Governance infrastructure for
AI-agent teams
System Kit stops your AI coding agents from destroying each other's work. Download the kit, paste one prompt into any agent thread, and get instant multi-thread coordination (four-mutex model with scoped parallel CODE), atomic machine-checked claims, claim/lock task queues, verification gates, push discipline, key isolation, and institutional memory. No technical background needed — setup is one pasted prompt plus plain-English questions, and your agent does all the file work. Works with or without git. Built from real production incidents across 26+ concurrent AI-agent sessions. Plain markdown methodology — works with opencode, Claude Code, Cursor, Codex, and any LLM agent.
100% Free & Open Source (MIT). Zero telemetry, zero network calls, zero data collection.
AI Agents Are Productive & Completely Ungoverned
Run two agents in parallel without a system and these failures are not hypothetical — every rule in System Kit exists because one of them actually happened:
Two agents edit the same files
Lost work and broken builds from unchecked parallel edits.
Invented APIs
Agents confidently call functions that don't exist anywhere in your codebase.
Untested syntax errors ship
Dead pages reach production because nothing forced a local verify.
Context compaction mid-task
The same problem gets researched twice because nothing was checkpointed.
Cumulative data shown as daily
Users catch reporting bugs before you do.
Dead models eat your requests
Quotas burn on routing errors no one probed for.
Keys read into chat context
Secrets leak to external servers through agent transcripts.
Four Steps to Governed Agent Sessions
Download the kit
Green Code button on GitHub → Download ZIP, then unzip it anywhere. No terminal, no install.
Paste SETUP_PROMPT.md into your agent
Open opencode, Claude Code, Cursor, or any agent in your project — paste the prompt as your first message.
Answer plain-English questions
The agent scans your project, asks about it in plain English, then builds everything: it auto-detects what your environment supports (shell? git? CI?), copies the kit's docs/ folder in, installs the atomic-claim and scope-check scripts, and fills every file. Already have docs/? Governance gets a dedicated subfolder.
Governance is live
Task board, thread registry, build log, checkpoints, laws. Every future thread starts at docs/START_HERE.md.
Nine Capabilities, Zero Runtime Dependencies
Four-Mutex Model
CODE (scoped — parallel holders with disjoint scopes, including glob scopes like src/**/*.test.ts), LEDGER, DB-CF, and MERGE lock separately — docs threads never wait behind code threads, and merge-backs are serialized.
Claim / Lock / Release
Single entry point, priority queue, and conflict detection so every thread knows exactly which task is safely its own.
Atomic, Machine-Checked Claims
A shared-lock register script makes claims race-free — two threads racing for one task, exactly one wins. Overlapping scopes are refused at claim time and rejected at commit time (git projects). Stale threads auto-flag at 2h+ without a heartbeat.
Verification Gates
Local-first testing discipline with a five-step order — owner verifies last, never first.
Institutional Memory
Append-only ledgers plus a checkpoint/resume system mean context loss never costs re-research.
Owner-Gated Deploys
Ledger currency is required before pushing; deployments stay owner-gated, always.
Machine-Checked Security
Credentials are handled internally and never exposed to agent contexts or logs — and a CI security scan now enforces it: credential patterns, prompt-injection markers, and untracked secret stores are flagged file:line, with matched values never echoed.
Universal Deployment
Same system in a plain folder with no git, a git repo, or hosted with CI — capabilities are auto-detected at setup with zero extra questions.
Plain-Language Decisions
Decisions arrive in simple English — the owner is interrupted only when genuinely needed.
11 Patterns, Each Backed by Real Incidents
Concurrency Protocol
The four-mutex model that makes parallel agent sessions safe.
Verification Standard
Local-first testing discipline; five-step order of operations.
Security Checklist
Universal + AI-specific requirements, including key isolation.
Failure Classes
Twelve documented bug classes, each backed by real incidents.
Cost-Zero Operation
Free-tier rotation and resource management without spend.
Model Rotation
Surviving weekly free-catalog churn with live probing rules.
Prompt Injection Defense
Four defense layers for untrusted data inside agent contexts.
Context Window Management
Checkpoint-before-compaction discipline with five triggers.
Worktree Parallel Coding
Git worktree isolation for truly parallel code threads (git projects).
Folder-Copy Parallel Coding
Per-thread folder isolation merged back under MERGE — true parallelism without git.
Non-VCS Mutex
Filesystem lock-files with PID atomicity replace git branches on projects with no version control.
Release History Lives on GitHub
Every release, documented as it ships
The live badge above always shows the current release — no stale version numbers here. Full history, from the initial release through every added pattern, fix, and breaking change, is maintained in the repository's changelog.
Read CHANGELOG.md ↗Machine-Checked Where It Counts. Not Magic.
Claims are atomic and race-free wherever a POSIX shell exists, and (in git projects) commits outside a thread's declared scope are rejected by hook. But machine enforcement covers claims and commits — it is not runtime surveillance. Where no shell exists, the kit degrades to a fully documented manual protocol: correct behavior stays explicit, checkable, and recoverable, just not automatic. Known constraints: filesystem locking assumes one shared filesystem, and templates are English-only. We think honesty about this beats pretending otherwise.
Stop losing work to ungoverned agents
Star the repo, download the kit, paste the setup prompt. Governance live in under five minutes.
⭐ View System Kit on GitHub ↗