Action Safety

Every Neptune Analytics graph created by this skill MUST carry these tags:

• created_by=neptune-skill
• generation_model=<model-id> (e.g., claude-sonnet-4-5)

Always issue tags in TWO steps. Inline --tags on create-graph is unreliable through some MCP serialization paths — do NOT rely on it as the sole tagging mechanism.

Step 1: Create the graph (inline tags as best-effort)

Default to --deletion-protection. Use --no-deletion-protection ONLY for ephemeral/CI graphs that an automated harness must tear down.

Step 2: Tag explicitly via tag-resource (REQUIRED)

The CreateGraph response includes the graph ARN. Pass it to tag-resource. A graph still in CREATING may reject the call, so use a SHORT bounded retry (a few attempts, a few seconds apart) rather than polling for AVAILABLE — long AVAILABLE-polling times out the MCP connection and some harness tasks explicitly forbid it.

This second call is what the validator’s list_tags_for_resource check relies on. The eval framework will fail any test that omits it. Apply this two-step pattern to every Neptune Analytics creation call — create-graph, create-graph-using-import-task, create-graph-snapshot, private endpoints.

Neptune Database uses a DIFFERENT tagging API — do not use neptune-graph tag-resource against a Database ARN:

Safety semantics for destructive and high-impact Neptune operations. Every action that can cause data loss, downtime, or irreversible changes must follow these safeguards.

Safety Levels

High Risk Actions

The agent does NOT execute anything in this section. SKILL.md lists these operations under “Do NOT execute” — the agent refuses, explains why, and redirects to the user’s change-control process or the AWS Console. The commands below document what the USER runs themselves, and the safeguards the agent should walk them through first.

Delete Neptune Database Cluster

Reversibility: Irreversible. All data permanently deleted unless final snapshot taken.

Required safeguards:

1. Always suggest a final snapshot before deletion.
2. Require explicit user confirmation with the cluster identifier.
3. Warn that all data, endpoints, and connections will be destroyed.
4. Confirm the cluster is not referenced by active applications.

Delete Neptune Analytics Graph

Reversibility: Irreversible. All in-memory data lost immediately.

Required safeguards:

1. Warn that Neptune Analytics is ephemeral — data cannot be recovered.
2. Suggest exporting results to S3 before deletion.
3. Require explicit user confirmation with the graph identifier.

Drop All Vertices/Edges (g.V().drop() / g.E().drop())

Reversibility: Irreversible. All graph data permanently deleted.

Required safeguards:

1. Warn about complete data loss.
2. Suggest taking a cluster snapshot before dropping.
3. Require explicit user confirmation.
4. Confirm this is intentional and not a targeted delete scenario.

Medium Risk Actions

Modify Cluster (Instance Class, Engine Version)

Reversibility: Varies. Instance class changes are reversible; major engine upgrades are not.

Required safeguards:

1. Flag that modification may cause brief failover.
2. Recommend applying during maintenance window for production.
3. Create snapshot before major engine version upgrades.
4. For engine upgrades, check client library compatibility.

Enable/Disable IAM Authentication

Reversibility: Reversible, but may disconnect active clients.

Required safeguards:

1. Warn that enabling IAM auth requires all clients to sign requests with SigV4.
2. Warn that disabling IAM auth on a public-endpoint cluster removes authentication.
3. Recommend testing in non-production first.

Enable/Disable Public Endpoints

Reversibility: Reversible.

Required safeguards:

1. Warn that public endpoints expose the cluster to internet traffic.
2. Confirm IAM auth is enabled (required for public endpoints).
3. Recommend restricting security group inbound rules to specific IPs.

Scale Down (Instance Class or Reader Replicas)

Reversibility: Reversible (scale back up).

Required safeguards:

1. Verify remaining capacity handles current data size and query load.
2. Warn about brief failover during instance class change.
3. Check BufferCacheHitRatio — scaling down may push working set out of memory.

Low Risk Actions

Create Snapshot — Low risk. No impact on running cluster.

Describe/List Operations — Read-only. No state change. No safeguards needed.

Add Read Replicas — Reversible. Brief sync impact on primary for large graphs.

Modify Tags — Fully reversible. May affect cost allocation and IAM tag policies.

Never-Auto-Execute List

The agent MUST NEVER execute these, in any context or automation. This list is the superset of SKILL.md’s “Do NOT execute” section — the two must stay in sync.

Mirrors SKILL.md’s refuse list:

• Delete cluster / instance / graph — delete-db-cluster, delete-db-instance, delete-graph
• Reset Analytics graph data — reset-graph
• Failover — failover-db-cluster
• Major engine version upgrade — modify-db-cluster --engine-version across majors
• Reboot — reboot-db-instance, reboot-db-cluster
• Cancel long-running work — cancel-import-task, cancel-export-task

Additionally never auto-execute:

• g.V().drop() or g.E().drop() (full data wipe)
• Snapshot deletion (especially the last/only snapshot)
• IAM auth toggle on production clusters
• Public endpoint toggle on production clusters
• Security group rule removal
• Migration cutover execution