AI Agent features

All Writer ships a single AI Agent for long-form work. It uses a ReAct loop (reason → act) with tools, respects your @system rules, and applies changes through a Diff queue so you stay in control.

For the full reference (long-form Chinese doc), see AI Agent 智能助手.

What you can do

• Draft, continue, rewrite, and structure novels, papers, docs, posts.
• Review for facts, logic, tone, and style.
• Search the project (grep), read/write files, and update structured notes.
• Use @-commands to attach files, search the web, query your knowledge base, or turn URLs into Markdown context.
• Delegate broad read-only work to subagents: explore_project, review_project, and research_context.
• Add your own OpenAI-compatible custom Chat models and pick them from the grouped model selector.
• Use Mermaid diagrams when explaining flows, structure, dependencies, or timelines.

ReAct workflow

1. Understand the goal
2. Plan steps
3. Call tools (list/read/edit/search)
4. Stream results and iterate

Phases: planning, writing, review

Phase	Focus
planning	Outline, structure, task breakdown, and next-step planning
writing	Produce and expand content
review	Check accuracy, structure, consistency, and style

Recommended: plan → write → review → accept diffs. system-extraction and document-import are one-off tasks launched from the creation wizard, project list, or Agent panel; they are not everyday input-stage options.

Safety and confirmations

• Deletes require confirmation.
• Writes/edits go through a Diff queue: accept to keep them, or reject to roll them back.
• Tool actions are shown as structured steps, so you can see when the Agent reads files, searches, edits, or handles an error.

Read-only subagents

For broad research, multi-file inspection, or evidence-based review, the main Agent can call a read-only subagent in its own sub-conversation.

Tool	Use it for	Notes
explore_project	Mapping folders, reading several files, compressing context before writing	Read-only exploration; may save a report
review_project	Reviewing project content for issues, evidence paths, and gaps	No writes, deletes, web, or memory
research_context	Collecting project and external context with sources	Web and URL-to-Markdown may be used; no writes or memory

Subagent cards show execution statistics, outer iterations, and a button to open the subagent conversation. When a reference report is saved, it appears under @system/@subagent/. These reports are reference material only; they do not override project rules or user instructions.

@system folder

Put project metadata in @system/ so the Agent reads your style, characters, world rules, or doc conventions:

MyProject/
├── @system/
│   ├── style.md
│   └── glossary.md
├── Chapter1/
│   └── intro.md

@-commands (overview)

• @filename — attach file content to the prompt.
• @web — web search when fresh information is needed; it costs ¥0.01 per call.
• @base — semantic / knowledge-base search over your materials.
• URL-to-Markdown — paste a URL and ask the Agent to summarize, rewrite, or save the page as project context; it costs ¥0.01 per call.

Streaming and edits

• Responses stream in the panel.
• Edit history messages to branch the conversation (see Chinese doc for full UX).
• Long messages use Markdown block caching and streaming throttling.
• Mermaid diagrams render lazily when they enter the viewport.
• Tool-call display errors degrade locally instead of breaking the full AI panel.
• Messages, thinking blocks, tool results, and final replies use stable server-side timeline sequence values, so refresh and pagination do not depend on local timestamps.

Custom Chat models

In Settings → AI Features, add an OpenAI-compatible Chat model with a Base URL, API key, and model ID.

• The model selector shows official models first, then groups custom models by provider.
• Provider/model tests can verify /models and streaming Chat Completions before use.
• Failed tests show a redacted request summary, HTTP status, content type, and response preview to help diagnose Base URL, key, or protocol issues.

Custom model calls use your configured upstream provider. If the same Agent run uses web search or URL-to-Markdown, each call costs ¥0.01 from your All Writer balance; knowledge-base search, AI Word export, and other platform abilities still consume account balance based on actual usage.

Billing

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

finish_agent is an internal silent control tool used to close a run. It is not shown as a user-facing tool.

Long-term memory (remember)

The Agent can persist user-scoped preferences with the remember tool. In the All Writer desktop or web app, open Settings → AI Features and use the Long-term memory section to view, edit, or delete entries. This documentation site does not store or edit memory data.

Next steps

• AI assistant — panel basics
• Getting started — install and first project
• Download — web or Windows