Introduction

The system of record for AI-assisted engineering — Git-native provenance, AI-aware analytics, and the forecasting your CFO will accept.

Section

Introduction Rift CLI v1.0

# What is Rift?

AI coding assistants — Claude Code, Codex, Cursor, Copilot, Gemini CLI, and dozens of internal agents — are now a structural part of how engineering teams ship software. Every leadership team is investing in these tools, but almost nobody can answer the basic questions a CTO should be able to answer about them: which PRs were meaningfully AI-assisted, what they cost, which models perform best for which workflows, whether AI-generated code is making delivery faster or just noisier, and what the bill will look like next quarter.

Git tells teams what changed. Rift tells teams why it changed, who or what changed it, what it cost, whether it worked, and what it implies for the future of the engineering organization.

Git tracks code. Rift tracks the AI work behind it.

# The three pillars

Rift is two products and one engine, organised around three primitives that build on each other.

Provenance

Every prompt, transcript, tool call, and agent edit linked to a commit, a PR, a ticket, and a person — stored Git-natively on a dedicated rift/checkpoints/v1 orphan branch, pushed alongside code or split to a dedicated audit repo, and synced encrypted to Rift Cloud in parallel.

Intelligence

AI-aware metrics: cost per PR, accepted-agent-output rate, AI rework rate, review burden delta, model ROI — DORA-discipline applied to AI-assisted work.

Forecasting

Monte Carlo cost forecasts, model-switch scenarios, new-tool rollouts, headcount changes, with P10/P50/P90 ranges your CFO will accept.

# How Rift works

The whole product is three commands and an open browser tab.

1. Install in 60 seconds

The CLI installs into Git hooks and your AI assistant's hooks. No daemon. No proxy. No service in the data path.

bash

brew install rift
rift init
rift enable --agent claude-code

2. Use AI normally

Every session, prompt, tool call, edit, and token count is captured into local-only shadow branches during the session, redacted, and condensed into a signed checkpoint commit on rift/checkpoints/v1 at git commit time — tamper-evident via the commit graph and per-checkpoint Merkle leaves.

3. Open the dashboard

Cost per PR, model mix, AI rework rate, team adoption, forecasts — all populated automatically. Provider billing reconciled in the background. PR cards posted to GitHub.

No daemon, no proxy

Rift never sits in the data path between you and your AI assistant. The CLI is a short-lived Go binary that runs on hook events and exits. If it crashes or is uninstalled, your workflow keeps working — you just lose the metadata.

# Key concepts

A handful of terms appear everywhere in these docs. Skim them here, link back when you need to.

Checkpoint branch: rift/checkpoints/v1. An orphan Git branch with a sharded layout where every AI session lands as a signed, redacted commit. The canonical durable store.
Shadow branch: Local-only Git branch (rift/<commit>-<worktreeHash>) that captures raw working-tree state while a live AI session is active. Never pushed.
Attribution: The human-vs-agent breakdown of committed lines, computed at multiple granularities — session, commit, PR, file, hunk, line — with an explicit confidence score at every layer.
Identity graph: Cross-system mapping of a developer across SSO, Git, GitHub, Jira, and provider workspaces. Every linkage carries a confidence score so dashboards never silently inflate numbers.
Reconciliation delta: The difference between Rift's local token-table cost estimate and the authoritative cost pulled from provider admin APIs. Surfaced as a reconciliation health score.

# Next

Continue to the Rift CLI v1.0 reference for installation, commands, and supported agents.

Rift CLI v1.0