AI writing assistant

The AI Agent lives in the right-hand panel of the three-pane layout: file tree → editor → AI.

What it does

• One assistant for drafting, expanding, rewriting, and reviewing.
• ReAct-style loop: plan → call tools → stream results.
• Tools for listing and reading files, searching, applying edits, and more—see AI features.
• Read-only subagents for broad exploration, evidence-based project review, and research context.
• Phases: planning (outline/structure), writing (produce content), and review (check facts, tone, structure). You can switch per message.
• Research context: @web, @base, and pasted URLs help the Agent bring web pages and project knowledge into the draft.
• Custom Chat models: add OpenAI-compatible models in settings, then select them from the grouped model picker before sending.

Basics

1. Type your goal in the input box at the bottom of the panel.
2. Press Enter to send.
3. Watch streaming output and tool status.
4. For destructive actions (e.g. delete), confirm when prompted. Edits usually go through a Diff queue so you can accept or revert.

Writing vs review

Phase	Best for
planning	Outlines, structure, task breakdown
writing	First drafts, continuation, expansion
review	Fact checks, consistency, style, line-level fixes

A typical flow: draft in writing → polish in review → accept diffs.

@system and memory

Place project rules under @system/ (style, characters, world-building). The Agent uses them to stay on-brand. User-scoped preferences can be stored with the remember tool so they persist across sessions (see changelog).

Research and URL-to-Markdown

• Use @web when you need fresh external facts.
• Use @base when you want the Agent to search your project knowledge base.
• Paste a URL and ask the Agent to summarize, rewrite, or save useful points. The page can be converted to Markdown context.

For long technical explanations, the Agent may also return Mermaid diagrams when a flow, dependency, or timeline is easier to inspect visually.

Subagents and reports

When the task is broad, ask the Agent to investigate first:

• explore_project maps structure and compresses multi-file context.
• review_project reviews without editing and returns issues plus evidence paths.
• research_context collects project and external context with source notes.

The result appears as a tool card with execution statistics, outer iterations, a link to the read-only sub-conversation, and optional report details. Saved reports are placed under @system/@subagent/ and can be reused as reference material.

Custom Chat models

Open Settings → AI Features → Custom Chat models to add a Base URL, API key, and model ID. Run the test before saving when possible. If it fails, check whether the Base URL reaches /v1, the key has access, and the model supports streaming Chat Completions.

After saving, use the model selector in the Agent input area. Official models are listed first; custom models are grouped by provider.

Long-form stability

Recent builds improve long conversations with Markdown block caching, streaming throttling, lazy Mermaid rendering, local error boundaries for tool-call displays, and stable server-side timeline ordering for streamed messages.

Internal run-control tools such as finish_agent are silent and do not appear as user-facing tools.

Billing

AI features, web search, URL-to-Markdown, knowledge-base search, and AI Word export use your account balance (CNY). Web search and URL-to-Markdown each cost ¥0.01 per call. Custom model bills are handled by your upstream provider. See Pricing.

Learn more

• AI features (detailed)
• Getting started
• Writing Agent, Review Agent