Archie runs its own agent loop only. Removes Kit resolution and launch, the harness runner and adapters, the setup terminal and sandbox view, harness secrets and OAuth capture (table dropped in migration 0075), the task-tools contract that only Kit containers reached, and the sandbox-kit-spec dependency. Agent profiles carry image and tools only. The egress proxy keeps the model-session path; the sandbox relay code stays for #1867.
6.9 KiB
Running this distribution
This archive holds the linux/amd64 binary for each service. Each service owns its data and reaches the others only through their gRPC contracts (or NATS, for task actions):
| Service | Owns | Reaches |
|---|---|---|
archie-state-store |
tasks, events, control-plane settings, identities and access policies, installed packages (PostgreSQL) | nothing |
archie-gateway |
conversations (PostgreSQL, its own tables) and every model call: chat turns, curators, MCP tools and sampling | State Store; daemon over NATS for task actions |
archied |
dispatch: forge polling, the task queue, the embedded NATS broker, agent containers | State Store; Gateway |
archie-messaging |
the Telegram, email and webhook channel connections | Gateway; State Store for channel settings |
archie-ui |
the dashboard and the webhook capture receiver | Gateway; State Store |
Without archie-messaging the channels are simply absent, so it needs a unit
of its own (deployments/systemd-user-service.md).
archie-agent (the per-task sandboxed runtime) has no unit of its own. The
agent image is a separate thing, pulled for the containers archied spawns
per task: [containers].image defaults to
git.catlow.cloud/sam/archie-agent:latest, while this archive's release.json
pins the digest the release published. Set [containers].registry_auth when
that reference is on a registry that needs a login -- see
deployments/systemd-user-service.md#6-pinning-the-archie-agent-image.
Quickstart
-
Pick a starting configuration. The full set of annotated templates lives in this repository's
deployments/directory on GitHub (https://offloaded.dev) —standalone.toml(no forge, the simplest start),single-forge-github.tomlanddocker-nats-stack.tomlcover the common cases. Every one needs PostgreSQL 18 with thearchie_migratorandarchie_runtimeroles created once byarchie-state-store provision(seedeployments/README.md);docker compose up -d postgresplus that command gives the local default. Copy whichever fits to:mkdir -p ~/.config/archie cp <template>.toml ~/.config/archie/config.toml -
Edit
config.toml(forge token, repos,[services.state]target, model providers) and place any secrets in~/.config/archie/envor referencedsecret://locations. -
Start the five processes. Order does not matter: a service whose peer is not up yet waits for it and reports it as degraded on its health surface meanwhile. The Gateway waits for the daemon's NATS endpoint when the broker is embedded.
./archie-state-store -config ~/.config/archie/config.toml -listen 127.0.0.1:9090 & ./archie-gateway -config ~/.config/archie/config.toml -listen 127.0.0.1:8585 & ./archie-messaging -config ~/.config/archie/config.toml & ./archie-ui -config ~/.config/archie/config.toml & ./archied -config ~/.config/archie/config.tomlOpen
[web].listen(default127.0.0.1:8484) to configure the organisation, workspace and first administrator account. Remote access requires an HTTPS reverse proxy. Complete setup before exposing the dashboard publicly. -
For a persistent, 24/7 deployment instead of a foreground run, follow the systemd user-unit runbook: https://offloaded.dev.
Version
This archive's binaries are built from the archied release tag in its
filename; archie-ui, archie-gateway, archie-state-store and
archie-messaging ship at that same version rather than
being independently tagged (see RELEASING.md in the source repository for why
only archied and archie-agent are independently versioned).
archie-state-store, archie-gateway, archie-messaging, archie-ui and
archied are services that must be running and addressable; archie-messaging
carries the Telegram, email and webhook channels since the v1.30.0 extraction, so
an install (or an upgrade across v1.30.0) needs its unit as well -- see
deployments/systemd-user-service.md.
Native self-update
Install on Settings → Status runs archied update install -version X against
the release source in the update settings. It fetches release vX, refuses it
unless the source's key signed its statement, and downloads the Linux/amd64 zip
only if it hashes to the digest that statement names. Every host binary in the
zip must report X. It then pulls the agent image the statement pins by digest, backs up the database to
db.pre-update.dump in the state directory, stages the new binaries beside the
live ones and hands the transaction to archied update watchdog in a process of
its own (a transient systemd user unit). Nothing that runs changes until the
watchdog takes over.
archied update check finds the newest GitHub release with a linux-amd64 zip on
the channel chosen under Settings → Status → Updates: stable, next (includes
prereleases) or a pinned version. From a shell it reads
ARCHIE_UPDATE_RELEASE_CHANNEL and ARCHIE_UPDATE_PIN, or -channel and -pin.
The notification destination remains separate from the release channel.
release.json supplies the service order and required config sections to both
first-install and self-update. A missing unit or config section produces an
ordered migration plan and refuses installation. Apply that plan before retrying.
Automatic topology edits are intentionally unsupported.
The watchdog stops the units in reverse release.json order, switches the
staged executables by rename, switches the control-plane image policy (only the
image field, with conflict detection), and restarts the units in order through
systemctl --user. archied must report the new version and stay healthy for
120 seconds. Otherwise the watchdog restores the previous binaries and image
policy. Data is not rolled back unless the new release migrated the database:
then the previous binaries cannot serve it, so the pre-update backup is restored
and the outcome report says that every write since the update began is gone.
Without a systemd user manager the watchdog switches binaries and image policy
but restarts nothing, and the install reports that the services need a manual
restart.
Transactions live in ${XDG_STATE_HOME:-~/.local/state}/archie/update, outside
the binary directory: transaction/journal.json, transaction/previous/, and
transaction/archied, the frozen copy of the archied that wrote the journal. On
daemon boot, an interrupted transaction is
handed to that frozen watchdog with -recover before normal startup: a
transaction that never started is discarded, one that was switching or observing
is rolled back, and a decided one is finished. To recover manually, run
transaction/archied update watchdog -recover -bin-dir ~/.local/bin from that
state directory. A failed recovery keeps the journal for another attempt.
The database backup stays in the state directory after a committed update for
manual recovery with archie-state-store restore.