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.
Remove controlplanerpc.NewClient (superseded by NewRPCClient), Loader.Overlay
(-config-overlay reaches Resolve directly), and setup.Run (RunParams is
the live entry point). Move ResolverFunc into a test file, since only
tests adapt functions to the egress Resolver.
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.
Remove PRD/doc paths, bead IDs, phase notes and deleted-test references
from comments; drop test-only hooks left by the test deletion. Includes
in-progress archied bootstrap, gateway and query changes.
The sandbox kit spec lets one credential@1 declare apiKey and oauth
together, meaning "whichever the host has bound": the published
docker/claude-code-kit does exactly that for anthropic, and codex for
openai. archie decided the kind from the Kit's declaration instead
(egress.IsOAuthManaged), so a dual-declared service always came out
OAuth-bound however the org had bound it:
- kitrun.resolveCredentials granted an API-key binding's service the
empty string, so the proxy had no key to inject and every request to
a domain the Kit named carried an empty credential. The reference
Claude Code Kit could not authenticate with an API key at all.
- an API-key-only service whose binding named no secret resolved to ""
through the zero SecretRef, presenting an empty credential rather than
staying unbound.
- the proxy fired an apiKey inject rule for an OAuth-bound service,
putting "Bearer " on the wire, and intercepted the token endpoint of
an API-key-bound one, whose only outcomes were a 502 on the missing
token set, a capture the org never bound, or -- for a required
dual-declared credential -- a refused launch when no token store was
configured.
- the container's SBX_CRED_*_MODE said "apikey" for an OAuth-bound
dual-declared service, and kitrun demanded a captured OAuth token for
an API-key-bound one whose Kit renders a credential file.
One root cause, fixed once: the org's binding decides the kind. It is
now a first-class fact (egress.CredentialKind: apikey or oauth, absent
meaning unbound), computed from the binding in resolveCredentials and
carried to every consumer of it -- the resolver grant, the container's
credential state (mode variable and credential file), and the proxy
session, whose rule sets are built from it so a rule can only fire for a
service the run carries as that same kind. A service the run does not
carry at all keeps both rule sets, so resolving it still refuses a
required credential and skips an optional one.
Declared and bound stay two independent facts, as the reported
declared-and-granted intersection requires: neither widens the other.
No secret value crosses into kit (Bound carries kinds only) and none is
logged or written to a task row.
A Kit's credentialFile cannot render {{.Scopes}} because nothing records the
granted scopes: the token response's `scope` field (RFC 6749 section 5.1,
space-delimited) was never read, and the token set had nowhere to hold it.
capture now parses `scope` into []string and keeps the previous set when the
field is absent or empty -- the same rule the refresh token follows, so a
refresh that does not repeat the scopes preserves what the login captured.
The set is carried end to end: Scopes on harnesssecret.Secret, in
harnessSecretPayload (JSON inside the existing encrypted envelope, so no
migration), on message HarnessSecret as `repeated string scopes = 8`, and
both ways through staterpc values.go. The proto field is load-bearing: the
daemon reads the token set over gRPC, so a local-struct-only field is
silently lost in production -- TestHarnessSecretConformance proves the round
trip in both its local and grpc modes.
The egress proxy now holds a Kit's OAuth tokens on the host. A refresh
carrying the sentinel refresh token goes upstream with the org's stored
one; a successful response is written to the State Store and returned to
the container with sentinels only. Requests to the Kit's resource hosts
carrying the sentinel access token get the stored one.
The org's token set is read or written only for a run whose credential
carries the service, so a Kit declaring a service never reaches the org
secret by itself. The launcher grants a bound OAuth service without a
config secret, and Register refuses a required OAuth credential when no
harness secret store is wired. A provider that does not rotate refresh
tokens keeps the stored one.
Part of archie-core-egkf.7 (#1208). Credential-file rendering and
per-binding concurrency remain.
docs/prds/external-agent-harness.md (Credentials) and orgs-and-access.md:
egress.Resolver had no production implementation, and no run carried a
credential naming the org secrets its identity is granted, so a Kit's
required credential@1 was refused at the proxy and no harness run
authenticated (archie-core-egkf.11).
Two things, as scoped (this does not attempt #1026's full run-credential
consolidation replacing WorktreeGrantIssuer/StateStoreGrantIssuer -- that
PRD item is real, separate, and still open; this is deliberately narrower):
1. A run credential naming granted secrets: config.CredentialBinding maps a
Kit credential@1 service to a secret; config.Config/IdentityConfig gain
Org and GrantedCredentials (a per-identity allowlist of service names,
reusing the existing per-identity config-override pattern DiffCapLines
already uses, not a new Cedar/Authorizer integration -- KindSecret policy
objects exist in internal/domain/access but assembling a full Principal
is #1026's job, not a two-line addition here). config.ContainerConfig.
BoundCredentials computes the declared (Kit) ∩ granted (identity) ∩
org-matched intersection Sam's own rule requires: neither side widens
the other, proven by a mutation-checked test flipping each of the three
independently.
2. egress.GrantResolver: the production Resolver, holding only what a
caller explicitly Grant-ed one run, replaced wholesale (not merged) on
a repeated Grant, gone entirely on RevokeGrant. kitrun.Launcher.Launch
computes the intersection per launch and resolves real values through
the existing secret registry (SecretResolver), Grants it before the
proxy session opens, and Release/error-path rollback (a new shared
Launcher.release) always revokes it.
Course-corrected mid-implementation (flagged by Sam): credential bindings
first landed as a boot-time config.ContainerConfig.Credentials field
captured once into Launcher.Bindings -- exactly the "parses but is not a
resource" trap CLAUDE.md warns about, and inconsistent with agent-profiles
(#1207) making a Kit profile apply without a restart. Rebuilt as its own
live control-plane resource, CredentialBindingsKind ("credential-bindings"),
following the agent-profiles/container-runtime-policies split precisely:
config.ContainerConfig.Credentials is json:"-", [containers].credentials
is only its seed, and kitrun.Launcher now holds *config.Holder (not a
captured slice) and reads Containers.Credentials fresh on every Launch.
Org/GrantedCredentials were already read fresh per dispatch (d.configFor
inside runKitTask, never cached), so no equivalent fix was needed there.
Deliberately not built: identities and grants as their own State Store
resource (orgs-and-access.md's "Org owns ... identities and their grants").
Identities are entirely config.toml-defined today with no control-plane
resource at all; moving GrantedCredentials alone into one would put half
an identity in two places. That migration is identities' own scope, not
credential bindings'.
A Kit task's worker shares the sandbox network with its harness, so it
reaches NATS and the State Store through the relay like the harness
reaches the proxy. archie-agent relay takes repeated -forward
listen=target pairs, binds them all before serving so a port conflict
fails at start, and stops every forward when one fails. RelayArgs builds
the command from the fixed relay ports.
A session carries its run and its Kit's credential@1 requests. The proxy
applies apiKey inject rules only in the credential's phase and only to
hosts its rule names, after the network policy has admitted the request,
so a key never reaches a host its Kit did not declare. The real value
comes from a Resolver that returns only secrets the run credential
carries: a required credential that is unbound or unresolvable stops the
request before anything is sent, and an optional one passes only when it
is genuinely unbound. Named keys appear in the sandbox as a sentinel;
inject-only keys have no presence.
Each sandbox gets an internal bridge on which the host holds no address,
so host services listening on every interface are unreachable; the relay
is attached under a fixed alias and is the one way out. ProxyEnv gives a
sandbox container its proxy URL and CA trust. The Docker integration test
proves allowed egress works and that the proxy listener, host services and
the internet are unreachable directly; it fails if the host-address
inhibition is removed.
archie-agent relay listens on the sandbox's isolated network and joins
every connection to one fixed target, the proxy. The relay holds no
secrets and can reach nothing else, so the host needs no address on a
sandbox network at all.
CONNECT tunnels and absolute-form http requests authenticate to a
per-container session, are checked against the session's phase policy,
and are forwarded with verified upstream TLS. A tunnel serves only its
admitted host: mismatched SNI and Host headers are refused. Upstream dials
refuse loopback, and refuse internal addresses unless the policy names the
exact IP, so an allowed name cannot resolve back into the host.