Sapcc Compute

SAP CC Compute (Nova)

MCP Tools

Read Tools

Write Tools (requires MCP_READ_ONLY=false)

Admin Tools (requires MCP_ADMIN_TOOLS=true)

Gotchas

1. Check quota before creating. Nova returns generic 403 or 409 when quota is exhausted — no helpful message. Always call `limes_get_project_quota` for `compute` resources (instances, cores, ram) before any create/resize operation. The error "Quota exceeded" is often not in the Nova response at all.

2. SHUTOFF still consumes quota. A stopped server (`SHUTOFF` status) continues to consume `instances` and `cores` quota. Only `DELETED` servers release quota. Users who "stopped" servers expecting freed capacity will be confused — clarify this distinction.

3. Ports are not in the Nova response. The `addresses` field on a server shows IP addresses grouped by network name, but does NOT include port UUIDs, MAC addresses, or security group assignments. To get port details, call `neutron_list_ports` with `device_id=`. This is the only way to find which security groups apply to a server.

4. SAP CC flavor naming conventions. Flavors follow the pattern `_` (e.g., `m2_xlarge`). Families: `m` = general purpose, `r` = memory-optimized, `c` = compute-optimized. Generations increment (1, 2, 3...). See `references/flavor-families.md` for the full matrix. Do not guess flavor names — always call `nova_list_flavors` to confirm availability in the current region.

5. Server actions are asynchronous. Calling `nova_server_action` returns 202 immediately. The server transitions through intermediate states (e.g., `REBOOT` → `ACTIVE`). You must poll with `nova_get_server` to confirm the action completed. Typical transitions take 5-30 seconds but can take minutes for large instances.

6. HARD reboot is destructive. A `HARD` reboot is equivalent to pulling the power cord — in-flight I/O is lost, filesystems may corrupt. Always attempt `SOFT` reboot first (sends ACPI shutdown signal). Only escalate to `HARD` if the guest OS is unresponsive. Confirm with the user before issuing HARD reboot.

7. ERROR state requires admin intervention. Servers in `ERROR` status cannot be recovered via `nova_server_action`. Common causes: host failure, scheduler error, failed live-migration. The user must contact their cloud admin or file a support ticket. Do not attempt repeated actions on ERROR servers.

8. Addresses field structure varies by network. The `addresses` response is keyed by network name, with each entry containing `addr` (IP), `version` (4/6), and `OS-EXT-IPS:type` (`fixed` or `floating`). Multiple networks produce multiple keys. Do not assume a single-network structure.

9. Server metadata is not automatically populated. Nova metadata is user-supplied key-value pairs. Do not expect metadata to contain project info, cost center, or ownership unless the user's automation sets it. The `host_id` is an opaque hash — it identifies co-location but is not a hostname.

10. Name filter is not exact match. `nova_list_servers` with a `name` filter uses regex-style matching. Searching for `name=web` returns `web-1`, `web-prod`, `my-web-server`, etc. For exact matches, filter results client-side after retrieval.

Common Workflows

List and Inspect Servers

1. nova_list_servers (optionally filter by status or name)
2. For each server needing detail: nova_get_server with its UUID
3. For network info: neutron_list_ports with device_id=<server-uuid>

Get Full Server + Network Picture

1. nova_get_server → note addresses (IPs) and server UUID
2. neutron_list_ports with device_id=<server-uuid> → port UUIDs, MAC, security groups
3. For each security group ID: neutron_list_security_groups for rules

This gives the complete picture: server → IPs → ports → security groups → rules.

Debug: Server Won't Start

1. nova_get_server → check current status
   - SHUTOFF: try nova_server_action start
   - ERROR: inform user, admin required
   - BUILD: still provisioning, wait
   - PAUSED/SUSPENDED: unpause/resume first
2. If start fails with 409: limes_get_project_quota → check compute quota
3. If quota ok: hermes_list_events with target.id=<server-uuid> → recent errors
4. If no audit clues: maia_query for host-level issues

Perform Server Action Safely

1. nova_get_server → confirm current status allows the action
   Valid transitions:
   - start: SHUTOFF → ACTIVE
   - stop: ACTIVE → SHUTOFF
   - reboot (SOFT): ACTIVE → ACTIVE (via REBOOT)
   - reboot (HARD): any running state → ACTIVE (destructive)
   - pause: ACTIVE → PAUSED
   - unpause: PAUSED → ACTIVE
   - suspend: ACTIVE → SUSPENDED
   - resume: SUSPENDED → ACTIVE
2. For destructive actions (stop, HARD reboot): confirm with user
3. nova_server_action with action and server_id
4. Poll nova_get_server until status reaches target (or timeout after 2 min)

Select a Flavor

1. nova_list_flavors → get available flavors in region
2. Match requirements to flavor family:
   - Balanced workload → m-series (general purpose)
   - Database/cache → r-series (memory-optimized)
   - Batch/CI → c-series (compute-optimized)
3. Check limes_get_project_quota to ensure cores/ram headroom
4. If flavor not found: may not be available in this region or project

Troubleshooting

Stuck in BUILD

Server has been in `BUILD` status for >10 minutes.

• Typical cause: Scheduler couldn't place it (no host with capacity), or image download is slow.
• Diagnostic: `hermes_list_events` filtered to `target.id=` — look for `compute/server/create` with outcome `pending` or `failure`.
• Resolution: If no events after 15 min, likely a scheduler issue — admin intervention required. Do not delete-and-retry without checking quota first.

ERROR After Resize

Server entered ERROR after a resize or migrate operation.

• Typical cause: Target host ran out of disk during resize, or live-migration timed out.
• Diagnostic: `hermes_list_events` for recent `compute/server/resize` or `compute/server/migrate` events. Check outcome field.
• Resolution: Admin must reset the server state. User cannot self-service from ERROR.

Can't Reach Server via Network

Server is ACTIVE but unreachable.

• Diagnostic steps:
  1. `nova_get_server` → confirm status is ACTIVE (not PAUSED/SUSPENDED)
  2. Check addresses: does it have a floating IP? Fixed IPs are only reachable from within the VPC/network.
  3. `neutron_list_ports` with `device_id` → check port `status` (should be `ACTIVE`, not `DOWN` or `BUILD`)
  4. Check security groups on the port → ensure ingress rules allow the traffic (SSH=22, ICMP, etc.)
  5. If port is DOWN: may be a binding failure — check `hermes_list_events` for port-related events

Security Considerations

• Confirm destructive actions: Always ask user confirmation before `stop`, `HARD reboot`, or any action that interrupts service. State what will happen: "This will immediately power off the instance, dropping all connections."
• Metadata visibility: Server metadata is visible to anyone with `compute:server:show` permission in the project. Do not store secrets, credentials, or PII in metadata.
• Host ID is semi-sensitive: While opaque, `host_id` reveals co-location (same hash = same hypervisor). Avoid exposing it in shared contexts without need.
• Audit trail: All server actions generate Hermes events. Inform users that actions are logged with their credential identity.
• Cross-project access: Nova operations are scoped to the authenticated project. You cannot see or act on servers in other projects without re-scoping credentials.

Cross-Service References

Routing