Open-source AI teammates in Slack, Teams, Discord, Telegram and Mattermost. Self-host it, with any agent harness and any model.
Python agent-orchestrationaiai-agentschatopsdiscordlangchainmcpmicrosoft-teamsmulti-agentopen-sourceopenaiself-hostedslack
Run Strands Decider 4B as the chart's default decision model
agent-runner3536098450aee5f
diff --git a/.github/workflows/switch-release.yml b/.github/workflows/switch-release.ymlindex abb3597a0..98d9fb28b 100644--- a/.github/workflows/switch-release.yml+++ b/.github/workflows/switch-release.yml@@ -1,7 +1,8 @@name: switch release# Builds and publishes the Switch core stack — the switch-core, gateway, setup,-# switch-decision and switch-hosted-controller container images, the switch and+# switch-decision, switch-decision-strands and switch-hosted-controller+# container images, the switch and# switch-hosted-controller Helm charts, and the standalone Docker Compose file# (as an OCI artifact) — so adopters can deploy without building from source.# See RELEASING.md for the full procedure.@@ -268,9 +269,20 @@ jobs:PYgrep -m1 '^version = ' core/pyproject.toml+ - name: Free disk space+ # A large image (the Strands Decider model server: CUDA and PyTorch,+ # several GB) does not fit in what the runner leaves free.+ if: ${{ matrix.large }}+ run: |+ df -h /+ sudo rm -rf /usr/share/dotnet /usr/local/lib/android /opt/ghc /opt/hostedtoolcache/CodeQL+ sudo docker image prune --all --force+ df -h /+- name: Set up QEMU# Enables arm64 emulation so buildx can produce a multi-arch manifest# from an amd64 runner.+ if: ${{ contains(matrix.platforms, 'linux/arm64') }}uses: docker/setup-qemu-action@c7c53464625b32c7a7e944ae62b3e17d2b600130 # v3.7.0- name: Set up Docker Buildx@@ -289,7 +301,9 @@ jobs:with:context: ${{ matrix.context }}file: ${{ matrix.dockerfile }}- platforms: linux/amd64,linux/arm64+ # From scripts/pin_chart_images.py: both, or amd64 alone for the+ # CUDA model server.+ platforms: ${{ matrix.platforms }}push: ${{ needs.prep.outputs.push == 'true' }}# Read by Dockerfile.switch only; the other images ignore them.build-args: |@@ -298,8 +312,10 @@ jobs:tags: |${{ env.REGISTRY }}/${{ needs.prep.outputs.namespace }}/${{ matrix.image }}:${{ needs.prep.outputs.version }}${{ needs.prep.outputs.latest == 'true' && format('{0}/{1}/{2}:latest', env.REGISTRY, needs.prep.outputs.namespace, matrix.image) || '' }}- cache-from: type=gha,scope=${{ matrix.image }}- cache-to: type=gha,scope=${{ matrix.image }},mode=max+ # A large image stays out of the GitHub Actions cache: its layers alone+ # would fill the repository's 10 GB and evict every other image's.+ cache-from: ${{ !matrix.large && format('type=gha,scope={0}', matrix.image) || '' }}+ cache-to: ${{ !matrix.large && format('type=gha,scope={0},mode=max', matrix.image) || '' }}# The digest of the multi-arch index, which is what a pod pulls by. The# chart job writes it into values.yaml, so it has to leave this job.diff --git a/RELEASING.md b/RELEASING.mdindex f8b449517..ca5e4b0c1 100644--- a/RELEASING.md+++ b/RELEASING.md@@ -1,8 +1,8 @@# Releasing SwitchThis describes how to cut a release of the **Switch core stack** — the-`switch-core`, `gateway`, `setup`, `switch-decision` and-`switch-hosted-controller` container images, the Helm charts, and the standalone Docker Compose file. The Switch Console desktop app releases separately+`switch-core`, `gateway`, `setup`, `switch-decision`,+`switch-decision-strands` and `switch-hosted-controller` container images, the Helm charts, and the standalone Docker Compose file. The Switch Console desktop app releases separately(see `.github/workflows/switch-console-release.yml` and `console/docs/INSTALL.md`).## Versioning@@ -83,7 +83,9 @@ peer still on it, and it can never be raised past what is running in the field.5. The **`switch-release`** workflow (`.github/workflows/switch-release.yml`)then, on that tag:- builds multi-arch (`linux/amd64`, `linux/arm64`) images for `switch-core`,- `gateway`, `setup` and `switch-decision` (and the hosted controller) and+ `gateway`, `setup` and `switch-decision` (and the hosted controller), and+ an amd64-only `switch-decision-strands` (the Strands Decider model server,+ CUDA and PyTorch, several GB; built without the Actions layer cache), andpushes them to the container registry tagged`<version>` and `latest`;- builds, tests and packs the agent controller at the same version and@@ -134,7 +136,7 @@ environment runs main from published, digest-pinned artifacts instead ofbuilding from source at deploy time.- **Separate packages.** Dev builds go to `ghcr.io/<owner>/dev/` (`dev/switch-core`,- `dev/gateway`, `dev/setup`, `dev/switch-decision`, `dev/switch-hosted-controller`, `dev/charts/switch`,+ `dev/gateway`, `dev/setup`, `dev/switch-decision`, `dev/switch-decision-strands`, `dev/switch-hosted-controller`, `dev/charts/switch`,`dev/charts/switch-hosted-controller`), so the release packages' tag listshold releases and RCs only. Nothing is ever tagged `latest`.- **The dev charts carry a moving `main` tag.** Once both charts are pushed,@@ -180,7 +182,9 @@ switchCore:and the same for `gateway`, `setup` and `decision` (the `switch-decision`image, rendered only with `decision.enabled`; the CI render with the optional-workloads on checks its pin). The hosted controller chart+workloads on checks its pin), and for `decision.modelServer.strands.image`+(`switch-decision-strands`, rendered with `decision.enabled` and the default+`decision.modelServer.kind: strands`). The hosted controller chart(`charts/switch-hosted-controller`) is pinned through `image.repository` and`image.digest` instead, so its image renders as`ghcr.io/<owner>/switch-hosted-controller@sha256:…`, by digest with no tag.diff --git a/artifacts.yaml b/artifacts.yamlindex d5ec83842..bb0e42581 100644--- a/artifacts.yaml+++ b/artifacts.yaml@@ -122,6 +122,14 @@ artifacts:version_from: switch-corepublished_as: switch-decision+ switch-decision-strands:+ description: >-+ The model server for Strands Decider, the decision service's default+ model in the Helm chart. NVIDIA GPU, amd64 only; its weights download on+ first start, never baked in.+ version_from: switch-core+ published_as: switch-decision-strands+helm-chart:description: >-The Kubernetes chart. Its values.yaml is a consumer-facing interface —diff --git a/console/packages/shared/src/artifacts.ts b/console/packages/shared/src/artifacts.tsindex e13f7266d..183a741eb 100644--- a/console/packages/shared/src/artifacts.ts+++ b/console/packages/shared/src/artifacts.ts@@ -27,6 +27,7 @@ export const ARTIFACT_VERSIONS = {gateway: '0.31.0',setup: '0.31.0','switch-decision': '0.31.0',+ 'switch-decision-strands': '0.31.0','helm-chart': '0.31.0','hosted-controller': '0.31.0','hosted-controller-chart': '0.31.0',diff --git a/console/packages/switch-agent-runtime/src/artifacts.ts b/console/packages/switch-agent-runtime/src/artifacts.tsindex e13f7266d..183a741eb 100644--- a/console/packages/switch-agent-runtime/src/artifacts.ts+++ b/console/packages/switch-agent-runtime/src/artifacts.ts@@ -27,6 +27,7 @@ export const ARTIFACT_VERSIONS = {gateway: '0.31.0',setup: '0.31.0','switch-decision': '0.31.0',+ 'switch-decision-strands': '0.31.0','helm-chart': '0.31.0','hosted-controller': '0.31.0','hosted-controller-chart': '0.31.0',diff --git a/core/switch_core/artifacts.py b/core/switch_core/artifacts.pyindex 2b1f6fc3a..483dcbd78 100644--- a/core/switch_core/artifacts.py+++ b/core/switch_core/artifacts.py@@ -27,6 +27,7 @@ ARTIFACT_VERSIONS: Final[dict[str, str]] = {"gateway": "0.31.0","setup": "0.31.0","switch-decision": "0.31.0",+ "switch-decision-strands": "0.31.0","helm-chart": "0.31.0","hosted-controller": "0.31.0","hosted-controller-chart": "0.31.0",diff --git a/core/tests/test_pin_chart_images.py b/core/tests/test_pin_chart_images.pyindex e87b156cc..31b5e1dd7 100644--- a/core/tests/test_pin_chart_images.py+++ b/core/tests/test_pin_chart_images.py@@ -27,9 +27,16 @@ DIGESTS = {"gateway": "sha256:" + "b" * 64,"setup": "sha256:" + "c" * 64,"switch-decision": "sha256:" + "e" * 64,+ "switch-decision-strands": "sha256:" + "9" * 64,"switch-hosted-controller": "sha256:" + "f" * 64,}-SWITCH_IMAGES = ("switch-core", "gateway", "setup", "switch-decision")+SWITCH_IMAGES = (+ "switch-core",+ "gateway",+ "setup",+ "switch-decision",+ "switch-decision-strands",+)def _load(name: str) -> ModuleType:@@ -52,7 +59,11 @@ def test_the_image_table_matches_the_dockerfiles_and_values() -> None:values = yaml.safe_load((REPO_ROOT / chart.path / "values.yaml").read_text())assert (REPO_ROOT / image.dockerfile).is_file(), nameassert (REPO_ROOT / image.context).is_dir(), name- assert image.values_key in values, name+ node = values+ for key in image.values_key.split("."):+ assert key in node, (name, image.values_key)+ node = node[key]+ assert "image" in node or "repository" in node, nameassert set(pin_chart_images.images_of("switch")) == set(SWITCH_IMAGES)@@ -113,6 +124,13 @@ def test_pins_every_first_party_image_in_the_real_values_file() -> None:pinned["decision"]["image"]== f"switch-decision:1.2.3@{DIGESTS['switch-decision']}")+ # Nested: the model server's image, not llama.cpp's beside it.+ model_server = pinned["decision"]["modelServer"]+ assert (+ model_server["strands"]["image"]+ == f"switch-decision-strands:1.2.3@{DIGESTS['switch-decision-strands']}"+ )+ assert model_server["llamacpp"]["image"] == "ghcr.io/ggml-org/llama.cpp:server-cuda"def test_changes_only_the_pinned_lines() -> None:@@ -124,7 +142,7 @@ def test_changes_only_the_pinned_lines() -> None:for a, b in zip(original.splitlines(), pinned.splitlines(), strict=True)if a != b]- assert len(changed) == 5 # the registry, then one image line per image+ assert len(changed) == 6 # the registry, then one image line per image# Pull policy is left to the chart, which derives it from the digest.assert yaml.safe_load(pinned)["switchCore"]["imagePullPolicy"] == ""@@ -207,6 +225,45 @@ def _render(*extra: str) -> str:)+def test_pins_a_nested_key_only_inside_its_own_block() -> None:+ values = (+ "a:\n"+ " image: top\n"+ " b:\n"+ " image: b\n"+ " c:\n"+ " # a comment\n"+ " image: old\n"+ "\n"+ " other:\n"+ " image: other\n"+ " c:\n"+ " image: not-this-c\n"+ )+ lines = values.splitlines(keepends=True)+ pin_chart_images._set_child(lines, "a.b.c", "image", '"new"')+ assert yaml.safe_load("".join(lines)) == {+ "a": {+ "image": "top",+ "b": {"image": "b", "c": {"image": "new"}, "other": {"image": "other"}},+ "c": {"image": "not-this-c"},+ }+ }+ with pytest.raises(pin_chart_images.PinError, match="no key 'd'"):+ pin_chart_images._set_child(lines, "a.d", "image", "x")+ with pytest.raises(pin_chart_images.PinError, match="a.b.other.tag"):+ pin_chart_images._set_child(lines, "a.b.other", "tag", "x")+++def test_the_strands_model_server_is_built_for_amd64_only() -> None:+ matrix = {+ image.dockerfile: (image.platforms, image.large)+ for image in pin_chart_images.IMAGES.values()+ }+ assert matrix["decision/model-servers/strands/Dockerfile"] == ("linux/amd64", True)+ assert matrix["decision/Dockerfile"] == ("linux/amd64,linux/arm64", False)++def test_verify_accepts_a_fully_pinned_render() -> None:assert len(pin_chart_images.verify("switch", _render(), "ghcr.io/acme")) == 3@@ -252,7 +309,15 @@ def test_verify_accepts_a_render_without_an_optional_image_but_covered_does_not(with_decision = pin_chart_images.verify("switch", _render(_pinned("switch-decision")), "ghcr.io/acme")- pin_chart_images.covered("switch", "\n".join([*refs, *with_decision]))+ with pytest.raises(pin_chart_images.PinError, match="switch-decision-strands"):+ pin_chart_images.covered("switch", "\n".join([*refs, *with_decision]))++ with_model_server = pin_chart_images.verify(+ "switch",+ _render(_pinned("switch-decision"), _pinned("switch-decision-strands")),+ "ghcr.io/acme",+ )+ pin_chart_images.covered("switch", "\n".join([*refs, *with_model_server]))def test_record_lists_every_image_and_chart() -> None:diff --git a/deploy/remote/helm/render-check-optional-values.yaml b/deploy/remote/helm/render-check-optional-values.yamlindex 8e303138e..f287ef594 100644--- a/deploy/remote/helm/render-check-optional-values.yaml+++ b/deploy/remote/helm/render-check-optional-values.yaml@@ -24,6 +24,5 @@ decision:bedrock:enabled: truemodelServer:- model:- hfTokenSecret:- name: render-check-hf-token+ hfTokenSecret:+ name: render-check-hf-tokendiff --git a/deploy/remote/helm/switch/README.md b/deploy/remote/helm/switch/README.mdindex 0939b99d4..c155707b2 100644--- a/deploy/remote/helm/switch/README.md+++ b/deploy/remote/helm/switch/README.md@@ -291,16 +291,41 @@ pods to port 8230. It sets `DECISION_ENDPOINT_URL` on switch-core to theService, so every workspace's Decision routing settings in Switch Console showit as the server default with no further setup.-- **GPU.** The default model (Qwen3.5-4B, 4-bit, on llama.cpp's CUDA server)- requests `nvidia.com/gpu: 1`, so the cluster needs the NVIDIA device plugin- and a GPU node. The pod tolerates the usual `nvidia.com/gpu` taint; set- `decision.nodeSelector` (and `decision.tolerations`) for your nodes. One- NVIDIA L4 answers in well under a second.-- **The model.** `decision.modelServer.image`, `.args` and `.model` (a download- URL, the file name, an optional SHA-256, and an optional Secret with a- Hugging Face token for a gated model) are all values, so another server with- the same API can be swapped in. `decision.model` is the name the service- reports. The image is pinned like the chart's other first-party images.+- **The model.** `decision.modelServer.kind` picks the model server:+ - `strands` (default): **Strands Decider 4B** (the E4B checkpoint) on its own+ server image, `switch-decision-strands`, which the release builds and pins+ like the chart's other first-party images. On one NVIDIA L4 it scored+ 92/100 on our routing scenarios with 2 wrong picks, at 0.27 s median.+ - `llamacpp`: **Qwen3.5-4B** (4-bit) on llama.cpp's CUDA server: 88/100 with+ 9 wrong picks, at 0.16 s median, in a smaller footprint (8 GiB of RAM).++ The decision service's backend and model name follow the kind+ (`systemone` and `strands-decider-E4B`, or `llamacpp` and+ `qwen3.5-4b-q4_k_m`); `decision.backend` and `decision.model` override them.+ Each kind's image, arguments, probes and resources are under+ `decision.modelServer.strands` or `decision.modelServer.llamacpp`.+- **GPU and memory.** Either model server requests `nvidia.com/gpu: 1`, so the+ cluster needs the NVIDIA device plugin and a GPU node. The pod tolerates the+ usual `nvidia.com/gpu` taint; set `decision.nodeSelector` (and+ `decision.tolerations`) for your nodes. Strands Decider 4B also requests+ 24Gi of memory to load its 16 GB of bf16 weights, so give it a node with+ 32 GiB (an AWS g6.2xlarge, one L4). The image needs an NVIDIA driver for+ CUDA 13.0 (580 or later).+- **Weights and the first start.** Nothing is baked into an image. Strands+ Decider downloads its checkpoint and base model (about 17 GB) from Hugging+ Face into the PVC (`decision.persistence.size`, 40Gi) on its first start,+ then loads them; the port opens only once the model is loaded. Its startup+ probe allows 30 minutes (`decision.modelServer.strands.startupProbe`), and+ the decision service reports not ready until then. Later starts only load+ the weights, a few minutes. For Qwen, an init container downloads the GGUF+ file (`decision.modelServer.llamacpp.model`: URL, file name, optional+ SHA-256) first. `decision.modelServer.hfTokenSecret` names a Secret with a+ Hugging Face token, for a gated checkpoint or higher Hub rate limits.+- **Licence.** Strands Decider is Apache-2.0, but its 4B checkpoint is built on+ Google's Gemma 4 (`google/gemma-4-E4B-it`), whose terms apply to anyone+ deploying it. The chart downloads the weights at deploy time from Hugging+ Face; Switch does not redistribute them. Choose `llamacpp` if those terms+ don't suit you.- **Haiku on Bedrock.** `decision.bedrock.enabled` lets workspaces choose ClaudeHaiku on Amazon Bedrock (`DECISION_BEDROCK_ENABLED` on the service and onswitch-core). The service calls Bedrock with the pod's credentials: give itdiff --git a/deploy/remote/helm/switch/templates/_helpers.tpl b/deploy/remote/helm/switch/templates/_helpers.tplindex 975bc2b0f..ff9828919 100644--- a/deploy/remote/helm/switch/templates/_helpers.tpl+++ b/deploy/remote/helm/switch/templates/_helpers.tpl@@ -1195,12 +1195,66 @@ Usage: {{ include "switch.waitForHttp" (dict "name" "gateway" "url" "http://host{{- end }}{{/*-The weights file name inside /models: decision.modelServer.model.file.+The weights file name inside /models: decision.modelServer.llamacpp.model.file.*/}}{{- define "switch.decisionModelFile" -}}-{{- $file := required "decision.modelServer.model.file is required" .Values.decision.modelServer.model.file -}}+{{- $file := required "decision.modelServer.llamacpp.model.file is required" .Values.decision.modelServer.llamacpp.model.file -}}{{- if contains "/" $file -}}-{{- fail (printf "decision.modelServer.model.file must be a file name, not a path: %q" $file) -}}+{{- fail (printf "decision.modelServer.llamacpp.model.file must be a file name, not a path: %q" $file) -}}{{- end -}}{{- $file -}}{{- end }}++{{/*+Which model server runs beside the decision service: decision.modelServer.kind,+strands or llamacpp.+*/}}+{{- define "switch.decisionModelServerKind" -}}+{{- $kind := .Values.decision.modelServer.kind -}}+{{- if not (has $kind (list "strands" "llamacpp")) -}}+{{- fail (printf "decision.modelServer.kind must be strands or llamacpp, not %q" $kind) -}}+{{- end -}}+{{- $kind -}}+{{- end }}++{{/*+DECISION_BACKEND: decision.backend, or else the API the model server speaks+(systemone for strands, llamacpp for llamacpp). Without a model server there is+nothing to derive it from, so it must be set (anthropic or bedrock).+*/}}+{{- define "switch.decisionBackend" -}}+{{- $d := .Values.decision -}}+{{- if $d.backend -}}+{{- $d.backend -}}+{{- else if $d.modelServer.enabled -}}+{{- ternary "systemone" "llamacpp" (eq (include "switch.decisionModelServerKind" .) "strands") -}}+{{- else -}}+{{- fail "decision.backend is required when decision.modelServer.enabled is false: anthropic or bedrock, say." -}}+{{- end -}}+{{- end }}++{{/*+DECISION_MODEL: decision.model, or else the model server's modelName. Empty+(without a model server) keeps the service's default for its backend.+*/}}+{{- define "switch.decisionModel" -}}+{{- $d := .Values.decision -}}+{{- if $d.model -}}+{{- $d.model -}}+{{- else if $d.modelServer.enabled -}}+{{- (get $d.modelServer (include "switch.decisionModelServerKind" .)).modelName -}}+{{- end -}}+{{- end }}++{{/*+HF_TOKEN env entry from decision.modelServer.hfTokenSecret, or nothing.+*/}}+{{- define "switch.decisionHfToken" -}}+{{- with .Values.decision.modelServer.hfTokenSecret.name }}+- name: HF_TOKEN+ valueFrom:+ secretKeyRef:+ name: {{ . }}+ key: {{ $.Values.decision.modelServer.hfTokenSecret.key }}+{{- end }}+{{- end }}diff --git a/deploy/remote/helm/switch/templates/decision/deployment.yaml b/deploy/remote/helm/switch/templates/decision/deployment.yamlindex c41b980c7..48de7090b 100644--- a/deploy/remote/helm/switch/templates/decision/deployment.yaml+++ b/deploy/remote/helm/switch/templates/decision/deployment.yaml@@ -1,5 +1,11 @@{{- if .Values.decision.enabled }}{{- $d := .Values.decision }}+{{- $ms := $d.modelServer }}+{{- $kind := "" }}+{{- if $ms.enabled }}+{{- $kind = include "switch.decisionModelServerKind" . }}+{{- end }}+{{- $backend := include "switch.decisionBackend" . }}apiVersion: apps/v1kind: Deploymentmetadata:@@ -27,7 +33,8 @@ spec:{{- end }}serviceAccountName: {{ include "switch.decisionHost" . }}securityContext:- # The decision image's user; the downloader writes the weights as it.+ # The decision images' user; the model server or the downloader writes+ # the weights as it.fsGroup: 10001fsGroupChangePolicy: OnRootMismatchseccompProfile:@@ -40,12 +47,13 @@ spec:tolerations:{{- toYaml . | nindent 8 }}{{- end }}- {{- if $d.modelServer.enabled }}+ {{- if eq $kind "llamacpp" }}+ {{- $llama := $ms.llamacpp }}initContainers:# Downloads the weights into the volume unless they are already there,# resuming a download a restart interrupted.- name: download-model- image: {{ $d.modelServer.model.downloaderImage }}+ image: {{ $llama.model.downloaderImage }}command: ["sh", "-c"]args:- |@@ -67,17 +75,13 @@ spec:fienv:- name: MODEL_URL- value: {{ required "decision.modelServer.model.url is required" $d.modelServer.model.url | quote }}+ value: {{ required "decision.modelServer.llamacpp.model.url is required" $llama.model.url | quote }}- name: MODEL_FILEvalue: {{ include "switch.decisionModelFile" . | quote }}- name: MODEL_SHA256- value: {{ $d.modelServer.model.sha256 | quote }}- {{- with $d.modelServer.model.hfTokenSecret.name }}- - name: HF_TOKEN- valueFrom:- secretKeyRef:- name: {{ . }}- key: {{ $d.modelServer.model.hfTokenSecret.key }}+ value: {{ $llama.model.sha256 | quote }}+ {{- with include "switch.decisionHfToken" . }}+ {{- . | trim | nindent 12 }}{{- end }}securityContext:runAsNonRoot: true@@ -109,14 +113,17 @@ spec:- name: PORTvalue: {{ $d.port | quote }}- name: DECISION_BACKEND- value: {{ $d.backend | quote }}- {{- with $d.model }}+ value: {{ $backend | quote }}+ {{- with include "switch.decisionModel" . }}- name: DECISION_MODELvalue: {{ . | quote }}{{- end }}- {{- if $d.modelServer.enabled }}+ {{- if eq $kind "strands" }}+ - name: SYSTEMONE_URL+ value: "http://127.0.0.1:{{ $ms.strands.port }}/v1/systemone"+ {{- else if eq $kind "llamacpp" }}- name: LLAMACPP_URL- value: "http://127.0.0.1:{{ $d.modelServer.port }}"+ value: "http://127.0.0.1:{{ $ms.llamacpp.port }}"{{- end }}{{- with $d.confidenceThreshold }}- name: DECISION_CONFIDENCE_THRESHOLD@@ -161,11 +168,65 @@ spec:volumeMounts:- name: tmpmountPath: /tmp- {{- if $d.modelServer.enabled }}+ {{- if eq $kind "strands" }}+ {{- $st := $ms.strands }}+ - name: model-server+ image: {{ include "switch.image" (dict "global" .Values.global "image" $st.image) }}+ imagePullPolicy: {{ include "switch.imagePullPolicy" (dict "policy" $st.imagePullPolicy "image" $st.image) }}+ {{- with $st.args }}+ # Appended to the image's `strands-decider serve` command.+ args:+ {{- range . }}+ - {{ . | quote }}+ {{- end }}+ {{- end }}+ env:+ - name: STRANDS_CHECKPOINT+ value: {{ required "decision.modelServer.strands.checkpoint is required" $st.checkpoint | quote }}+ - name: STRANDS_MODEL_NAME+ value: {{ required "decision.modelServer.strands.modelName is required" $st.modelName | quote }}+ {{- with include "switch.decisionHfToken" . }}+ {{- . | trim | nindent 12 }}+ {{- end }}+ {{- with $st.env }}+ {{- toYaml . | nindent 12 }}+ {{- end }}+ ports:+ - name: model+ containerPort: {{ $st.port }}+ protocol: TCP+ # The port opens only once the checkpoint is downloaded (first start)+ # and loaded, which takes minutes.+ startupProbe:+ httpGet:+ path: /health+ port: model+ periodSeconds: {{ $st.startupProbe.periodSeconds }}+ failureThreshold: {{ $st.startupProbe.failureThreshold }}+ livenessProbe:+ httpGet:+ path: /health+ port: model+ periodSeconds: 20+ timeoutSeconds: 5+ failureThreshold: 3+ securityContext:+ runAsNonRoot: true+ allowPrivilegeEscalation: false+ capabilities:+ drop: ["ALL"]+ resources:+ {{- toYaml $st.resources | nindent 12 }}+ # Hugging Face's cache (HF_HOME): written on the first start.+ volumeMounts:+ - name: weights+ mountPath: /models+ {{- else if eq $kind "llamacpp" }}+ {{- $llama := $ms.llamacpp }}- name: model-server- image: {{ $d.modelServer.image }}- imagePullPolicy: {{ $d.modelServer.imagePullPolicy }}- {{- with $d.modelServer.args }}+ image: {{ $llama.image }}+ imagePullPolicy: {{ $llama.imagePullPolicy }}+ {{- with $llama.args }}args:{{- range . }}- {{ . | quote }}@@ -176,15 +237,15 @@ spec:value: "/models/{{ include "switch.decisionModelFile" . }}"ports:- name: model- containerPort: {{ $d.modelServer.port }}+ containerPort: {{ $llama.port }}protocol: TCP# llama.cpp's /health is 503 until the model is loaded.startupProbe:httpGet:path: /healthport: model- periodSeconds: 5- failureThreshold: 120+ periodSeconds: {{ $llama.startupProbe.periodSeconds }}+ failureThreshold: {{ $llama.startupProbe.failureThreshold }}livenessProbe:httpGet:path: /health@@ -196,7 +257,7 @@ spec:capabilities:drop: ["ALL"]resources:- {{- toYaml $d.modelServer.resources | nindent 12 }}+ {{- toYaml $llama.resources | nindent 12 }}volumeMounts:- name: weightsmountPath: /models@@ -205,7 +266,7 @@ spec:volumes:- name: tmpemptyDir: {}- {{- if $d.modelServer.enabled }}+ {{- if $ms.enabled }}- name: weightspersistentVolumeClaim:claimName: {{ include "switch.decisionHost" . }}-weightsdiff --git a/deploy/remote/helm/switch/values.yaml b/deploy/remote/helm/switch/values.yamlindex 5eae1c238..1706bed5a 100644--- a/deploy/remote/helm/switch/values.yaml+++ b/deploy/remote/helm/switch/values.yaml@@ -865,14 +865,16 @@ mattermost:## decision the switch-decision image: POST /decide and GET /health on# `port`. CPU only. The Service exposes this port and no other.-# model-server the self-hosted decision model (llama.cpp's server by-# default), reached by the decision container on localhost. Its-# weights are downloaded at first start into a PVC, never baked-# into an image.+# model-server the self-hosted decision model (modelServer.kind: Strands+# Decider 4B by default, or Qwen3.5-4B on llama.cpp), reached+# by the decision container on localhost. Its weights are+# downloaded at first start into a PVC, never baked into an+# image.#-# The default model needs one NVIDIA GPU (an L4 answers in well under a-# second): the model server requests nvidia.com/gpu, so the cluster needs the-# NVIDIA device plugin and a GPU node; point nodeSelector/tolerations at it.+# The model server needs one NVIDIA GPU (an L4 answers in well under a+# second): it requests nvidia.com/gpu, so the cluster needs the NVIDIA device+# plugin and a GPU node; point nodeSelector/tolerations at it. Strands Decider+# 4B also requests 24Gi of memory, so the node needs 32 GiB (a g6.2xlarge).# Turn on the "decision_routing" and "follow_the_conversation" feature flags# (switchCore.featureFlags.enabled) for auto-route to be offered in rooms.decision:@@ -882,13 +884,14 @@ decision:# chart), Always otherwise. Set it to override.imagePullPolicy: ""port: 8230- # DECISION_BACKEND: llamacpp reads the model server's probabilities over the- # candidate agents. systemone, anthropic and bedrock are the others; with- # anthropic or bedrock, turn modelServer.enabled off and pass their settings- # in `env`.- backend: llamacpp+ # DECISION_BACKEND. Empty: the API the model server speaks (systemone for+ # modelServer.kind strands, llamacpp for llamacpp). To run without a GPU,+ # turn modelServer.enabled off and set anthropic or bedrock (their settings+ # in `env`).+ backend: ""# DECISION_MODEL, the name /health reports and the decision log records.- model: qwen3.5-4b-q4_k_m+ # Empty: the model server's modelName.+ model: ""# DECISION_CONFIDENCE_THRESHOLD; empty keeps the service's default (0.6).confidenceThreshold: ""# Claude Haiku on Amazon Bedrock as a per-workspace choice. Sets@@ -910,54 +913,99 @@ decision:limits:memory: "1Gi"cpu: "1000m"- # The model server, a sidecar in the decision pod. Any server with llama.cpp's- # OpenAI-compatible API and GET /health works; swap image and args for- # another (a SystemOne server for Strands Decider or Clef, say, with- # backend: systemone and SYSTEMONE_URL in env).+ # The model server, a sidecar in the decision pod on one GPU, reached on+ # localhost. Its weights live in the PVC below, downloaded on the first start.modelServer:enabled: true- image: ghcr.io/ggml-org/llama.cpp:server-cuda- imagePullPolicy: IfNotPresent- port: 8300- # $(MODEL_PATH) is /models/<model.file>, expanded by Kubernetes.- args:- - -m- - $(MODEL_PATH)- - --host- - 0.0.0.0- - --port- - "8300"- - -ngl- - "99"- - -c- - "16384"- - -np- - "4"- - --jinja- # The weights, downloaded into the PVC by an init container when the file- # is not there yet. Changing them downloads the new file on the next start.- model:- url: https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/resolve/main/Qwen3.5-4B-Q4_K_M.gguf- file: Qwen3.5-4B-Q4_K_M.gguf- # Optional hex SHA-256 of the file, checked after the download.- sha256: ""- # A Secret holding a Hugging Face token, for a gated model (one whose- # licence the token's account has accepted). Empty name: no token.- hfTokenSecret:- name: ""- key: HF_TOKEN- downloaderImage: curlimages/curl:8.11.1- resources:- requests:- memory: "8Gi"- cpu: "2"- nvidia.com/gpu: 1- limits:- memory: "14Gi"- nvidia.com/gpu: 1+ # strands: Strands Decider 4B (E4B), the default. 92/100 on our routing+ # scenarios with 2 wrong picks, 0.27 s median on one NVIDIA L4. Needs a+ # 24 GB GPU and about 24 GiB of RAM to load (a g6.2xlarge, say).+ # llamacpp: Qwen3.5-4B in 4-bit on the llama.cpp server. 88/100 with 9+ # wrong picks, 0.16 s; a smaller footprint (8 GiB of RAM).+ kind: strands+ # A Secret holding a Hugging Face token (key below), passed to whichever+ # model server downloads: for a gated checkpoint (one whose licence the+ # token's account has accepted) or higher Hub rate limits. Empty name: none.+ hfTokenSecret:+ name: ""+ key: HF_TOKEN+ strands:+ # The release pins this image (switch-decision-strands) by digest.+ image: switch-decision-strands:latest+ # Empty: IfNotPresent when pinned by digest, Always otherwise.+ imagePullPolicy: ""+ port: 8400+ # The Hugging Face checkpoint, downloaded into the PVC (HF_HOME) on the+ # first start with its base model (google/gemma-4-E4B-it, about 16 GB,+ # under the Gemma terms). Later starts only load it.+ checkpoint: StrandsAgents/strands-decider-E4B-gemma4-v1-2610+ # --model-name, and DECISION_MODEL unless `model` above is set.+ modelName: strands-decider-E4B+ # Extra arguments for `strands-decider serve` (e.g. --max-batch 8).+ args: []+ # Extra container env, as Kubernetes env entries.+ env: []+ # The port opens only once the model is loaded: the first start+ # downloads about 16 GB and then loads it (a few minutes with 32 GiB of+ # RAM). 10 s x 180 allows 30 minutes before the container is restarted.+ startupProbe:+ periodSeconds: 10+ failureThreshold: 180+ resources:+ requests:+ memory: "24Gi"+ cpu: "2"+ nvidia.com/gpu: 1+ limits:+ memory: "28Gi"+ nvidia.com/gpu: 1+ llamacpp:+ image: ghcr.io/ggml-org/llama.cpp:server-cuda+ imagePullPolicy: IfNotPresent+ port: 8300+ # DECISION_MODEL unless `model` above is set.+ modelName: qwen3.5-4b-q4_k_m+ # $(MODEL_PATH) is /models/<model.file>, expanded by Kubernetes.+ args:+ - -m+ - $(MODEL_PATH)+ - --host+ - 0.0.0.0+ - --port+ - "8300"+ - -ngl+ - "99"+ - -c+ - "16384"+ - -np+ - "4"+ - --jinja+ # The weights, downloaded into the PVC by an init container when the+ # file is not there yet. Changing them downloads the new file on the+ # next start.+ model:+ url: https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/resolve/main/Qwen3.5-4B-Q4_K_M.gguf+ file: Qwen3.5-4B-Q4_K_M.gguf+ # Optional hex SHA-256 of the file, checked after the download.+ sha256: ""+ downloaderImage: curlimages/curl:8.11.1+ # llama.cpp's /health is 503 until the model is loaded (seconds).+ startupProbe:+ periodSeconds: 5+ failureThreshold: 120+ resources:+ requests:+ memory: "8Gi"+ cpu: "2"+ nvidia.com/gpu: 1+ limits:+ memory: "14Gi"+ nvidia.com/gpu: 1# The weights volume. Empty storageClass uses global.storageClass.+ # 40Gi holds Strands Decider's Hugging Face cache (about 17 GB) with room+ # for a second checkpoint; Qwen's file is 2.7 GB.persistence:- size: 20Gi+ size: 40GistorageClass: ""nodeSelector: {}tolerations:diff --git a/scripts/check_chart_decision.py b/scripts/check_chart_decision.pyindex ecce9a79f..3c378cf2c 100755--- a/scripts/check_chart_decision.py+++ b/scripts/check_chart_decision.py@@ -2,7 +2,8 @@"""Render the switch chart with its decision service off and on, and check it.The decision service (auto-route) is optional in the chart. Off, nothing of it-renders and switch-core gets no endpoint. On, the chart runs it, points+renders and switch-core gets no endpoint. On, the chart runs it beside its+model server (Strands Decider by default, or Qwen on llama.cpp), pointsswitch-core at it and admits only switch-core to it. An endpoint run elsewherecan be named instead, but not as well as the chart's own. These are thepromises deploy/remote/helm/switch/README.md makes about `decision.*`; a render@@ -118,12 +119,7 @@ def test_on_runs_the_service_and_points_switch_core_at_it() -> None:} in pod["tolerations"]by_name = containers(deployment)assert set(by_name) == {"decision", "model-server"}- assert env_of(by_name["decision"])["LLAMACPP_URL"] == "http://127.0.0.1:8300"- assert env_of(by_name["decision"])["DECISION_BACKEND"] == "llamacpp"- gpu = by_name["model-server"]["resources"]- assert gpu["limits"]["nvidia.com/gpu"] == 1assert "nvidia.com/gpu" not in by_name["decision"]["resources"].get("limits", {})- assert [c["name"] for c in pod["initContainers"]] == ["download-model"]assert {v["name"]: v["persistentVolumeClaim"]["claimName"]for v in pod["volumes"]@@ -152,6 +148,105 @@ def test_on_runs_the_service_and_points_switch_core_at_it() -> None:]+def test_strands_decider_is_the_default_model_server() -> None:+ docs = render("--set", "decision.enabled=true", *FLAGS)+ objects = decision_objects(docs)+ pod = objects["Deployment"]["spec"]["template"]["spec"]+ by_name = containers(objects["Deployment"])++ env = env_of(by_name["decision"])+ assert env["DECISION_BACKEND"] == "systemone"+ assert env["SYSTEMONE_URL"] == "http://127.0.0.1:8400/v1/systemone"+ assert env["DECISION_MODEL"] == "strands-decider-E4B"+ assert "LLAMACPP_URL" not in env++ server = by_name["model-server"]+ # Its own first-party image, which the release pins by digest.+ assert server["image"] == "switch-decision-strands:latest"+ assert env_of(server) == {+ "STRANDS_CHECKPOINT": "StrandsAgents/strands-decider-E4B-gemma4-v1-2610",+ "STRANDS_MODEL_NAME": "strands-decider-E4B",+ }+ assert [p["containerPort"] for p in server["ports"]] == [8400]+ resources = server["resources"]+ assert resources["limits"]["nvidia.com/gpu"] == 1+ assert resources["requests"]["memory"] == "24Gi"++ # It downloads its own weights into the Hugging Face cache on the volume,+ # so no init container and a writable mount; the first start downloads+ # and loads about 16 GB, so the startup probe waits at least 20 minutes.+ assert "initContainers" not in pod+ assert server["volumeMounts"] == [{"name": "weights", "mountPath": "/models"}]+ probe = server["startupProbe"]+ assert probe["httpGet"] == {"path": "/health", "port": "model"}+ assert probe["periodSeconds"] * probe["failureThreshold"] >= 20 * 60+++def test_qwen_on_llama_cpp_is_the_alternative() -> None:+ docs = render(+ "--set", "decision.enabled=true", "--set", "decision.modelServer.kind=llamacpp"+ )+ objects = decision_objects(docs)+ pod = objects["Deployment"]["spec"]["template"]["spec"]+ by_name = containers(objects["Deployment"])+ env = env_of(by_name["decision"])+ assert env["DECISION_BACKEND"] == "llamacpp"+ assert env["DECISION_MODEL"] == "qwen3.5-4b-q4_k_m"+ assert env["LLAMACPP_URL"] == "http://127.0.0.1:8300"+ assert "SYSTEMONE_URL" not in env++ server = by_name["model-server"]+ assert server["image"] == "ghcr.io/ggml-org/llama.cpp:server-cuda"+ assert env_of(server) == {"MODEL_PATH": "/models/Qwen3.5-4B-Q4_K_M.gguf"}+ assert server["resources"]["limits"]["nvidia.com/gpu"] == 1+ assert server["volumeMounts"] == [+ {"name": "weights", "mountPath": "/models", "readOnly": True}+ ]+ assert [c["name"] for c in pod["initContainers"]] == ["download-model"]+ assert "PersistentVolumeClaim" in objects+++def test_the_hugging_face_token_reaches_whichever_server_downloads() -> None:+ token = "decision.modelServer.hfTokenSecret.name=hf"+ secret_ref = {"secretKeyRef": {"name": "hf", "key": "HF_TOKEN"}}++ def hf_token(container: dict[str, Any]) -> Any:+ return next(+ e.get("valueFrom") for e in container["env"] if e["name"] == "HF_TOKEN"+ )++ strands = decision_objects(render("--set", "decision.enabled=true", "--set", token))+ assert hf_token(containers(strands["Deployment"])["model-server"]) == secret_ref++ llama = decision_objects(+ render(+ "--set",+ "decision.enabled=true",+ "--set",+ "decision.modelServer.kind=llamacpp",+ "--set",+ token,+ )+ )+ init = llama["Deployment"]["spec"]["template"]["spec"]["initContainers"][0]+ assert hf_token(init) == secret_ref+++def test_backend_and_model_can_be_overridden() -> None:+ docs = render(+ "--set",+ "decision.enabled=true",+ "--set",+ "decision.model=my-decider",+ "--set",+ "decision.modelServer.strands.checkpoint=acme/my-decider",+ )+ by_name = containers(decision_objects(docs)["Deployment"])+ assert env_of(by_name["decision"])["DECISION_MODEL"] == "my-decider"+ assert env_of(by_name["decision"])["DECISION_BACKEND"] == "systemone"+ assert env_of(by_name["model-server"])["STRANDS_CHECKPOINT"] == "acme/my-decider"++def test_bedrock_and_the_pod_role() -> None:docs = render("--set",@@ -186,6 +281,10 @@ def test_without_a_model_server_there_is_no_gpu_or_volume() -> None:pod = objects["Deployment"]["spec"]["template"]["spec"]assert [c["name"] for c in pod["containers"]] == ["decision"]assert "initContainers" not in pod+ env = env_of(pod["containers"][0])+ assert env["DECISION_BACKEND"] == "anthropic"+ assert "DECISION_MODEL" not in env+ assert "SYSTEMONE_URL" not in env and "LLAMACPP_URL" not in envdef test_network_policy_can_be_left_out() -> None:@@ -231,13 +330,25 @@ def test_refuses_contradictions() -> None:"--set","decision.enabled=true","--set",- "decision.modelServer.model.file=../x.gguf",+ "decision.modelServer.kind=llamacpp",+ "--set",+ "decision.modelServer.llamacpp.model.file=../x.gguf",+ )+ assert "must be strands or llamacpp" in render_error(+ "--set", "decision.enabled=true", "--set", "decision.modelServer.kind=vllm"+ )+ assert "decision.backend is required" in render_error(+ "--set", "decision.enabled=true", "--set", "decision.modelServer.enabled=false")TESTS: list[Callable[[], None]] = [test_off_by_default_renders_nothing_and_sets_no_endpoint,test_on_runs_the_service_and_points_switch_core_at_it,+ test_strands_decider_is_the_default_model_server,+ test_qwen_on_llama_cpp_is_the_alternative,+ test_the_hugging_face_token_reaches_whichever_server_downloads,+ test_backend_and_model_can_be_overridden,test_bedrock_and_the_pod_role,test_without_a_model_server_there_is_no_gpu_or_volume,test_network_policy_can_be_left_out,diff --git a/scripts/pin_chart_images.py b/scripts/pin_chart_images.pyindex 613139207..cbdeb5fed 100644--- a/scripts/pin_chart_images.py+++ b/scripts/pin_chart_images.py@@ -11,7 +11,8 @@ it, and what runs is reproducible from that one pin.Two charts are published:* `switch` (deploy/remote/helm/switch): switch-core, gateway, setup and the- optional decision service (switch-decision).+ optional decision service (switch-decision), with its Strands Decider model+ server (switch-decision-strands).* `switch-hosted-controller` (deploy/hosted/chart): the controller that runsSwitch cloud machines, and its one image.@@ -28,8 +29,9 @@ Subcommands:* `images` prints the image build matrix as JSON; `charts` prints the charts.* `pin --chart <name>` rewrites that chart's `values.yaml` in place, in thechart's `pin_style`. The switch chart gets `global.imageRegistry` and- `<name>:<version>@sha256:…` per image (it pulls a digest-pinned image- `IfNotPresent` on its own); the hosted controller chart gets+ `<name>:<version>@sha256:…` per image, at `<values_key>.image`, where+ `values_key` may be a dotted path to a nested mapping (it pulls a+ digest-pinned image `IfNotPresent` on its own); the hosted controller chart gets`image.repository` and `image.digest`, so its image renders as`<registry>/<name>@sha256:…`, with no tag. The edit isline-based so the file's comments, which are the chart's documentation under@@ -77,12 +79,15 @@ from typing import Any, Literalimport yaml+MULTI_ARCH = "linux/amd64,linux/arm64"+@dataclass(frozen=True)class Image:chart: str- # Top-level values key holding the image; what goes under it is the- # chart's `pin_style`.+ # Values key holding the image: a top-level key, or a dotted path to a+ # nested mapping ("decision.modelServer.strands"). What goes under it is+ # the chart's `pin_style`.values_key: strdockerfile: strcontext: str@@ -90,6 +95,12 @@ class Image:# say), so a render may leave it out; `covered` still requires it in at# least one of the chart's renders.optional: bool+ # What the release builds it for.+ platforms: str = MULTI_ARCH+ # A multi-GB image (CUDA and PyTorch): the release frees the runner's disk+ # before building it and keeps it out of the GitHub Actions layer cache,+ # whose 10 GB per repository it would fill on its own.+ large: bool = False@dataclass(frozen=True)@@ -152,6 +163,18 @@ IMAGES: dict[str, Image] = {"switch-decision": Image("switch", "decision", "decision/Dockerfile", "decision", True),+ # Strands Decider's model server, beside the decision service when+ # decision.modelServer.kind is strands (the default). NVIDIA GPUs on x86+ # only, so amd64 alone.+ "switch-decision-strands": Image(+ "switch",+ "decision.modelServer.strands",+ "decision/model-servers/strands/Dockerfile",+ "decision/model-servers/strands",+ True,+ platforms="linux/amd64",+ large=True,+ ),"switch-hosted-controller": Image("switch-hosted-controller","image",@@ -170,43 +193,58 @@ def images_of(chart: str) -> dict[str, Image]:DIGEST = re.compile(r"^sha256:[0-9a-f]{64}$")PLACEHOLDER_DIGEST = "sha256:" + "0" * 64-TOP_LEVEL_KEY = re.compile(r"^([A-Za-z0-9_]+):\s*(#.*)?$")class PinError(Exception):pass+def _block_end(lines: list[str], start: int, indent: int) -> int:+ """The line after the block opened by `lines[start]`, a key at `indent`."""+ for i in range(start + 1, len(lines)):+ stripped = lines[i].strip()+ if not stripped or stripped.startswith("#"):+ continue+ if len(lines[i]) - len(lines[i].lstrip(" ")) <= indent:+ return i+ return len(lines)++def _set_child(lines: list[str], parent: str, child: str, value: str) -> None:- """Set `parent.child` to `value`, where `parent` is a top-level mapping key.+ """Set `parent.child` to `value`; `parent` is a top-level mapping key or a+ dotted path to a nested mapping (two-space indent per level).- Exactly one `child` line at two-space indent must sit inside `parent`'s- block; anything else means the values file changed shape under the pinner,- which must fail rather than pin the wrong line or nothing at all.+ Each key on the path, and the `child` line, must appear exactly once at its+ indent inside its parent's block; anything else means the values file+ changed shape under the pinner, which must fail rather than pin the wrong+ line or nothing at all."""- start = None- for i, line in enumerate(lines):- m = TOP_LEVEL_KEY.match(line)- if m and m.group(1) == parent:- if start is not None:- raise PinError(f"top-level key {parent!r} appears more than once")- start = i- if start is None:- raise PinError(f"no top-level key {parent!r}")-- end = len(lines)- for i in range(start + 1, len(lines)):- line = lines[i]- if line and not line[0].isspace() and not line.startswith("#"):- end = i- break-- pattern = re.compile(rf"^ {re.escape(child)}:(\s.*)?$")+ start, end = -1, len(lines)+ for depth, key in enumerate(parent.split(".")):+ pattern = re.compile(rf"^{' ' * (2 * depth)}{re.escape(key)}:\s*(#.*)?$")+ hits = [+ i for i in range(start + 1, end) if pattern.match(lines[i].rstrip("\n"))+ ]+ where = ".".join(parent.split(".")[:depth]) or "the top level"+ if not hits:+ raise PinError(f"no key {key!r} in {where}")+ if len(hits) > 1:+ raise PinError(f"key {key!r} appears more than once in {where}")+ start, end = hits[0], _block_end(lines, hits[0], 2 * depth)++ indent = " " * (2 * len(parent.split(".")))+ pattern = re.compile(rf"^{indent}{re.escape(child)}:(\s.*)?$")hits = [i for i in range(start + 1, end) if pattern.match(lines[i].rstrip("\n"))]if len(hits) != 1:raise PinError(f"expected one {parent}.{child} line, found {len(hits)}")newline = "\n" if lines[hits[0]].endswith("\n") else ""- lines[hits[0]] = f" {child}: {value}{newline}"+ lines[hits[0]] = f"{indent}{child}: {value}{newline}"+++def _lookup(values: Any, path: str) -> Any:+ for key in path.split("."):+ values = values[key]+ return valuesdef _check_digests(digests: dict[str, str]) -> None:@@ -265,7 +303,7 @@ def pin(loaded = yaml.safe_load(pinned)for (parent, child), value in expected.items():- if loaded[parent][child] != value:+ if _lookup(loaded, parent)[child] != value:raise PinError(f"pinned values do not read back {parent}.{child}")return pinned@@ -467,6 +505,8 @@ def main(argv: list[str]) -> int:"image": name,"dockerfile": image.dockerfile,"context": image.context,+ "platforms": image.platforms,+ "large": image.large,}for name, image in IMAGES.items()]