[Randomname653]

Why: start.bat checks and builds the models before the desktop server starts. The container starts the server directly, so a fresh NAS install came up with no bakes until someone opened a shell. Don, rightly: that's nonsense for a container.

services/model_autobuild.ensure_main_models

• When: with AUTO_BUILD_MODELS (set in the Dockerfile), the lifespan starts it in the background on a configured install.
• What: builds what the main host is missing, from host_plan("main"). The server answers right away; a first build is 20+ GB, far past any health check.
• Logging: what it builds, and what failed (ERROR). A split install's PC bakes that are missing are only named, never pulled onto the gaming PC unasked.
• Embedding identity: reports it, and logs ERROR on a weights mismatch.
• Never raises. Model calls fail closed until the bakes exist, as before.

Also

• setup_wizard.build_ollama_models(embed_model=): callers on the event loop pass the model, so no database read happens on the loop.
• The Settings "build models" button (/api/setup/build-models) follows host_plan("main"): no 17.7 GB pitcher on a split install's NAS, and the host's own window.
• docs/NAS_MOVE.md step 6: start the container and watch the log, instead of a manual build.

Tests: new tests/test_model_autobuild.py (13 checks):

• the four outcomes: Ollama down, nothing missing, build with the plan (embed model passed), failed build
• the split PC only named
• mismatch is an ERROR; never raises
• the wiring: lifespan, flag default off with the Dockerfile on, the button

test_agent_only follows the reworded runbook. The full battery is green (228/228).

🤖 Generated with Claude Code