jobsiroh

README

jobs-iroh

A small, self-contained build system: one server, N runners, and a client, connected only by iroh QUIC. Builds are hermetic (rootless Linux namespace sandbox, no network inside), identified by content (canonical-CBOR definitions and a content-addressed amber store), and described by Starlark recipes. The server embeds NATS JetStream for scheduling — there is no other infrastructure: no HTTP, no database, no container runtime.

Full design: docs/architecture/architecture.md.

Binaries & ALPNs

Three core binaries: jobs-server, jobs-runner, jobs-client — plus an optional jobs-registry (an OCI registry serving build outputs as pullable images). One server endpoint, five ALPNs:

ALPN	Who	What
jobs-build/1.0	client	submit builds, watch, logs, cancel
jobs-runner-nats/3.0	runner	NATS tunnel to the embedded scheduler
jobs-runner-amber/1.0	runner	store sync (objects + refs)
jobs-admin/1.0	client (TUI)	observe builds, stats, fleet, refs, cancel/delete, diagnose
jobs-amber-admin/1.0	client / registry	push source trees up, pull outputs home

A build file

A build is a BUILD.jobs Starlark file next to your source. Minimal shape:

def build():
    toolchain = imp(fetcher = "tarball+https",
                    params = {"url": "https://go.dev/dl/go1.24.linux-amd64.tar.gz"})
    return struct(
        inputs       = {"toolchain": toolchain},
        env          = {"GOFLAGS": "-mod=mod"},
        script       = "go build -o $out/app .",
        runtime_deps = [],
    )

Imports (fetches) are the only steps with network access; the script runs in a sandbox with its inputs mounted read-only, the source at $SRC, and the output tree collected from $out. Language plugins (Go, npm, PyPI, cargo, RubyGems, …) expand lockfiles into per-dependency cached imports.

Quickstart: local build

No server needed — builds run hermetically against an embedded store under --data-dir (default ~/.local/share/jobs-iroh):

jobs-client build --source ./myapp                  # build → prints F + output key
jobs-client run   --source ./myapp -- arg1          # build, then exec JOBS.entrypoint
jobs-client develop --source ./myapp                # interactive shell in the build sandbox
jobs-client image  --source ./myapp -o app.tar --tag myapp:dev
docker load -i app.tar

Quickstart: server, runner, remote build

# 1. Server (prints its endpoint ID on startup):
jobs-server --data-dir /var/lib/jobs-iroh

# 2. Runner(s), anywhere that can reach the server over iroh:
jobs-runner --server <endpoint-id> [--addr host:port] [--size c1-m2]

# 3. Client: push source, build remotely, pull the output home:
jobs-client remote-build --server <endpoint-id> --source ./myapp
jobs-client watch  --server <endpoint-id> --request-id <id>   # re-attach
jobs-client status --server <endpoint-id>                     # one-shot tables
jobs-client tui    --server <endpoint-id>                     # interactive admin

remote-build streams the running steps' build output alongside the progress block by default (--no-logs opts out). When something fails, jobs-client diagnose --server <id> --request <id> prints the durable failure trail — every failed attempt with its captured output, surviving retries and server restarts (--json for the machine-friendly shape).

Local and remote builds share the same canonical definitions, so the same source yields the same build identity K/F everywhere — remote outputs pulled home join the local cache and vice versa.

Registry: docker pull a build

jobs-registry serves build outputs straight to docker/containerd/k8s — no export/load/push hop:

jobs-registry --server <endpoint-id> [--addr host:port] --listen :5000
docker pull my-registry-host:5000/jobs:<build-K>

Images live in a single jobs repository whose tags are build keys K (the value remote-build prints); /v2/jobs/tags/list enumerates every submitted build (a local build's bare F key also pulls, but is not listed). Images have two layers — the runtime closure under /jobs/store/<key> plus the platform shell (script entrypoints need their #!/bin/sh shebang; --no-shell opts out), and the build artifact at the root — assembled on the fly from the registry's own store, which syncs from the jobs-server on demand. Layers are uncompressed (…image.layer.v1.tar) and generated straight from that store on every request rather than stored as blobs, so serving costs CPU per pull instead of disk, and --cache-ttl (default 24h) now expires only the small manifest/config blobs. Size the volume for the store, which holds every synced build and is never reclaimed. The registry is pull-only and speaks plain HTTP: terminate TLS at an ingress or mark it insecure in the container runtime. A sample k8s Deployment lives in deploy/jobs-registry/.

Dev setup

Go toolchain comes from the Nix devShell:

direnv allow          # sets up the flake shell + GOPRIVATE
# or without direnv:
nix develop -c go test ./...

GOPRIVATE=github.com/fables-for-robots/* is required for module fetches (exported by .envrc).

Tests

nix develop -c go build ./... && nix develop -c go vet ./...
nix develop -c go test ./...

License

AGPL-3.0-or-later.

Documentation

Overview

Package jobsiroh holds repo-root embeds. go:embed can only reference files in the embedding package's own directory, so the checked-in fetchers.toml is embedded here and consumed via bootstrap.LoadEmbedded.

Index

• Variables

Variables

var FetchersTOML []byte

FetchersTOML is the checked-in fetcher manifest (fetchers.toml), embedded so every binary carries the fetcher pins it was built with and a standalone binary can provision fetchers from any working directory.