Only the daemon owns containers, and the dashboard consumes a harness
session over its gRPC contract. Go imports are atomic, so a single
reference relinks a runtime, and the boundary holds only while something
re-measures it. cmd/archie-ui/architecture_test.go enforced this until it
was purged with the test suite; nothing enforced it since.
The restored check runs go list -deps . and fails if the dashboard links
internal/container, internal/app/archied or internal/agentexec. It is
transitive, which is what "links" means, and runs in the existing test
step.
forward intercepted a declared OAuth credential's token endpoint only when
the run was granted the service. For an optional credential it fell
through to the reverse proxy, so the provider's real access and refresh
tokens reached the container un-sentinelized, contradicting the harness
contract's "the real credential never enters the container".
A matched token endpoint is now refused when the service is not granted
(403 for an optional credential; a required one keeps its existing 502).
The table test drives forward against a recording upstream and asserts
both the status and that the upstream was never contacted.
The dependency bump moved sandbox-kit-spec from v3.0.0-m.6 to m.8 without
adapting the code, so the tree no longer builds: CredentialCapability.Phase
is now spec.Phases, not a string (internal/infrastructure/egress), and
Client.Assemble became a package-level fetch.Assemble that takes a Loader
and a LayerValidator instead of spec.MergeOptions (internal/infrastructure/
kitrun).
Migrating to m.8 is not a version-string change: m.8 pushes kit integrity,
safe layer extraction and resource limits onto the caller, so it is
security-relevant work, not an unblock. Pin back to m.6, which the code is
written against, and leave the migration to a deliberate change.
Verified: go build ./..., go vet ./..., go test -short ./... all pass with
this pin. The gate's ui:check and lint steps fail for unrelated
toolchain reasons recorded separately.
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.
29 of 59 directives did nothing. Removing each one individually and
re-running the gate revealed no issue from any of them: 20 named gosec,
which is not enabled in .golangci.yml and never has been, plus 4 nilerr
in functions that return no error, 2 dupl where dupl reports nothing
repo-wide, and one each of contextcheck, nestif and unused. The funlen
name on settings.go:147 is gone for the same reason; its gocyclo and
cyclop names are both live.
The unused directive in webui/authorize.go sat on `var _ = slog.Default`,
a statement that only existed to keep the log/slog import alive. Both go.
The 30 remaining directives all suppress a real issue.
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.
A scheduled job runs unattended, so a payload that restarts the daemon,
formats a disk or powers the machine down is refused when the job is
created rather than when it fires. scheduling.RefuseDangerousJob is the
one rule every creation path reaches: the dashboard, any control-plane
client and the config seed all replace the schedules document through the
same validator, so a new token is added in one place.
Delete the hand-rolled MiniMax HTTP/poll client; ai-sdk v0.1.43 already
ships provider/minimax implementing video.Provider with the same endpoints,
defaults and poll semantics. The tool now takes a video.Provider and the
composition root builds it, matching the transcription and embedding
capabilities.
A drain pass no longer waits for the tasks it submitted, so -once (the
systemd timer and test invocation) returned while their containers were
still running and the process then tore them down. Wait for the pass's
tasks here, restoring what the flag means.
A wait:true workflow.call run held its global slot and its repo's place for
the callee's whole life, and the drain pass waited for its own tasks, so a
callee queued by that pass was never claimed. With containers.max_concurrency
saturated, or a repo that does not allow concurrency, the wait could not end.
A drain pass now returns once the stream and the claim fallback are empty, and
shutdown waits for in-flight work once. Suspend/Resume give a run's slot and
repo place back while it waits on a callee, driven by the workflow call events
on the daemon's bus; a hold records one slot and one place, so a second
re-acquire gives back what it took.
golangci-lint failed on the change and had not before: gofumpt wants
StartCall's same-typed parameters grouped, and runWorkflowCall reached a
cyclomatic complexity of 16. The unwaited call's early close moves into
closeUnwaitedCall on *TaskContext -- it needs only the call step, the
callee and the settings -- which leaves runWorkflowCall below the limit
with its behaviour unchanged.
The walk's id-less call case is deleted: ParseDefinition now refuses that
definition before validateWorkflowCalls runs, so it passed for the wrong
reason. The refusal is pinned where it happens instead, by the compile
case in TestWorkflowCallKeyIsTheDeclaredPath, now asserting the whole
message so it still names the step it refused.
A call's key was the step's declared path alone, so one task running a
second definition at a colliding path -- workflow.handoff, human.approve
and a definition re-pin all change it -- was handed the first callee's
child, waited on the wrong workflow and published its outputs. The key
now carries the callee too, while the step row and the call's events stay
named and staged for the call site, so a rail pairing an event with its
step still attaches them.
A child that reached dead or closed_wont_do can never do the work and
holds no result the caller can use, so it leaves the call key's index and
frees the key for the caller's retry to start a fresh child. parked and
rejected keep theirs: a parked child is recoverable by retrying it, and a
rejected child ran.
The id a call step must declare is now refused where the definition is
compiled, which is also the path a pinned definition takes, so an id that
spells a step type is accepted again and no stage needs a heuristic.
A call site had no durable identity, so any retry of it -- a lost
EnqueueCallTask reply, a failed call-step record, a requeued attempt that
re-runs the stage -- started a second child and its work happened twice.
The call step's declared path within the run is that identity: wrapStep
sets it (branch-qualified, restored on return), a workflow.call step must
declare an id so two call sites cannot share one, and the key crosses
EnqueueCallTaskRequest.call_key into tasks.call_key, which EnqueueCallTask
upserts on so the retry returns the child already started.