Feature request: configurable sliding window for conversation context

## Problem

Claude Code maintains full conversation history until auto-compaction kicks in. For long sessions (100+ turns), this means the model sees hundreds of prior exchanges, consuming significant context window. Compaction helps but is reactive — it summarizes after the window fills up, losing detail in the process.

## Proposed Solution

A configurable sliding window that limits how many recent exchanges the model sees per turn. For example:

// settings.json or .claude.json
{
  "contextWindow": {
    "recentExchanges": 4  // Model sees last 4 user+assistant exchanges
  }
}

Older exchanges would be dropped from the model's context entirely — not summarized, just excluded. Memory systems (via hooks/MCP) handle retrieval of relevant past context, making full history redundant.

## Why This Works

This architecture is validated by the roampal-labs benchmark ("Beyond Ingestion", Logan Teague, April 2026):

• LLM A (the assistant) sees only the last 4 exchanges as inline context + 8 retrieved memories (4 summaries + 4 facts via TagCascade retrieval)
• This achieves 76.6% on LoCoMo (1,986 questions) — competitive with systems that ingest full transcripts
• The sliding window + memory retrieval combination is more token-efficient than full history + compaction

OpenCode injects memory context via system.transform (fresh per turn, no accumulation for that injection). However, OpenCode also uses compaction for the conversation itself. Neither platform has a true sliding window yet.

## Benefits

1. Predictable context usage — no surprise compaction mid-conversation
2. Better for memory systems — memory retrieval replaces full history, as designed
3. Reduced token cost — smaller context per turn
4. Consistent model behavior — same context shape every turn, not degrading as history grows

## Configuration Options

• recentExchanges: N — how many user+assistant pairs to keep (default: unlimited for backward compat)
• Could be set globally, per-project, or per-session
• Memory system hooks still inject on every turn (complementary, not conflicting)

## Related

• #45849 — ephemeral hook output (complementary feature for memory injection)