Doca Container Deployment

DOCA container deployment — Tasks

Where to start: The order is configure → build → modify → run → test → debug. The ## test verb is an iterative loop, not a one-shot pass — see the eval-loop overlay in ## test below. For container deployment, build and modify are about deployment configuration (per-service image selection, pod-spec YAML, mount contracts, per-service config-file mount), not about compiling source.

These verbs cover the in-scope cross-cutting deployment workflows for an external operator deploying one of the four supported service overlays on the BlueField — Argus, DMS, Firefly, or UROM service. Flow-Inspector and OS-Inspector are outside this bundle's scope and must route through doca-public-knowledge-map/SKILL.md, not through this deployment workflow. Every step assumes the operator has consulted the live public DOCA Container Deployment Guide and the matching per-service guide (both reachable through doca-public-knowledge-map ## DOCA services (doca-public-knowledge-map/SKILL.md#doca-services)) and is using them as the authoritative reference; this file prescribes the order and what to look up where, not a copy-paste runbook.

Non-goals (externally-productized NVIDIA services not in the DOCA SDK): BlueMan, HBN, SNAP, Virtio-net, DOCA Telemetry Service (DTS, as productized), and any other NVIDIA networking software that is not a DOCA SDK service. A user asking "how do I deploy BlueMan?" (or any of the others) is routed to the public NVIDIA docs at docs.nvidia.com/doca/sdk/ for that specific product, NOT silently extrapolated from this skill's contract. The strict-to-DOCA invariant is documented at AGENTS.md ## Non-goals row 7.

configure

Preparing the BlueField target, confirming every precondition the deployment will rely on, and picking the per-service skill that owns the overlay BEFORE any pod-spec is written. This is also the verb where the smoke-before-bulk posture is established up front — every later verb assumes the operator has read it here.

1. Confirm the env is healthy first. This skill expects DOCA installed on the BlueField Arm and the BlueField OS image shipping kubelet standalone + the documented container runtime. If install health is unverified, run doca-setup ## test (doca-setup/TASKS.md#test) first. If the user has no install yet, route to doca-setup ## no-install (doca-setup/TASKS.md#no-install) for the public NGC DOCA container path. Confirm the BFB version on the BlueField is one the public DOCA Container Deployment Guide certifies for the operator's DOCA release.

2. Pick the per-service skill and load it in parallel. Container deployment is a cross-cutting runtime; the per-service overlay (config schema, paired workload, "healthy" definition) lives in the matching per-service skill. Load both skills together:

   • Argus → doca-argus/SKILL.md

   • DMS → doca-dms/SKILL.md

   • Firefly → doca-firefly/SKILL.md

   • UROM service → doca-urom-svc/SKILL.md

   Quote the per-service config schema and "healthy" signal from the per-service skill; quote the runtime contract (kubelet standalone, static-pod manifests directory, pod-spec shape, mount contract) from this skill.

3. Confirm the BlueField preconditions. Per the precondition table in CAPABILITIES.md ## Capabilities and modes, close every precondition the per-service skill flags as required BEFORE writing a pod spec:

   • DOCA installed on the BlueField Arm and healthy (per doca-setup ## test (doca-setup/TASKS.md#test)).

   • Container runtime + kubelet standalone present and running on the BlueField, as documented in the public Container Deployment Guide.

   • BFB version compatible with the per-service container tag.

   • Per-service firmware slot — does NOT apply to any of the four supported service overlays (none of them emulate a host-facing PCIe device); see the firmware-slot disclaimer in CAPABILITIES.md ## Capabilities and modes. If the operator is deploying an externally-productized NVIDIA service that does emulate a PCIe device, that is out of scope here per the Non-goals callout above and the operator's source of truth is the public NVIDIA docs at docs.nvidia.com/doca/sdk/.

   • Image-pull reachability from the BlueField to nvcr.io (corporate proxy / firewall / air-gap concerns close here with the operator's network team).

   • Host-OS permissions to write into the documented static-pod manifests directory and to read kubelet's log surface and the container runtime's log surface.

4. Identify the per-service image source. Per the per-service public guide, the container image is published to the NGC catalog. The agent MUST route through doca-public-knowledge-map ## DOCA services (doca-public-knowledge-map/SKILL.md#doca-services) for the current image string + tag for the operator's DOCA release; do NOT invent an image name from memory. Wrong image string is the most common first-app failure for any DOCA service deployment, and is the load-bearing first-app failure for this shared runtime.

5. Plan the per-service config-file mount. Every DOCA service container is configured by a documented config file mounted into the container at the path the per-service guide names. Plan, in this order: (a) where the config file will live on the BlueField filesystem (a host path the kubelet runs as can read); (b) what mount path the per-service guide expects inside the container; (c) the per-service config field names the per-service guide quotes — quote from there, do NOT infer from generic Kubernetes / config intuition.

6. Plan the rollback path. For any in-bundle service the per-service skill flags as HIGH-STAKES (Firefly on a production time-sensitive plant; DMS replacing the BlueField's gNMI control surface; UROM service driving an in-flight paired host↔DPU workload), every deploy on a live BlueField must have: (a) the pre-deploy BlueField state captured (PHC state before Firefly; gNMI session inventory before DMS; paired-workload checkpoint before UROM service); (b) the previous-known-good pod spec (or a no-service baseline) ready to re-apply; (c) an out-of-band way to reach the BlueField if the deploy disrupts host connectivity (BlueField console, redundant management path, IPMI to the host); (d) a maintenance window agreed with whoever uses the host. For Argus, which has no HIGH-STAKES per-service overlay by default, the rollback bar is lower but the "be able to revert" rule still applies.

7. Identify the three version anchors. Per CAPABILITIES.md ## Version compatibility, record BOTH the host DOCA install version on the BlueField Arm (via doca-version TASKS.md ## configure (doca-version/TASKS.md#configure)), the BFB version on the BlueField, AND the per-service container tag pulled. A mismatch among the three is the silent "the docs say this should work" trap; align them explicitly.

build

Container deployment is a service runtime, not a library. There is no application artifact for the operator to build — the per- service container image ships from the public NGC catalog and the operator pulls it; clients of any per-service surface (gNMI clients for DMS and PTP followers for Firefly) are tools the operator already has.

If the user is asking how to build a per-service config bundle (e.g. a Firefly PTP config, a DMS gNMI endpoint bundle, or an Argus policy bundle), that is the per-service skill's ## modify workflow, not a build. Route to the matching per-service skill.

If the user is asking how to build a DOCA application that will be deployed by the same kubelet-standalone + pod-spec-drop pattern — that is, a user-authored container image rather than an NVIDIA-shipped one — the build is the application's normal DOCA build (per doca-programming-guide ## build (doca-programming-guide/TASKS.md#build)) plus a container build the user owns. The shared runtime contract in CAPABILITIES.md ## Capabilities and modes still applies to the deploy step; the build step is library- shaped, not service-shaped.

If the user is asking how to build the per-service container itself from source, that is not an external-operator workflow — the container ships pre-built from NGC and rebuilding it is out of scope. Route to the per-service public guide via doca-public-knowledge-map ## DOCA services (doca-public-knowledge-map/SKILL.md#doca-services).

modify

Container deployment does not have a "modify a sample program" workflow analogous to DOCA libraries; there is no shared sample the user starts from. The deployment-side analog of "modify" is adapt the documented pod-spec recipe from the public DOCA Container Deployment Guide plus the matching per-service guide to the operator's environment:

1. Start from the documented recipe. Identify, in the public DOCA Container Deployment Guide for the BFB the operator is on, the pod-spec recipe that matches the operator's service. If the per-service guide ships its own pod-spec recipe (e.g. a Firefly or DMS YAML), use that as the per-service overlay on top. Quote both; do not author a new pod-spec shape from scratch.

2. Diff against the operator's environment. Note the substitutions the operator MUST make: the per-service container image tag pulled from NGC, the host path that holds the per-service config file, the volume-mount points the per-service guide names, any per-service envvars / security-context the per-service guide quotes, and the per-service paired-workload coordinates (PHC device node for Firefly, gNMI client endpoint for DMS, paired host↔DPU socket for UROM service). Capture each substitution explicitly; the substitution list is the operator's edit log for the deploy.

3. Apply minimum-change. Change only what the operator's environment forces. Every additional deviation from the documented recipe widens the surface for an unintended mount-path / network-policy / image-pull / version failure the operator will have to debug later.

4. Re-validate against the precondition table. Each substitution is a chance to accidentally break a BlueField precondition (firmware slot, BFB version, image-pull reachability, host-path permission). Walk CAPABILITIES.md ## Capabilities and modes precondition table after every substitution.

5. Re-validate against the per-service safety overlay. Per- service skills add their own safety overlays (Firefly's PTP-aware-path-required rule, DMS's gNMI-credential rotation rule, UROM service's paired-workload-restart-order rule, and Argus's policy-rollback rule). The modify step is the right place to surface the overlay, not debug time.

6. Stage the pod spec outside the watched directory. Write, review, and validate the adapted YAML in a staging location that kubelet does not watch. ## modify performs no live manifest write. The sole routine deployment write into the documented static-pod manifests directory occurs in ## run step 2, after the staged file and rollback snapshot are ready. The conditional destructive smoke in ## test step 2 is a non-routine exception with stricter approval and rollback gates.

The agent's anti-pattern alert: a "start from a generic Kubernetes pod-spec template and adapt" is almost always slower than starting from the public Container Deployment Guide plus per-service guide, because the documented BlueField pod-spec fields and the static-pod manifests directory are not 1:1 with upstream Kubernetes intuition.

run

Bringing up the per-service pod, confirming the runtime layer reaches a healthy state, and confirming the per-service liveness signal before layering any real workload on top. Every step here assumes the prerequisites in ## configure are done.

1. Confirm the per-service image is pullable from NGC. Per ## configure step 4, the image string + tag comes from the per-service public guide reached through doca-public-knowledge-map/SKILL.md; the BlueField has reach to nvcr.io; the documented image-list command on the BlueField container runtime can pull a fresh copy of the image and report the digest. A pull failure here is a layer-3 symptom in CAPABILITIES.md ## Error taxonomy and must be resolved before the pod-spec is dropped.

2. Drop the documented pod spec into the static-pod manifests directory — the sole routine deployment write. Per ## modify step 6, the pod-spec YAML was authored and validated outside the watched directory from the public Container Deployment Guide plus the per-service guide. After capturing the prior live manifest (or no-manifest baseline) for rollback, the operator performs the workflow's routine deployment write at the documented watched path. Adding the file triggers kubelet's documented reconcile.

3. Verify kubelet has scheduled the pod. Use the BlueField container manager's documented status command (or its structured-status surface per doca-structured-tools-contract/SKILL.md) to confirm the pod is in Running with no recent restarts. A pod stuck in Pending is layer 2 of CAPABILITIES.md ## Error taxonomy; a pod in ImagePullBackOff is layer 3; a pod in CrashLoopBackOff is layer 4. Drop to ## debug BEFORE proceeding if the pod is not Running.

4. Read the container's ENTRYPOINT log. Use the BlueField container manager's documented log-stream command for the running container. Confirm the per-service guide's documented bring-up lines appear and that no documented error / warning lines repeat. A "container Running but ENTRYPOINT log shows a config-parse error" is the canonical layer-4 (runtime / ENTRYPOINT) trap — the pod is "running" in kubelet's view but the service inside is not actually serving its purpose.

5. Verify the per-service liveness signal. Each per-service skill names the documented liveness signal that proves the service inside the container is actually ready (event-stream output for Argus, gNMI session up + per-RPC counter for DMS, ports state + PHC offset for Firefly, or per-operation counter for UROM service). Read it now, per the matching per-service skill. If the per-service signal is NOT healthy, drop to ## debug layer 4 (runtime / ENTRYPOINT) or to the per-service skill's debug ladder — not to "restart the pod and hope".

6. Smoke before bulk (next: ## test step 1). Before driving any per-service workload (real PTP-locked workload for Firefly, gNMI traffic for DMS, paired UROM workload for UROM service, policy-driven monitoring for Argus), walk ## test step 1 once to confirm end-to-end readiness; only then layer the workload on top.

For the runtime version + container-tag cross-checks that underlie "my service behaves differently from what the docs say", see doca-version TASKS.md ## run (doca-version/TASKS.md#run) and apply the three-version-anchor overlay from CAPABILITIES.md ## Version compatibility.

test

Container deployment has no "compile and unit-test" workflow — testing is operational and end-to-end.

## test is an iterative loop, not a one-shot pass. Every mutation (pod-spec field edit, image-tag bump, mount path change, config-file edit, per-service config knob, BFB upgrade) re-opens the smoke sweep. Skipping the re-run after a mutation is the failure mode this loop replaces.

Iteration identity: compare the tuple of mutation class, staged manifest identity, kubelet pod state, container exit signature, ENTRYPOINT result, and per-service liveness result. Two consecutive iterations with the same complete tuple are unchanged and trigger escalation; unchanged is never success.

The eval-loop overlay (rows apply to every DOCA service deployment, not just one service):

The agent's rule: every mutation re-opens the sweep. A pod-spec edit followed by "it probably still works" is exactly the failure mode the iterative loop is here to prevent — and on a HIGH-STAKES in-bundle service (Firefly on a production plant, DMS replacing a live gNMI control surface, UROM service driving an in-flight paired workload) the cost of that failure mode is host disruption, not just "weird traffic".

1. End-to-end smoke (the recommended deployment smoke). With the pod scheduled, confirm in this order:

   1. Pod is in Running per kubelet's documented status surface; restart count stable (NOT in CrashLoopBackOff, NOT in ImagePullBackOff).

   2. Container ENTRYPOINT log shows the per-service config parsed cleanly and the documented bring-up lines complete.

   3. The per-service liveness signal (per the matching per- service skill) is healthy. This is the load-bearing end-to-end proof; "pod Running" alone is NOT the same as "service ready".

   4. A trivial liveness probe against the service's documented endpoint succeeds (per-service skill names what "trivial" means — one event for Argus, one gNMI Get for DMS, one PTP frame for Firefly, or one UROM operation for UROM service). Only after all four pieces pass is the deployment ready for bulk per-service workload.

2. CONDITIONAL DESTRUCTIVE SMOKE (non-routine) — pod-spec smoke. Confirm the negative case: temporarily edit a non-critical pod-spec field that the public Container Deployment Guide flags as kubelet-validated (e.g. the resource request, the documented label set, the documented pod-spec metadata field) and confirm kubelet's status surface reports the expected reconcile. A live-manifest mutation smoke is allowed only on a target with no active paired workloads and a confirmed maintenance window. Before the mutation, all three gates are mandatory:

   • [ ] Explicit operator approval is recorded.

   • [ ] A pre-mutation manifest snapshot is saved.

   • [ ] Rollback is tested and verified.

   If any gate is unchecked, SKIP this destructive smoke. Restore the correct value immediately afterwards and verify against the snapshot. This is the operator's evidence that the deploy contract is in fact reconcile-driven on their specific BlueField.

3. Precondition smoke. Independently of the pod, confirm each precondition the per-service skill flags is still satisfied: nvcr.io is reachable from the BlueField; the per-service firmware slot (if any) is still enabled; the BFB version matches the Container Deployment Guide; the host path holding the per-service config file is still in place. A divergence between "precondition healthy independently" and "pod is not healthy" is a layer-4 / layer-5 / layer-6 symptom in CAPABILITIES.md ## Error taxonomy; convergence is a per-service-config symptom routed to the per-service skill.

4. Capability + deploy snapshot. Save the as-deployed answer to: which per-service container tag is running, which BFB and host DOCA versions it pairs with, which pod-spec YAML was dropped, where the per-service config file lives, what kubelet reports for the pod, what the container ENTRYPOINT log's bring-up looks like, what the per-service liveness signal reads. This snapshot is the artifact that lets future debug sessions skip rediscovery — and on a HIGH-STAKES service it is the rollback baseline.

5. Multi-pod smoke (only when the service has paired peers). For supported services with paired pods or paired host-side workloads (host-side time follower for Firefly, gNMI client for DMS, paired host↔DPU workload for UROM service), bring the paired peer up after this pod and confirm the per-service smoke the per-service skill names (host's chronyc tracking follows the PHC for Firefly, gNMI Get returns documented schema for DMS, or a paired UROM operation completes for UROM service). One peer at a time; multi-peer bring-up that batches everything together makes ingest-failure attribution much harder.

Loop termination: stop iterating once two consecutive iterations have the same mutation class, staged manifest identity, pod state, exit signature, ENTRYPOINT result, and liveness result. This matching evidence across both iterations requires escalation, never success; the cause is below the deployment runtime (per-service config, paired peer, BlueField OS, host). Escalate to the per-service skill's debug ladder plus doca-debug TASKS.md ## debug (doca-debug/TASKS.md#debug) with the captured layer evidence.

debug

Layered diagnosis. Walk the layers in this order; do not skip down without clearing the layer above. The eight layers match CAPABILITIES.md ## Error taxonomy.

1. Pod-spec syntax layer (layer 1). Is the YAML pod-spec file even parsing? Symptoms: kubelet's log surface shows a parse / schema error; the pod never appears in kubelet's status output even though the YAML is in the documented static-pod manifests directory. Resolution: re-quote the pod-spec shape from the public DOCA Container Deployment Guide for the BFB the operator is on; do NOT infer field names from generic Kubernetes intuition. A common trap is a pod-spec field name that exists in upstream Kubernetes but is rejected by the kubelet build the BlueField OS image ships.

2. Pod scheduling layer (layer 2). Pod-spec parses, but the pod never reaches Running. Symptoms: pod stuck in Pending; kubelet reports a scheduling refusal. Resolution: re-walk the BlueField precondition table in CAPABILITIES.md ## Capabilities and modes (BFB version, host-OS permissions, image-pull reachability; none of the four supported service overlays require a firmware emulation slot per the firmware-slot disclaimer in CAPABILITIES.md ## Capabilities and modes); confirm the resource requests in the pod spec are within the BlueField Arm's budget per the public guide.

3. Image-pull layer (layer 3). Pod scheduled but the container runtime cannot pull the image. Symptoms: ImagePullBackOff; the container runtime's image-pull log surface shows a 404 / 403 / timeout against nvcr.io. Resolution: re-quote the image string + tag from the per-service public guide reached through doca-public-knowledge-map/SKILL.md; confirm BlueField reach to nvcr.io independently of the pod (a documented image-pull probe from the BlueField CLI is the cheapest check); engage the operator's network team when the BlueField does not have egress to nvcr.io. NEVER "fix" this by inventing an image tag from memory; that path only deepens the failure.

4. Container-runtime / ENTRYPOINT layer (layer 4). Image pulled, container starts, ENTRYPOINT exits non-zero or the pod is in CrashLoopBackOff. Symptoms: the container's ENTRYPOINT log shows a config-parse error, a missing dependency, or a failed internal precondition; the pod restart count climbs. Resolution: read the container's last full ENTRYPOINT log FIRST (the BlueField container manager's documented log-stream command); route the per-service config-parse error to the matching per-service skill's debug ladder. A pod that crashes twice in a row with the same exit signature is HIGH-STAKES per CAPABILITIES.md ## Safety policy — STOP the retry loop, clear the root cause, only then re-enable scheduling.

5. Volume-mount layer (layer 5). Pod runs but the container cannot read / write a path the pod-spec names. Symptoms: ENTRYPOINT log shows permission / not-found errors on a mounted path; the per-service config file is mounted but empty / missing. Resolution: confirm the host path exists on the BlueField; confirm the container user can read / write it per the per-service guide; re-quote the volume-mount shape from the public Container Deployment Guide. A directory bind-mounted where a file was expected (or vice versa) is the canonical trap.

6. Network-policy / host-firewall layer (layer 6). Pod runs, the service inside is up, but a paired peer cannot reach it (or the service cannot reach a paired peer). Symptoms: liveness probe times out; an external client cannot connect; outbound to NGC / upstream PTP master / TOR / S3 endpoint hangs. Resolution: confirm host network reachability independently of the pod; confirm the pod's hostNetwork / port mapping matches what the per-service guide expects; engage the operator's network team when a corporate firewall is in the picture.

7. Version layer (layer 7). The pod is healthy, the service answers, but the behavior contradicts what the per-service guide on the screen describes. Symptoms: a documented field name is rejected; a documented "healthy" output line never appears. Resolution: walk doca-version TASKS.md ## debug (doca-version/TASKS.md#debug) layer 2 and re-confirm the three version anchors (host DOCA on the BlueField Arm, BFB, per-service container tag) per CAPABILITIES.md ## Version compatibility. A per-service guide on a different DOCA release than the container tag is the canonical trap.

8. Cross-cutting host layer (layer 8). Deployment looks healthy at every layer above but a cross-cutting host issue (kernel version, driver loaded / not loaded, hugepage allocation, PCIe link state, BFB log surface) breaks the service's downstream behavior. Resolution: drop to doca-debug TASKS.md ## debug (doca-debug/TASKS.md#debug) for the cross-cutting debug ladder; engage the operator's BlueField OS team.

Command appendix

Container-deployment commands the verbs above reach for, grouped by purpose so the agent picks the right family without searching prose. Every row is a class — the agent must not invent flags beyond what the row names; flag and command discovery is --help on the installed tool or the public guide, not prose recall.

Infra-aware preamble (every row below). Per the bundle's detect → prefer → fall back → report contract documented in doca-structured-tools-contract ## The agent behavior contract (doca-structured-tools-contract/SKILL.md#the-agent-behavior-contract), the agent should:

1. Probe for the matching structured helper FIRST (doca-env --json for version + devices + libraries + drivers + hugepages in one shot; doca-capability-snapshot for per-device capability flags; version-matrix.json for "available since" lookups).

2. If the probe succeeds, the structured tool's output is the authoritative answer and the agent SHOULD NOT also run the manual command in the row below. Report "using structured <tool>".

3. If the probe fails, fall back to the manual command in the row. Report "falling back to manual chain".

4. The schemas the structured tools emit are defined in doca-structured-tools-contract ## Schemas (doca-structured-tools-contract/SKILL.md#schemas); the version-handling semantics (four-way match, NGC, headers-win) are owned by doca-version/SKILL.md.

Three cross-cutting rules for this appendix:

• Never invent an image string, tag, kubelet flag, pod-spec field name, or the static-pod manifests directory path. The public DOCA Container Deployment Guide and the per-service guide reached through doca-public-knowledge-map/SKILL.md are the contract; prose-derived strings are the most common hallucination failure for this skill.

• Kubelet status before container logs; container logs before per-service liveness. When triaging, read kubelet's pod- status surface first (does kubelet think the pod is up?); only then read the container runtime's log stream (did the ENTRYPOINT parse the config and run?); only then read the per-service liveness signal (does the service actually serve its purpose?). Reading the per-service signal against a pod that kubelet has reported as Pending is meaningless.

• Cross-link instead of duplicate. Cross-cutting env commands (port-state, devlink, ip link, ethtool on the BlueField, lspci on the host) live in doca-setup TASKS.md ## Command appendix (doca-setup/TASKS.md#command-appendix) and doca-debug TASKS.md ## debug (doca-debug/TASKS.md#debug); this appendix names only the container-deployment-specific ones. Per-service liveness commands live in the matching per-service skill's Command appendix.

Deferred task verbs

• Per-service config-schema authoring (Argus runtime-security policy; DMS gNMI / gNOI endpoint config; Firefly four-axis PTP config; UROM service operations-endpoint config) — out of scope here. Route to the matching per-service skill's ## configure and ## modify workflows; this skill names that a config file must be mounted, not what the per-service field names are.

• Installing DOCA on the BlueField target — out of scope here. Route to doca-setup ## configure (doca-setup/TASKS.md#configure) for env preparation and doca-setup ## test (doca-setup/TASKS.md#test) for install health verification, or doca-setup ## no-install (doca-setup/TASKS.md#no-install) for the public NGC DOCA container path if there is no DOCA install yet.

• Flipping a BlueField firmware emulation slot — out of scope here for two reasons. First, none of the four supported service overlays emulate a host-facing PCIe device, so none of them require an emulation slot (see the firmware-slot disclaimer in CAPABILITIES.md ## Capabilities and modes). Second, externally-productized NVIDIA services that do require a slot are themselves out of scope per the Non-goals callout at the top of this file; the firmware-tool workflow itself is owned by doca-setup ## configure (doca-setup/TASKS.md#configure) and the public BlueField firmware-configuration documentation.

• Building a paired-workload application (host-side time-sync follower behind Firefly, gNMI client behind DMS, custom analyzer against Argus, paired host↔DPU workload behind UROM service) — not a container-deployment question. Route to the matching per-service skill's deferred-verbs block, which itself routes to the per-library skill (doca-telemetry-exporter/SKILL.md, doca-flow/SKILL.md, doca-rmax/SKILL.md, …) when the paired workload is a DOCA-using application.

• Operating a full Kubernetes cluster (cluster API, kubectl, Deployment / Service / Ingress objects, cluster-wide observability, CNI overlays) — out of scope here. This skill covers kubelet standalone mode on the BlueField, not a full Kubernetes control plane. For Kubernetes-scale DPU provisioning, see DOCA Platform Framework on GitHub via doca-public-knowledge-map ## Public source code: GitHub (doca-public-knowledge-map/SKILL.md#public-source-code-github).

• Authoring or rebuilding a DOCA service container image — out of scope here. The container ships pre-built from NGC at the tag the per-service public guide names; rebuilding it from source is not an external-operator workflow. Route to the per-service public guide via doca-public-knowledge-map ## DOCA services (doca-public-knowledge-map/SKILL.md#doca-services).

• Cross-cutting host-side debug (BlueField kernel, driver loaded / not loaded, hugepage allocation, PCIe link state) — out of scope here. Route to doca-debug TASKS.md ## debug (doca-debug/TASKS.md#debug).

Cross-cutting

• The public DOCA Container Deployment Guide is the single source of truth for the kubelet-standalone-mode runtime, the static-pod manifests directory path, the pod-spec shape, the mount contract, the image-pull procedure, and the lifecycle commands. Per-service guides own the per-service overlay.

• The runtime is uniform across every DOCA service on the BlueField. The agent's job is to walk the shared runtime here, then route to the per-service skill for the overlay — not to re-state the runtime inside the per-service skill.

• The three version anchors (host DOCA on BlueField Arm, BFB version, per-service container tag) must all align. A mismatch among any two is the canonical "the docs say this should work but it does not" failure mode.

• Smoke before bulk. The documented smoke (pod reaches Running; ENTRYPOINT log shows config parsed and bring-up lines complete; per-service liveness signal healthy; trivial liveness probe answered) goes BEFORE any per-service workload on top, never after.

• Failed pod restart is HIGH-STAKES. A pod that recurringly crashes is no longer evidence the deployment can self-heal; stop the retry loop, clear the root cause, only then re-enable scheduling. Letting kubelet loop a known-broken pod is delayed diagnosis, not resilience.

• The agent does NOT invent pod-spec field names, kubelet flags, image tags, or the static-pod manifests directory path. Quote each from the public DOCA Container Deployment Guide plus the per-service guide reached through doca-public-knowledge-map/SKILL.md.

• For URL routing to the Container Deployment Guide, the per-service guides, and the NGC catalog, see doca-public-knowledge-map ## DOCA services (doca-public-knowledge-map/SKILL.md#doca-services).