A7-RT v0.1.0beta
Transactional Build System
for LLM Agents

AI coding assistants write code quickly. They break production systems just as fast.
A7-RT is a build harness you run locally. It treats LLM agents as unreliable compilers and verifies every artifact before it reaches your working tree.

2-stage syscall boundary (Core to Integration)
Adversarial verification. Builder never sees tests.
Hermetic ShadowFS execution
Stateless manager. Zero context drift between turns.
Append-only ledger. Resume after months with zero catch-up.

The Post-Exploration Gap

AI coding assistants excel at exploration. They generate the 80% solution in minutes. Then they accumulate broken state: hallucinated imports, tests that pass only in the model's imagination, architectural decisions that collapse under their own weight.

A7-RT occupies the gap between exploratory AI and production software. It hardens the output of Cursor, Copilot, and Claude, operating as a pre-commit gate that validates every artifact before it touches your working tree. Unlike CI, which validates after commit, A7-RT enforces contracts before code is written.

The Syscall Boundary

A7-RT enforces a two-stage ratchet. Stage 1 is pure logic. No open(), no socket(), no print(). Stage 2 handles I/O, CLI, config, and process lifecycle. The harness catches integration drift without re-grounding your core. Nodes in Stage 2 cannot dispatch until every Stage 1 dependency is grounded and sealed.

How It Works

Adversarial Verification
Test Author → Builder → Harness
Core

Three agents in an adversarial loop. The Test Author writes tests and defines behavioral contracts. The Builder implements against those contracts, blind to the test files. The Harness validates through hard-path execution before any commit. No self-declared status. The contract is the oracle.

near
Initial State
provisional
Schema Valid
grounded
Tests Pass
sealed
Stage Complete
Stateless Manager Blind Builder Hard-Path Validation
Executable Contracts
exports, assumptions, guarantees
Core

Every node declares its interface: function signatures, dependencies, behavioral guarantees, and raised exceptions. Mode A — exports are pre-declared; the builder implements to a locked surface. Mode B: exports emerge from implementation, and the test author becomes the architect. Both require concreteness. Vague descriptions do not ground.

Mode A
Contract-First
Mode B
Test-First
DEP:
Structural Dep
ANTI:
Anti-Guarantee
Mode A Mode B Dependency Graph
ShadowFS
Hermetic Execution
Core

Every agent works in an isolated filesystem sandbox. Failed iterations stay in shadow — visible in telemetry, invisible to git. Only passing code reaches your working tree. No partial implementations, no broken states, no infection cascades.

Isolated
Agent Workspaces
Atomic
Commit Model
Auto
Rollback on Fail
Zero
Drift to Git
Isolated Workspaces Atomic Commit Automatic Rollback
Stateless Manager
Zero Context Drift
Core

The manager LLM receives a computed board view, emits exactly one action, and terminates. No memory between turns. No rotting conversation history. Maximum observed context: 8k tokens, discarded immediately after use. Interrupt today, resume in six months, pay zero catch-up cost.

8K
Typical Manager Token Use
1
Action per Turn
0
Memory Between Turns
Resume Window
Stateless Append-Only Ledger BoardView → Action

$1 for 13 Nodes. Fully Tested.

A complete API mock factory built end-to-end under the harness. Type definitions, pattern matching, templating, session management, OpenAPI loading, HTTP server, admin API, CLI runner, and system bootstrap. Built using Kimi K2.5.

$0.38
Stage 1 — 8 nodes, 32 turns
$0.62
Stage 2 — 5 nodes, 16 turns
211
Tests generated
95%
Contract compliance

Three bugs found and attributed precisely. Each failure localized to minor contract deficiency.

Read the full run log →

Start with the Contract Cookbook

The Contract Cookbook contains patterns for writing contracts A7-RT can ground efficiently.

Read CONTRACTS.md → README.md