Skip to content

Pin the public OpenShell 0.1.6 interceptor and record rollout prerequisites - #160

Draft
mtalvi wants to merge 7 commits into
validatedpatterns-sandbox:mainfrom
mtalvi:codex/issue-157-public-interceptor
Draft

mtalvi wants to merge 7 commits into
validatedpatterns-sandbox:mainfrom
mtalvi:codex/issue-157-public-interceptor

Conversation

@mtalvi

@mtalvi mtalvi commented Nov 12, 2026 •

Copy link
Copy Markdown
Collaborator

Pins the published OpenShell 0.1.6 governance interceptor so new workspaces use an interceptor built from the same release as their gateways. Runtime corrections remain separate in #162 and #164; this PR changes only the interceptor pin and release documentation. Cluster users, endpoints and credentials are excluded.

The main-only publication run, built from repository fe19b46c29928e5d7bd6ad62440e591bb4907d3f, published NVIDIA/OpenShell v0.1.6 at ac7f615cd12d5cc49cb16eb1f5f10ffeb1e90f47. It includes the provider-annotation workaround and driver-config admission guard. Both native builds passed all six admission tests, followed by the multi-platform merge and anonymous public-pull verification.

  • Public index: sha256:19f145ac163822bca6d406898975b5b770b75b196fa0bee0bc86b108fad213b3.
  • linux/amd64: sha256:118827092bfa9074c1fd1a2ca9fa60d9711025a24e574414295f463213f6f5a4.
  • linux/arm64: sha256:309f0381e45d47efdc4579d5ddb18312709896f2b75888c56c453bacafeb86f4.

Anonymous v0.1.6 and repository-commit tags resolve to this index. The live acceptance pod uses the chart-default public pin and expected amd64 child imageID, with no interceptor override.

Initial clean README Option A installation on OpenShift 4.22.22/amd64 passed from shipping 7d8c8440ac76381e277a56e05121c05317ef58aa, separate deployment bd7ce216c1a436af51cab10422df571214e10f52. All sixteen applications became Synced/Healthy. NetworkPolicy was enabled from first boot; web search/GPU hardware were disabled. The operator requested skipping prereqs-check; topology was verified directly.

Completed initial evidence includes real default/custom TUI/GUI inference, owner access and peer HTTP 403; all three sandboxes' real workspace write/edit and rejected absolute/traversal/symlink/outside-edit calls; governance approval/rejection audit; credential/admin-CA isolation and cleanup; peer-required two-way NetworkPolicy 36/36; and both reapply operations preserving files/passwords/policy/identity/confinement. Local suite: 2097 passed, 3 skipped; chart/rendered-shell lint, OIDC 50/50 and offline network 4/4 passed. Known local Rust skips were covered by publication admission builds; the other skip needs an APF bundle.

The independent reboot investigation found stale credential placeholders and obsolete model auth-profile bindings (#162), plus a host-triggered shutdown stall. Longer grace alone did not resolve that stall. #164 now makes the documented restart request guest poweroff through normal SSH, wait for a replacement, and refuse forced fallback. Its final documented command completed naturally on both controls at 14:52:34 / 14:52:38 UTC, with filesystem unmount and power-down evidence and no grace expiry/Destroyed. All six #164 CI checks pass at 8001cc7c6454bf5398a6ea83065174af0c087a06; it is ready for review.

Final combined deployment: cbccc0f9a0c316cdf4316c2f9f1d7bee9c07a33f, credential code ab114e6980382f2e8e5d963880e247a5e0994873, installer SHA256 e34a7401e4036c0bdb2ca4efce184c28fae837e32fd919b98fb129e7349c143c. Both live ConfigMaps and both guests’ staged installers match that hash. A final alias-lifetime correction fixes native reapply after replacement activation; the exact generated native dry-run and 28 custom-provider regressions pass. Both documented guest restarts again shut down naturally at 15:21:31 / 15:22:16 UTC. Both controls completed boot and automatic reapply with current inputs and healthy consistent catalogs. All three gateways answered real inference without fallback; C and D’s original notebook passed workspace tool/preservation receipts, and C’s four representative outside operations were rejected. D’s original CUDA workspace failed preservation: the existing installer automatically deleted/recreated it at14:56 UTC; its old volume/sentinel are absent and its password changed. The replacement’s read-only inference success is not preservation acceptance. This is handed off in #165. The final peer check was not run after that failure, at the operator’s scope decision. Require CI on this PR's current head and #162's final PR head 5af7718161b7c9befad7df1f4d36e2344470cb03 (test-only correction after runtime source ab114e6).

Deployment hold: merge #162/#164 and resolve the separate data-preservation blocker #165 with actual acceptance before rollout. Keep this PR draft and keep #157 open. CI for this PR’s final 869d314 head was reported all green by the operator; #162’s test-only 5af7718 CI remains with the operator to confirm. Initial security evidence is reused; no additional fresh VM/cluster installation or repeated full security matrix is claimed. One original Podman catalog remains damaged and preserved privately, with both workspace sentinels intact. Its cause/onset and any relationship to this pin remain unproven; no package/journal-mode/database reset or repair was used. #161 remains open for that finding and the generic host-triggered termination stall. Nothing has been merged; #157 stays open until the verified follow-up merges.

@mtalvi mtalvi changed the title Pin the public OpenShell 0.1.6 interceptor and validate a clean install Pin the public OpenShell 0.1.6 interceptor and record rollout prerequisites Nov 12, 2026

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant