applyContainerDefaults no longer fills containers.image with a hard-coded
private-registry reference. An unset image now fails configuration
validation with 'containers.image is required for autonomous workflow
workers', so a fresh instance is told at startup (and via the readiness
probe at runtime) instead of at the first failed container pull.
Configs that relied on the default now set it explicitly: dev.toml (the
local task image:agent build, which e2e also overlays), plus commented
[containers] blocks in config.example.toml and standalone.toml. The
deploy docs no longer describe a code-level default.
Task containers log in with their run credential on an external broker and
reach only their own subjects, decided by the embedded broker's authenticator.
The broker token is never put in a container environment.
The task session carries containers.egress_hosts and the profile's remote MCP
hosts, and the container gets ProxyEnv with the relay exempt from the proxy so
the State Store and NATS dials bypass it. A host the session lists exactly may
resolve to a private address; wildcard and any-host matches may not.
egress_hosts moves from openshell to containers, as it applies to both runtimes.
Runtimes now take a task launch: image, mounts, env, the daemon's NATS
and State Store with the task's credentials, its model providers and the
hosts its tools may reach. Docker addresses them through the egress relay
as before. An OpenShell sandbox reaches NATS and the State Store over
plain TCP, the model providers with their credential bound, and the
task's remote MCP hosts and the allowed package registries; services
addressed by IP or loopback are named by the configured services host,
since sandbox policy admits only DNS names.
Container runtime policies gain a runtime (docker or openshell) and the
OpenShell gateway's address, workspace and token, set on the container
runtime page and applied on restart. The OpenShell backend creates a
sandbox per task with the worktree and volumes bound in, reports when
the agent exits or outlives its maximum uptime, deletes the sandbox on
release, and sweeps this instance's leftovers at boot. Storage volumes
are created attachable to OpenShell sandboxes.
Deletes the hand-written internal/tools/mcp client and transports. The
provider talks to an SDK ClientSession over CommandTransport,
StreamableClientTransport or SSEClientTransport, and reconnects when the
server exits. The http transport now keeps a real streamable session
instead of stateless POSTs. Sampling goes through one adapter, since the
SDK marks it deprecated (SEP-2577). message_endpoint is removed: the
SSE transport takes the POST URL from the server's endpoint event.
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.
Every service connected to PostgreSQL as the database owner, which is a
superuser, so anything that compromised an app process owned the cluster, and
every connection reached the server with an empty application_name, so
pg_stat_activity could not say which service a session belonged to.
Split the DB credential into two login roles and name every pool:
- archie_migrator owns public and its objects and runs goose. Migrations run
on a short-lived pool opened from the new migration_database_url, before the
serving pool opens. It is not a superuser.
- archie_runtime serves requests with SELECT, INSERT, UPDATE, DELETE, sequence
USAGE/SELECT and function EXECUTE only. It has no CREATE privilege, so it
cannot change the schema.
- Every pool sets application_name: archie-state-store, archie-gateway, and a
-migrate / -recovery / -schema / -provision suffix on the transient
administrative connections. Existing one-URL installs keep working: an
absent migration_database_url falls back to database_url and the migration
runs on the serving pool exactly as before.
Provisioning is a new idempotent `archie-state-store provision` subcommand: it
needs an admin URL the services never hold (database owner or superuser),
creates or reconciles both roles, hands every public-schema object to the
migration role, revokes PUBLIC's CREATE on public, sets row-level grants, and
sets the migration role's default privileges so tables future migrations
create are usable without a hand-run grant. It stores SCRAM verifiers, refuses
superuser roles, and verifies the boundary before committing.
install.sh now provisions the bundled path and separates the credentials: the
superuser password the Compose database is created with lives only in a
compose-only postgres.env, while the env file the units load holds the
non-superuser PGPASSWORD. A bundled install that predates the split is
upgraded in place: the roles are created, the superuser password is rotated
out of the env file, and config.toml is moved to the runtime and migration
URLs. External databases keep their operator-managed credential; the
deployments README documents the one-shot provision command.
setup writes both bundled URLs, dev:postgres provisions the dev database, the
e2e stack provisions its throwaway server, and the deployment templates point
at the two roles.
Drop the bot_user-derived root identity and the empty-identity sentinel. System
polls forge connections and owns instance-level work; a polled task stores the
polling identity, task_queued and work_request_submitted carry task id, actor
and principal, and an empty task identity is refused. Migration 0072 moves
empty tasks.identity, the identity-grants root key and the old root's
connection permits to System so a re-poll dedups.
Add [access].denial_retention (default 90 days, matching the audit-retention
seed) wired through config defaults and validation; the State Store sweeps
denial records older than the configured age hourly via the new
access.DenialPruner / PruneDenials store query. The dashboard gains
GET /api/access/denials (read_audit-gated, so org admins/owners see their
org's records; the store scopes by the credential's org) and an
/org/denials page backed by a Pinia store.
A container task parked when a provider carried an api_key_env but no
credential binding named after it, even though the task's models never
reached that provider. Egress pre-bound every provider the instance
configures and ignored api_key_ref, so the credentials an operator stores in
provider-settings were never used.
A task's config now carries only the providers its model aliases resolve to,
each with the variable its credential is read from: the operator's
api_key_env, or a name derived from the provider when its credential is a
secret reference. modelEgress resolves that reference once for the session
and the proxy substitutes the session's key for the sentinel, so the
container still holds no key. A provider that only declares an env var keeps
resolving through its org credential binding.
A connection is an organisation's own resource: the forge host, the
credential reference and the external account it belongs to. It records a
reference, never a token value, and names per execution identity the
operations that identity may perform through it.
The file's [forge] sections seed it: the root's becomes the root connection
and each identity's own forge becomes a connection only that identity may
use, so the migration keeps exactly the grants the file expressed.
Replace the service bearer tokens with [tls] ca_file and cert_dir. Caller
identity comes from the verified certificate, not metadata. Without TLS,
services listen and dial on loopback only. Task containers get only the CA
and keep their grant.
* feat(state): durable operator alerts owned by the State Store
Alerts dedup on org+kind+subject, resolve themselves when a condition
clears, and are raised, resolved, dismissed and listed through the state
contract. Task transitions raise needs-human and failed alerts in the same
transaction; the store checks service health and prunes closed alerts.
Each raise and resolve queues a webhook delivery whose attempts are kept.
* feat(notify): deliver operator alerts to the webhook; raise config and release alerts
notify.webhook now receives raised and resolved alerts above a severity
threshold with HMAC-signed payloads, bounded retry and every attempt recorded
on the alert. archied raises model alias, provider and credential-binding
alerts from the effective configuration, an update alert from the release
channel, and a credential-rejected alert from the egress resolver. The
dashboard API lists and dismisses the caller's org's alerts.
* feat(ui): alert bell in the top bar and webhook alert settings
The bell lists the org's open alerts by severity with dismiss and a link per
alert, and refreshes from the event stream. The task settings page gains the
webhook severity threshold and signing secret. ui/dist rebuilt.
archied update install verifies the release zip, checks every binary
reports the release, pulls the agent image by digest, backs up the
database in-process and stages the binaries before handing off to
archied update watchdog in its own process. The watchdog restarts units
in release order, requires archied healthy for 120s and otherwise rolls
back binaries and image policy, restoring the database only when a
migration ran. Boot recovery resumes the frozen watchdog.
Deletes scripts/archie-update-install, scripts/archie-update-watchdog,
the updater's --from-source and --auto, CommandInstaller,
ScriptInstaller and the update-image and update-topology subcommands.
The system org sets each org's allowed models and whether it may add its
own providers (org-model-policy); an org with no entry may use nothing.
An org's aliases and providers are checked against that policy on save
and filtered by it at resolution. Tasks get their org's providers, keyed
by the org's credential binding through the egress proxy. Org Settings
gains a Models tab; instance Models gains Org access.
Agent steps, implement, pr-review, MCP sampling, chat, embedding and
transcription resolve a model alias (empty means the purpose's default)
and fail on an unknown alias instead of falling back. The step editor and
chat pick aliases; workflow definitions naming an unknown alias are
refused on save.
The dashboard's setup terminal was a 503: it had a WebSocket route, an
authorizer and an xterm.js surface, but nothing owned a container. The
daemon owns containers, so the session contract lands there and the
dashboard consumes it over a new gRPC contract.
One session concept, two targets. A setup session is addressed by Kit
profile and is ephemeral; a run session (archie-core-qi91) is addressed by
task and attempt and is bounded by the task. This slice serves setup and
answers Unimplemented for run.
The contract, the authorisation, the lifetime and the egress phase are
settled in docs/prds/external-agent-harness.md, "Session contract". The
short version: the daemon serves harness.v1.HarnessService; the dashboard
authorizes the operator as an update on the secret and then dials the
daemon with the harness service token, required even on loopback because a
session is a shell holding an org's credential; the setup container runs
no worker, mounts no worktree, holds a keepalive, and carries an empty
container environment. Install hooks run at the install phase and the
session enters runtime before the shell, so a runtime-phase credential's
token endpoint is the one the proxy intercepts.
Capture is gated by the egress proxy on a granted service, and a setup
session has no task, so the daemon registers a session grant: the
session's proxy token resolves to the org's bindings for the Kit's
declared services. The real token never enters the container: the token
endpoint response is sentinelized and stored as the org's harness secret.
Every teardown path releases the pool slot, revokes the egress session and
the grant, removes the sandbox network, and removes the session's volumes.
The listener requires HARNESS_TOKEN and stops with a 10s grace before
Stop, so an open session cannot hold shutdown to the watchdog leash.
disableProvider already cleared a failed provider's credential, but that
made it indistinguishable from a provider nobody configured, so
usableProvider fell back to the catalog's ambient env var (e.g.
OPENAI_API_KEY) and sent turns with a key the operator did not choose.
Mark the provider Disabled when its key fails to resolve, refuse it in
usableProvider, and carry the flag through mergeProviders so a catalog
refresh cannot resurrect it.