Live smoke test
Live smoke test
This is the release check for the management plane: a person signs a real client in, and an app with redirects, an enterprise organization with SSO, a claimed domain, an Admin Portal link and a webhook get set up without anybody touching the console except to approve. It uses the released clients, not test doubles:
| Client | Version used | What it proves |
|---|---|---|
cbox CLI (PHAR) |
0.1.0 | Device sign-in at the platform root, cbox id use, the action catalogue as commands, approvals |
| MCP Inspector CLI | @modelcontextprotocol/[email protected] |
/mcp on the root (a person's token) and on an environment host (a management key), a tool held for approval and finished with approval_id |
| JS SDK | @cboxdk/[email protected] |
EnvironmentClient with a key: list, create with Idempotency-Key, replay |
| Python SDK | cbox-id-client==0.10.0 |
The same, plus one root token driving an environment with environment= |
It was first run on 2026-10-08 against v2.0.0 on a laptop (section 1). Every step below records that historical run. Re-run it against the candidate commit and record the current CLI/SDK versions; the old result does not certify a newer release. Against staging, skip section 1 and set the variables in section 2 to staging's hosts.
1. A production-shaped deployment on one machine
What runs, and why each piece:
| Piece | How | Why |
|---|---|---|
| App | php artisan serve on :8000, PHP_CLI_SERVER_WORKERS=8 |
The CLI polls while the browser approves, so it needs more than one worker |
| TLS | Caddy (tls internal) on :443 for id.localhost and *.id.localhost |
Environment issuers are always https://{slug}.{base} with no port, and the CLI refuses http for anything but loopback. Real TLS is the only way the issuers match |
| Database | Postgres 18 (Docker) | What production runs (cbox.yaml) |
| Cache, sessions, queue | Valkey 8 with noeviction (Docker) |
As production |
| Mailpit (Docker) | Real SMTP; invitations and portal links are readable at http://127.0.0.1:8025 |
|
| Queue | php artisan queue:autoscale |
Without it webhooks and mail never leave |
| Scheduler | php artisan schedule:work |
Without it the event relay never runs |
*.localhost resolves to the loopback address on macOS and most Linux resolvers, so no
/etc/hosts edits are needed.
compose.yml (in a scratch directory, next to the Caddyfile):
name: cboxid-live
services:
postgres:
image: postgres:18-alpine
environment: { POSTGRES_DB: cbox_id, POSTGRES_USER: cbox_id, POSTGRES_PASSWORD: secret }
ports: ["127.0.0.1:55433:5432"]
valkey:
image: valkey/valkey:8-alpine
command: ["valkey-server", "--requirepass", "secret", "--maxmemory-policy", "noeviction"]
ports: ["127.0.0.1:56380:6379"]
mailpit:
image: axllent/mailpit:latest
ports: ["127.0.0.1:1025:1025", "127.0.0.1:8025:8025"]
caddy:
image: caddy:2-alpine
ports: ["443:443"]
volumes: ["./Caddyfile:/etc/caddy/Caddyfile:ro", "caddydata:/data"]
extra_hosts: ["host.docker.internal:host-gateway"]
volumes:
caddydata:
Caddyfile:
{
local_certs
skip_install_trust
}
https://id.localhost, https://*.id.localhost {
tls internal
reverse_proxy host.docker.internal:8000
}
docker compose up -d
docker cp cboxid-live-caddy-1:/data/caddy/pki/authorities/local/root.crt ./caddy-root.crt
The development CA is handed to each client explicitly (CBOX_CA_BUNDLE,
NODE_EXTRA_CA_CERTS, httpx.Client(verify=…), curl --cacert). It is never added to the
system trust store.
.env (the parts that differ from .env.example):
APP_ENV=production
APP_DEBUG=false
APP_URL=https://id.localhost
DB_CONNECTION=pgsql
DB_HOST=127.0.0.1
DB_PORT=55433
DB_DATABASE=cbox_id
DB_USERNAME=cbox_id
DB_PASSWORD=secret
REDIS_HOST=127.0.0.1
REDIS_PORT=56380
REDIS_PASSWORD=secret
CACHE_STORE=redis
SESSION_DRIVER=redis
QUEUE_CONNECTION=redis
SESSION_SECURE_COOKIE=true
MAIL_MAILER=smtp
MAIL_HOST=127.0.0.1
MAIL_PORT=1025
TRUSTED_PROXIES=*
CBOX_ID_ENVIRONMENT_BASE_DOMAINS=id.localhost
HEALTH_TOKEN=<openssl rand -hex 16>
TRUSTED_PROXIES=* is safe here only because the app listens on a port nothing but Caddy
talks to. Install, then start the three processes:
php artisan key:generate
php artisan cbox-id:install --no-interaction \
[email protected] --name="Live Operator" \
--multi-tenant --console-host=id.localhost \
--environment=Acme --organization="Acme Workspace" \
--issuer=https://id.localhost
PHP_CLI_SERVER_WORKERS=8 php artisan serve --host=0.0.0.0 --port=8000
php artisan queue:autoscale
php artisan schedule:work
php artisan cbox-id:doctor
Result:
- The install migrated, minted the crypto key, created the platform root
platform, the operator, the workspace Acme Workspace with its environment (nameAcme, slugacme-workspace: the first environment's slug comes from the workspace name), both CLI clients and printedcbox login --issuer https://acme-workspace.id.localhost. The generated operator password is printed once. cbox-id:doctoronce the processes run: Ready, with 2 warnings. One isLOG_CHANNELwriting to disk (expected locally). The other, Signing keys: no active signing key yet, is wrong: both environments have an active key. The framework's doctor counts keys outside any environment, where the tenancy scope answerswhere 1 = 0. Harmless; reported againstcboxdk/laravel-id.- Probes:
https://id.localhost/up200;/.well-known/openid-configuration404 on the root (the root is not an IdP) and 200 onacme-workspace.id.localhost;/.well-known/cbox-cli200 on both, naming the root's and the environment's CLI client.
2. Variables
export ROOT=https://id.localhost # the platform root
export CA=$PWD/caddy-root.crt # omit on staging (public certificates)
alias cbox='CBOX_CA_BUNDLE=$CA php ./cbox' # the released PHAR
gh release download v0.1.0 -R cboxdk/cbox-cli -p cbox -p cbox.sha256 and compare the
checksum before running it. To keep the run away from your own CLI profile, run it with
HOME pointed at a scratch directory.
3. The cbox CLI, signed in at the root
| # | Command | Result |
|---|---|---|
| 3.1 | cbox login --issuer $ROOT --no-browser |
Prints a user code and $ROOT/device?user_code=…. Approved by the operator, signed in at $ROOT/login, on /device. The CLI then reports the root's scopes, resource: $ROOT/mcp and a refresh token, and fetches the catalogue: 249 actions |
| 3.2 | cbox id whoami |
Person Live Operator, workspace Acme Workspace (owner), operator yes, environments acme-workspace |
| 3.3 | cbox id environments create --name=acme --type=sandbox --yes |
Held: "Check the code there matches: 8D16". The operator approved the request with that code on $ROOT/approvals; the CLI finished by itself: environment acme, slug acme-workspace-acme, issuer https://acme-workspace-acme.id.localhost |
| 3.4 | cbox id use acme-workspace-acme |
"Commands on id.localhost now act in acme-workspace-acme" |
| 3.5 | cbox id apps create --name="Acme Web" --type=web --redirect-uris=https://app.acme.test/callback --yes |
Held, approved on /approvals, then the app with client_id and its client_secret (shown once) |
| 3.6 | cbox id apps update <app> --redirect-uris=https://app.acme.test/callback --redirect-uris=https://app.acme.test/auth/callback --post-logout-redirect-uris=https://app.acme.test/ |
Not held (danger: write). Both redirect URIs and the sign-out URI |
| 3.7 | cbox id apps secrets rotate <app> --grace-seconds=3600 --yes |
Held, approved; the new secret once, previous_expire_at an hour out |
| 3.8 | cbox id organizations create --name="Globex Corporation" --slug=globex |
Organization globex |
| 3.9 | cbox id sso connections create --organization-id=<org> --name="Globex Okta" --type=saml --pending-idp |
A draft and its service_provider block (entity ID, ACS URL, metadata URL) for the IT admin to paste into their IdP, on https://acme-workspace-acme.id.localhost/sso/saml/{id}/…. The first run answered them on the root, which 404s there (found) |
| 3.10 | cbox id sso domains create --organization-id=<org> --domain=globex.example |
The TXT record _cbox-id-challenge.globex.example to publish |
| 3.11 | cbox id sso domains verify <domain> |
verified: false. Expected: nobody published the record. On staging, use a domain you can publish TXT records for |
| 3.12 | cbox id organizations portal_links create <org> --intents=sso --intents=domain_verification [email protected] --yes |
Held, approved; a one-time link on https://acme-workspace-acme.id.localhost/setup/… (on the root in the first run, where it said "expired"). Opening it and pressing Open setup lands on the portal for Globex Corporation. The mail was not sent: a sandbox environment suppresses outbound mail (logged as Suppressed outbound email from sandbox environment). The answer says so: emailed_to: null, email_suppressed: true |
| 3.13 | cbox id webhooks create --url=https://example.com/cbox-webhook --event-types=organization.updated --organization-id=<org> --yes |
Held, approved; the endpoint and its signing secret (once) |
| 3.14 | cbox id webhooks pause <webhook> |
active: false. Paused so nothing is delivered to a host we do not own |
| 3.15 | cbox id invitations send <org> --env=acme-workspace [email protected] --role=admin |
In the production environment, so the mail is sent: Mailpit shows "Live Operator invited you to join Initech" with an accept link on https://acme-workspace.id.localhost/invitations/… (on the root in the first run, which redirected the invitee to the root's sign-in) |
Every held step printed the binding code, and the request on /approvals ended with the
same code. That match is what the person checks; nothing else is approved.
4. MCP, with the MCP Inspector CLI
export NODE_EXTRA_CA_CERTS=$CA
inspect() { npx -y @modelcontextprotocol/[email protected] --cli "$1" --transport http \
--header "Authorization: Bearer $2" "${@:3}"; }
TOKEN=<the access token `cbox login` stored> # profiles.default.oidc.access_token
| # | Call | Result |
|---|---|---|
| 4.1 | inspect $ROOT/mcp $TOKEN --method tools/list |
252 tools: whoami, list_actions, approval_status, the workspace's, every environment tool with a required environment argument, the account's and, for an operator, the platform's. The Inspector also prints "Schema portability: 185 warnings across 89 tools": nullable fields are written "type": ["string", "null"], which is valid JSON Schema but which single-type dialects (Gemini function declarations) do not accept |
| 4.2 | … --method tools/call --tool-name keys_create --tool-arg environment=acme-workspace-acme --tool-arg "name=MCP agent key" --tool-arg 'scopes=["apps:read","apps:write","webhooks:read","webhooks:write","organizations:read","organizations:write","sso:read","sso:write","keys:read"]' --tool-arg 'require_approval={"min_danger":"critical"}' |
approval_pending, code 5717, and the next step in words |
| 4.3 | Approve on $ROOT/approvals, then --tool-name approval_status --tool-arg approval_id=<id> |
approved |
| 4.4 | 4.2 again plus --tool-arg approval_id=<id> |
The key, its token once, require_approval: {min_danger: critical} |
| 4.5 | inspect https://acme-workspace-acme.id.localhost/mcp $KEY --method tools/list |
60 tools: exactly what the key's scopes allow |
| 4.6 | … --tool-name whoami |
environment_key, the key's id, environment, issuer and scopes |
| 4.7 | … --tool-name apps_list |
Acme Web |
| 4.8 | … --tool-name sso_connections_create --tool-arg organization_id=<org> --tool-arg "name=Globex Entra" --tool-arg type=oidc --tool-arg pending_idp=true |
An OIDC draft with its redirect_uri on the environment's host. Not held: write is below the key's critical |
| 4.9 | … --tool-name apps_create --tool-arg "name=Agent-made SPA" --tool-arg type=spa --tool-arg 'redirect_uris=["https://spa.acme.test/callback"]' --tool-arg idempotency_key=smoke-spa-1 |
approval_pending (code 9AEC), shown on the key owner's $ROOT/approvals as Key "MCP agent key" wants to run apps.create |
| 4.10 | Approve, approval_status → approved, then 4.9 again with approval_id |
The public SPA client |
To connect Claude Code itself: claude mcp add --transport http cbox-id $ROOT/mcp and sign
in when it asks (the root serves MCP OAuth with dynamic client registration), or
--header "Authorization: Bearer $KEY" against an environment host.
5. The SDKs
Both scripts take the environment host, the key from 4.4 and the organization. The key
holds critical actions, so the first create waits for the approval on $ROOT/approvals.
smoke.mjs (npm i @cboxdk/[email protected]; run with NODE_EXTRA_CA_CERTS=$CA):
import { readFileSync } from 'node:fs';
import { EnvironmentClient } from '@cboxdk/id-js/management';
const env = new EnvironmentClient({
baseUrl: process.env.CBOX_ID_BASE_URL,
apiKey: readFileSync(process.env.CBOX_ID_ENV_KEY_FILE, 'utf8').trim(),
approvalPollIntervalMs: 1000,
onApprovalRequired: (approval, { action }) =>
console.log(` ${action} needs approval. Check the code there matches: ${approval.binding_code}`),
});
const apps = [];
for await (const app of env.apps.listAll()) apps.push(app);
console.log(`apps.list: ${apps.length} app(s): ${apps.map((a) => a.name).join(', ')}`);
const body = { url: 'https://example.com/js-sdk-webhook', event_types: ['organization.updated'],
organization_id: process.env.CBOX_ID_ORG_ID };
const idempotencyKey = `js-sdk-smoke-${Date.now()}`;
const first = await env.webhooks.create(body, { idempotencyKey });
const second = await env.webhooks.create(body, { idempotencyKey });
console.log(first.status, first.data.id, first.replayed, Boolean(first.data.secret));
console.log(second.status, second.data.id, second.replayed, Boolean(second.data.secret));
if (first.data.id !== second.data.id || second.replayed !== true) process.exit(1);
await env.webhooks.pause(first.data.id);
Result:
apps.list: 2 app(s): Acme Web, Agent-made SPA
webhooks.create needs approval. Check the code there matches: A1D1
webhooks.create #1: status 201, id 01M4EEQB7AJKNPBJXG49NSZYGD, replayed false, secret returned once
webhooks.create #2 (same Idempotency-Key): status 201, id 01M4EEQB7AJKNPBJXG49NSZYGD, replayed true, secret null
webhooks.pause: active false
smoke.py (pip install cbox-id-client==0.10.0) is the same calls through
cbox_id.management.EnvironmentClient(base_url=…, api_key=…, http_client=httpx.Client(verify=CA))
with idempotency_key=, and gave the same result: one approval, 201 twice, the same id,
replayed False then True, the secret only the first time.
root_token.py drives the environment with the CLI's root token instead of a key:
env = EnvironmentClient(base_url=ROOT, access_token=token, environment="acme-workspace-acme",
http_client=httpx.Client(verify=CA))
connection = env.sso.connections.create({"organization_id": org, "name": "Root-token SAML draft",
"type": "saml", "pending_idp": True}).data
link = env.organizations.portal_links.create(org, {"intents": ["sso"]}).data
# every URL in connection["service_provider"] and link["url"] must be on the environment's host
env.sso.connections.delete(connection["id"])
Result: the service-provider URLs and the portal link on
https://acme-workspace-acme.id.localhost, two approvals (the portal link and the delete:
a person's token waits for every critical action), and the SDK's approval poll answered
on the root. Before the fix below, all four URLs were on the root.
6. What this run found
Fixed in this application:
- URLs minted for an environment from the platform root pointed at the root. A SAML
connection created with
cbox id …(or the root/mcp, or an SDK withenvironment=) answered an ACS URL, entity ID and metadata URL on the root, which serves no SAML; the Admin Portal link opened "This link has expired" on the root while the same token opened on the environment's host; and an invitation mailed from the root had its accept link on the root. The action now runs with URLs pinned to the environment's issuer, mailed links included. The approvalpoll_urlstays on the host the request came to, where the token that polls it is valid. - MCP tool descriptions called every critical action destructive.
apps_createandkeys_createtold the agent they "remove or revoke something". A critical tool now says what critical means. - A refusal knowable from the input came after the approval. A webhook URL that does
not resolve publicly was checked once the person had approved, so the repeat failed with
unsafe_url. Actions now refuse what they can before the approval gate: an unsafe webhook, hook, log stream or SCIM URL, an unsafe SAML metadata URL or OIDC issuer, a foreign id, a missing owner, an incomplete SSO config, a malformed SSO domain. - The portal link said
emailed_tofor a mail a sandbox dropped. The answer is nowemailed_to: nullwithemail_suppressed: true. - The MCP server said version
1.0.0. It now reports the application's version. - A key's approval request did not name its environment. The binding message now says
Key "…" in <environment> wants to run …. - MCP tool schemas wrote nullable fields as
type: ["string", "null"], which the Inspector flagged 185 times and some clients refuse. Tool input schemas now useanyOf; the REST OpenAPI (3.1) is unchanged.
Reported elsewhere, not fixed here:
cbox-id:doctorwarns no active signing key on every multi-environment install (framework: the count runs outside any environment).cbox id environments(CLI 0.1.0) prints an emptyhostcolumn: the API answersissueranddomain.
Not possible locally, and replaced:
- Push approvals on a phone. Nothing here can reach Cbox Authenticator; every approval
was given on the console's
/approvalspage, which is the same request and the same binding code. - The in-app browser. It does not trust the development CA, so the operator's console
steps (signing in, the device page,
/approvals) were made with the console's own HTTP requests and a real session cookie. On staging use a browser. - Webhook delivery. The SSRF guard (
CBOX_ID_WEBHOOKS_VERIFY_URL) refuses a receiver on this machine, and turning it off is not part of a production-shaped run. Endpoints were created againsthttps://example.com/…and paused. On staging, point one at a receiver you own and check theX-Cbox-Signatureon a delivery. - Domain verification. Needs a TXT record in public DNS (step 3.11).