LLM Agent for <perspective-viewer>

[texodus]

This PR adds a first-party LLM agent to <perspective-viewer>: a chat sidebar tab and a JavaScript API that let a model drive the viewer through its own public API: reading the schema, writing the ViewerConfig, choosing a plugin, authoring ExprTK expressions and managing panels.

The feature is opt-in at runtime. The Chat tab stays hidden and no network request is ever made until the host calls agentConfig().

import { providers } from "@perspective-dev/viewer";
import docs from "@perspective-dev/viewer/dist/docs/perspective-docs.json" with { type: "json" };

viewer.agentConfig({ ...providers.anthropic, apiKey: "sk-ant-...", docs });
await viewer.agentPrompt("Show me Sales by Region as a bar chart");

Added

Element API agentConfig(config), agentPrompt(text) (resolves with the final answer once tool calls have been applied), and agentReset(). Tool activity fires perspective-agent-tool CustomEvents.

Chat sidebar tab Streams tokens (with automatic fallback), streams reasoning-model reasoning_content into a collapsible display-only block that follows its own tail, renders assistant markdown, and shows tool calls.

OpenAI compatible A "provider" is a plain spreadable Object; src/ts/providers.ts ships presets for anthropic, gemini, openai, openrouter, lmstudio and ollama, and anything OpenAI-compatible works without one. engine accepts an in-page engine (WebLLM's MLCEngine), so the agent can run entirely on-device or be proxied to a private model.

Documentation bundle — the build emits dist/docs/perspective-docs.json (~300 KB) containing a BM25-searchable corpus built from docs/md plus authored plugin-config reference files, and generated JSON schemas for ViewerConfigUpdate / ViewerConfigInitial. Passing it as docs powers the search_docs tool and upgrades the config tools' parameter schemas from permissive objects to the real generated ones. Hosts can append their own
{title, text} entries to teach the agent additional context about the data itself, or override/audit the documentation bundle entirely.

Docs and New Examples

• docs/md/how_to/javascript/agent.md — new guide covering connection, in-page engines, the docs bundle, teaching the agent about your data.
• Substantial updates to viewer.md, events.md, save_restore.md, expressions.md, advanced.md, schema.md, FAQ.md and a new windows.md - these are both user-facing improvements and the agent's retrieval corpus, so gaps in them are gaps in the agent.
• examples/blocks/src/agent/ — a runnable demo with a provider modal (OpenAI-compatible, Anthropic, Gemini, WebLLM), dataset-specific corpus entries, and notes on the WebLLM caveats (Hermes-only tool calling).