Add PreCompact hook to preserve original goal and hard constraints #27

Part of #26.

Problem

As conversations get longer, Claude Code compacts older messages to stay within the context window. The compaction summary frequently drops:

• The user's original ask (the why)
• Hard constraints declared early ("don't touch the database layer", "preserve behavior")
• CLAUDE.md rules that were loaded once and never re-loaded

Result: Claude drifts from the original intent. anthropics/claude-code#22638 documents this exact failure mode — Claude ignored CLAUDE.md rules mid-session and executed a destructive git command. Community writeups (Fortune, The Register, dev.to "Feb-Mar 2026 regression") cite instruction drop-off as one of the top complaints.

Proposed solution

Add a PreCompact hook (released v2.1.105, Apr 13 2026) that:

1. Reads the original user prompt(s) and any explicit constraints from the conversation
2. Re-injects them as additionalContext immediately after compaction so they survive into the post-compaction window
3. Optionally returns {"decision":"block"} if the constraints would be lost and the user has opted into strict mode

File changes

• New: scripts/preserve-context.py — PreCompact hook handler
• Update: hooks/hooks.json — register the new event
• Update: README.md — document the new behavior + opt-out
• New tests: tests/test_precompact_hook.py

Implementation sketch

# scripts/preserve-context.py
import json, sys

data = json.load(sys.stdin)
# Walk transcript_path; find first user message + any "constraint markers"
# (lines containing "must", "do not", "preserve", "only", "never")
preserved = extract_constraints(data["transcript_path"])  # truncated to budget

print(json.dumps({
    "hookSpecificOutput": {
        "hookEventName": "PreCompact",
        "additionalContext": f"<original-goal-and-constraints>\n{preserved}\n</original-goal-and-constraints>"
    }
}))

Token budget

• Total injected on compaction: ≤200 tokens, comprising:
  • Original first user prompt: ≤120 tokens (truncate with ellipsis if longer)
  • Detected hard constraints: ≤80 tokens (keep highest-priority first; drop overflow)
• Zero tokens injected when no compaction is occurring (hook fires only on PreCompact event)
• Zero tokens injected if no constraints detected and original goal already lives in retained context (skip injection entirely)
• Constraint extraction runs in <50ms; no LLM calls

Acceptance criteria

• PreCompact hook fires on compaction events
• Original first user prompt is re-injected post-compaction (within budget)
• Detected hard constraints (negation patterns, "must"/"never"/"only") are re-injected
• Total injection ≤200 tokens (truncate cleanly with … markers)
• Hook completes in <100ms (no heavy parsing)
• Bypass: users can disable via env var (PROMPT_IMPROVER_PRESERVE=0)
• Tests cover: no constraints, multi-constraint, malformed transcript, over-budget truncation

References

• PreCompact docs: https://code.claude.com/docs/en/hooks
• Changelog v2.1.105: https://code.claude.com/docs/en/changelog.md
• Drop-off complaint: [MODEL] Claude repeatedly ignored CLAUDE.md rules, executed destructive git command causing data loss anthropics/claude-code#22638
• Feb-Mar 2026 regression analysis: https://dev.to/shuicici/claude-codes-feb-mar-2026-updates-quietly-broke-complex-engineering-heres-the-technical-5b4h