v0.11.3 - context usage and local runtimes

What changed

• Added a context usage meter to the web chat header with used, maximum, and remaining token counts.
• Added runtime and catalog-based context-window and model-capability discovery, plus explicit admin overrides for local caps and unknown models.
• Persisted assistant usage metadata so the meter survives reloads, device changes, and export/import round trips.
• Corrected multi-step tool-loop accounting so repeated prompt input is not double-counted in the context meter.
• Added first-class vLLM, llama.cpp, and SGLang presets with conventional local endpoints, optional API keys, editable connection details, discovery support, and provider icons.
• Fixed title generation when a model emits a reasoning trace before its title text.
• Updated the default Compose app image to 0.11.3.

Upgrade note

Migration 0006 runs automatically on startup. It adds nullable message-metadata and model-discovery columns; existing chats and model configurations remain valid, and no manual backfill is required.

The context meter is enabled by default and can be disabled per device under Settings → General. A context-window override is available under Settings → Models → Advanced.

The local runtime defaults are http://localhost:8000/v1 for vLLM, http://localhost:8080/v1 for llama.cpp, and http://localhost:30000/v1 for SGLang. Docker deployments may need host.docker.internal instead of localhost when inference runs on the host.

Validation

• 230 web tests and 8 site tests
• Typecheck, lint, and production build
• Migration generation and schema review
• Docker Compose configuration validation
• Five GitHub validation and Playwright smoke jobs on the release PR

Image

• ghcr.io/yoloyash/overtchat-app:0.11.3
• ghcr.io/yoloyash/overtchat-app:latest
• Platforms: linux/amd64 and linux/arm64

Full Changelog: v0.11.2...v0.11.3