uapb/ai

Drop the in-container Ollama install and its 2GB per-boot model pull

Serve the site when weight downloads fail instead of exiting

Disable Ollama and Jem on HF Spaces to avoid memory issues

Add error handling for Ollama and Jem startup failures

Remove unused asset downloads from startup

Add zstd dependency for Ollama installation

Make the build context match the package architecture

Thin the panel tint so the glass actually reads as glass

Clear the tints off the head and foot

Fold the title bar into the panel, leaving one curved edge

Recents: keep the list still, page it at ten, and rename it

Slide the chat list in from the left, on glass, and leave it open

Stop Jem volunteering the tool, the lab, and Ramena's address

Type the empty panel's greeting on a loop, and take the gold off the glass

Centre her picture on the title bar

Steady the compose box, and strip the header back to her picture

Chats drawer, gold-edged panel, and answers that finish without you

Finish the package refactor: the parts still pointing at the old names

Keep many chats, named after what you opened them with

Centre the opener chips, and clear them on the first message

Never let a failed turn strand the Send button

Make the cloud line an invitation again

Reword the cloud line to reach out

Let the cloud fall instead of unwriting, and drop the opening greeting

Say the same line in the speech cloud as in the chat

Pull the speech cloud in closer to her

Open the speech cloud rightward from a fixed left edge

Let the cloud grow out of nothing, and drop it on hover

Shorten Jem opening line to her extension-assistant intro

Give canned answers the same dots and Typing… as an inferred one

Shrink the speech cloud and tuck it in, instead of hiding it

Demarcate the chat header with a black rule

Hush her speech cloud over the AquaVision deck

Match the chat header hairline to the glass instead of gold

Glass over the chat panel, lift her speech cloud, and stop her going soft

Type every reply out, and swap Delivered to Read the way Messages does

Show Delivered / Read under the last sent message, and ring Jem in gold

Shorten Jem's reading beat to two seconds

Make Jem read before she types, and fix the agent's launch path

Add 5-second delay to Jem chat responses

Fix: Clarify aquavision_tool is a library, not a backend

Update Dockerfile documentation to reflect modular architecture

Move all aquavision detection logic to aquavision_tool

Update Dockerfile to reflect new modular architecture

Revert "Remove __init__.py files"

Remove __init__.py files

Fix: restore cv2 import to server/main.py

Rename Triton to Jem throughout, restore Flask app

Restore Jem persona to jem_agent

Remove old hyphenated directories (jem-agent, aquavision-tool)