Files
agentic-mobile-control/docs/CLOUD_DEPLOYMENT.md
T
q792602257 1107ace89c
Tests / Test passed: 626
Default-enable AI Planner in Host Agent; propose cloud-planner-proxy
- Host Agent now defaults AI_PLANNER_ENABLED=true (opt-out via env),
  scoped to apps/device-host-agent/host_agent/execution.py only; the
  shared runtime.planner_config default (disabled) is unchanged.
- Add openspec proposal for cloud-planner-proxy: centralize LLM
  provider config/credentials on the Cloud Control Plane and let the
  Host Agent proxy AI Planner decisions through it instead of holding
  provider API keys locally. Proposal only, no implementation yet.
2026-07-13 20:50:35 +08:00

18 KiB

Cloud Control Plane Deployment

This guide covers the deployable Cloud API and outbound Device Host Agent. Run commands from the repository root after synchronizing the uv workspace:

uv sync --locked --all-packages

Local SQLite

SQLite is intended for local development and tests with one Cloud API process. The Cloud API uses persistent user sessions for human access and a Host-generated secret for later Host operations; no static bearer credential configuration is required.

$env:CLOUD_ENVIRONMENT = "local"
$env:CLOUD_DATABASE_URL = "sqlite:///cloud/cloud.sqlite3"
$env:CLOUD_SESSION_COOKIE_SECURE = "false"
uv run --package device-cloud-api device-cloud-api --host 127.0.0.1 --port 8001

Create the first administrator interactively, then open the Console and sign in with that account:

uv run --package device-cloud-api device-cloud-admin users create --username admin --display-name "Local Administrator" --role admin

In a second terminal, configure the local Host Agent to reach this Cloud API:

$env:HOST_AGENT_CONTROL_PLANE_URL = "http://127.0.0.1:8001"
uv run --package device-host-agent device-host-agent setup
uv run --package device-host-agent device-host-agent

At startup, the Host Agent loads device registrations from tasks/device_config.sqlite3, the same DeviceConfigStore used by the local Runtime console API. Register or update devices before starting the Host Agent, then restart it to reload changes. In Compose, HOST_AGENT_TASKS_PATH selects the host directory mounted at /app/tasks; it defaults to ./tasks.

The Host Agent only initiates outbound HTTP requests. It does not expose an inbound port.

Direct Edge Enrollment

The Host Agent's default HOST_AGENT_CONTROL_PLANE_URL is https://amcp.home.jerryyan.top. This is a single-operator home deployment default; override the environment variable for local/dev/test runs pointed at a different Cloud API.

When no cached identity exists, the Host Agent creates a random instance identifier and Host secret, registers directly with the Cloud API, persists the assigned host_id, and uses that secret for all later device enrollment, heartbeat, claim, renewal, and result calls. The Cloud stores only the secret digest. There is no configured enrollment-token or static Host-credential path.

This trades away any approval step: any caller that can reach the control plane can register itself as a new Host. There is no rate limiting or throttling on this path by design; restrict access at the firewall or reverse proxy before exposing the endpoint.

Before the Host Agent's first unattended start, create the one-time local operator account interactively:

uv run --package device-host-agent device-host-agent setup

This prompts for a username/password and writes a PBKDF2-hashed credential file to HOST_AGENT_LOCAL_ACCOUNT_PATH (default tasks/host_local_account.json), gating only this first-run bootstrap step — it is not re-checked on subsequent unattended restarts. Running the daemon's default command without a controlling terminal before this file exists fails fast with a message naming the setup step, instead of hanging on a prompt no one can answer.

Running under Compose, create the account once before docker compose up:

docker compose run --rm host-agent device-host-agent setup
docker compose up -d

Keep HOST_AGENT_IDENTITY_PATH, the local-account file, and tasks/device_config.sqlite3 on persistent storage with permissions limited to the service account. The identity file is a bearer secret: do not put it in an image, repository, log, or general backup.

Local Web Console

The Host Agent can optionally serve a small local-only web console on the edge machine: heartbeat/enrollment status, registered local devices, current assignment progress, local device add/edit/remove, a local account password change, and recent assignment/heartbeat history. It authenticates with the same local account created by device-host-agent setup above — there is no separate console credential.

HOST_AGENT_CONSOLE_ENABLED=false
HOST_AGENT_CONSOLE_BIND_HOST=127.0.0.1
HOST_AGENT_CONSOLE_PORT=8765
HOST_AGENT_CONSOLE_ALLOW_NON_LOOPBACK=false
HOST_AGENT_CONSOLE_SESSION_TTL_SECONDS=43200
HOST_AGENT_CONSOLE_HISTORY_LIMIT=200
  • HOST_AGENT_CONSOLE_ENABLED — starts the console when true; disabled by default, so existing deployments see no new listening port.
  • HOST_AGENT_CONSOLE_BIND_HOST — the address the console binds to; defaults to loopback-only.
  • HOST_AGENT_CONSOLE_PORT — the TCP port the console listens on.
  • HOST_AGENT_CONSOLE_ALLOW_NON_LOOPBACK — required opt-in before HOST_AGENT_CONSOLE_BIND_HOST may be a non-loopback address; the Host Agent refuses to start otherwise.
  • HOST_AGENT_CONSOLE_SESSION_TTL_SECONDS — sliding idle timeout, in seconds, for an authenticated console session.
  • HOST_AGENT_CONSOLE_HISTORY_LIMIT — number of recent assignment/heartbeat entries the console retains before pruning older ones.

Treat HOST_AGENT_CONSOLE_ALLOW_NON_LOOPBACK as an explicit, operator-accepted risk: the console has no built-in TLS and no rate limiting, so a non-loopback bind exposes an unencrypted login form to whatever network can reach that port. To reach the console from another machine instead, keep it bound to loopback and open an SSH local port-forward to the edge machine:

ssh -L 8765:127.0.0.1:8765 user@edge-host
# then open http://127.0.0.1:8765 from the local browser

PostgreSQL Deployment

Start from .env.example, replace every change-me-* value, and keep the resulting .env file outside version control. The Compose stack contains PostgreSQL, the Cloud API, and one Host Agent:

docker compose up --build -d
docker compose ps

The Cloud API container waits for PostgreSQL, runs the committed Alembic migrations, then starts the API on port 8001. Readiness is available at:

GET http://127.0.0.1:8001/health/ready

Deploying A Jenkins-Built Image

docker compose up --build above builds the image from source on the target host. For environments that pull a pre-built image instead, Jenkinsfile runs the pytest suite (-m "not integration"), builds the image from the same Dockerfile, smoke-tests both console scripts (device-cloud-api --help, device-host-agent --help), and pushes <REGISTRY>/<IMAGE_NAME>:<BUILD_NUMBER>-<git short sha> plus :latest to the configured registry.

compose.deploy.yaml contains only PostgreSQL and the Cloud API; Host Agents run at their edge sites rather than beside the Cloud API. It references the fixed Jenkins registry image and interpolates only IMAGE_TAG (see .env.example). Deploy without a local build step:

docker compose -f compose.deploy.yaml pull
docker compose -f compose.deploy.yaml up -d
docker compose -f compose.deploy.yaml ps

For a deployment that manages processes outside Compose, apply migrations before starting the new Cloud API version:

$env:CLOUD_DATABASE_URL = "postgresql+psycopg://USER:PASSWORD@HOST:5432/device_cloud"
uv run alembic -c packages/cloud-platform/cloud/migrations/alembic.ini upgrade head
uv run --package device-cloud-api device-cloud-api --host 0.0.0.0 --port 8001

Production startup requires CLOUD_ENVIRONMENT=production, a current schema, and HTTPS before browser sessions are exposed. It does not require credential JSON. Create the first administrator interactively after startup, before opening the Console to operators.

Authentication And Host Identities

Cloud users are the human authorization boundary. Their viewer, operator, and admin roles map to the existing API scopes, and the browser sends an HttpOnly session cookie plus CSRF proof for unsafe operations. The deployment does not accept CLOUD_PUBLIC_CREDENTIALS_JSON, CLOUD_HOST_CREDENTIALS_JSON, or CLOUD_ENROLLMENT_TOKENS_JSON.

A fresh Host sends its generated candidate secret only during direct registration. The Cloud stores its digest and returns a host_id; later Host operations use that secret and are strictly bound to the returned host_id. Protect the persisted Host identity file as a bearer secret.

Dynamically enrolled Host credentials are stored as digests in the cloud database. This release exposes repository-level revocation rather than a public administration endpoint. An operator with database deployment access can revoke a Host without deleting its task history:

export HOST_ID="host-..."
uv run --package device-cloud-platform python - <<'PY'
import os
from cloud.database import CloudDatabase
from core.models import utc_now

database = CloudDatabase(os.environ["CLOUD_DATABASE_URL"], create_schema=False)
try:
    changed = database.repository.revoke_enrolled_host(
        os.environ["HOST_ID"],
        revoked_at=utc_now(),
    )
    print("revoked" if changed else "not an enrolled host")
finally:
    database.close()
PY

Cloud Console (Web UI)

The repository ships an independent Vue 3 + Vite SPA at cloud-console/ that renders task history, devices, hosts, and plugins. Human operators sign in with a username and password; the Cloud API creates an expiring, revocable HttpOnly session cookie and uses a separate CSRF cookie/header for writes. The Console has no bearer-token fallback or user-directory view.

HTTPS and session configuration

CLOUD_ENVIRONMENT=production requires CLOUD_SESSION_COOKIE_SECURE=true. Terminate TLS at a reverse proxy and open the same-origin Console through HTTPS, for example https://cloud.example.com/console/. Direct http://host:8001 access is for local/test mode only; it cannot retain production login cookies.

The defaults are an 8-hour idle session TTL, 7-day absolute TTL, and a temporary block after five failed logins in a 15-minute username/client-address window:

CLOUD_USER_SESSION_IDLE_SECONDS=28800
CLOUD_USER_SESSION_ABSOLUTE_SECONDS=604800
CLOUD_LOGIN_FAILURE_LIMIT=5
CLOUD_LOGIN_FAILURE_WINDOW_SECONDS=900
CLOUD_LOGIN_BLOCK_SECONDS=900
CLOUD_SESSION_COOKIE_SECURE=true
CLOUD_TRUST_PROXY_HEADERS=false

Set CLOUD_TRUST_PROXY_HEADERS=true only when a trusted proxy overwrites X-Forwarded-For before requests reach the Cloud API. This affects the client-address bucket used for login throttling; the default is safe when the application is reached directly.

Create and recover administrator accounts

After migrations and Cloud API startup, create the first account interactively:

docker compose exec cloud-api \
  device-cloud-admin users create \
  --username admin --display-name "Cloud Administrator" --role admin

The command prompts twice for the password, so it does not enter shell history, Compose configuration, process arguments, logs, or container inspection output. Recovery commands are also interactive:

docker compose exec cloud-api device-cloud-admin users reset-password --username admin
docker compose exec cloud-api device-cloud-admin users enable --username admin
docker compose exec cloud-api device-cloud-admin users revoke-sessions --username admin

Roles are fixed: viewer can read tasks/pool/plugins; operator additionally submits tasks; admin has unrestricted Cloud API access. The administration CLI creates users, resets passwords, changes roles, enables accounts, and revokes sessions. New and reset users must change their temporary password before accessing other resources, and the API will not disable or demote the last enabled administrator.

Configure the CORS allow-list

The Cloud API has no CORS middleware by default. Before a browser can call it cross-origin, set CLOUD_CONSOLE_CORS_ORIGINS to a comma-separated allow-list that includes the exact origin (scheme + host + port, no trailing slash) the operator's browser will load the console from:

# Allow a local Vite dev server
export CLOUD_CONSOLE_CORS_ORIGINS="http://127.0.0.1:5173"
# Or a deployed origin
export CLOUD_CONSOLE_CORS_ORIGINS="https://console.example.com"

Restart the Cloud API after changing this env. The middleware is added only when the allow-list is non-empty — existing deployments see no behavior change until an operator opts in. Blanket allow_origins=["*"] is intentionally not supported because browser sessions are credentialed.

Run the console

cd cloud-console
cp .env.example .env.local
# Edit .env.local if your Cloud API is not at http://127.0.0.1:8001
npm install
npm run dev

Vite prints a local URL (default http://127.0.0.1:5173). That exact origin must be in CLOUD_CONSOLE_CORS_ORIGINS on the Cloud API. For local development set CLOUD_SESSION_COOKIE_SECURE=false, then open the dev URL and sign in with a user account. The Console sends credentialed requests and attaches CSRF proof to writes.

For a production build, run npm run build and serve the resulting dist/ behind any static file server or CDN, with VITE_CLOUD_API_BASE_URL baked in at build time. The deployed origin must be in CLOUD_CONSOLE_CORS_ORIGINS.

Same-origin deployment (baked into the Cloud API image)

The Jenkins-built Docker image already carries the SPA at /app/console-static and sets CLOUD_CONSOLE_STATIC_DIR in the image itself. The Cloud API serves the Console at /console/ (visiting / 307-redirects there). Put that origin behind an HTTPS reverse proxy, then open for example https://cloud.example.com/ directly — no separate dev server, static host, or CORS allow-list is needed because the SPA and API share one origin.

The SPA shell (index.html, JS, CSS) is served without credentials by design so it can render the login page. All /v1/* resource calls remain scope-gated, and unsafe cookie-authenticated calls require CSRF proof. The browser receives only the non-secret CSRF value; it never receives the HttpOnly session secret.

Source-based local development leaves CLOUD_CONSOLE_STATIC_DIR unset; the mount remains conditional on that image-provided setting.

Jenkins build args (NODE_IMAGE, NPM_REGISTRY, UV_IMAGE, APT_MIRROR, UV_INDEX_URL) default to CN mirrors so builds don't time out pulling from Docker Hub / ghcr.io / npmjs.org / deb.debian.org. Blank any of them to fall back to the upstream.

Runtime AI Planner

The Host Agent reuses the local Runtime planner. Unlike the shared Runtime library (whose own default is the deterministic stub planner), the Host Agent defaults AI_PLANNER_ENABLED to on -- it is the actual device-control path, so goal assignments use a model unless an operator explicitly opts out. Provide provider credentials before deploying:

AI_PLANNER_PROVIDER=anthropic
AI_PLANNER_MODEL=claude-sonnet-5
AI_PLANNER_TIMEOUT_SECONDS=30
ANTHROPIC_API_KEY=<secret manager reference>

Without a valid API key, every planning step raises immediately and the task fails on its first step (no silent fallback to the stub planner). Set AI_PLANNER_ENABLED=false to opt back out to the deterministic stub planner (e.g. for offline/dev hosts with no provider credentials).

For OpenAI, set AI_PLANNER_PROVIDER=openai, choose the deployed model through AI_PLANNER_MODEL, and provide OPENAI_API_KEY. Provider credentials belong only on the Host Agent; the Cloud API does not need them.

Operational Limitations

Run exactly one scheduler-enabled Cloud API process. SQLite supports only the documented single-control-plane development mode. PostgreSQL row locking makes assignment and claim transactions safe if requests overlap, but this release does not implement scheduler leader election or claim active-active scheduler operation. Starting multiple Cloud API replicas would start one scheduler loop per replica and is outside the supported deployment topology.

Device execution provides at-least-once side-effect semantics, not exactly-once semantics. A device action can succeed immediately before the Host Agent loses its lease or its result response, after which the control plane may retry the task. Lease renewal and cooperative stop checks prevent later interruptible actions where possible, but they cannot roll back an action already sent to a device or safely terminate an in-progress synchronous driver call.

Use bounded attempts, inspect task attempt and failure metadata, and design device workflows to tolerate repeated actions when the target operation allows it. Do not use this release for operations that require a transactional exactly-once guarantee across the cloud database and an external device.

Shutdown And Rollback

For a normal shutdown, stop Host Agents first so they stop polling, interrupt later cooperative actions, finish terminal reporting where the lease remains valid, and attempt a final heartbeat. Stop the Cloud API after Host Agents have exited, then stop PostgreSQL only if the database itself is being maintained:

docker compose stop host-agent
docker compose stop cloud-api
docker compose stop postgres

For rollback:

  1. Stop all Host Agents and the Cloud API.

  2. Back up PostgreSQL or the SQLite database file.

  3. Before rolling back to an image that still requires static credentials, restore that image's matching deployment configuration and provision the required legacy credentials outside this release's Compose contract.

  4. If the previous application version cannot use the current schema, run the tested downgrade while no application process is connected:

    uv run alembic -c packages/cloud-platform/cloud/migrations/alembic.ini downgrade -1
    
  5. Restore the previous application image or checkout and start the Cloud API.

  6. Verify /health/ready, then restart Host Agents with identities compatible with the restored Cloud API.

Downgrading revision 0002 removes dynamic credential bindings and durable device enrollment mappings. It retains the revision-0001 Host heartbeat rows, pooled devices, queued tasks, attempts, and plugins.

Do not remove the PostgreSQL volume during an application rollback. Queued and attempt history are durable database state and should remain available to the restored or forward-deployed control plane.