sentinel

Sentinel

Self-improving development methodology for Claude Code.

Sentinel is a Claude Code plugin that gives your AI assistant institutional memory, quality enforcement, and continuous learning — so every session builds on the last.

What It Does

Autonomous Execution — Claude Code often asks users to run commands instead of running them itself. Sentinel fixes this with three layers: (1) a behavioral rule loaded into every session that instructs Claude to execute, never suggest, (2) auto-configured tool permissions in .claude/settings.json that eliminate permission prompts for routine dev commands, and (3) CLAUDE.md instructions that survive context compaction. The result: Claude runs tests, lints, builds, and git commands without asking.

Git Autopilot — You never touch git. Sentinel auto-creates branches when sessions start and auto-commits when they end. No branch management, no commit messages, no git knowledge required.

Concurrent Session Isolation — Run multiple Claude Code agents on the same repo simultaneously. Sentinel detects concurrent sessions, auto-creates git worktrees for isolation, and auto-merges changes back when sessions end. No conflicts, no coordination needed.

Team Collaboration — Multiple developers share vault knowledge through git. A custom merge driver prevents conflicts on vault files. A daily activity feed logs what each team member's sessions did. New members get guided onboarding via /sentinel-onboard.

Verification, Not Trust — Claude can claim "tests pass" without running them, or say "all done" with tasks still pending. Sentinel catches both. An evidence log records every test/lint/build command with its actual exit status — Claude can't retroactively claim success. A todo mirror tracks task state independently — if tasks are incomplete at session end, they're listed. The stop hook audits evidence against what should have happened: "5 Python files modified, 0 test executions found."

Verification Gap Detection — Claude fixes a narrow symptom and writes a narrow test — but the user finds new bugs in the browser. Sentinel catches this with three checks: (1) test scope breadth — warns when only a single test function was run but multiple files changed, (2) adjacent test detection — finds test files that import modified modules and warns if they weren't executed, (3) bug-fix mode — detects bug-fix tasks and enforces reproduce-first verification (a failing test should precede the fix).

Loop, Batch & Autoresearch — Some tasks are too large for one context window or need many iterations to converge. /sentinel-loop runs a task repeatedly until a condition is met (fix all lint errors, get tests passing, tune prompts). /sentinel-batch breaks a massive task into work items and processes each with isolated sub-agents — generate codemaps for a 500K-line repo, migrate hundreds of files, bulk-add documentation. /sentinel-autoresearch runs a score-driven optimization loop — give it a task and a shell command that returns one number, and it iterates: propose an edit, score it, commit to a run branch if it improved or git reset --hard if it didn't, logging every attempt to an append-only TSV ledger. Inspired by Karpathy's autoresearch. All three track progress to disk and are resumable.

Context Optimization — Sentinel minimizes its own context footprint. Workflow references use progressive disclosure (loaded on demand, not eagerly). The session-start loader operates within a configurable token budget, loading vault entries in priority order and filtering gotchas by relevance to recently changed code. /sentinel-context audits all context sources (CLAUDE.md, rules, MCP servers, plugins, hooks, vault) with token estimates and actionable recommendations.

Documentation Drift Detection — Architecture docs and CLAUDE.md go stale as code changes. Sentinel detects this automatically. At session end, it scans architecture docs for dead file references. At session start, it verifies CLAUDE.md numerical claims against actual counts. Stale docs get flagged so Claude updates them.

Memory — A vault system that persists investigations, gotchas, decisions, and patterns across sessions. When a fix attempt fails, it's logged. When a non-obvious constraint is discovered, it's recorded. The next session reads these before starting work.

Discipline — Hooks that enforce quality gates, TDD workflow, and code review standards automatically. A stop hook verifies all gates pass before work is declared complete. Pre-tool hooks surface relevant gotchas before you repeat a known mistake.

Growth — Pattern extraction that identifies recurring solutions and promotes them to reusable knowledge. Stale vault entries are flagged and cleaned. The system gets smarter over time.

Install

Step 1: Add the marketplace (one-time setup)