claude-context-optimizer

Claude Code sessions tend to forget earlier decisions as conversations grow long.
This tool hooks into Claude Code's lifecycle to automatically record decisions, inject them back into context, and version-control them — so nothing gets lost.

How it works

SessionStart  →  Read decisions.json → Inject into CLAUDE.md  (every session)
     ↓
  Claude works...
     ↓
Stop          →  Parse transcript → Detect decisions → Append to decisions.json
     ↓
SessionEnd    →  Save versioned snapshot → Update context_summary.md

On every session start, your active decisions are injected into CLAUDE.md automatically, so Claude always knows what has been decided — even in a brand-new session.

Quick start

# Set up in your current project
npx claude-context-optimizer

# Or for all projects at once (~/.claude)
npx claude-context-optimizer --global

That's it. Open Claude Code and the hooks are active.

What gets installed

.claude/
├── settings.json              ← Hook registrations (merged if exists)
├── hooks/
│   ├── session_start.py       ← Injects decisions → CLAUDE.md
│   ├── on_stop.py             ← Auto-detects decisions from transcript
│   ├── session_end.py         ← Snapshots + summary update
│   └── decision_manager.py    ← Manual CLI
└── context/
    ├── decisions.json          ← Versioned decision log
    ├── decisions_history/      ← Snapshots (latest 30 kept)
    └── context_summary.md      ← Human-readable summary
CLAUDE.md                       ← Optimizer section appended (merged if exists)

Existing files are never destroyed. settings.json hooks are merged, CLAUDE.md gets a new section appended.

Automatic decision detection

Just speak naturally in your session. Patterns like these are auto-detected:

• 「〜を採用します」「〜に決めました」「〜方針で進めます」（Japanese）
• "decided to ...", "going with ...", "we'll use ..." (English)

Detected decisions are saved to decisions.json with a version number and category.

Manual management (CLI)

# Add a decision
python3 .claude/hooks/decision_manager.py add "Use PostgreSQL" \
  --content "RLS-based multi-tenancy" \
  --category tech_stack

# List active decisions
python3 .claude/hooks/decision_manager.py list
python3 .claude/hooks/decision_manager.py list --category architecture

# Version-up an existing decision (old one becomes 'superseded')
python3 .claude/hooks/decision_manager.py update dec_0001 \
  --title "Use PostgreSQL + PGVector" \
  --content "Added vector search requirement"

# Mark a decision as no longer valid
python3 .claude/hooks/decision_manager.py supersede dec_0002 \
  --reason "Requirements changed"

# View snapshot history
python3 .claude/hooks/decision_manager.py history

decisions.json schema

{
  "version": "1.0.0",
  "last_updated": "2026-03-25T12:00:00",
  "decisions": [
    {
      "id": "dec_0001",
      "version": "v1",
      "title": "Use PostgreSQL",
      "content": "RLS-based multi-tenancy",
      "category": "tech_stack",
      "status": "active",
      "timestamp": "2026-03-25T12:00:00",
      "supersedes": null,
      "tags": []
    }
  ]
}

Status values

| status | meaning | |---|---| | active | Currently valid decision | | superseded | Replaced by a newer version | | archived | Manually retired |

Categories

architecture · tech_stack · api · ui_ux · infrastructure · policy · business · general

Customization

Add detection patterns (on_stop.py)

DECISION_PATTERNS = [
    (r"(.{5,80})(?:で行く|でいく)", "decision"),  # add your own
    ...
]

Change what's injected into CLAUDE.md (session_start.py)

Edit format_decisions_for_context() to control how decisions appear in every session.

Requirements

• Claude Code (any recent version)
• Node.js ≥ 18 (for npx)
• Python ≥ 3.8 (stdlib only, no pip installs needed)

Contributing

See CONTRIBUTING.md for ideas, dev setup, and test procedures.

License

MIT © 2026 gajumaru4444