oci-provision
Create Always Free Oracle Cloud infrastructure with the oci CLI — pre-flight the free allowance, then create VCN, internet gateway, route table, subnet, NSG and SSH rules, and launch a VM.Standard.A1.
SKILL.mdskills/oci-provision/SKILL.md
GitHubOCI provision
Translate the user's infrastructure request into a bounded plan before creating
anything. Show requested resources, sizing, region, and free-tier usage; ask for
confirmation when the request is ambiguous or could bill. Create resources only
after the free allowance has been checked. Procedures:
${CLAUDE_PLUGIN_ROOT}/docs/runbooks/create-the-network.md and
${CLAUDE_PLUGIN_ROOT}/docs/runbooks/launch-an-instance.md. Read
${CLAUDE_PLUGIN_ROOT}/docs/staying-free.md first. Configuration and
authentication come from the oci-setup skill — run it first if config is
missing. Requires an authenticated session — see the oci-session skill.
Resolve runtime context
Discover tenancy, compartment, home region, operator CIDR, and SSH key at runtime. If any value cannot be discovered, ask the user rather than guessing or asking them to paste an OCID.
# Discover OCI_TENANCY_ID, OCI_COMPARTMENT_ID, OPERATOR_CIDR, and
# SSH_PUBLIC_KEY_FILE before creating anything. Do not persist them.
RESOURCE_PREFIX="af"
MANAGED_BY_TAG="oracle-cloud-cli-ops"
Never substitute a placeholder, a guessed OCID, or a default CIDR — that creates a real resource in the wrong place, or opens a rule to the wrong network.
Take a tagged inventory snapshot before the first create. Repeat it after all creates and compare the snapshots for the final report:
oci search resource structured-search --query-text \
"query all resources where (freeformTags.key = 'managed-by' && freeformTags.value = '$MANAGED_BY_TAG')" \
--query 'data.items[].{id:identifier,name:"display-name",type:"resource-type",state:"lifecycle-state"}' \
--output table
Pre-flight check — mandatory, every launch
Moving off Terraform deleted the variables.tf validation that rejected an
oversized shape before any API call; this check is the only guardrail left. Show
the user the numbers and refuse to launch if it fails.
- Intended sizing must fit 2 OCPUs / 12 GB RAM / 200 GB total storage.
- Add existing tenancy usage. The allowance is per tenancy, not per instance: two 1-OCPU instances exhaust it as thoroughly as one 2-OCPU instance, and each extra launch looks individually harmless.
oci compute instance list --compartment-id "$OCI_COMPARTMENT_ID" --all \
--query "data[?\"lifecycle-state\"=='RUNNING'].{name:\"display-name\",shape:shape,ocpus:\"shape-config\".ocpus,mem:\"shape-config\".\"memory-in-gbs\"}" --output table
- Confirm the configured region equals the home region, both discovered at runtime — never hardcoded. Free compute exists only in the home region.
HOME_REGION=$(oci iam region-subscription list \
--query 'data[?"is-home-region"]."region-name" | [0]' --raw-output)
CURRENT=$(grep -A8 "^\[$OCI_CLI_PROFILE\]" ~/.oci/config | grep '^region' | cut -d= -f2 | tr -d ' ')
[ "$HOME_REGION" = "$CURRENT" ] || echo "STOP: $CURRENT is not home region $HOME_REGION"
Network, in dependency order
VCN → internet gateway → route table → subnet → NSG → NSG rules
Each step needs the OCID of the one before, so create them one at a time with
--wait-for-state AVAILABLE. Nothing here bills. Flags are in
${CLAUDE_PLUGIN_ROOT}/docs/runbooks/create-the-network.md: --cidr-blocks
(JSON array) on the VCN,
--is-enabled true on the gateway, --route-table-id on the subnet or it
silently gets no internet route.
Never create an SSH ingress rule with "source":"0.0.0.0/0". Require
$OPERATOR_CIDR as a /32, refreshed with
export OPERATOR_CIDR="$(curl -s https://checkip.amazonaws.com)/32" if the user's
IP may have rotated. Leave 80/443 closed until something listens on them.
Launching compute
--shape-config "{\"ocpus\":2,\"memoryInGBs\":12}"is required for.Flexshapes. Omitted, you inherit shape defaults that can exceed the allowance.- Discover the image at runtime:
oci compute image list ... --shape "VM.Standard.A1.Flex" --sort-by TIMECREATED --sort-order DESC.--shapefilters out x86 images. Never paste an image OCID from a blog — region-specific, and deprecated over time. --ssh-authorized-keys-file "$SSH_PUBLIC_KEY_FILE"takes the public key. A private or wrong key makes the instance permanently unreachable — no password login, no reset.--wait-for-state RUNNINGon every create; without it the command returns before provisioning is known to have succeeded.- Tag everything:
--freeform-tags "{\"managed-by\":\"$MANAGED_BY_TAG\"}". - Apply the same tag to every created resource wherever the OCI command supports freeform tags. Never create a managed resource without the tag.
Out of host capacity is normal
Out of host capacity on A1 is contested free capacity, not a mistake. Retry the
same launch against each availability domain from
oci iam availability-domain list, then retry later or drop to 1 OCPU / 6 GB.
Never resolve it by switching to a non-A1 shape — that is the billable "fix".
Verify, then persist
Verify the launched shape against what was requested — do not assume the API honoured it:
oci compute instance get --instance-id "$INSTANCE_ID" --query 'data.{state:"lifecycle-state",
shape:shape, ocpus:"shape-config".ocpus, memory:"shape-config"."memory-in-gbs", region:region}'
Wrong shape, OCPUs, memory or region: terminate immediately, billing is hourly.
oci compute instance terminate --instance-id "$INSTANCE_ID" \
--preserve-boot-volume false --force --wait-for-state TERMINATED
Keep returned OCIDs only in the current shell. The final report must come from the post-create tagged snapshot, not intended commands: list each new resource's truncated OCID, name, type, lifecycle state, region, and free-tier status, plus anything skipped, failed, or already present.
Report what was created, the verified shape, and the home region, then point at
${CLAUDE_PLUGIN_ROOT}/docs/staying-free.md for the budget alert.