feat(omni): Slack, WhatsApp, RCS, SMS and iMessage as Stream Chat messages

[tschellenbach]

Why

Sprint 4: an agent should answer Slack, WhatsApp, RCS, SMS and iMessage from one Stream channel per conversation. This adds vision-agents-plugins-omni (from vision_agents.plugins import omni).

Each provider reads its webhook into an OmniMessage, and to_stream turns that into a Stream MessageRequest with attachments and custom.omni. In the other direction, from_stream reads the agent's Stream reply back, and the provider's render builds the bodies for its send API.

Changes

• Providers per channel. A channel can be carried by more than one provider. ProviderRegistry holds the providers by name and dispatches parse(name, payload) and render(message) on message.provider. It rejects a channel that provider doesn't carry.

  Provider	Channels
  slack	Slack
  whatsapp (Meta Cloud API)	WhatsApp
  google_rbm	RCS
  twilio	SMS, plus WhatsApp and RCS when the sender has a whatsapp: or rcs: prefix
  linq	iMessage, RCS, SMS, taken from each message's service

  To add a provider, subclass OmniProvider with a unique name and register it.

• Message model. OmniMessage / OmniAttachment with kinds image, video, audio, file and location. Messages record their provider, which is also stored in custom.omni.

• Stream mapping.

  • Message text stays the Stream text.
  • Files become image (image_url) or video / audio / file (asset_url) attachments, with mime_type, file_size and media_id on the attachment.
  • A shared place becomes a location attachment with latitude and longitude.
  • The readable ids (channel, provider, conversation, sender, account, thread, reply_to, sent_at) go in custom.omni.
  • from_stream(reply, route=inbound) sends a reply back the way the inbound message came, and skips ai_reasoning and ai_tool_call attachments.

• Stream ids.

  • stream_channel_id is {provider}_{first 24 hex chars of a sha256 of account and conversation}. That's one channel per conversation per business number or account, and always at most 64 characters.
  • stream_user_id is {provider}_{digest of the sender}.

• Slack. The provider remembers recent (channel, ts) pairs. A message delivered as both message and app_mention, or retried by Slack, is read once.

• Workspace. The package is registered in the workspace, with a vision-agents[omni] extra. Its uv.lock entries are written by hand: uv lock doesn't resolve on accelerate because of an existing moondream/roboflow pillow conflict. UV_FROZEN=1 uv sync installs it.

Only pure translation is included. There are no HTTP senders, no credentials, and no webhook signature checks.

UV_FROZEN=1 uv run --no-sync dev.py check: ruff, the extras validation, and mypy for the core package and plugins all pass. Unit tests: 1689 passed, including the 45 omni tests. 44 tests error, and they error the same way without this change: the Redis testcontainer fixture needs Docker, which this VM doesn't have.