WorldScript Studio's Global AI Copilot is an always-present ambient-intelligence layer that surfaces proactive manuscript insights and lets you chat with an AI writing assistant — all without leaving the editor.
Settings → Experimental → Enable Global AI Copilot (enableGlobalCopilot). On by default.
Click the ✦ AI Copilot FAB in the bottom-right corner (or press the registered keyboard shortcut). The panel can run in two modes:
| Mode | Description |
|---|---|
| Dialog (default) | Floating panel above the page content, bottom-right |
| Sidebar | Fixed right edge, full viewport height (desktop ≥ 768 px only) |
Toggle between modes with the dock/float icon in the panel header. The preference is saved in localStorage (copilot.mode).
The Copilot continuously analyses your project with the Heuristic Engine and surfaces findings in the Manuscript Insights section of the panel (collapsed by default). Findings are colour-coded by severity:
- 🔵 Info — suggestions for improvement
- 🟡 Warning — patterns that commonly weaken manuscripts
- 🔴 Error — structural issues likely to confuse readers
Clicking an insight's Tell me more button pre-fills the chat with a context-aware question. Clicking Open view navigates directly to the relevant editor view.
Clicking an inline annotation badge in the manuscript editor opens the panel and auto-expands the Insights section so you immediately see the relevant findings.
Toggle the Heuristics Only switch (brain icon in the panel header) to disable all AI calls. In this mode the Copilot replies using only the local heuristic engine — useful when you want to keep data on-device or have no API key configured.
The Copilot is context-aware: it knows the current view, your project title, word count, character and world counts, and outline completeness. The system prompt is rebuilt on every message so suggestions stay relevant as you navigate.
Dynamic suggestion pills adapt to your current view and project state (e.g. Plot Board shows tension-gap suggestions; Manuscript with >500 words shows pacing and opening suggestions).
When the ProForge flag is also enabled, typing a phrase like "run a diagnostic" triggers the ProForge Intake stage and returns a quality score summary. Each review item in the ProForge Review Panel also has an ✦ Ask Copilot chip that pre-fills the chat with the item's context.
When the last assistant message contains a fenced code block and a manuscript chapter is open, an Apply to chapter button appears. Clicking it rewrites the active chapter content with the code-block text — undoable via Ctrl+Z (action dispatched into redux-undo). The button is only shown when the block covers ≥ 70 % of the existing chapter length to prevent accidental overwrites from partial snippets.
CopilotLauncher.tsx FAB + panel mount
└── CopilotPanel.tsx dialog/sidebar shell, focus trap, mode toggle
├── InsightSection.tsx heuristic findings, collapsed by default
│ └── InsightCard.tsx per-finding card with Tell me more / Open view
├── CopilotMessageList.tsx conversation transcript, markdown rendering
└── CopilotComposer.tsx textarea + send button, draft pre-fill
hooks/useGlobalCopilot.ts all business logic (insights, streaming, apply)
services/copilot/
heuristicEngine.ts 8 built-in rules, pure TypeScript
insightGenerator.ts 400 ms debounce, LRU cache (10 entries)
copilotContextService.ts system prompt + CopilotContext builder
copilotActions.ts intent detection (chat vs. diagnostic)
actionApplier.ts safe manuscript mutation (applyTextEdit)
features/copilot/copilotSlice.ts ephemeral chat state (Redux)
app/transientUiStore.ts panel overlay state (Zustand)
See docs/HEURISTIC-RULES.md for the full rule catalogue.