An action judgement holds at propose and wait is no longer a dead end. The
new approvals module owns the request: it binds to the decision and the
action's digest, so an answer covers the action the person was shown and
nothing else, and it settles once. The answer wakes the run, which judges the
action again and runs it only as the approval allows: a refusal still refuses,
an altered input runs nothing, and a judgement that came back worse runs
nothing on that approval.
The composer carries the request as a message with the request id and its
deadline, and its answer is a message too, so the trail shows what was asked
and what was answered. A request nobody answers inside its deadline is settled
as expired and nothing runs.
How long a request waits is an organisation setting, edited in the dashboard
and stored in milliseconds, because a Postgres interval scans as text. A run
pins the deadline it started with, like its model.
An agent that proposes the same action twice is asked about it once, and the
notification says when the investigation proposed nothing at all.
A new judgement module assesses one concrete action: hard rules refuse before
any guard runs; every guard sets a ceiling and the action runs at the lowest;
a guard that fails, panics or gives no ceiling holds it at propose and wait;
an advisory verdict is recorded and sets no ceiling. Reversibility and
Precedent (the fresh-install class default) are the first guards. Every
decision is recorded with its verdicts and class, and the run's trail lists
them. Agent tools now offer connector actions too: each call is assessed and
executed only as decided and only with the judged input, and a held action
comes back to the agent as a failed call naming its level and findings.
Tools gain Resolve, so an MCP server's declared effect reaches judgement.
GET /api/runs/{id} assembles a run's trail from rows already kept: the event
and trigger that started it, each workflow step, and each model and tool call
from the agent's journal. Investigations end by notifying their findings on
the composer, whose messages are read from a cursor with a live stream that
only says when to read. The binary serves its own SPA entry, archie.html,
calling only the new API. The investigate prompt now treats the event as a
claim to check against what it measures.
Credentials are typed (machine, age), write-only through the API and sealed
to the age identity in the bootstrap config. A field holds a value or names a
secret engine key; connectors and engines are accepted with credential slots
that resolve into their Configure secrets. The secret engine contract gains
secrets so an engine can receive its own key.
The agent model is an organisation setting in the new governance module.
Every model and tool call is a journal step, so a killed run resumes without
repeating either. Agents are offered connector lookups and read-only actions
only, until judgement exists.
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.
ListOpenPRs dropped org_id, so the reconciler resolved the forge and
worktrees in the default org instead of the task's own and skipped the
task silently. Select the org, log a skipped task, and let the Merge
action succeed when the forge reports the PR already merged.
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.
Each task container now gets an internal bridge with no host address,
shared only with the egress relay. NATS, the State Store and the model
proxy are reached through the relay's ports; a loopback State Store is
forwarded to the host gateway. The network is removed with the task, and
boot sweeps any a previous daemon left. The per-task State Store relay on
the host bridge is gone.