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
Pin the decision model weights and recover from a bad download
agent-runner3538068934256a3
diff --git a/decision/README.md b/decision/README.mdindex 53b6f0dcf..df186f4e3 100644--- a/decision/README.md+++ b/decision/README.md@@ -137,8 +137,12 @@ downloads the checkpoint and its base model (about 17 GB) into `/models`.`strands-decider serve`: it merges the checkpoint's LoRA adapter into the baseweights once at load, which on an L4 gives the same answers in 0.26 s insteadof 0.42 s for a typical room and about 3.9 decisions a second instead of 2.4.-It's configured by env only (`STRANDS_CHECKPOINT`, `STRANDS_MODEL_NAME`,-`STRANDS_MAX_BATCH`):+It's configured by env only (`STRANDS_CHECKPOINT`, `STRANDS_REVISION`,+`STRANDS_BASE_REVISION`, `STRANDS_MODEL_NAME`, `STRANDS_MAX_BATCH`). The image+pins the checkpoint and its base model to Hugging Face commits+(`STRANDS_REVISION`, `STRANDS_BASE_REVISION`); the base revision must be the one+the checkpoint records in its `provenance.json`, or the server refuses to+start:```docker build -t switch-decision-strands decision/model-servers/strands/@@ -152,8 +156,8 @@ docker run -d -p 8230:8230 \The image pins `strands-decider==0.2.0` and every dependency by hash(`model-servers/strands/requirements.txt`, compiled from `requirements.in`-with the command at its top). `STRANDS_CHECKPOINT` and `STRANDS_MODEL_NAME`-choose another checkpoint.+with the command at its top). `STRANDS_CHECKPOINT`, the two revisions and+`STRANDS_MODEL_NAME` choose another checkpoint.Qwen3.5-4B on the llama.cpp server image:diff --git a/decision/model-servers/strands/Dockerfile b/decision/model-servers/strands/Dockerfileindex f086600fb..d0c7d7ee5 100644--- a/decision/model-servers/strands/Dockerfile+++ b/decision/model-servers/strands/Dockerfile@@ -11,7 +11,8 @@## No weights in the image. The checkpoint and its base model download from# Hugging Face on the first start into HF_HOME=/models, which is meant to be a-# volume, so later starts only load them. The E4B checkpoint is built on+# volume, so later starts only load them. Both are pinned to commits+# (STRANDS_REVISION, STRANDS_BASE_REVISION; serve.py). The E4B checkpoint is built on# google/gemma-4-E4B-it (about 16 GB in bf16), whose Gemma terms apply; the# weights are fetched by whoever deploys, never redistributed in this image.# Set HF_TOKEN for a gated checkpoint or higher Hub rate limits.@@ -51,6 +52,8 @@ ENV PATH="/opt/strands/bin:$PATH" \HF_HOME=/models \HF_HUB_DISABLE_TELEMETRY=1 \STRANDS_CHECKPOINT=StrandsAgents/strands-decider-E4B-gemma4-v1-2610 \+ STRANDS_REVISION=1344d44397c457f3be34b8ff15c593b41248c6bf \+ STRANDS_BASE_REVISION=ee0ef6023621cff504d758262d4e04895a5af4a2 \STRANDS_MODEL_NAME=strands-decider-E4BUSER 10001:10001WORKDIR /home/decisiondiff --git a/decision/model-servers/strands/serve.py b/decision/model-servers/strands/serve.pyindex c3cc753a6..00530bf9e 100644--- a/decision/model-servers/strands/serve.py+++ b/decision/model-servers/strands/serve.py@@ -3,9 +3,17 @@ adapter merged into the base weights.python serve.py-Reads STRANDS_CHECKPOINT, STRANDS_MODEL_NAME, STRANDS_PORT (default 8400),-STRANDS_HOST (default 0.0.0.0), STRANDS_DEVICE (default cuda) and-STRANDS_MAX_BATCH (default 32, requests evaluated together).+Reads STRANDS_CHECKPOINT, STRANDS_REVISION, STRANDS_BASE_REVISION,+STRANDS_MODEL_NAME, STRANDS_PORT (default 8400), STRANDS_HOST (default+0.0.0.0), STRANDS_DEVICE (default cuda) and STRANDS_MAX_BATCH (default 32,+requests evaluated together).++STRANDS_REVISION is the Hugging Face commit of the checkpoint to download, and+STRANDS_BASE_REVISION the commit of its base model, so a new push to either+repo changes nothing that runs. The loader takes the base model's revision+from the checkpoint itself (its config, else its provenance.json), so the+server checks that the checkpoint names STRANDS_BASE_REVISION and refuses to+start otherwise, then downloads the base at that commit before loading.`strands-decider serve` keeps the adapter separate and applies it on everyforward pass. Merging it once at load gives the same answers (15/15, 36/40 and@@ -23,10 +31,19 @@ import uvicornfrom fastapi import FastAPI, HTTPExceptionfrom fastapi.responses import JSONResponsefrom strands_decider.infer import EngineConfig, SystemOneEngine-from strands_decider.modeling import StrandsDeciderModel+from huggingface_hub import snapshot_download+from strands_decider.modeling import (+ StrandsDeciderConfig,+ StrandsDeciderModel,+ base_revision,+ checkpoint_dir,+ config_path,+)from strands_decider.schema import SystemOneRequestCHECKPOINT = os.environ["STRANDS_CHECKPOINT"]+REVISION = os.environ["STRANDS_REVISION"]+BASE_REVISION = os.environ["STRANDS_BASE_REVISION"]MODEL_NAME = os.environ["STRANDS_MODEL_NAME"]PORT = int(os.environ.get("STRANDS_PORT", "8400"))HOST = os.environ.get("STRANDS_HOST", "0.0.0.0")@@ -34,9 +51,27 @@ DEVICE = os.environ.get("STRANDS_DEVICE", "cuda")MAX_BATCH = int(os.environ.get("STRANDS_MAX_BATCH", "32"))+def pinned_checkpoint() -> str:+ """The checkpoint at REVISION, downloaded, with its base model downloaded at+ BASE_REVISION: the local directory to load."""+ path = checkpoint_dir(CHECKPOINT, revision=REVISION)+ config = StrandsDeciderConfig.from_json(config_path(path))+ recorded = base_revision(path, config)+ if recorded != BASE_REVISION:+ raise RuntimeError(+ f"{CHECKPOINT}@{REVISION} loads {config.base_model} at revision "+ f"{recorded or 'main (none recorded)'}, but STRANDS_BASE_REVISION is "+ f"{BASE_REVISION}. Set STRANDS_BASE_REVISION to the revision the "+ "checkpoint records in its provenance.json, or pin a checkpoint "+ "revision that records the base you want."+ )+ snapshot_download(config.base_model, revision=BASE_REVISION)+ return path++def load_engine() -> SystemOneEngine:started = time.monotonic()- model = StrandsDeciderModel.load(CHECKPOINT)+ model = StrandsDeciderModel.load(pinned_checkpoint())if not hasattr(model.torso, "merge_and_unload"):raise RuntimeError(f"{CHECKPOINT} has no LoRA adapter to merge; this server expects a "@@ -45,7 +80,8 @@ def load_engine() -> SystemOneEngine:model.torso = model.torso.merge_and_unload()model.torso.to(DEVICE, dtype=torch.bfloat16)print(- f"loaded {CHECKPOINT} as {MODEL_NAME} on {DEVICE} with the adapter merged, "+ f"loaded {CHECKPOINT}@{REVISION} (base at {BASE_REVISION}) as "+ f"{MODEL_NAME} on {DEVICE} with the adapter merged, "f"in {time.monotonic() - started:.0f}s",flush=True,)@@ -72,6 +108,8 @@ def health() -> dict:"status": "ok","model": MODEL_NAME,"checkpoint": CHECKPOINT,+ "revision": REVISION,+ "base_revision": BASE_REVISION,"adapter": "merged",}diff --git a/deploy/local/docker-compose.yml b/deploy/local/docker-compose.ymlindex fab947943..c173fdcf2 100644--- a/deploy/local/docker-compose.yml+++ b/deploy/local/docker-compose.yml@@ -158,9 +158,15 @@ services:profiles: ["decision", "decision-gpu"]# The fresh named volume is root's.user: "0:0"+ # Pinned to a Hugging Face commit and checked against its SHA-256, as in+ # the Helm chart. A file that fails the check is deleted (and, if it was+ # already on the volume, downloaded again), so a bad file never sticks.+ # To try another model, set DECISION_MODEL_URL, _FILE and _SHA256+ # together (an empty _SHA256 skips the check).environment:- MODEL_URL: ${DECISION_MODEL_URL:-https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/resolve/main/Qwen3.5-4B-Q4_K_M.gguf}+ MODEL_URL: ${DECISION_MODEL_URL:-https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/resolve/e87f176479d0855a907a41277aca2f8ee7a09523/Qwen3.5-4B-Q4_K_M.gguf}MODEL_FILE: ${DECISION_MODEL_FILE:-Qwen3.5-4B-Q4_K_M.gguf}+ MODEL_SHA256: ${DECISION_MODEL_SHA256-00fe7986ff5f6b463e62455821146049db6f9313603938a70800d1fb69ef11a4}# Only for a gated model (one whose licence your Hugging Face account# has accepted).HF_TOKEN: ${HF_TOKEN:-}@@ -170,12 +176,22 @@ services:- |set -eutarget="/models/$$MODEL_FILE"- if [ -s "$$target" ]; then echo "have $$MODEL_FILE"; exit 0; fi+ verify() { [ -z "$$MODEL_SHA256" ] || echo "$$MODEL_SHA256 $$target" | sha256sum -c -; }+ if [ -s "$$target" ]; then+ if verify; then echo "have $$MODEL_FILE"; exit 0; fi+ echo "$$MODEL_FILE on the volume fails its checksum; deleting it and downloading it again" >&2+ rm -f "$$target"+ fiecho "downloading $$MODEL_URL"set -- -fsSL --retry 5 --retry-delay 10 -C - -o "$$target.partial"if [ -n "$$HF_TOKEN" ]; then set -- "$$@" -H "Authorization: Bearer $$HF_TOKEN"; ficurl "$$@" "$$MODEL_URL"mv "$$target.partial" "$$target"+ if ! verify; then+ echo "the downloaded $$MODEL_FILE fails its checksum; deleted it, so the next start downloads it again" >&2+ rm -f "$$target"+ exit 1+ fivolumes:- decision-models:/models@@ -258,6 +274,10 @@ services:profiles: ["decision-strands-gpu"]environment:HF_TOKEN: ${HF_TOKEN:-}+ # The checkpoint's and its base model's Hugging Face commits, as in the+ # Helm chart (decision.modelServer.strands.revision / baseRevision).+ STRANDS_REVISION: 1344d44397c457f3be34b8ff15c593b41248c6bf+ STRANDS_BASE_REVISION: ee0ef6023621cff504d758262d4e04895a5af4a2volumes:- decision-hf:/modelsdeploy:diff --git a/deploy/remote/helm/switch/README.md b/deploy/remote/helm/switch/README.mdindex 76ef1c371..05b49777a 100644--- a/deploy/remote/helm/switch/README.md+++ b/deploy/remote/helm/switch/README.md@@ -333,9 +333,24 @@ no further setup.probe allows 30 minutes (`decision.modelServer.strands.startupProbe`), andthe decision service reports not ready until then. Later starts only loadthe 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.+ file (`decision.modelServer.llamacpp.model`: URL, file name, SHA-256) first.+ `decision.modelServer.hfTokenSecret` names a Secret with a Hugging Face+ token, for a gated checkpoint or higher Hub rate limits.+- **Pinned weights.** Every model is downloaded at a fixed Hugging Face+ commit, so a new push upstream changes nothing that runs:+ - Qwen's URL names a commit (`…/resolve/<commit>/…`), and the init container+ checks the file's SHA-256 (`model.sha256`). A file already on the volume+ that fails the check is deleted and downloaded again; a fresh download+ that fails it is deleted before the init container exits non-zero, so the+ next start downloads it again rather than failing on every restart.+ - Strands Decider's checkpoint is downloaded at+ `decision.modelServer.strands.revision` and its base model at+ `baseRevision` (`STRANDS_REVISION`, `STRANDS_BASE_REVISION`). The base+ revision must be the one the checkpoint records in its `provenance.json`,+ the one it was trained on; the server refuses to start otherwise.++ To move to a newer model, set the URL, file and SHA-256 together (or both+ Strands revisions), from the model's Hugging Face page.- **Licence.** Strands Decider is Apache-2.0, but its 4B checkpoint is built onGoogle's Gemma 4 (`google/gemma-4-E4B-it`), whose terms apply to anyonedeploying it. The chart downloads the weights at deploy time from Huggingdiff --git a/deploy/remote/helm/switch/templates/decision/deployment.yaml b/deploy/remote/helm/switch/templates/decision/deployment.yamlindex db6e49c9a..5236003b1 100644--- a/deploy/remote/helm/switch/templates/decision/deployment.yaml+++ b/deploy/remote/helm/switch/templates/decision/deployment.yaml@@ -50,8 +50,11 @@ spec:{{- 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.+ # Downloads the weights into the volume unless they are already there+ # and pass their checksum, resuming a download a restart interrupted.+ # A file that fails the checksum is deleted: one already on the volume+ # is downloaded again at once; a fresh download that fails exits+ # non-zero, and the next start downloads it again.- name: download-modelimage: {{ $llama.model.downloaderImage }}command: ["sh", "-c"]@@ -59,19 +62,28 @@ spec:- |set -eutarget="/models/$MODEL_FILE"+ verify() {+ [ -z "$MODEL_SHA256" ] || echo "$MODEL_SHA256 $target" | sha256sum -c -+ }if [ -s "$target" ]; then- echo "have $MODEL_FILE"- else- echo "downloading $MODEL_URL"- set -- -fsSL --retry 5 --retry-delay 10 -C - -o "$target.partial"- if [ -n "${HF_TOKEN:-}" ]; then- set -- "$@" -H "Authorization: Bearer $HF_TOKEN"+ if verify; then+ echo "have $MODEL_FILE"+ exit 0fi- curl "$@" "$MODEL_URL"- mv "$target.partial" "$target"+ echo "$MODEL_FILE on the volume fails its checksum; deleting it and downloading it again" >&2+ rm -f "$target"fi- if [ -n "$MODEL_SHA256" ]; then- echo "$MODEL_SHA256 $target" | sha256sum -c -+ echo "downloading $MODEL_URL"+ set -- -fsSL --retry 5 --retry-delay 10 -C - -o "$target.partial"+ if [ -n "${HF_TOKEN:-}" ]; then+ set -- "$@" -H "Authorization: Bearer $HF_TOKEN"+ fi+ curl "$@" "$MODEL_URL"+ mv "$target.partial" "$target"+ if ! verify; then+ echo "the downloaded $MODEL_FILE fails its checksum; deleted it, so the next start downloads it again" >&2+ rm -f "$target"+ exit 1fienv:- name: MODEL_URL@@ -177,8 +189,14 @@ spec:env:- name: STRANDS_CHECKPOINTvalue: {{ required "decision.modelServer.strands.checkpoint is required" $st.checkpoint | quote }}+ - name: STRANDS_REVISION+ value: {{ required "decision.modelServer.strands.revision is required: the checkpoint's commit on Hugging Face" $st.revision | quote }}+ - name: STRANDS_BASE_REVISION+ value: {{ required "decision.modelServer.strands.baseRevision is required: the base model's commit the checkpoint records" $st.baseRevision | quote }}- name: STRANDS_MODEL_NAMEvalue: {{ required "decision.modelServer.strands.modelName is required" $st.modelName | quote }}+ - name: STRANDS_PORT+ value: {{ $st.port | quote }}{{- with include "switch.decisionHfToken" . }}{{- . | trim | nindent 12 }}{{- end }}diff --git a/deploy/remote/helm/switch/values.yaml b/deploy/remote/helm/switch/values.yamlindex b5d248615..08ad5867e 100644--- a/deploy/remote/helm/switch/values.yaml+++ b/deploy/remote/helm/switch/values.yaml@@ -939,6 +939,14 @@ decision:# 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+ # The commits of the checkpoint (STRANDS_REVISION) and of its base model+ # (STRANDS_BASE_REVISION) to download, so a new push to either repo+ # changes nothing here. The base revision must be the one the checkpoint+ # records in its provenance.json (the one it was trained on): the server+ # refuses to start otherwise. Bump both together, from the checkpoint's+ # Hugging Face page.+ revision: 1344d44397c457f3be34b8ff15c593b41248c6bf+ baseRevision: ee0ef6023621cff504d758262d4e04895a5af4a2# The name the server answers as, and DECISION_MODEL unless `model`# above is set.modelName: strands-decider-E4B@@ -982,13 +990,15 @@ decision:- "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.+ # file is not there yet, from a fixed Hugging Face commit. A file that+ # fails its checksum is deleted and downloaded again; a download that+ # fails it is deleted and the pod restarts, so a bad file never sticks.+ # To change model, set url, file and sha256 together.model:- url: https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/resolve/main/Qwen3.5-4B-Q4_K_M.gguf+ url: https://huggingface.co/unsloth/Qwen3.5-4B-GGUF/resolve/e87f176479d0855a907a41277aca2f8ee7a09523/Qwen3.5-4B-Q4_K_M.gguffile: Qwen3.5-4B-Q4_K_M.gguf- # Optional hex SHA-256 of the file, checked after the download.- sha256: ""+ # Hex SHA-256 of the file. Empty skips the check (not recommended).+ sha256: 00fe7986ff5f6b463e62455821146049db6f9313603938a70800d1fb69ef11a4downloaderImage: curlimages/curl:8.11.1# llama.cpp's /health is 503 until the model is loaded (seconds).startupProbe:diff --git a/docs/old/LOCAL_DEVELOPMENT.md b/docs/old/LOCAL_DEVELOPMENT.mdindex f791e69a2..c0aaad018 100644--- a/docs/old/LOCAL_DEVELOPMENT.md+++ b/docs/old/LOCAL_DEVELOPMENT.md@@ -230,7 +230,9 @@ Auto-route (the `decision_routing` feature flag) asks a decision service whoshould answer an untagged message. `just decision-up` starts one beside thestack: the decision service on `:8230` (`DECISION_PORT` to change it) and allama.cpp server for its model, Qwen3.5-4B in 4-bit, whose weights (about-2.7 GB) download once into the `decision-models` volume. It is the compose+2.7 GB) download once into the `decision-models` volume, from a fixed+Hugging Face commit and checked against their SHA-256 (a file that fails is+deleted and downloaded again). It is the compose`decision` profile; `just decision-up gpu` is the `decision-gpu` profile, forLinux with an NVIDIA GPU and the NVIDIA Container Toolkit.diff --git a/scripts/check_chart_decision.py b/scripts/check_chart_decision.pyindex 79dec2852..208d150cb 100755--- a/scripts/check_chart_decision.py+++ b/scripts/check_chart_decision.py@@ -5,7 +5,8 @@ The decision service (auto-route) is optional in the chart. Off, nothing of itrenders and switch-core gets no endpoint. On, the chart runs it beside itsmodel server (Strands Decider by default, or Qwen on llama.cpp), pointsswitch-core at it and admits only switch-core to it, both by NetworkPolicy and-by a bearer token the chart generates. An endpoint run elsewhere+by a bearer token the chart generates. Its model weights are pinned to fixed+revisions, the same in docker-compose. 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 renderthat breaks one fails here, before a cluster finds out.@@ -18,6 +19,7 @@ Needs `helm` on PATH and PyYAML. Run from the repo root:from __future__ import annotationsimport base64+import reimport subprocessimport sysfrom collections.abc import Callable@@ -26,6 +28,8 @@ from typing import Anyimport yamlCHART = "deploy/remote/helm/switch"+COMPOSE = "deploy/local/docker-compose.yml"+COMMIT = re.compile(r"^[0-9a-f]{40}$")BASE = "deploy/remote/helm/render-check-values.yaml"RELEASE = "rel"FLAGS = [@@ -188,10 +192,14 @@ def test_strands_decider_is_the_default_model_server() -> None:# pinned it: either way, the strands model-server image.repository = server["image"].split("@")[0].rsplit(":", 1)[0]assert repository.rsplit("/", 1)[-1] == "switch-decision-strands", server["image"]- assert env_of(server) == {+ env = env_of(server)+ assert env == {"STRANDS_CHECKPOINT": "StrandsAgents/strands-decider-E4B-gemma4-v1-2610",+ "STRANDS_REVISION": env["STRANDS_REVISION"],+ "STRANDS_BASE_REVISION": env["STRANDS_BASE_REVISION"],"STRANDS_MODEL_NAME": "strands-decider-E4B",- }+ "STRANDS_PORT": "8400",+ }, envassert [p["containerPort"] for p in server["ports"]] == [8400]resources = server["resources"]assert resources["limits"]["nvidia.com/gpu"] == 1@@ -389,6 +397,94 @@ def test_an_existing_token_secret_replaces_the_generated_one() -> None:assert secret_refs_of(core_container(docs))["DECISION_API_KEY"] == ref+def _llama_init(docs: list[dict[str, Any]]) -> dict[str, Any]:+ pod = decision_objects(docs)["Deployment"]["spec"]["template"]["spec"]+ return pod["initContainers"][0]+++def test_the_qwen_weights_are_pinned_and_a_bad_file_is_deleted() -> None:+ docs = render(+ "--set", "decision.enabled=true", "--set", "decision.modelServer.kind=llamacpp"+ )+ init = _llama_init(docs)+ env = env_of(init)+ match = re.fullmatch(+ r"https://huggingface\.co/unsloth/Qwen3\.5-4B-GGUF/resolve/([0-9a-f]{40})/"+ r"Qwen3\.5-4B-Q4_K_M\.gguf",+ env["MODEL_URL"],+ )+ assert match, env["MODEL_URL"]+ assert re.fullmatch(r"[0-9a-f]{64}", env["MODEL_SHA256"]), env["MODEL_SHA256"]+ script = init["args"][0]+ # A file on the volume that fails is deleted and downloaded again; a+ # download that fails is deleted before the container exits non-zero.+ assert script.count('rm -f "$target"') == 2, script+ after_download = script.split('mv "$target.partial" "$target"', 1)[1]+ assert "if ! verify; then" in after_download+ assert after_download.index('rm -f "$target"') < after_download.index("exit 1")+++def test_the_strands_checkpoint_and_base_are_pinned() -> None:+ server = containers(+ decision_objects(render("--set", "decision.enabled=true"))["Deployment"]+ )["model-server"]+ env = env_of(server)+ assert COMMIT.match(env["STRANDS_REVISION"]), env["STRANDS_REVISION"]+ assert COMMIT.match(env["STRANDS_BASE_REVISION"]), env["STRANDS_BASE_REVISION"]+ assert env["STRANDS_REVISION"] != env["STRANDS_BASE_REVISION"]+ assert "decision.modelServer.strands.revision is required" in render_error(+ "--set",+ "decision.enabled=true",+ "--set",+ "decision.modelServer.strands.revision=",+ )+ assert "decision.modelServer.strands.baseRevision is required" in render_error(+ "--set",+ "decision.enabled=true",+ "--set",+ "decision.modelServer.strands.baseRevision=",+ )+++def test_compose_pins_the_same_weights_as_the_chart() -> None:+ with open(COMPOSE, encoding="utf-8") as fh:+ services = yaml.safe_load(fh)["services"]++ def compose_default(value: str) -> str:+ # ${NAME:-default} or ${NAME-default}+ found = re.fullmatch(r"\$\{[A-Z0-9_]+:?-(.*)\}", value)+ assert found, value+ return found.group(1)++ llama = env_of(+ _llama_init(+ render(+ "--set",+ "decision.enabled=true",+ "--set",+ "decision.modelServer.kind=llamacpp",+ )+ )+ )+ download = services["decision-model-download"]+ assert compose_default(download["environment"]["MODEL_URL"]) == llama["MODEL_URL"]+ assert (+ compose_default(download["environment"]["MODEL_SHA256"])+ == llama["MODEL_SHA256"]+ )+ script = download["entrypoint"][-1]+ assert script.count('rm -f "$$target"') == 2, script++ strands = env_of(+ containers(+ decision_objects(render("--set", "decision.enabled=true"))["Deployment"]+ )["model-server"]+ )+ compose_strands = services["decision-model-strands"]["environment"]+ for name in ("STRANDS_REVISION", "STRANDS_BASE_REVISION"):+ assert compose_strands[name] == strands[name], name++def test_refuses_contradictions() -> None:assert "Clear one of them" in render_error("--set",@@ -431,6 +527,9 @@ TESTS: list[Callable[[], None]] = [test_an_endpoint_elsewhere,test_the_service_requires_a_token_that_switch_core_sends,test_an_existing_token_secret_replaces_the_generated_one,+ test_the_qwen_weights_are_pinned_and_a_bad_file_is_deleted,+ test_the_strands_checkpoint_and_base_are_pinned,+ test_compose_pins_the_same_weights_as_the_chart,test_refuses_contradictions,]