Ops Assistant

Chat control (assistant add-on)

Chat control lets users type what a mower should do ("mäh morgen früh nur hinten", "send it to the station"). The optional service assistant turns the request into a work order of validated actions; the user confirms it in the web app, which then sends the actions through the normal command and schedule endpoints. Design and safety rules: ADR 0075. For users: guide.

It is off unless you enable it. Without it nothing changes and no chat box appears.

What it needs

• No secret, no database, no NATS user. The assistant only calls MowDeck's HTTP API (Identity, Fleet, Telemetry, Vision) with the token of the user who asked.
• An AI provider for the default interpreter provider: each user's own provider (Settings, AI provider) or the installation's model (VISION_DEFAULT_*, plan feature ai_recover_hosted), the same as AI assist (ADR 0033). Users without one see a link to the provider settings. Vision limits the requests per user and hour (VISION_STRUCTURED_PER_HOUR, default 30) and records them in its cost accounting with kind structured.
• Or the interpreter phrases: a built-in phrase parser for a few German and English requests, without any AI. Demo mode always uses it.

Docker Compose

In infra/compose/.env:

COMPOSE_PROFILES=assistant            # comma separated with other profiles, e.g. relay,assistant
IDENTITY_EXTRA_MODULE_URLS=http://assistant:8000
ASSISTANT_INTERPRETER=provider        # or phrases
ASSISTANT_REQUESTS_PER_MINUTE=20
ASSISTANT_AI_TIMEOUT_SECONDS=60

Then docker compose up -d --wait. Identity lists the assistant's UI area within a minute (it polls missing module URLs every 30 s); users reload the app once to see the chat box on the mower page.

To remove it, take assistant out of COMPOSE_PROFILES and its URL out of IDENTITY_EXTRA_MODULE_URLS, then docker compose up -d --wait --remove-orphans.

Helm

services:
  assistant:
    enabled: true
config:
  ASSISTANT_INTERPRETER: provider

The chart adds the assistant to Identity's module list and to the Ingress (/api/v1/assistant) only when it is enabled.

Settings

Variable	Default	Meaning
ASSISTANT_INTERPRETER	provider	provider: the user's AI provider through Vision; phrases: the phrase parser without AI. Demo mode forces phrases.
ASSISTANT_REQUESTS_PER_MINUTE	20	Proposals per user and minute (per replica).
ASSISTANT_AI_TIMEOUT_SECONDS	60	Timeout of one structured answer from the AI provider.
VISION_STRUCTURED_PER_HOUR	30	Vision: structured answers per user and hour (the AI cost bound).

The service-internal URLs (MOWDECK_ASSISTANT_IDENTITY_URL, ..._FLEET_URL, ..._TELEMETRY_URL, ..._VISION_URL) default to the Compose service names; the chart sets its own.

Demo

A public demo (ADR 0068) takes no Compose profiles, so the demo stack does not run the assistant. If you run it next to a demo yourself, it uses the phrase parser regardless of ASSISTANT_INTERPRETER and calls no external AI.

---

Mirrored from docs/ops/assistant.md at commit ea8b3955b591ac0393dfcfda93ecacdeb6b8f1c3. Edit the file in the repository, not this wiki page.