Restore Codex source acquisition and reliable CI image preparation

.downstream/native-domains.md

@@ -96,7 +96,11 @@ groups, quota or scheduling state. Normal credential refresh remains available.
 Source acquisition shares one process-local `__oailb` candidate. Its maximum
 lease is 230 seconds, capped by an earlier upstream expiry; receiving the same
 value during a live lease does not extend it. Source preparation is sequential
-and has a 90-second budget. Each target/model is validated with two completed
+and uses the ordinary OAuth account-test payload (default instructions, fixed
+short user prompt, medium effort), with its own configured normal connection
+policy. It does not inherit target-probe session/window headers, routing hints,
+encrypted-output options or connection-close behavior.
+It has a 90-second budget. Each target/model is validated with two completed
 HTTP 200 streams, at most 45 seconds each: the first must return STATE, and
 the second may omit STATE or return the same value. Qualification includes
 the target identity, actual model, exit, client headers, STATE and TLS profile.

.github/buildkitd.toml

@@ -0,0 +1,4 @@
+# Google-operated cache of public Docker Hub images. BuildKit keeps the original
+# image references and validates their content digests; no credentials change.
+[registry."docker.io"]
+  mirrors = ["mirror.gcr.io"]

.github/scripts/prepare_test_images.py

@@ -0,0 +1,27 @@
+#!/usr/bin/env python3
+"""Preload unchanged CI image names from their publishers' alternate registries.
+
+The OCI index digests below were checked against the matching Docker Hub tags.
+Postgres/Redis are Docker Official Images on ECR Public; Ryuk is published by
+Testcontainers on GHCR. No Docker Hub credentials or pull-limit workarounds are
+used. Docker verifies immutable digests; existing tests use these local aliases.
+"""
+
+import subprocess
+
+
+IMAGES = (
+    ("public.ecr.aws/docker/library/postgres@sha256:aa6eb304ddb6dd26df23d05db4e5cb05af8951cda3e0dc57731b771e0ef4ab29", "postgres:18.1-alpine3.23"),
+    ("public.ecr.aws/docker/library/redis@sha256:0514fa59e3d84e2f3be66731c5401fc2a85b8ea71b49c0116c850f6a558076f4", "redis:8.4-alpine"),
+    ("ghcr.io/testcontainers/ryuk@sha256:31b31269d06603366cbfd0284708dcd2e281e8a4188e53fce3d3304439d0df3d", "testcontainers/ryuk:0.13.0"),
+)
+
+
+def prepare(run=subprocess.run):
+    for source, local_name in IMAGES:
+        run(["docker", "pull", "--platform", "linux/amd64", source], check=True)
+        run(["docker", "tag", source, local_name], check=True)
+
+
+if __name__ == "__main__":
+    prepare()

.github/workflows/backend-ci.yml

@@ -45,7 +45,9 @@ jobs:
         run: make test-unit
       - name: Integration tests
         working-directory: backend
-        run: make test-integration
+        run: |
+          python3 ../.github/scripts/prepare_test_images.py
+          make test-integration
 
   frontend:
     runs-on: ubuntu-latest

.github/workflows/downstream-release.yml

@@ -94,6 +94,9 @@ jobs:
           python3 .github/scripts/release_preflight.py --stage before-build \
             --tag "$RELEASE_TAG" --source-sha "$SOURCE_SHA"
       - uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f # v3
+        with:
+          driver-opts: image=mirror.gcr.io/moby/buildkit@sha256:cec9f139f45e93c5c69c60f8b07cfad9f43f4ef6b6a6cd917527fea5ff2e3dea
+          buildkitd-config: .github/buildkitd.toml
       - name: Build release image without publication
         id: build
         uses: docker/build-push-action@10e90e3645eae34f1e60eeb005ba3a3d33f178e8 # v6

.github/workflows/downstream-verify.yml

@@ -45,6 +45,7 @@ jobs:
       - name: Native gateway borrowing contracts
         working-directory: backend
         run: |
+          python3 ../.github/scripts/prepare_test_images.py
           go test -p=1 -tags unit ./internal/service ./internal/repository ./internal/handler ./internal/handler/admin -run '^Test(CodexGatewayBorrow|Pelican|AccountObservation|ObservedAccountLease)' -count=1
           go test -p=1 -tags integration ./internal/repository -run '^TestCodexGatewayBorrowTestRepositoryIntegrationLifecycle$' -count=1
       - name: Codex preparation coordination race tests
@@ -194,6 +195,9 @@ jobs:
     steps:
       - uses: actions/checkout@d23441a48e516b6c34aea4fa41551a30e30af803 # v6
       - uses: docker/setup-buildx-action@8d2750c68a42422c14e847fe6c8ac0403b4cbd6f # v3
+        with:
+          driver-opts: image=mirror.gcr.io/moby/buildkit@sha256:cec9f139f45e93c5c69c60f8b07cfad9f43f4ef6b6a6cd917527fea5ff2e3dea
+          buildkitd-config: .github/buildkitd.toml
       - name: Build candidate
         uses: docker/build-push-action@10e90e3645eae34f1e60eeb005ba3a3d33f178e8 # v6
         with:

backend/internal/service/codex_gateway_borrow_probe.go

@@ -105,8 +105,11 @@ func (s *CodexGatewayBorrowService) acquireSource(ctx context.Context, id int64)
 	if err != nil {
 		return nil, err.Error(), err
 	}
-	shot, err := s.fireObservation(ctx, template.Header, a, codexGatewayBorrowSourceModel, "Reply with OK only.", "medium", proxy, nil,
-		"", nil, HTTPUpstreamProfileCodexBorrowSource, codexGatewayBorrowProbeMaxBody)
+	req, err := borrowSourceObservationRequest(ctx, template.Header)
+	if err != nil {
+		return nil, err.Error(), err
+	}
+	shot, err := s.sendObservation(req, a, proxy, nil, codexGatewayBorrowSourceModel, codexGatewayBorrowProbeMaxBody)
 	if err != nil {
 		return nil, shot.errorText, err
 	}
@@ -125,6 +128,36 @@ func (s *CodexGatewayBorrowService) acquireSource(ctx context.Context, id int64)
 	return candidate, "", nil
 }
 
+// The pinned ranxi source acquisition uses the ordinary OAuth account-test
+// payload, not the target's two-shot STATE-probe protocol. Preserve the shared
+// account identity and observation isolation, but do not manufacture probe
+// sessions, a window or a routing hint, or force this normal source connection
+// closed. The source keeps its own configured OpenAI transport and proxy.
+func borrowSourceObservationRequest(ctx context.Context, headers http.Header) (*http.Request, error) {
+	ctx = WithCodexGatewayBorrowObservation(WithHTTPUpstreamRedirectsDisabled(WithHTTPUpstreamProfile(ctx, HTTPUpstreamProfileCodexBorrowSource)))
+	payload := createOpenAITestPayload(codexGatewayBorrowSourceModel, true, "Reply with OK only.")
+	payload["reasoning"] = map[string]string{"effort": "medium"}
+	body, err := json.Marshal(payload)
+	if err != nil {
+		return nil, err
+	}
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, chatgptCodexURL, bytes.NewReader(body))
+	if err != nil {
+		return nil, err
+	}
+	req.Host, req.Header = "chatgpt.com", headers.Clone()
+	if req.Header == nil {
+		req.Header = make(http.Header)
+	}
+	for _, name := range []string{"Cookie", "Content-Length", "Content-Encoding", "Content-MD5", "X-Codex-Turn-State", "X-Codex-Turn-Metadata", "X-OpenAI-Internal-Codex-Responses-Lite", "conversation_id", "session_id", "session-id", "thread-id", "x-client-request-id", "x-codex-window-id", openAICodexRoutingHintHeader} {
+		deleteOpenAIHeaderEqualFold(req.Header, name)
+	}
+	req.Header.Set("Accept", "text/event-stream")
+	req.Header.Set("Content-Type", "application/json")
+	req.Header.Set("OpenAI-Beta", "responses=experimental")
+	return req, nil
+}
+
 func borrowCookieScoped(cookie *http.Cookie, path string) bool {
 	if cookie == nil || (cookie.Domain != "" && !strings.EqualFold(strings.TrimPrefix(cookie.Domain, "."), "chatgpt.com")) {
 		return false

backend/internal/service/codex_gateway_borrow_source_contract_test.go

@@ -0,0 +1,37 @@
+package service
+
+import (
+	"context"
+	"net/http"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+	"github.com/tidwall/gjson"
+)
+
+func TestCodexBorrowSourceUsesNormalAccountTestContract(t *testing.T) {
+	headers := http.Header{"Authorization": {"Bearer synthetic-source"}, "User-Agent": {"synthetic-persisted-identity"}, "Originator": {"codex-tui"}, "Version": {"0.162.1"}}
+	for _, name := range []string{"session-id", "thread-id", "session_id", "x-codex-window-id", "x-codex-turn-state", "x-codex-turn-metadata", "x-codex-routing-hint", "Cookie", "Content-Encoding"} {
+		headers.Set(name, "target-only-value")
+	}
+	original := headers.Clone()
+	req, err := borrowSourceObservationRequest(context.Background(), headers)
+	require.NoError(t, err)
+	require.False(t, req.Close, "source acquisition keeps normal connection policy")
+	require.Equal(t, original, headers)
+	require.Equal(t, "Bearer synthetic-source", req.Header.Get("Authorization"))
+	require.Equal(t, "synthetic-persisted-identity", req.Header.Get("User-Agent"))
+	for _, name := range []string{"session-id", "thread-id", "session_id", "x-codex-window-id", "x-codex-turn-state", "x-codex-turn-metadata", "x-codex-routing-hint", "Cookie", "Content-Encoding"} {
+		require.Empty(t, req.Header.Get(name), name)
+	}
+	body := borrowCoreBody(t, req)
+	want := createOpenAITestPayload("gpt-6-astra", true, "Reply with OK only.")
+	require.Equal(t, want["instructions"], gjson.GetBytes(body, "instructions").String())
+	require.Equal(t, "Reply with OK only.", gjson.GetBytes(body, "input.0.content.0.text").String())
+	require.Equal(t, "medium", gjson.GetBytes(body, "reasoning.effort").String())
+	require.False(t, gjson.GetBytes(body, "parallel_tool_calls").Exists())
+	require.False(t, gjson.GetBytes(body, "include").Exists())
+	require.Equal(t, HTTPUpstreamProfileCodexBorrowSource, HTTPUpstreamProfileFromContext(req.Context()))
+	require.True(t, IsCodexGatewayBorrowObservation(req.Context()))
+	require.True(t, HTTPUpstreamRedirectsDisabled(req.Context()))
+}