V:/Program Files/Git/compact auto-read control — post-compact context management

Feedback: /compact Auto-Read Control — Post-Compact Context Management

Filed: 2026-04-04
Product: Claude Code (CLI)
Severity: MEDIUM — efficiency + token waste, not data loss
Version tested: 2.1.91, Opus 4.6 (1M context)
Related: Cache replay cost bomb (sibling feedback)

---

The Problem

After /compact (auto-compact or manual), Claude Code automatically re-reads recently-used files back into context. This creates two issues:

1. Double injection with hooks

Users who have UserPromptSubmit hooks that inject context — project status, configuration, task queues, team notifications, custom prompts — get double content on the first post-compact turn:

• Compact's auto-read of recently-used files (configs, source files, docs)
• Hook injections of overlapping content (the same configs, status files, project state)

Example: a hook injects the project's TODO.md on every prompt. Compact also auto-reads TODO.md because it was recently used. The model receives TODO.md twice. Scale this across multiple hooks (status dashboards, linter configs, test results, team chat) and the first post-compact turn adds 10-20K tokens of redundant content.

2. Auto-reads don't create useful context

The model receives the file contents but doesn't know WHY they were being used. Post-compact, the summary says "we were editing server.py" but the auto-read dumps the entire file without the model knowing it was mid-refactor on the authentication middleware. The model typically re-reads the file anyway to orient itself — making the auto-read a pure token waste.

---

Proposed Options

Option 1: No auto-reads — prioritized file list only

Post-compact, inject ONLY:

• The compact summary (already generated)
• A prioritized list of recently-used files with one-line context: server.py (editing, auth middleware refactor), config.yaml (read for DB credentials)
• Let the model (or user) decide what to re-read

Pro: Minimal post-compact bloat. User control.
Con: Model is cold-booted with a list it may not interpret well without more context. Requires good summary quality.

Option 2: Summary-only to separate file

Write the compact summary to a session-scoped file. This file is the ONLY auto-injection post-compact. No file re-reads. The summary is the complete orientation document.

Pro: Clean architecture. Single source of truth. Small injection.
Con: Single point of failure — if the summary dropped important context (active task, mid-edit state), there's no fallback. Summary quality is model-dependent.

Option 3: Current behavior (keep as-is)

Auto-read everything that was recently used.

Pro: Model has working context immediately. Files serve as ground truth even when summary is incomplete.
Con: Redundant with hook injections. Wastes tokens. Doesn't respect the user's hook investment.

Recommended: User-configurable behavior

Expose the auto-read behavior as a setting:

{
  "compact": {
    "autoReadFiles": "none" | "list" | "full",
    "injectSummary": true,
    "suppressHooksPostCompact": false
  }
}

• "none": summary only, no file re-reads
• "list": summary + file list with context hints (default)
• "full": current behavior (backward compatible)
• "suppressHooksPostCompact": if true, skip UserPromptSubmit hooks for the first turn after compact (prevents double injection)

---

Why This Matters

Users who invest in hook systems — custom dashboards, project state injection, notification feeds, automated context management — are penalized by /compact because hooks and compact's auto-reads fight over the same context space. The first post-compact turn should be the leanest turn in the session. Instead, it's the most bloated.

This compounds with the cache replay issue (sibling feedback): if the post-compact turn is large, it's also more likely to break the prompt prefix cache, causing a full replay on the next message.

---

Based on observed double-injection patterns in a production hook system with 8+ UserPromptSubmit hooks.