Artifactize

The AI-native collaboration layer for one-of-a-kind teammates and their agents.

Artifactize splits a project into Artifacts (code, docs, designs, images), each with its own evals: tests, LLM reviews, human sign-offs and derived dependency checks. Every teammate, human or agent, works their own way on their part, and integrating that work reuses each review while its eval definition, covered Artifact kinds and fingerprints match. Earlier GREEN or RED results are reused. Artifactize is open source under the Apache License 2.0 and runs on Linux, macOS and Windows.

Why I built it

Building games with AI, working alone got fast, but working with others still felt like hitting a wall. There had to be a sweet spot between traditional development and AI-native work. So I looked for a way for every teammate, human or agent, to keep their own style and still add up to one coherent result, and I structured the project around evaluation.

Artifactize is that collaboration layer: it splits a project into modules that each carry their own evaluation criteria, so everyone can work their way, and it cuts the evaluation cost wasted during integration.

See it in your terminal

Change one file: status predicts one execution and names the file, and verify re-reviews only that Artifact.

How it helps a team

• Each part carries its own bar. An Artifact declares the evals that judge it: runtime evals run commands; Agent evals ask one exact model (OpenAI or Anthropic API key, or a ChatGPT/Codex sign-in) with the read-only tools you declare; Human evals wait for a sign-off; dependency evals derive readiness from other Artifacts' required evals without executing anything. RED blocks what depends on it. Whoever works on a part, in whatever style, meets the same bar.
• Integration reuses every review that still holds. You define each Artifact's fingerprint; by default, artifactsum hashes the Artifact's own files. fingerprint = false disables reuse. While the eval and the kinds and fingerprints of the Artifacts it depends on are unchanged, verify reuses the earlier verdict, from any profile, and reports what it executed, what it reused and the tokens reuse saved. Review cost follows the size of a change.
• Centralize the reviews, not the repo. One artifactize server shares verdicts across teammates' checkouts and CI, so a branch's reviews carry over to the merge and a person signs off a change once.
• Know what a change will cost before you run it. status predicts what verify will execute, reuse or wait for, and names the files and dependencies that changed.
• Human review in the terminal. artifactize monitor drills down from Scope to Runs, the Run tree and Detail: explicitly claim a sign-off, run tools and submit without leaving the terminal.

Install

Every install provides artifactize, including its current-directory tools as artifactize tools. Non-Store channels also put artifactize-tools, the same built-in tools as a standalone CLI, on your PATH. No Rust toolchain is needed.

Linux (x86_64 or aarch64): the script installs the static binaries from the latest GitHub release into ~/.local/bin. It checks their SHA-256 and needs no sudo:

curl -fsSL https://artifactize.dev/install.sh | sh
artifactize --version

With a Rust toolchain, use cargo binstall artifactize artifactize-tools for the prebuilt binaries, or build from crates.io with cargo install artifactize artifactize-tools --locked.

macOS (Apple silicon or Intel): use Homebrew, or the same install script.

brew install lhj6102/tap/artifactize

Windows (x64): use winget, the Microsoft Store or irm https://artifactize.dev/install.ps1 | iex in PowerShell. With Smart App Control on, use the Store version. In WSL 2, follow the Linux steps.

winget install lhj6102.Artifactize

To update, use the same channel again (brew upgrade, winget upgrade, or run the script again); the Store updates by itself. Your state is kept. Install covers every channel, updating and uninstalling, the prerequisites and the Agent backends.

A tiny example

A folder with an index.artf is an Artifact. A sidecar such as page.md.artf declares the neighboring file. Both use TOML and require a name. This folder Artifact, from examples/runtime-relations, checks that a page has a heading. Omitting fingerprint uses artifactsum, so the result is reusable while the inputs are unchanged:

name = "usage"

[evals.heading]
title = "The page has a top-level heading"
profile = { kind = "runtime", command = "grep", args = ["-m", "1", "^# ", "page.md"], timeout_ms = 5000 }
payload.instruction = "Check that {usage} starts with a top-level heading."

No model or API key is needed to try it:

git clone https://github.com/lhj6102/artifactize   # the example projects
cd artifactize/examples/runtime-relations
export ARTIFACTIZE_STATE_HOME=$(mktemp -d)   # keep the tour's state apart
artifactize verify --all    # three GREEN results, exit 0
artifactize verify --all    # exit 0 and nothing executes: all three results are reused

The Quick start continues with status, file Artifacts, dependency evals and a Human sign-off.

A dependency eval can collect readiness without another review:

# player/index.artf
name = "player"
tags = ["scope:player"]

[evals.ready]
title = "Movement and art are ready"
profile = { kind = "dependency", depends_on = ["player-movement", "hero-art"] }

artifactize verify player/ready includes the required dependency evals without --recursive. The result is derived from current evidence, not cached or published to the team store. Tags appear in status, graph and monitor output; they do not change reuse keys. See Declarations (.artf).

Upgrading to 0.9

Declarations move to TOML index.artf and <file>.artf, with snake_case keys; the ignore file becomes .artfignore. There is no converter: have an agent rewrite the declarations and commit them. Families are removed; projects that want templates generate and commit their own .artf files.

State schema 6 requires a new state directory; there is no migration. Reuse-key v2 matches no earlier result, including in a team review store or with a script fingerprint. The first full verify reviews every executable eval once, Human sign-offs included; dependency evals derive their current verdicts.

The terminal UIs share Human Detail and require an explicit c claim. In standalone review, s no longer opens a verdict popup and Enter no longer submits: use g/r, then Ctrl-S. Esc stops editing before leaving Detail and never quits. See Upgrading to 0.9 for the removed keys and state/declaration changes.

Reuse

verify reviews once; the next verify reuses every result whose eval and Artifact kinds and fingerprints are unchanged and says where each came from.

Fingerprints and reuse · Runs, status and validation

Human review in the terminal

verify waits by default while a Human eval needs a sign-off. The monitor drills down through Scope → Runs → Run tree → Detail. Its immediate context and preview panes remain at every width, except at the ends of the drill-down. Side panes take 16, 28 or 36 columns at widths below 100, 100–139 or 140 and above; the focused pane fills the rest, with sides clipped on extremely small terminals. j/k selects an eval and reveals its outcome in the peek; Enter or o opens Detail. Press c to claim a waiting Human request, run the tools its owner declared, choose GREEN or RED with g/r, and submit the schema-backed form with Ctrl-S. Standalone artifactize review uses the same Human Detail and keys. i expands the instruction and t opens Technical while not editing. Esc first stops editing, keeping the draft, then leaves Detail. A submission resumes the waiting Run's dependents.

? shows help and ! finds the next error, RED or waiting Human eval. Esc steps back and never quits; q quits from idle views. Ctrl-C quits or cancels a running Human tool first. The tree names what an eval waits for; b jumps upstream and Backspace returns. Validation stays at Run end in the headline; a dim * marks rows changed by later request updates.

The monitor follows local state changes without a permanent daemon. Reads run in background jobs, so navigation and cancellation remain available during a refresh; unchanged selected evidence keeps its tree and Detail. The Run list holds one bounded page plus an older selected Run when newer entries push it off the page. j/k at a page edge moves to the next/previous page without retaining the earlier pages. Agent Detail shows live local sessions with scrolling and expandable tool activity groups; reused results keep their original evidence when it is available.

Human reviews

Team review store

Each machine keeps its own checkout and state. One artifactize server keeps the verdicts by reuse key and returns the latest, so a review done on one laptop is reused on the next and in CI. A read-only CI token reuses but never publishes.

Team review store · CI · Walkthrough

Documentation

Everything else is at artifactize.dev/docs:

• Concepts: Artifacts, evals, fingerprints, Runs and status.
• Guides: Agent evals and backends, Human reviews, the team review store and CI.
• Reference: every command and flag, index.artf, tools, state, limits and the review store protocol.
• Examples: runtime relations, Agent and Human tools, file Artifacts and dependency evals.

The recordings above are reproducible: their VHS tapes and demo projects are in website/demo (how to record).

License

Licensed under the Apache License, Version 2.0.