fix(providers): restore supervisor-backed GCP metadata discovery by feloy · Pull Request #3973 · NVIDIA/OpenShell

feloy · 2026-09-30T16:49:42Z

🏗️ build-from-issue-agent

Summary

Restore the Google Cloud metadata emulator removed during the sandbox/supervisor split. SDK requests to the injected 127.0.0.1:8174 address now reach a supervisor-owned emulator through the existing authenticated network relay, allowing Application Default Credentials discovery and repeated token refresh.

Related Issue

Closes #3860

Changes

Adapt the metadata handler removed in feat(isolation): implement the RFC 0012 sandbox architecture #2942 to the supervisor's live provider state, serving project/account metadata and token placeholders while preserving non-secret classification and credential expiry.
Reserve the exact metadata loopback destination in the sandbox broker and dispatch direct TCP and HTTP proxy requests locally, with verified workload identity and bounded request parsing.
Add Google SDK discovery and repeated-refresh E2E coverage, plus the expected 503 response without an explicit credential binding.
Update Google provider documentation and the sandbox-policy skill to explain supported Linux runtimes, the reserved address, and credential-binding requirements.

Deviations from Plan

Local E2E validation uses Podman because Docker is unavailable on this development host. The regression tests remain in the existing Python E2E suite.

Testing

mise run pre-commit passes, including the commit hook.
mise run ci passes (lint, compile/type checks, Rust/Python/TypeScript/Go tests, and dependency policy checks).
Provider E2E suite passes against an isolated Podman gateway: 18 passed, 0 skipped in 134.02s.
Unit tests added/updated: metadata responses and headers, credential refresh/removal/expiry, configuration classification, account aliases, exact destination matching, identity verification, and malformed/oversized/incomplete requests.
E2E tests added/updated: pinned google-auth==2.40.3, ADC metadata detection, project/account discovery, repeated refresh, placeholder isolation, and missing credential bindings.
Manual Podman reproduction: metadata probe returns 200 with a bound Google Cloud provider and 503 without a binding.
Linux-only broker tests cover TCP relay, ordinary loopback connections, and rejection of UDP to the reserved address; these tests are not executable on the macOS host.

E2E command:

OPENSHELL_GATEWAY_BIN="$PWD/target/debug/openshell-gateway" \
OPENSHELL_BIN="$PWD/target/debug/openshell" \
SUPERVISOR_IMAGE=localhost/openshell/supervisor:dev \
SANDBOX_IMAGE=localhost/openshell/sandbox:dev \
OPENSHELL_E2E_PODMAN_SANDBOX_IMAGE=localhost/openshell/e2e-python:dev \
UV_NO_SYNC=1 PYTHONPATH=python \
e2e/with-podman-gateway.sh uv run pytest \
  -o 'python_files=test_*.py *_test.py' -vv e2e/python/test_sandbox_providers.py

Checklist

Follows Conventional Commits
Commits are signed off (DCO)
Documentation updated in docs/how-it-works/providers/google.mdx; related public policy skill reviewed and updated.

Relay the reserved metadata endpoint to the supervisor and restore project, account, and placeholder token responses from live provider state. Cover Google SDK discovery and repeated refresh with provider E2E tests. Closes NVIDIA#3860 Signed-off-by: Philippe Martin <phmartin@redhat.com>

copy-pr-bot · 2026-09-30T16:49:46Z

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

feloy · 2026-09-30T16:50:07Z

🏗️ build-from-issue-agent

E2E Test Attestation

Local provider E2E tests passed against an isolated Podman gateway.

Field	Value
Commit	`ed234d71ee5f9677e6bce4fed21d93b3b01e21b4`
Gateway mode	Podman on macOS, Linux arm64 runtime images
Result	18 passed, 0 skipped in 134.02s
SDK	google-auth 2.40.3, requests 2.32.5

Command

OPENSHELL_GATEWAY_BIN="$PWD/target/debug/openshell-gateway" \
OPENSHELL_BIN="$PWD/target/debug/openshell" \
SUPERVISOR_IMAGE=localhost/openshell/supervisor:dev \
SANDBOX_IMAGE=localhost/openshell/sandbox:dev \
OPENSHELL_E2E_PODMAN_SANDBOX_IMAGE=localhost/openshell/e2e-python:dev \
UV_NO_SYNC=1 PYTHONPATH=python \
e2e/with-podman-gateway.sh uv run pytest \
  -o 'python_files=test_*.py *_test.py' -vv e2e/python/test_sandbox_providers.py

Runtime images

Supervisor: localhost/openshell/supervisor@sha256:976c0ed75b7a0612517d347062ee504eba9b29a9e7670fa069b1a41c2c292758
Sandbox runtime image ID: 9c7561b5c640dfb892610763cd8f4a44d681c57c030309a89eec2569c27d1f14
Workload: localhost/openshell/e2e-python@sha256:f48c1464036a5fddc110d594697d6d390f55af8809520758c96ade1bc639a138

Tests Executed

All tests are in e2e/python/test_sandbox_providers.py.

test_provider_credentials_available_as_env_vars — PASSED
test_profileless_provider_creation_is_rejected — PASSED
test_endpointless_profile_credentials_fail_closed_without_policy_binding — PASSED
test_endpointless_profile_credentials_use_explicit_policy_binding — PASSED
test_google_metadata_sdk_discovery — PASSED
test_nvidia_provider_injects_nvidia_api_key_env_var — PASSED
test_attach_detach_updates_credentials_for_later_exec_launches — PASSED
test_imported_openai_profile_allows_native_endpoint_with_attached_provider — PASSED
test_imported_anthropic_profile_allows_native_endpoint_with_attached_provider — PASSED
test_create_sandbox_rejects_unknown_provider — PASSED
test_credentials_not_in_persisted_spec_environment — PASSED
test_update_provider_preserves_unset_credentials_and_config — PASSED
test_update_provider_empty_maps_preserves_all — PASSED
test_update_provider_merges_config_preserves_credentials — PASSED
test_update_provider_rejects_type_change — PASSED
test_github_provider_allows_https_git_clone — PASSED
test_provider_profile_platform_vs_workspace_isolation — PASSED
test_cross_workspace_profile_ids_do_not_collide — PASSED

Full local mise run ci and final mise run pre-commit also passed. Linux-only broker unit tests could not run on the macOS host; the Podman E2E exercised the Linux broker.

johntmyers

gator-agent

PR Review Status

The restoration is project-valid and the full initial patch review found one blocking correctness gap in the default documented provider setup.

Action required: @feloy, make metadata-based refresh work when service_account_email is omitted and add coverage for that documented path.

Blocking findings:

GATOR-ed234d71-01: the recursive account response can select an empty account name and send the SDK to an unsupported token route.

Carried findings:

None

Gator metadata

Validation: Implements the validated regression in #3860.
Docs: Updated, but the documented no-email setup exposes the blocking finding.
Checks: Current-head branch and Helm gates are pending; pipeline handoff waits for review feedback.
E2E: Required for provider credential and sandbox-network behavior; dispatch waits until the blocker is addressed.
Head SHA: ed234d71ee5f9677e6bce4fed21d93b3b01e21b4
Base SHA: 7caff12d3c9013e7063d22391a76f775f6ce93b5
Merge base SHA: 7caff12d3c9013e7063d22391a76f775f6ce93b5
Patch ID: c51d73bf9e9c8c269ff8710b12d2b5bc440af861
Gator payload: 9
Review mode: initial
Previous reviewed SHA: none
Review budget exhausted: no
Maintainer decision required: no
Next state: gator:in-review

Use the default account identifier when the optional service account email is missing or empty. Cover repeated SDK refresh for missing, empty, and configured email values with distinct providers for parallel E2E execution. Signed-off-by: Philippe Martin <phmartin@redhat.com>

github-actions · 2026-09-30T23:09:13Z

Label test:e2e applied, but pull-request/3973 does not exist yet. A maintainer needs to comment /ok to test c158fff6b7562b59a536420d23cc3574eba63640 to mirror this PR. Once the mirror exists, re-apply the label or re-run Branch E2E Checks from the Actions tab.

johntmyers · 2026-09-30T23:10:40Z

/ok to test c158fff

github-actions · 2026-09-30T23:11:09Z

Label test:e2e applied for c158fff. Open the existing run and click Re-run all jobs to execute with the label set. The run will execute the standard E2E suite after building the required gateway, sandbox, and supervisor images once. The matching required CI gate status on this PR will flip green automatically once the run finishes.

johntmyers

gator-agent

PR Review Status

The follow-up review confirms that the default-account fix and expanded SDK coverage resolve GATOR-ed234d71-01. No blocking findings remain, and the required current-head Branch Checks and E2E workflows are now running.

Blocking findings:

No blocking findings remain

Carried findings:

None

Gator metadata

Validation: Implements the validated GCP metadata regression in #3860.
Docs: Updated for optional or empty service_account_email behavior.
Checks: Current-head Branch Checks are queued or running; Helm dispatch is complete.
E2E: test:e2e applied; the current-head Branch E2E Checks workflow is running.
Head SHA: c158fff6b7562b59a536420d23cc3574eba63640
Base SHA: 7caff12d3c9013e7063d22391a76f775f6ce93b5
Merge base SHA: 7caff12d3c9013e7063d22391a76f775f6ce93b5
Patch ID: ea286f19a426e7fd6b2be7f6c31795406df95643
Gator payload: 9
Review mode: follow_up
Previous reviewed SHA: ed234d71ee5f9677e6bce4fed21d93b3b01e21b4
Review budget exhausted: no
Maintainer decision required: no
Next state: gator:watch-pipeline

Signed-off-by: Philippe Martin <phmartin@redhat.com>

johntmyers

gator-agent

PR Review Status

The follow-up review confirms that the test-only timeout and import cleanup do not reopen the resolved metadata-account finding or introduce a new blocker. No blocking findings remain; current-head CI and E2E still need dispatch.

Blocking findings:

No blocking findings remain

Carried findings:

None

Gator metadata

Validation: Implements the validated GCP metadata regression in #3860.
Docs: Updated for the restored Google Cloud metadata behavior.
Checks: Current-head workflows await contributor authorization and dispatch.
E2E: test:e2e is required; current-head dispatch is pending.
Head SHA: ec01442e51ff943679ab0e5b1647881b794bfa31
Base SHA: 7caff12d3c9013e7063d22391a76f775f6ce93b5
Merge base SHA: 7caff12d3c9013e7063d22391a76f775f6ce93b5
Patch ID: 514dbff67fcb4989bf62c31a966de9eae9263133
Gator payload: 9
Review mode: follow_up
Previous reviewed SHA: c158fff6b7562b59a536420d23cc3574eba63640
Review budget exhausted: no
Maintainer decision required: no
Next state: gator:in-review

johntmyers · 2026-09-30T23:48:26Z

/ok to test ec01442

github-actions · 2026-09-30T23:49:04Z

Label test:e2e applied for ec01442. Open the existing run and click Re-run all jobs to execute with the label set. The run will execute the standard E2E suite after building the required gateway, sandbox, and supervisor images once. The matching required CI gate status on this PR will flip green automatically once the run finishes.

feloy requested review from a team, derekwaynecarr, mrunalp and sjenning as code owners September 30, 2026 16:49

feloy mentioned this pull request Sep 30, 2026

bug: Restore GCP metadata server support #3860

Open

3 tasks

johntmyers reviewed Sep 30, 2026

View reviewed changes

Comment thread crates/openshell-supervisor-network/src/google_cloud_metadata.rs Outdated

johntmyers added the gator:in-review Gator is reviewing or awaiting PR review feedback label Sep 30, 2026

johntmyers added the test:e2e Requires end-to-end coverage label Sep 30, 2026

johntmyers added test:e2e Requires end-to-end coverage and removed test:e2e Requires end-to-end coverage labels Sep 30, 2026

johntmyers reviewed Sep 30, 2026

View reviewed changes

test(sandbox): fix metadata relay lint and timeout

ec01442

Signed-off-by: Philippe Martin <phmartin@redhat.com>

johntmyers reviewed Sep 30, 2026

View reviewed changes

johntmyers added test:e2e Requires end-to-end coverage and removed test:e2e Requires end-to-end coverage labels Sep 30, 2026

Navigation Menu

fix(providers): restore supervisor-backed GCP metadata discovery - #3973

fix(providers): restore supervisor-backed GCP metadata discovery#3973
feloy wants to merge 3 commits into
NVIDIA:mainfrom
feloy:fix-3860/restore-gcp-metadata

feloy commented Sep 30, 2026

Uh oh!

copy-pr-bot Bot commented Sep 30, 2026

Uh oh!

feloy commented Sep 30, 2026

Uh oh!

johntmyers left a comment

Uh oh!

Uh oh!

github-actions Bot commented Sep 30, 2026

Uh oh!

johntmyers commented Sep 30, 2026

Uh oh!

github-actions Bot commented Sep 30, 2026

Uh oh!

johntmyers left a comment

Uh oh!

johntmyers left a comment

Uh oh!

johntmyers commented Sep 30, 2026

Uh oh!

github-actions Bot commented Sep 30, 2026

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

feloy commented Sep 30, 2026

Summary

Related Issue

Changes

Deviations from Plan

Testing

Checklist

Uh oh!

copy-pr-bot Bot commented Sep 30, 2026

Uh oh!

feloy commented Sep 30, 2026

E2E Test Attestation

Command

Runtime images

Tests Executed

Uh oh!

johntmyers left a comment

Choose a reason for hiding this comment

PR Review Status

Uh oh!

Uh oh!

github-actions Bot commented Sep 30, 2026

Uh oh!

johntmyers commented Sep 30, 2026

Uh oh!

github-actions Bot commented Sep 30, 2026

Uh oh!

johntmyers left a comment

Choose a reason for hiding this comment

PR Review Status

Uh oh!

johntmyers left a comment

Choose a reason for hiding this comment

PR Review Status

Uh oh!

johntmyers commented Sep 30, 2026

Uh oh!

github-actions Bot commented Sep 30, 2026

Uh oh!

This branch has not been deployed

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants