Skip to content
feedback-loop-aiPublic

About

Deterministic delivery engine for autonomous multi-agent software delivery: an event-sourced phase state machine at the outermost layer, agent sessions as leaf effects.

Resources

Contributing

Stars

5 stars

Watchers

0 watching

Forks

Latest commit

 

History

691 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Brokkr — sealed anvil mark, terminal rail node pulsing

Brokkr

ci release license: MIT OR Apache-2.0 clippy · -D warnings coverage · literal 100% deps · permissive-only platforms rust

Coordination tools help agents work together. Brokkr proves what they did.

Brokkr is a deterministic delivery engine for autonomous multi-agent software delivery: you give it a feature and a reviewable recipe, its agents implement, verify, review and ship, and its event-sourced phase machine records every result and ruling in a journal you can replay and audit.

The v0.12.0 release makes every capability the realm's to grant, opens a terminal console that shows who must act, and adds gates that hold the code to its architecture.

Install

Pick one channel. Each row is the channel's published command; the availability labels and the checksum-first release path are explained in the quickstart.

Channel One command Availability
Release tarball (Linux x86_64) curl -LO https://github.com/feedback-loop-ai/brokkr/releases/latest/download/brokkr-linux-x86_64.tar.gz && curl -LO https://github.com/feedback-loop-ai/brokkr/releases/latest/download/SHA256SUMS && sha256sum --ignore-missing -c SHA256SUMS && tar xzf brokkr-linux-x86_64.tar.gz working today
cargo-binstall cargo binstall brokkr-cli live — the seven crates are on crates.io from v0.9.1, published by the release workflow at each tag
Nix nix profile install github:feedback-loop-ai/brokkr live from v0.9.0 — the release renders the flake digests and opens their pull request; nix reads the default branch
apt (repository configured) sudo apt-get install brokkr live from v0.9.0 — a signed repository on GitHub Pages; the one-time keyring and source lines are in packaging/README.md
dnf (repository configured) sudo dnf install brokkr live from v0.9.0 — same site, same signature; the repo file is in packaging/README.md
Homebrew brew install feedback-loop-ai/tap/brokkr live from v0.9.0 — the tap is bumped by the release and merged by the operator

On Windows, use WSL2: it is Linux, and every Linux row above serves it (decision 0063).

60-second bootstrap

Sixty seconds from a fresh machine to a lit run, then five minutes to a completed run. Both budgets are measured in CI on Ubuntu and macOS by scripts/bootstrap-bench.sh, which prints what it mocks. You need:

  • Every host: git, bash, a git repository you are willing to let an agent edit, and claude or codex on PATH. init scaffolds for the first of claude, codex or dsh it finds, or for the one --cli names. dsh alone is not enough: it cannot hold the review gate, because its adapter is untrusted and names no judges, so a dsh scaffold keeps the reviewer on claude.
  • Linux: bubblewrap (bwrap) 0.11 or newer on PATH, with unprivileged user namespaces allowed. The scaffold's verify and ship gates run boxed under the default namespace boundary. A codex scaffold's model seats run their commands in that box too.
  • macOS: nothing more. init declares the harness boundary, because namespace is Linux-only, so verify and ship run their pinned scripts unboxed.
flowchart LR
  subgraph a["≤ 60 s — measured"]
    direction LR
    install["1 · install<br/>the verified tarball above"] --> doctor["2 · brokkr doctor<br/>tools · drivers · models"]
  end
  subgraph b["≤ 5 min to the first effect — measured"]
    direction LR
    init["3 · brokkr init .<br/>a recipe for your stack"] --> run["4 · brokkr run<br/>implement → verify → review → ship"]
  end
  doctor --> init
  run --> inspect["5 · brokkr inspect<br/>the proof, from the journal"]
Loading
brokkr doctor                       # ok / warn / MISSING per tool and driver; executes no agent
cd your-repo && brokkr init .       # writes one .brokkr/ folder: bundle.json, policy.json, realms.json, agents/, adapters/, scripts/, .gitignore (and dialects/ for a spec repo) — open it
git add .brokkr/ && git commit -m "starter bundle"   # ship closes out on a clean tree; .brokkr/state/ stays ignored
brokkr run --feature "add one visible improvement" && brokkr inspect --run latest

From a clone of this repository skip init: brokkr run --recipe brokkr-fast --repo . --feature "…" resolves the library's own Rust recipe under ./recipes. fast and node are Brokkr's two portable recipes, which also install onto the shelf of any folder brokkr init made; the rest are Brokkr-only (decision 0087 ruling 5). Its exit code says whether the run completed, parked for you or stopped; every verb and flag is in the CLI reference.

The inspection is derived from the journal; it shows the reviewer's verdict, the exact rule that accepted it and the phase graph. The sample is abridged: a real trail lists every event, and each finished seat carries its duration.

run  add-one-visible-improvement-8bf6d692
     completed · phase done · seq 31
ruling  SHIP-COMPLETE  ship → done · shipped

seats
  participant status    attempts turns cost activity
  implement   succeeded 1        —     —    complete
  verify      succeeded 1        —     —    pass
  review      succeeded 1        —     —    clean
  ship        succeeded 1        —     —    shipped

trail
  22 transition/decided REVIEW-CLEAN-NO-FIXES review → ship · clean
  29 transition/decided SHIP-COMPLETE ship → done · shipped
  31 run/completed      completed

graph
  implement ×1 → verify ×1 → review ×1 → ship ×1 → done ×1  ←current

The full quickstart covers init per stack, parks, the read surfaces and the evidence commands.

Read next

  • Guides — the task map: first run, recipes, agents, adapters, secrets, journals and repository anatomy.
  • What works today — per harness: gates, hands, native powers, resume, telemetry, secrets and web search, with every known limitation linked to its issue.
  • Security model — what the box does and does not contain, how secrets flow, and how seat output is treated.
  • Decision record — the constitution: every semantic rule, its status and its enforcement binding.
  • Research registry — source-backed findings and comparisons, including GitHub peers for graphs, agents and charters.
  • Essays — the paradigm argued against the repository's own history and evidence.
  • Lore — why Brokkr works the bellows, and why story is commentary rather than specification.
  • Evidence shelf — redacted journal exports that let the project's claims be inspected.
  • Contributing — the mandatory sixty-second path from a fork to a pull request backed by a completed Brokkr run.
  • Architecture — the implemented crates, journal, effect discipline and verification layers.

Acknowledgments

The standing-overseer concept reached this product by way of the lieutenant in Robert C. Martin's SwarmForge, as decision 0019 ruling 7 records. The idea is credited here and nothing else is taken: SwarmForge carries no license, which means all rights reserved, so no code, scripts, prompts or prose from it has entered — or may enter — this tree.

Muninn is an independent design with an inverted authority model, described in decision 0020 (amended for the crossings a map draws by decision 0059) and built as brokkr muninn: it reads the journal and the world's crossings, proposes to the operator, and rules nothing.

License

Licensed under Apache-2.0 OR MIT, at your option; decision 0018 records why.

About

Deterministic delivery engine for autonomous multi-agent software delivery: an event-sourced phase state machine at the outermost layer, agent sessions as leaf effects.

Resources

Contributing

Stars

5 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages