openai/provision-droplet

Provision a DigitalOcean droplet as a Codex remote workspace

Follow these steps in order. Do not skip or reorder them. Only the installed Codex DigitalOcean app tools and the bundled Python scripts may be used. doctl, ad hoc integration configs, and any other DigitalOcean CLI tools are prohibited.

Before you start

• Prerequisites: a funded DigitalOcean account, the installed and authenticated Codex DigitalOcean app, a local ssh/ssh-keygen (OpenSSH), Python 3, and the Codex desktop app.
• Cost: the droplet bills hourly from creation until you delete it. Sizes in step 5 show approximate monthly rates. Remind the user to delete it when done (see Cleanup below).
• Time: end-to-end takes ~10–15 minutes — roughly 7 minutes waiting for the droplet to boot (steps 7-8) plus up to 7 minutes for cloud-init (step 9). This is normal; do not abort.
• Locate the bundled scripts first (do this before Step 2). The helper scripts live in the scripts/ folder next to this SKILL.md (i.e. provision-droplet/scripts/). Your current working directory is not the skill directory, so bare relative paths like scripts/keygen.py will fail. Resolve the absolute directory that contains this SKILL.md and call it <skill_dir>. If you don't already know it, find it — the installed plugin may nest it under a version folder (e.g. .../<version>/provision-droplet/), so locate the directory that actually contains scripts/keygen.py. Use <skill_dir>/scripts/<name>.py (an absolute path) in every command below.

Step 1 — Verify DigitalOcean app access

This plugin depends on the single Codex DigitalOcean app. Use it for all DigitalOcean operations; do not register or log in to separate plugin-owned app integrations.

The DigitalOcean app provides both:

• SSH key tools: key-create, key-list, key-delete.
• Droplet tools: droplet-create, droplet-get, droplet-delete.

Confirm that these tools are available before continuing. If the app's tools are missing or unauthenticated, stop and tell the user to install or authenticate the DigitalOcean app in Codex. Do not fall back to doctl, API tokens, or a local integration config.

Step 2 — Generate SSH key pair

python3 <skill_dir>/scripts/keygen.py

Parse the JSON output and keep these values for the steps below: prefix, name, key_name, key_path, pub_key.

How these relate (all derived from one random prefix like bright-hawk-a3f2):

• name = codex-<prefix> — the droplet name and the local SSH alias (they are identical).
• key_name = codex-key-<prefix> — the label for the key on DigitalOcean's side only.
• key_path — the local private key file.

Step 3 — Upload SSH public key

Call the DigitalOcean app tool key-create:

| Parameter | Value | |-----------|-------| | Name | key_name from step 2 | | PublicKey | pub_key from step 2 |

Extract ssh_key.id from the response — this is <key_id>.

If the call fails because a key with that name or fingerprint already exists (e.g. a previous run), do not create a duplicate: call the DigitalOcean app tool key-list, find the entry whose name matches key_name (or whose fingerprint matches the uploaded key), and use its id as <key_id>.

Step 4 — Choose a region

Ask the user, in chat:

Use the defaults — region nyc3 (New York, US) and size s-2vcpu-4gb (2 vCPU / 4 GB, ~$24/mo) — or customize them?

If they choose the defaults, use nyc3 as <region> and s-2vcpu-4gb as <size>, then skip step 5 and continue to step 6.

If they want to customize the defaults, present this region list first and ask them to reply with a slug:

| Slug | Location | |------|----------| | nyc3 (default) | New York, US | | sfo3 | San Francisco, US | | tor1 | Toronto, CA | | lon1 | London, UK | | fra1 | Frankfurt, DE | | ams3 | Amsterdam, NL | | sgp1 | Singapore, SG | | blr1 | Bangalore, IN | | syd1 | Sydney, AU |

Validate their reply against this table. If it is not one of these slugs, ask again — do not pass an unlisted value through. The chosen slug is <region>.

Step 5 — Choose a droplet size

Skip this step if <size> was already set to the default in step 4.

Otherwise, present this size list and ask the user to reply with a slug. Every size below is above the 1 vCPU / 2 GB floor required by the Codex Universal image. Prices are approximate — confirm in the DigitalOcean dashboard.

| Slug | vCPU | RAM | Tier | ~$/mo | |------|------|-----|------|-------| | s-2vcpu-4gb (default) | 2 | 4 GB | Shared basic | $24 | | s-4vcpu-8gb | 4 | 8 GB | Shared basic | $48 | | s-8vcpu-16gb | 8 | 16 GB | Shared basic | $96 | | c-2 | 2 | 4 GB | Premium CPU-optimized | $42 | | g-2vcpu-8gb | 2 | 8 GB | Premium general-purpose | $63 |

Validate their reply against this table. If it is not one of these slugs, ask again — do not pass an unlisted value through. The chosen slug is <size>.

Step 6 — Create droplet

Call the DigitalOcean app tool droplet-create:

| Parameter | Value | Notes | |-----------|-------|-------| | Name | name from step 2 | | | Region | <region> from step 4 | | | Size | <size> from step 5 | | | ImageID | 234061005 | DigitalOcean Codex Universal image | | SSHKeys | ["<key_id>"] | |

Extract droplet.id from the response — this is <droplet_id>.

If droplet-create fails, show the user the error and handle it by cause — do not blindly retry the same values:

• Size not available in this region (premium tiers like c-2 and g-2vcpu-8gb are not in every region): go back to step 4 or 5 and pick a different region/size combination.
• Payment / quota / limit errors: stop and tell the user to resolve it in the DigitalOcean dashboard, then re-run.
• Invalid image or any other error: stop and report it.

The uploaded SSH key from step 3 is harmless to leave, but if you abort here see Cleanup below.

Step 7 — Schedule delayed deployment check-in

After droplet-create succeeds, inspect the response before polling. If the droplet is not already active with a public IPv4 address in the create response, assume provisioning is still in progress.

Create a Codex heartbeat to resume this same thread in 5 minutes for the first status check. Use the Codex automation tool, not a shell sleep or local timer:

| Field | Value | |-------|-------| | mode | create | | kind | heartbeat | | destination | thread | | name | Check DigitalOcean droplet <name> | | rrule | FREQ=MINUTELY;INTERVAL=5 | | status | ACTIVE | | prompt | Resume provisioning DigitalOcean droplet <name> (<droplet_id>). Start at Step 8: check whether it is active and has a public IPv4 address, then continue the workflow. The droplet bills hourly until deleted. |

After creating the heartbeat, tell the user the droplet was created and Codex will check back in about 5 minutes. Stop the active turn here. Do not start 20-second polling until the heartbeat wakes the thread back up. This avoids busy-waiting during the normal 5-7 minute deployment window.

Keep the created heartbeat's automation id as <heartbeat_id> if the tool returns one. When the heartbeat resumes the thread, use step 8 to decide whether to keep checking or to delete/pause the heartbeat and continue.

If the create response already includes status == "active" and a public IPv4 address, skip the heartbeat and continue immediately to step 8.

Step 8 — Check whether the droplet is active

Call the DigitalOcean app tool droplet-get once with ID: <droplet_id>.

If the response has status == "active" and networks.v4 contains an entry with type == "public", extract ip_address from that entry — this is <ip>. Delete or pause <heartbeat_id> if it still exists, then continue to step 9.

If the droplet is still not active or does not yet have a public IPv4 address, schedule Codex to check this same thread again in 1 minute, then stop the active turn here. Keep checking back every minute until the droplet is ready; do not give up merely because the DigitalOcean deployment is slow. Prefer updating the existing <heartbeat_id> to a 1-minute interval; if that is not possible, create a replacement heartbeat and delete/pause the old one so there is only one active check-in.

Use these values for the 1-minute follow-up heartbeat:

| Field | Value | |-------|-------| | mode | create or update, depending on the available automation tool | | kind | heartbeat | | destination | thread | | name | Check DigitalOcean droplet <name> | | rrule | FREQ=MINUTELY;INTERVAL=1 | | status | ACTIVE | | prompt | Resume provisioning DigitalOcean droplet <name> (<droplet_id>). Start at Step 8: check whether it is active and has a public IPv4 address, then continue the workflow. If it is still not ready, schedule another 1-minute heartbeat. The droplet bills hourly until deleted. |

Keep all status checks with the DigitalOcean app — do not use doctl or any other tool to check droplet status.

Step 9 — Configure local SSH

python3 <skill_dir>/scripts/configure_ssh.py \
  --alias codex-<prefix> \
  --ip <ip> \
  --user root \
  --key-path <key_path>

⚠️ Do not interrupt this step. It waits for cloud-init to finish (up to 7 minutes) and prints a ⏳ status line every 5 seconds — that output means it is working normally. Do not run any other commands. Wait for the line DROPLET READY. If the script exits with an error instead, report it and offer to delete the droplet (see Cleanup).

Final step: adding it to Codex

Ask Codex to open the add-SSH-host flow by responding with a clickable Markdown link. Replace <ssh-alias> with name from step 2:

[Add <ssh-alias> to Codex SSH](codex://settings/connections/ssh/add?name=<ssh-alias>)

The SSH alias is the local host alias created by step 9. In this workflow, it is the same value as the droplet name: codex-<prefix>.

If the link does not open the flow, or if the user wants to check it manually, tell them to open: Codex App → Settings → Connections → Add SSH Host → pick the alias → choose the remote folder.

Cleanup (on failure or when done)

The droplet bills hourly until deleted. To tear down:

1. Delete the droplet — DigitalOcean app tool droplet-delete with ID: <droplet_id>.
2. Delete the SSH key (optional) — DigitalOcean app tool key-delete with the <key_id> from step 3.

Always confirm with the user before deleting. Do not use doctl for cleanup.