feat(supply-chain): sign published images, add a chart image inventory, and correct the self-hosting docs - #7432
Closed
waleedlatif1 wants to merge 14 commits into
Closed
feat(supply-chain): sign published images, add a chart image inventory, and correct the self-hosting docs#7432waleedlatif1 wants to merge 14 commits into
waleedlatif1 wants to merge 14 commits into
Conversation
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Collaborator
Author
Collaborator
Author
|
@cubic-dev-ai review this PR |
Contributor
@waleedlatif1 I have started the AI code review. It will take a few minutes to complete. |
Contributor
Greptile SummaryThe PR adds keyless signing, provenance, and architecture-specific SBOM publication for GHCR images, introduces a rendered Helm image inventory with CI drift checks, and updates self-hosting and enterprise documentation.
Confidence Score: 2/5The PR does not appear safe to merge until the disconnected-registry path mismatch and the successful-stale-read gap in latest-image attestation are fixed. The inventory directs operators to mirror Sim-owned images beneath Files Needing Attention: scripts/generate-image-manifest.ts, helm/sim/images.yaml, helm/sim/templates/_helpers.tpl, .github/workflows/ci.yml
|
| Filename | Overview |
|---|---|
| .github/workflows/ci.yml | Adds digest-based subject discovery and a matrix that signs images and publishes provenance and per-architecture SBOM attestations. |
| scripts/generate-image-manifest.ts | Renders the Helm chart, extracts unique container references, and generates or verifies the checked-in image inventory. |
| helm/sim/templates/gpu-device-plugin.yaml | Moves the NVIDIA device-plugin image through the shared configurable image-rendering helper. |
| helm/sim/values.yaml | Adds configurable device-plugin repository, tag, pull policy, and digest values. |
| helm/sim/images.yaml | Records the complete rendered image set for disconnected-install preparation. |
| .github/workflows/helm.yml | Adds the generated image-inventory drift check to Helm CI and expands workflow path triggers. |
Sequence Diagram
sequenceDiagram
participant Build as Architecture builds
participant Publish as create-ghcr-manifests
participant Resolve as attest-subjects
participant Attest as attest-images
participant Release as create-release
Build->>Publish: Immutable architecture tags
Publish->>Publish: Create SHA, version, and eligible latest indexes
Publish->>Resolve: Publication job succeeds
Resolve->>Resolve: Resolve tags and deduplicate digests
Resolve->>Attest: Dynamic subject matrix
Attest->>Attest: Generate per-architecture SBOM
Attest->>Attest: Publish provenance and signatures
Attest->>Release: Successful attestation gate
Reviews (10): Last reviewed commit: "fix(supply-chain): scan the plain manife..." | Re-trigger Greptile
![greptile-apps[bot]](https://avatars.githubusercontent.com/in/867647?s=60&v=4){kind=small}
![cubic-dev-ai[bot]](https://avatars.githubusercontent.com/in/1082092?s=60&v=4){kind=small}
waleedlatif1
force-pushed
the
feat/supply-chain-attestations
branch
from
September 3, 2026 07:06
d1e8fee to
4820548
Compare
Collaborator
Author
Collaborator
Author
|
@cubic-dev-ai review this PR |
Contributor
@waleedlatif1 I have started the AI code review. It will take a few minutes to complete. |
Collaborator
Author
Collaborator
Author
|
@cubic-dev-ai review this PR |
Contributor
@waleedlatif1 I have started the AI code review. It will take a few minutes to complete. |
Collaborator
Author
Collaborator
Author
|
@cubic-dev-ai review this PR |
Contributor
@waleedlatif1 I have started the AI code review. It will take a few minutes to complete. |
Collaborator
Author
Collaborator
Author
|
@cubic-dev-ai review this PR |
Contributor
@waleedlatif1 I have started the AI code review. It will take a few minutes to complete. |
Collaborator
Author
|
@cubic-dev-ai review this PR |
Contributor
@waleedlatif1 I have started the AI code review. It will take a few minutes to complete. |
…e inventory
Enterprise security reviews ask for three artifacts Sim did not publish: a
signature proving who built an image, SLSA provenance describing how, and an
SBOM listing what is inside. Add all three, plus the image inventory an
operator needs to mirror Sim into a disconnected registry.
attest-images runs after create-ghcr-manifests rather than inside the build.
buildx's own provenance/sbom attestations stay off because the extra manifests
they add to an index break the imagetools retagging promote-images depends on;
attaching attestations to the finished index leaves it untouched, since they
are stored as separate referrer manifests. The subject is the sha index digest
— imagetools create is deterministic, so the version and latest indexes built
from the same per-arch manifests share that digest and one attestation covers
every tag a customer can pull.
The digest comes from `{{json .Manifest}}` piped through jq rather than
`{{.Manifest.Digest}}`: the latter resolves the index to the runner's own
platform, which would attest one architecture and leave the pulled index
unsigned.
helm/sim/images.yaml is derived from the rendered chart, not from values.yaml,
because one image is written directly into a template — the NVIDIA device
plugin — and a values-derived list misses it in the case that is hardest to
notice, where the mirror succeeds and one pod still pulls from the internet.
Tags stay unresolved: digests belong to a release, and the chart's sim.image
helper already accepts a per-image digest for pinning at install time.
The check runs in the helm workflow, which has Helm set up; check:audits does
not.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015BwsJTEQRzWJaY4BRCkPZt
An audit of the 27 self-hosting and enterprise pages checked every concrete claim against the code it describes. This fixes what it found. Errors that made a documented path fail: - `openssl rand -hex 32` prints 64 hex characters, not the "32 hex chars" five entries claimed. Only ENCRYPTION_KEY and API_ENCRYPTION_KEY are length- validated; the rest are free-form minimums, so the fix differs per variable. - API_ENCRYPTION_KEY was absent from the whole Kubernetes path and listed as required elsewhere. It is optional, and unset means API keys are stored in plain text behind one warning — a silent failure documented nowhere. - The per-purpose S3 fallback was described backwards: knowledge-base, chat, copilot and profile-picture buckets resolve to an empty string, not a literal. The GCS table carried the mirror-image claim, and Azure had no warning at all. - Local uploads go to /app/uploads, not /uploads, and no compose file or chart mounts a volume there. - The remote-sandbox table omitted the immutable template and snapshot refs, so the documented configuration could not work. - The GKE BackendConfig snippet set a values key no template renders. - Retention deletion is scheduled by nothing in the chart, so a deployment that followed the docs deleted nothing. - Every "Settings → Enterprise/Security → X" path named nav groups that do not exist, across eleven enterprise pages. Claims that understated what ships: - The outbound-proxy section said no platform traffic can be proxied. The server runs on Bun, whose fetch honors the proxy variables, so provider SDK and Resend traffic already proxies; guarded egress, SMTP, object storage and OTLP do not. Replaced with a per-path table, including the constraint that the per-request proxyUrl refuses a private address. - Telemetry is off by default on Helm, not on, and enabling the collector collects nothing until NEXT_TELEMETRY_DISABLED is cleared. - The CronJob count was 18 in two places; there are 22 in both the chart and the crontab. - The permission-group table was stale throughout, named a deployment key that does not exist, and inverted the checkbox semantic. About fifteen governed toggles were undocumented. Also documents ~25 environment variables nothing described, the migration advisory lock and its operational knobs, the CloudWatch metrics that activate on any deployment with AWS credentials, and sim-setup doctor. Redundancy removed where a page restated itself: FAQ blocks that repeated their own body, four drifting copies of the same tables, and three cloud tabs identical but for a filename. The sandbox base-image runbook moves from the enterprise enablement page, where it was out of place, to its own self-hosting page rather than being lost. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015BwsJTEQRzWJaY4BRCkPZt
A second round of verification checked every added and removed line against source. It found errors the first pass introduced, which matter more than the ones it fixed — a wrong new claim is worse than the stale one it replaced. Corrections to the corrections: - The GCS bucket table was inverted. Every GCS context falls back to GCS_BUCKET_NAME; the rewrite claimed none did. The pre-existing text was right and was replaced with a falsehood. - S3 and Azure each have three fallback exceptions, not one: og-images and workspace-logos fall back to the general bucket, and execution-files falls back to a literal. - The proxy table had four wrong rows. Azure OpenAI, Azure Anthropic and vLLM are pinned only when the endpoint is user-supplied, so an operator-configured endpoint does proxy. Azure Blob and GCS honor the proxy variables; S3 and Bedrock do not. SES is the AWS SDK over HTTPS, not raw TCP. Adds the on-prem hazard that matters most here: Ollama, LiteLLM and vLLM are dialed over the default fetch, so NO_PROXY is required. - The CloudWatch warning described a bill that cannot happen — every hosted-key metric call site is gated on the deployment being Sim Cloud. - networkPolicy.egressExceptCidrs: [] keeps the metadata defaults rather than removing them, since Helm's `default` treats an empty list as unset. Verified by rendering the chart. The security guidance said the opposite. - The migrator retries far more than two error classes, so a database that is simply down backs off rather than failing immediately. - /api/v1/audit-logs authenticates with an API key, not a session, and the Enterprise-plan requirement does not apply self-hosted — the exact deployment the new paragraph addressed. - The local upload path is /app/apps/sim/uploads: the standalone server chdirs into its own directory. - The update feed answers 200, not 302. KB_OPENAI_MODEL_NAME is optional. OCR_PROVIDER inference is current behavior, not a legacy path. Restores six things the cleanup dropped that existed nowhere else: version-skip guidance, the workspace and last-member rejection rules, the gemma3:4b starter model, kubectl top for OOMKill diagnosis, the grandfathered-workspace case, and an explicit note that Compose telemetry is on by default. Also resolves contradictions only visible across pages: the sandbox prerequisites disagreed on three pages, and authentication.mdx kept a settings path the enterprise pages had corrected. The image inventory drops chartVersion. It made every chart-version bump fail a check whose message says an image changed, and appVersion already pins the tags. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015BwsJTEQRzWJaY4BRCkPZt
helm/sim/** changed, so the chart-version gate requires a SemVer bump. Patch: images.yaml is an added inventory file, with no template or values change. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015BwsJTEQRzWJaY4BRCkPZt
…eview Review found that the attestation covered only the multi-arch index while create-ghcr-manifests also publishes single-arch tags (:version-amd64, :latest-arm64, and so on) that resolve to different digests. Those were pullable but unsigned, so verification failed for anyone who pinned one. The matrix now attests all three subjects per image — the index and both per-arch manifests. That also makes the SBOMs truthful: the amd64 and arm64 images contain different packages, and one SBOM on the index cannot describe both. create-release now requires attest-images, so a release can no longer advertise images whose signing failed. The helm workflow's path filter now includes the inventory generator, which could previously change without the check that validates it ever running. The inventory header named global.imageRegistry alone as the redirect. That only rewrites the simstudioai/* images; third-party ones also need global.useRegistryForAllImages, and the NVIDIA device plugin is hardcoded in a template and cannot be redirected by any values key at all. Docs corrections from the same round: - The copilot workloads cannot be covered by networkPolicy.egress — it is only rendered into policies the chart creates, and it creates none for them. - Data drains and retention cleanup need Trigger.dev. Without it the work is claimed and enqueued to the database, but nothing drains that job type. - MIGRATION_DATABASE_URL has no chart values key: the migrations init container builds DATABASE_URL itself and reads only the database Secret. - Finished CronJob Jobs are deleted after ten minutes by ttlSecondsAfterFinished, so the history limits do not mean what the page implied. - lock_timeout is 5s for ordinary DDL, but a migration can set it to 0 for statements like CREATE INDEX CONCURRENTLY. - Losing Redis interrupts in-flight execution streams and cancellation, which "drops live updates and nothing else" understated. - NEXT_PUBLIC_BRAND_BACKGROUND_COLOR sets no background; it only tells Sim whether the brand ground is dark. - Pi uses its own image, not the Function base. - doctor skips the Postgres, migration, and Redis probes on the production Compose layout, and does not shape-check optional secrets. - The Daytona snapshot build needs the E2B parity manifest to exist first. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015BwsJTEQRzWJaY4BRCkPZt
Reading the final state of all 34 pages as prose, rather than diffing them, surfaced seams that per-claim review cannot see. The worst was mine: an earlier replacement produced "with a configured Trigger.dev project, and a configured Trigger.dev project" in three files. Also fixed: - A dangling colon in troubleshooting where the list it introduced had been removed, and a lead-in promising two deployment options above a list of three. - "Before the UI walkthrough" on a section that now sits after it, and a reference to "steps 3 and 4 below" that are above. - A rollback snippet whose comment said to set the previous tag but whose commands never did, and an LM Studio section telling operators to recreate a service using the development compose file. - A scaling page that recommended the 8 Gi default two lines above warning that peaks reach 12 GB. - An ADFS example setting SSO_ISSUER to Sim's own URL, which the callout four lines earlier says it must not be. - Table cells that were sentence fragments, prose glued onto the end of an <Image> tag, and a comma splice. - Paragraphs restating an earlier paragraph on the same page verbatim, in access-control, self-hosted, whitelabeling, object-storage and email. - "canvas" used for the visual builder in five places, competitor name-dropping, and two marketing-shaped passages on operator pages. - Title Case headings in troubleshooting, the only page not using sentence case. - type="warning" normalized to type="warn". Both resolve identically in fumadocs, so this is consistency rather than a rendering fix. Every in-scope link and anchor still resolves after the heading renames. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015BwsJTEQRzWJaY4BRCkPZt
…mage mirrorable Review was right on both blocking points, and both needed code rather than wording. `imagetools create` always writes an index, even from a single manifest — so `:<version>-amd64` is a single-entry index whose digest differs from the `:<sha>-amd64` manifest it wraps. Verified against a local registry: pushing a manifest at bd44eb13 and retagging it produced e857a74f. Attesting the manifest therefore left the tag people actually pin unverifiable, and a static matrix could not fix it: resolving the live tags for one image returns five distinct digests, and which of them exist depends on whether the run is a release and whether the latest guard passed. A new attest-subjects job resolves the published tags to digests, de-duplicates, and feeds attest-images as a dynamic matrix. Tags this run did not publish are skipped rather than failing it. The NVIDIA device plugin was hardcoded in the template, so the previous guidance — mirror it to the same path, or patch the DaemonSet — was not durable: copying to another registry does not redirect the pull, and a patch is reverted by the next helm upgrade. It now takes its image from ollama.gpu.devicePlugin.image and goes through the shared helper, so it honors global.imageRegistry with useRegistryForAllImages like every other third-party image. Default renders byte-identical; chart minor bumped for the new key. Docs corrections from the same round, several of them regressions from my own previous commit: - The ADFS example set SSO_ISSUER to the IdP identifier, and SSO_SAML_AUDIENCE defaults to SSO_ISSUER — so it silently broke audience validation. The example now sets both. - Compose does not pass MIGRATION_DATABASE_URL through from .env: its migrations service declares an explicit environment list. - CLI authentication requires Redis; it does not fall back, and redis.mdx already said so. - The rollback snippet assigned SIM_VERSION in a subshell that Compose never saw. - Custom blocks hide their internals only when tracing is off. - A new workspace is still governed by the organization's default group. - doctor is not a Kubernetes tool, so it cannot be the first step for Helm users. - ENTERPRISE_ENABLED needs its NEXT_PUBLIC twin. - The 100 MB response cap does not cover MCP's standalone SSE stream. - Generating CRON_SECRET in a shell configures nothing by itself. Renaming two headings broke published anchors that an in-repo library article links to; both now point at the new slugs. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015BwsJTEQRzWJaY4BRCkPZt
Retention's cleanup dispatcher runs its chunks inline when Trigger.dev is not configured, so deletion completes on a default self-host. Only data drains have no inline path. The previous wording told operators both would stay pending. Docker Compose also schedules the drain endpoint hourly through its cron service, so 'outside Helm, schedule it yourself' pointed operators at a redundant scheduler. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015BwsJTEQRzWJaY4BRCkPZt
The resolver added `latest` unconditionally. When create-ghcr-manifests' guard holds those tags back — it does that when the branch has moved on — they still point at an older build, so the job would have put this run's signature and provenance on an image it did not produce. A moving alias is now included only when it resolves to the index this run published; the per-arch aliases ship in the same guarded block, so one comparison gates all three. `digest_of` also aborted the step instead of reporting an absent tag: under `set -e` a failing inspect inside a command substitution takes the step down before the caller can skip it. Confirmed against a real registry, then fixed with an explicit `if !` rather than swallowing the status. Docs, all verified against source: - The General tab does not buffer everything: workspace scope, the default switch, and member changes apply immediately, so Discard does not undo them. - Both data-drain screenshots were declared at the wrong aspect ratio when I converted them to <Image>; they rendered distorted. - LM Studio needs VLLM_API_KEY when its API authentication is on. - The rollback snippet hardcoded a version that does not exist. - Docker Compose schedules the drain endpoint too, so the table implied Compose operators had to add their own. - Generating CRON_SECRET in a shell configures neither the app nor the caller. - SANDBOXES_ENABLED is required beyond the provider and Function base on a billing-free self-host, and Pi runs on its own image and fails closed without E2B_PI_TEMPLATE_ID or DAYTONA_PI_SNAPSHOT_ID. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015BwsJTEQRzWJaY4BRCkPZt
…ipping it `digest_of` treated every inspect failure as "tag absent", so a transient registry error would drop a published digest from the matrix while both jobs still went green — an unsigned image that verification would later reject, with nothing in the run to show for it. It now retries, reports absent only when the registry says the manifest is unknown, and fails the step otherwise. The three sha tags are required outright, and an alias group that is current must resolve all three of its tags. SBOMs no longer attach to index subjects. Syft resolves an index to a single platform, so the SBOM described amd64 while the index also serves arm64. The per-arch subjects each carry an accurate SBOM; the index keeps its signature and provenance. Docs: - The member-directory restriction does not apply to organization owners and admins — the roster routes exempt those roles. - Python's execution does not hinge on NEXT_PUBLIC_SANDBOXES_ENABLED; that flag gates the controls, the server-side provider and image decide whether it runs. - The security checklist demanded sslMode: require on a Compose database that ships without TLS, so it could not be satisfied as written. - CERT_HAS_EXPIRED is usually an expired certificate, an incomplete chain, or a drifted clock. Sending operators straight to NODE_EXTRA_CA_CERTS fixes none of those, and the section also said the same thing twice. - Four endpoints, not "both". Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015BwsJTEQRzWJaY4BRCkPZt
… for a missing one digest_of accepted "manifest unknown" on the first attempt. GHCR can report a just-published alias as unknown for a moment, and taking that at face value would skip a tag this run did publish — leaving it unsigned while the workflow went green, which is the failure the retry existed to prevent. Absence is now retried like any other failure and only accepted when the registry still says so on the final attempt. Docs: - The LM Studio steps recreated services with docker-compose.ollama.yml, a development stack a production install does not run. They now use the file that started the installation. - Retention is not external work: its cleanup runs inline in the app process, and only data drains hand off to a background job. The endpoints start both, which is what the sentence should have said. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015BwsJTEQRzWJaY4BRCkPZt
… index `latest` is allowed to lag — the monotonic guard holds it back when the branch has moved on, and it then belongs to an older run that already attested it. A version tag has no such carve-out: this run published it, so a mismatch now fails rather than skipping, and a release can no longer ship a version image nothing has attested. The device-plugin mirror path was ambiguous. `sim.image` prepends the registry to the repository verbatim, so the plugin lands at `<registry>/nvcr.io/nvidia/k8s-device-plugin` — an operator who mirrored it without the embedded host would get ImagePullBackOff. The inventory now states the destination rule and offers the direct repository override for registries that cannot nest that path. Docs: - doctor needs no root .env from a source checkout; that guidance was Compose-specific. - The LM Studio commands hardcoded the production Compose file after the last round, which breaks an install started from the Ollama stack. They now take the file as a variable. - NODE_EXTRA_CA_CERTS does fix an incomplete chain by supplying the missing intermediate; only expiry and clock drift are beyond it. The section had ruled it out for all three. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015BwsJTEQRzWJaY4BRCkPZt
…t all The SBOM step could not have succeeded. Half the per-arch subjects are single-entry indexes — `imagetools create` writes an index even from one manifest — and Syft resolves an index against the runner's platform, so an arm64-only index fails outright on an amd64 runner with "no child with platform linux/arm64". That is 5 of 25 legs failing on every push to main, taking attest-images red and, through its gate, preventing any release from being created. Scanning by the `<sha>-<arch>` tag fixes it: that is the plain manifest the alias index wraps, so no platform resolution happens. It also halves the pulls, since both subjects for an architecture now share one scan of identical content. An empty subject list would also have opened the release gate — a matrix built from an empty include produces no legs and still reports success — so the gate now checks a published count rather than trusting a sibling job's guard. Two smaller corrections: `digest_of` no longer sleeps after its final attempt, which was costing 18 seconds per legitimately absent tag, and the alias comment now says what the code proves (content identity with this run's index) rather than implying provenance. The release gate's comment also now says what it gates. CodePipeline fires from the ECR tags that promote-images moves, upstream of attestation, and only the GHCR mirrors are signed — so this gates the GitHub release, not the deploy. Docs, each verified against source: - Restricted MCP, custom-tool and skill types fail the block; they are not silently skipped, and the same page already said so 180 lines later. - NO_PROXY needs every internal destination, not just model endpoints — the app calls realtime, Presidio, and itself over the same default fetch. - The proxy table omitted Google/Gemini and every other OpenAI-compatible provider, and omitted Azure Communication Services from the email row. - activeDeadlineSeconds caps a failing CronJob at five minutes, so it gets about one attempt rather than six retries. - Docker Compose does not schedule the retention endpoints either; the callout blamed Helm alone while the table beside it was already right. - Import-free JavaScript does leave the local VM when a block selects a sandbox or references a file path. - The GCS CORS list named headers that do not exist; the real key is sent server-side and never by the browser. - CRON_SECRET is optional, and nothing enforces the 32-character floor at runtime because env validation is skipped. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015BwsJTEQRzWJaY4BRCkPZt
Collaborator
Author
waleedlatif1
force-pushed
the
feat/supply-chain-attestations
branch
from
September 3, 2026 16:08
f7ab61d to
2dc5476
Compare
Collaborator
Author
|
@cubic-dev-ai review this PR |
Contributor
@waleedlatif1 I have started the AI code review. It will take a few minutes to complete. |
Contributor
There was a problem hiding this comment.
1 issue found and verified against the latest diff
Confidence score: 3/5
- In
scripts/generate-image-manifest.ts, the fallback for registries that cannot nestnvcr.iocan prepend the global registry twice, causing the GPU device-plugin image pull to fail; documentnvidia/k8s-device-pluginas the required override.
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="scripts/generate-image-manifest.ts">
<violation number="1" location="scripts/generate-image-manifest.ts:110">
P2: When an operator follows this fallback for a registry that cannot nest `nvcr.io`, the chart prepends the global registry again and the GPU device-plugin pull fails. Document `nvidia/k8s-device-plugin` as the override while global rewriting remains enabled, then regenerate `helm/sim/images.yaml`.</violation>
</file>
You've manually re-run cubic several times on this PR. Each manual re-review checks the full PR again and counts toward your usage quota. To preserve your usage limits, we recommend letting cubic automatically review new commits.
Fix all with cubic | Re-trigger cubic
Contributor
There was a problem hiding this comment.
All reported issues were addressed across 46 files
Requires human review: Auto-approval blocked because this review re-detected 2 unresolved issues already reported by Cubic.
Re-trigger cubic
Contributor
There was a problem hiding this comment.
4 issues found across 46 files
Confidence score: 3/5
apps/docs/content/docs/platform/self-hosting/networking.mdxcontains two operationally significant examples: the ingress snippet omitscopilot.enabled: true, and the timeout command assumes the default Helm release name, so copied guidance can leave Copilot unrouted or preserve GKE's 30-second timeout. Update the examples to show enablement and resolve the realtime Service by Helm labels.apps/docs/content/docs/platform/self-hosting/index.mdxinaccurately promises HTTP 422 when a custom sandbox lacks a configured provider; the route returns 503 before file handling, so revise the documented failure behavior.apps/docs/content/docs/platform/self-hosting/networking.mdxuses contradictory wording about which providers bypass the proxy, leaving operators uncertain about model-call routing; explicitly identify the listed providers as exceptions or examples.
Prompt for AI agents (unresolved issues)
Check if these issues are valid — if so, understand the root cause of each and fix them. If appropriate, use sub-agents to investigate and fix each issue separately.
<file name="apps/docs/content/docs/platform/self-hosting/index.mdx">
<violation number="1" location="apps/docs/content/docs/platform/self-hosting/index.mdx:116">
P2: When a user selects a custom sandbox without a configured provider, the function route returns HTTP 503 before file handling; this sentence incorrectly promises a 422 for that case. Say that sandbox selection fails with 503 and reserve 422 for file mounts or exports.</violation>
</file>
<file name="apps/docs/content/docs/platform/self-hosting/networking.mdx">
<violation number="1" location="apps/docs/content/docs/platform/self-hosting/networking.mdx:86">
P2: Copying this snippet without separately setting `copilot.enabled: true` renders no Copilot route, even though the example appears to configure one. State the required enablement alongside the ingress example.</violation>
<violation number="2" location="apps/docs/content/docs/platform/self-hosting/networking.mdx:228">
P2: When the chart is installed under any release name other than the documented default, this command annotates a nonexistent Service and GKE keeps the 30-second timeout. Resolve the realtime Service from its Helm labels or show the release-derived name instead of hard-coding `sim-realtime`.</violation>
<violation number="3" location="apps/docs/content/docs/platform/self-hosting/networking.mdx:310">
P2: The “except the three below” wording contradicts the provider list and the `Yes` result, so operators cannot tell which model calls use the proxy. Replace it with wording that identifies the listed providers as examples, or enumerate the actual exceptions explicitly.</violation>
</file>
Requires human review: Auto-approval blocked because this review re-detected 1 unresolved issue already reported by Cubic.
You've manually re-run cubic several times on this PR. Each manual re-review checks the full PR again and counts toward your usage quota. To preserve your usage limits, we recommend letting cubic automatically review new commits.
Re-trigger cubic
| Both `ingress` and `ingressInternal` accept an optional `copilot` block, commented out in `values.yaml` because the `copilot` service is off by default. It follows the same host-sharing rule: | ||
|
|
||
| ```yaml | ||
| ingress: |
Contributor
There was a problem hiding this comment.
P2: Copying this snippet without separately setting copilot.enabled: true renders no Copilot route, even though the example appears to configure one. State the required enablement alongside the ingress example.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At apps/docs/content/docs/platform/self-hosting/networking.mdx, line 86:
<comment>Copying this snippet without separately setting `copilot.enabled: true` renders no Copilot route, even though the example appears to configure one. State the required enablement alongside the ingress example.</comment>
<file context>
@@ -56,6 +55,42 @@ Both hostnames need DNS records and TLS certificates.
+Both `ingress` and `ingressInternal` accept an optional `copilot` block, commented out in `values.yaml` because the `copilot` service is off by default. It follows the same host-sharing rule:
+
+```yaml
+ingress:
+ copilot:
+ host: sim.yourdomain.com
</file context>
Suggested change
| ingress: | |
| copilot: | |
| enabled: true | |
| ingress: |
| annotations: | ||
| cloud.google.com/backend-config: '{"default": "sim-realtime-backendconfig"}' | ||
| ```bash | ||
| kubectl annotate service sim-realtime -n simstudio \ |
Contributor
There was a problem hiding this comment.
P2: When the chart is installed under any release name other than the documented default, this command annotates a nonexistent Service and GKE keeps the 30-second timeout. Resolve the realtime Service from its Helm labels or show the release-derived name instead of hard-coding sim-realtime.
Prompt for AI agents
Check if this issue is valid — if so, understand the root cause and fix it. At apps/docs/content/docs/platform/self-hosting/networking.mdx, line 228:
<comment>When the chart is installed under any release name other than the documented default, this command annotates a nonexistent Service and GKE keeps the 30-second timeout. Resolve the realtime Service from its Helm labels or show the release-derived name instead of hard-coding `sim-realtime`.</comment>
<file context>
@@ -187,24 +222,22 @@ spec:
- annotations:
- cloud.google.com/backend-config: '{"default": "sim-realtime-backendconfig"}'
+```bash
+kubectl annotate service sim-realtime -n simstudio \
+ cloud.google.com/backend-config='{"default": "sim-realtime-backendconfig"}'
</file context>
</details>
```suggestion
kubectl annotate service "$(kubectl get service -n simstudio -l app.kubernetes.io/component=realtime -o jsonpath='{.items[0].metadata.name}')" -n simstudio \
…e for it The mirror instruction was wrong for the five first-party images. The inventory lists what you pull — `ghcr.io/simstudioai/app` — but `ghcr.io` is the chart's default registry, so a redirected install pulls `<registry>/simstudioai/app`. Following "mirror to <registry>/<reference as listed>" put them at `<registry>/ghcr.io/simstudioai/...` and every Sim image failed to pull. The device plugin is the opposite case: its `nvcr.io/` really is part of the repository and survives the rewrite. Rather than write that rule out and hope it stays true, the generator now renders the chart a second time against a sentinel registry and strips it back off, so each entry carries both the source reference and the exact path the chart resolves to. Verified by rendering with a real mirror and diffing: all 13 mirror paths match. Also retries the moving-alias read. A mismatch meant "the guard held this tag back", but it can equally mean GHCR is still serving the previous digest a moment after this run wrote it — and that reading dropped three subjects from the matrix without a word. Docs, each verified against source: - Chat drops a disallowed tool while generating a workflow and records the omission; a run that reaches one fails. Both statements were needed, and the page previously asserted one or the other in different places. - A quickly-failing CronJob pod is retried six times by Kubernetes; the five-minute deadline only preempts one that hangs. - Selecting a sandbox without a provider is a 503; the 422 is the file path. - Chat's public code tool is run_function, not the internal function_execute. - The parity step needs SANDBOX_PROVIDER=e2b, or an exported daytona provider makes step 1 exit before writing the manifest. - PII_URL is auto-wired by the chart and needs a service URL on Compose. - NO_PROXY should use the cluster suffix; the Service names depend on the release name. - The bucket loops now fail loudly instead of silently iterating nothing when SIM_BUCKETS is unset in a fresh shell. - Generating CRON_SECRET beside the curl suggested it was the token to use. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_015BwsJTEQRzWJaY4BRCkPZt
Merged
7 tasks
Collaborator
Author
|
Closing in favour of #7437, which combines this with the other self-hosting PR against current staging and folds in a pre-review audit pass. Everything here is carried over — nothing dropped. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
create-ghcr-manifestsbecause buildx's own attestations break theimagetools createretagging the deploy gate depends on.helm/sim/images.yaml— the complete set of images a full install pulls, generated from the rendered chart and drift-checked in the helm workflow. Derived from the render rather thanvalues.yamlbecause one image (the NVIDIA device plugin) is written directly into a template and no values key names it.What was wrong
openssl rand -hex 32prints 64 hex characters, not the "32 hex chars" five entries claimed.API_ENCRYPTION_KEYwas missing from the whole Kubernetes path and marked required elsewhere. It is optional, and unset means API keys are stored in plain text behind one warning.NEXT_TELEMETRY_DISABLEDis cleared.Settings → Enterprise/Security → Xpath named nav groups that do not exist, across eleven pages.Also documents ~25 environment variables nothing described, the migration advisory lock and its operational knobs, and
sim-setup doctor. Moves the sandbox base-image runbook out of the enterprise enablement page into its own self-hosting page rather than losing it. Removes FAQ blocks that restated their own page body and four drifting copies of the same tables.Type of Change
Testing
bun run check:audits(45 audits),bun run lint,bun run docs-manifest:check,bun run images:check,actionlint, and the script suite (271 tests) all pass. Chart image list verified againsthelm template. Every action pinned by SHA and resolved against the GitHub API.Checklist
🤖 Generated with Claude Code
https://claude.ai/code/session_015BwsJTEQRzWJaY4BRCkPZt