This release focuses on three areas: giving you control over local AI providers, making hooks visible in your conversation timeline, and handling large sub-agent sessions gracefully.

On-Device AI settings

If you’re running a local LLM — whether through Ollama, LM Studio, or any OpenAI-compatible server — the settings experience just got a lot better.

The previous implementation was tightly coupled to oMLX, Apple’s on-device framework. That worked for Apple Silicon users running specific models, but left everyone else out. v0.30.0 replaces this with a provider-agnostic architecture:

Model selector dropdown — pick from detected models instead of typing slugs
Any OpenAI-compatible endpoint — point it at localhost:11434, localhost:1234, or wherever your local inference runs
Redesigned settings card — cleaner layout that makes connection status and model selection obvious at a glance

The underlying integration no longer manages processes for you. It connects to whatever you’re already running and gets out of the way.

Hook events in your conversation

Claude Code hooks — the PreToolUse, PostToolUse, and Stop handlers you configure in settings.json — have always run invisibly. You knew they fired because you saw their side effects (a linter ran, a file was formatted), but the hooks themselves didn’t appear in your session timeline.

v0.30.0 renders hook events as conversation blocks alongside assistant messages, tool calls, and user input. Three delivery paths ensure you see them regardless of how you’re viewing the session:

Live sessions stream hook blocks over WebSocket in real time
Historical sessions merge hook events from the database into the block timeline
JSONL replay picks them up from the raw session file

This means when you scroll through a conversation, you can see exactly when your formatting hook ran, what it did, and how long it took — without checking terminal output separately.

Sub-agent pagination

Sessions with sub-agents can generate hundreds of blocks. Previously, opening a sub-agent panel loaded every block at once — slow for long-running agents and unusable for agents that ran thousands of tool calls.

Sub-agent block views now paginate. Blocks load incrementally as you scroll, with the same stable scroll anchoring used in the main conversation view. This also fixed a bug where sub-agent WebSocket connections failed for historical sessions (99% of sessions), because the handler only checked live sessions instead of falling back through the full session lifecycle.

Under the hood

The sidecar had a memory problem: listSessions() spawned a full Claude CLI process on every call, consuming 900MB+ of RSS. Removed entirely — the sidecar now scans the filesystem directly.
The Activity page moved from client-side aggregation to a dedicated rich_activity() database function. Faster queries, less data transferred, more accurate results.
CLI plugin commands that produced more than 64KB of stdout were silently truncated due to a Node.js pipe buffering limit. Fixed by routing output through a temp file.
Session mutations and the multiplexed WebSocket hook were centralized into shared modules, and pagination was extracted to the coordinator level so every session phase gets it automatically.

What’s next

Phase classification confidence tuning — the ML classifier is approaching the precision threshold for production display
Linux and Windows binary builds in the release pipeline
Search across all sessions with the Tantivy full-text index

Update now

npx claude-view@latest

Open your settings to try the new On-Device AI card, or check any session with hooks to see them rendered inline.