Night Mode

Safe unattended Claude Code operation overnight. Three components:

1. Command Safety Gate — PreToolUse hook that classifies every Bash command via local AI (Qwen 3B). Hard-blocks git push, logs dangerous commands, allows everything else.
2. RAM Watchdog — Polls system RAM every 30s. When >85%, gracefully kills Claude instances (biggest first, commits work before killing, never kills the last one).
3. Morning Report — Summary of overnight activity: dangerous commands, blocks, watchdog kills, branch status.

Anthropic has since shipped a first-party auto-permissions system in Claude Code itself, so this mostly lives on as a portfolio piece. Still works fine if you want a local, auditable version you own end-to-end.

Install

Clone into your Claude config dir (so logs and flag files live alongside Claude's own state):

cd ~/.claude   # or wherever your CLAUDE_CONFIG_DIR points
git clone https://github.com/<your-user>/nightmode.git
cd nightmode

python -m venv .venv
.venv/bin/pip install psutil requests pytest      # Unix
# .venv/Scripts/pip install psutil requests pytest  # Windows

Install Ollama and pull the classification model:

ollama pull qwen2.5-coder:3b

Register the hook in your Claude Code settings.json:

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          { "type": "command", "command": "python /absolute/path/to/nightmode/hook.py" }
        ]
      }
    ]
  }
}

Optional — add a shell alias:

alias nightmode="bash ~/.claude/nightmode/nightmode.sh"

Quick Start

# From a separate terminal (not inside Claude Code):
nightmode on      # starts Ollama, warms model (~60s first time), activates hook + watchdog
nightmode status  # check if active

# Run your Claude Code instances in full access mode. Sleep.

nightmode off     # deactivates everything, prints morning report

How It Works

Command Safety Gate (hook.py)

Fires on every Bash tool call via the PreToolUse hook.

• Night mode OFF: Instant pass-through (no output, no overhead)
• Night mode ON:
  • Checks command against blocked patterns (blocked-patterns.txt)
  • If blocked (e.g. git push): denied, logged
  • Otherwise: sends to Qwen 3B via Ollama for SAFE/DANGEROUS classification
  • DANGEROUS: allowed but logged with reason
  • SAFE: allowed, no log
  • Ollama timeout: allowed, logged as UNKNOWN (fail-open)

RAM Watchdog (watchdog.py)

Runs as a background process. Polls RAM every 30 seconds.

RAM %	Action
< 80%	Nothing
80-85%	Log warning
85%+	Begin graceful shutdown

Shutdown sequence: finds Claude instances sorted by memory (biggest first), runs git add -A && git commit in each working directory, then kills the process tree. Waits 10s between kills. Stops when RAM < 75% or only 1 instance remains. Never kills the last instance.

Morning Report (morning_report.py)

Run automatically by nightmode off, or manually:

nightmode report              # today + yesterday's logs
nightmode report 2026-04-03   # specific date

Configuration

config.json:

Key	Default	Description
ollama_url	http://localhost:11434	Ollama API endpoint
model	qwen2.5-coder:3b	Model for command classification
classification_timeout_ms	5000	Max time to wait for Ollama response
watchdog_poll_interval_s	30	RAM check interval
ram_warning_threshold	80	Log warning above this %
ram_kill_threshold	85	Start killing instances above this %
ram_safe_threshold	75	Stop killing when RAM drops below this %
min_surviving_instances	1	Never kill below this many instances

Blocked Patterns

blocked-patterns.txt — one regex per line, # for comments:

^git\s+push(\s|$)

Ollama binary

nightmode.sh looks for ollama on PATH. Override with OLLAMA_BIN=/path/to/ollama.

File Layout

Scripts resolve paths relative to their own location. Flag files and logs live one directory up from nightmode/:

<parent-dir>/
  nightmode.active              # exists = night mode on
  watchdog.lock                 # watchdog PID (runtime)
  safety-logs/YYYY-MM-DD.jsonl  # command classification logs
  watchdog-logs/YYYY-MM-DD.jsonl # RAM watchdog logs
  nightmode/
    hook.py            # PreToolUse hook
    watchdog.py        # RAM watchdog
    morning_report.py  # overnight summary
    nightmode.sh       # on/off toggle
    config.json        # thresholds and model config
    blocked-patterns.txt
    .venv/             # Python venv (psutil, requests, pytest)
    tests/

Running Tests

cd nightmode
.venv/bin/python -m pytest tests/ -v       # Unix
# .venv/Scripts/python -m pytest tests/ -v   # Windows

Hardware Notes

• Qwen 3B takes ~60s to cold-load into VRAM (tested on GTX 1660 Ti, 6 GB). nightmode on handles this automatically.
• Once warm, classification takes ~1.5-3.5s per command.
• The model uses ~2 GB VRAM. Ollama unloads it after inactivity, so daytime VRAM is free for other things.