435 — Configurable autocompact buffer size