Sentinel — 24/7 Autonomous Code Review

Sentinel is a built-in Ouro Loop module that runs continuous, unattended code review loops on any project. It scans your codebase partition-by-partition, finds real issues across 7 dimensions, fixes what it can safely, and records everything.

Production-validated

Sentinel's architecture was extracted from a system that ran 292 rounds over 13 hours on a production blockchain codebase, finding 438 issues (17 CRITICAL) with 100% fix success rate and zero abandoned fixes.

---

Quick Start

pip install ouro-loop

cd your-project/
ouro-sentinel init .       # Scan -> detect commands -> generate partitions + config
ouro-sentinel install .    # Generate runner + dashboard scripts
make sentinel-start        # Start the 24/7 review loop
make sentinel-dashboard    # Watch live progress

---

Prerequisites

• Python 3.10+ with pip install ouro-loop
• Claude Code CLI (claude) in PATH
• Git -- used for activity scoring, worktree-based fixes, and PR creation

Security

The runner uses --permission-mode bypassPermissions for unattended operation. Only run in trusted, sandboxed environments.

---

What init Does

When you run ouro-sentinel init ., Sentinel:

1. Scans your project (languages, file count, LOC)

2. Detects build/test/lint commands automatically:

   Marker File	Build	Test	Lint
   go.mod	go build ./...	go test ./...	go vet ./...
   Cargo.toml	cargo build	cargo test	cargo clippy
   package.json	npm run build	npm test	npx eslint .
   pyproject.toml	--	python -m pytest	ruff check .
   pom.xml	mvn compile	mvn test	mvn checkstyle:check
   Makefile	make build	make test	make lint

3. Generates partitions -- directories scored by risk:

   • high -- overlaps with DANGER ZONES from your CLAUDE.md
   • medium -- high git activity or file count
   • low -- default

4. Renders a Sentinel-specific CLAUDE.md with your BOUND rules inherited

5. Creates .ouro/sentinel/ with all state files

---

The Review Loop

Each Claude session executes this 6-step loop:

1. MAP -- Select Review Target

Priority formula: recency x 0.30 + criticality x 0.25 + staleness x 0.25 + density x 0.10 + gap x 0.10

Review rhythm: every iteration = 1 file, every 10 = module-level, every 30 = cross-module.

2. SCAN -- 7-Dimension Analysis

Dimension	What to Look For
Security	Injection, auth bypass, secrets, unsafe crypto
Quality	Dead code, complexity, naming, error handling
Performance	N+1 queries, unnecessary allocations, blocking I/O
Test Coverage	Missing tests, edge cases, test quality
Architecture	Coupling, SRP violations, dependency direction
Doc Sync	Stale comments, missing docs, misleading names
Project Rules	BOUND violations, IRON LAW compliance

Only findings with confidence > 0.8 are recorded.

3. FIX -- Isolated Repair

Each fix runs in an isolated git worktree -> build -> test -> commit -> PR. Requires confidence >= 0.9, not in DANGER ZONE, blast radius <= 3 files.

4. VERIFY -- Three-Layer Check

L1: ouro verify . | L2: BOUND compliance | L3: Lint command

5. REMEDIATE -- Autonomous Decision

3 retries max. After 3 failures -> human-intervention-required. Never asks for help.

6. LOOP -- Record and Rotate

Update state, log iteration, write learnings every 10 rounds. At ~70% context -> ROTATE. If 5 consecutive zero-finding iterations -> DONE.

---

Runner — The Immortal Daemon

The runner is designed to survive everything short of a reboot. When you run make sentinel-start, the process detaches from your terminal and becomes immortal:

make sentinel-start
  └→ nohup sentinel-runner.sh & disown
       └→ Terminal closes? SIGHUP absorbed by nohup
            └→ macOS launchd (PID 1) adopts the orphan
                 └→ Sleep/wake? launchd children survive
                      └→ Result: sentinel lives until you kill it

No systemd service files. No launchd plists. No Docker. Just Unix primitives — nohup, &, disown — and the OS doing what it was designed to do.

Battle-tested

This mechanism was validated in production: two runner processes survived 33+ hours across multiple macOS sleep/wake cycles, terminal closures, and SSH disconnects — accumulating 331 review rounds and 469 findings without interruption.

.ouro/sentinel/sentinel-runner.sh start    # Start immortal daemon
.ouro/sentinel/sentinel-runner.sh stop     # Graceful shutdown (kills orphans too)
.ouro/sentinel/sentinel-runner.sh status   # Check if running
.ouro/sentinel/sentinel-runner.sh restart  # Stop + start

Features:

• Immortal process: nohup + disown = survives terminal close, SSH disconnect, sleep/wake
• Orphan-proof stop: PID file + pgrep fallback — no zombie processes even if PID file is lost
• Duplicate prevention: cmd_start checks both PID file and process table before launching
• Session rotation: ROTATE/DONE/BLOCKED signals from Claude sessions
• Crash recovery: validates state, backs up, cleans up worktrees between sessions
• Log rotation: gzip at 10MB, interruptible sleep between sessions

---

Dashboard

.ouro/sentinel/sentinel-dashboard.sh           # One-shot display
.ouro/sentinel/sentinel-dashboard.sh --watch   # Auto-refresh every 5s

Shows: runner status, severity breakdown, coverage progress bar, recent iteration history.

---

Configuration

Key settings in sentinel-config.json:

Setting	Default	Description
confidence_threshold	0.8	Minimum confidence to record a finding
fix_confidence_threshold	0.9	Minimum confidence to attempt a fix
blast_radius_limit	3	Maximum files changed per fix
max_fix_attempts	3	Retries before marking human-required
auto_pr	false	Automatically create PRs for fixes
model	claude-opus-4-6	Claude model for review sessions
max_turns	200	Maximum tool calls per session
cooldown_seconds	30	Pause between sessions

---

CLI Reference

ouro-sentinel init <path>        # Initialize sentinel for a project
ouro-sentinel partition <path>   # Regenerate partitions
ouro-sentinel status <path>      # Show iteration count, findings, coverage
ouro-sentinel install <path>     # Install runner + dashboard + Makefile targets