PreCompact hooks need output control: replaceCompactSummary and compactNumber

Problem

The PreCompact hook event fires before context compaction, but the hook's output has no way to control what gets injected into the post-compact context. Specifically:

1. No way to replace the default summary. Custom extraction systems (external LLMs, database-backed state preservation) produce better-structured context than the default narrative summary, but their output is additive — the default summary still consumes tokens. There's no replaceCompactSummary or skipDefaultSummary output field.

1. No compact counter. There's no way to know how many times a session has compacted. This is critical for multi-compact sessions where you need to weight recent decisions higher and archive older ones. A simple compactNumber: N in the hook input or post-compact context would enable this.

1. Duplicate of locked issue. This was previously requested in #13170, which was auto-closed as a duplicate of #3349, which is itself now closed and locked with zero Anthropic response. The feature request has nowhere to live.

Proposed Solution

1. replaceCompactSummary output field

{
  "additionalContext": "...",
  "replaceCompactSummary": true
}

When replaceCompactSummary: true, the hook's additionalContext becomes the only context injected post-compact, replacing the default auto-generated summary. This lets power users control their own compaction strategy.

2. compactNumber in hook input

{
  "hook_event": "PreCompact",
  "compactNumber": 3,
  "transcript_path": "..."
}

Sequential compact counter for the current session. Trivial to implement (increment a session-scoped integer), transformative for multi-compact context management.

3. compactSummary in hook input

{
  "hook_event": "PreCompact",
  "compactSummary": "The default summary that would be injected..."
}

Give the hook access to the default summary so it can augment, restructure, or selectively include parts of it.

Why This Matters

I've built and maintained a production PreCompact system for 6 weeks (100+ sessions). The results are dramatic:

• Task continuity after compact: ~40% (default) → ~95% (structured extraction)
• CLAUDE.md rule retention: Degrades after 3+ compacts → 100% retention
• Recovery turns needed: 3-8 → 0-1

But the system wastes tokens because the default summary is always injected alongside our structured context. With replaceCompactSummary, we could save 2-4K tokens per compaction — significant in sessions that compact 3-5 times.

Accessibility Impact

Context loss after compaction disproportionately affects neuroatypical users (ADHD, executive function differences) who rely on Claude as a cognitive scaffold. When Claude forgets mid-task, these users bear the highest cognitive load to recover. Better compaction control makes Claude Code more accessible. (Related: #24949 on table rendering accessibility.)

Environment

• Claude Code version: latest (2026-02)
• OS: macOS 15.3
• Shell: zsh

Workaround

Currently using external LLM extraction + PostgreSQL persistence + additionalContext injection. Works but wastes tokens due to additive-only output.