"""AI Labs — the employee-facing playground over the org's Qwen tenancy.

Helix already pays for a DashScope / Bailian tenancy (the AI/LLM connector
that drives the support copilot, the intelligence briefs and the AI-spending
summaries). This domain wraps that same tenancy in a surface every employee
can open, so the capability is available to all ~500 people rather than only
to the flows that happen to embed it.

Surfaces
--------

* **Chat** — streaming multi-turn conversation over the catalogue's chat
  models, with optional image and document attachments. Server-Sent Events,
  because a thinking model's answer takes tens of seconds to complete and
  buffering it reads as a hang.
* **Generations** — a durable job per artifact: text-to-image, image
  editing, text/image-to-video, text-to-speech and speech-to-text. Every
  job is executed in the background and its result is re-homed into Helix's
  own S3 bucket, because DashScope's result links expire ~24h out.
* **Translate** — a single synchronous call against the dedicated
  translation model.

Two product promises are load-bearing and constrain the code
------------------------------------------------------------

1. **Helix does not read what employees write here.** There is no admin
   endpoint anywhere in this domain that returns message content, a prompt,
   or a generated artifact. Admin surfaces read
   :class:`~app.domains.ai_labs.models.AiLabUsageDaily`, which holds
   counters only. Audit rows record who / when / which capability / which
   model / how many tokens — never the text. Deleting a conversation is a
   hard delete, cascading to its messages and to the S3 objects.
2. **Helix does not touch the employee's prompt.** No system message is
   injected, no rewriting, no grounding, no tool calls. What the employee
   types is what the vendor receives, verbatim. This is asserted by a test
   (``test_prompt_passthrough_contract``) rather than left to reviewer
   vigilance, because it is the kind of invariant a later "small
   improvement" erodes silently.

Where the code lives
--------------------

* :mod:`app.domains.ai_labs.catalog` — every model id, parameter option and
  limit, in code. Each entry carries the date it was verified against the
  live endpoint; ``server/scripts/verify_ai_labs_catalog.py`` re-checks the
  whole catalogue against the configured key.
* :mod:`app.domains.ai_labs.models` — four tables (``ai_labs_*``).
* :mod:`app.domains.ai_labs.service` — conversation + generation lifecycle.
* :mod:`app.domains.ai_labs.executor` — the background job runner shared by
  the request path and the reaper cron.
* :mod:`app.domains.ai_labs.quota` — per-person daily ceilings, unlimited by
  default.
* :mod:`app.domains.ai_labs.router` — employee + admin HTTP surfaces.

The vendor clients themselves are shared platform pieces and live with the
Bailian connector: :mod:`app.domains.bailian.llm_client` (chat, streaming)
and :mod:`app.domains.bailian.media_client` (images, video, speech).
"""