Conversation

Overview

Work's conversation is driven by the Agent Loop. After the user sends a message, the Agent enters an autonomous reasoning-execution loop: analyze the requirement → call a tool → obtain the result → continue reasoning, until the task is complete or an exit condition is met. The entire process is presented on the interface in real time via streaming events.

Message Types

The conversation message flow includes the following types:

Message Type	Rendering	Description
User Message	Right-aligned bubble	Text and attachments sent by the user
Assistant Text	Left-aligned Markdown rendering	The Agent's text reply, with support for code highlighting, math formulas, and tables
Tool Call	ToolCallGroup collapsible panel	Tool name + input parameters + execution status + output result
Thinking Process	Thinking collapsible block	The Agent's reasoning process (native thinking models such as Claude)
Task Management	TaskPanel task panel	Structured task list (To Do / In Progress / Done)
Subagent Activity	SubagentToolRow	Subagent execution process, with special markers
Question Interaction	QuestionWizard	Structured questions the Agent poses to the user

Tool Call Display

When the Agent calls a tool, the interface displays the ToolCallGroup component:

Tool Call Display-CN

Display Element	Description
Tool Name	The identifier of the tool being called (such as Bash, Read File)
Input Parameters	The parameters received by the tool (collapsible for viewing)
Execution Status	loading (spinning icon) / success (green check) / error (red cross)
Output Result	The content returned by the tool (long content is collapsible)

Multiple tool calls are grouped and displayed in execution order. Tools executed concurrently are shown in the same group.

Streaming Output

The Agent's replies are pushed to the interface in real time via streaming events:

Event Type	Description
`text_delta`	Text fragment, appended character by character to the message area
`thinking`	Thinking process fragment (supports Anthropic native thinking)
`tool_use_start`	Tool call starts, displays the tool name and parameters
`tool_result`	Tool execution result, updates the status icon
`error`	Error message
`done`	This round is complete

Users can see the Agent's real-time reasoning and operations during streaming output, providing a transparent execution experience.

Session Management

Creating a New Session

Click the "New Conversation" button in the top-left corner or use the keyboard shortcut to create a new session. A new session starts from the welcome page, or you can enter a message directly to start the conversation.

Session List Operations

Operation	Description
Search	Filter session titles by keyword
Pin	Pin important sessions to the top of the list
Folder Grouping	Organize sessions into custom folders
Delete	Remove a session (unrecoverable)
Rename	Modify the session title

Session Snapshots and Recovery

The APP supports session snapshot persistence to prevent conversation loss caused by accidental interruptions:

Snapshot Strategy	Description
`every_tool_round` (default)	Save after each tool call
`every_round`	Save after each round of LLM interaction
`manual`	Save only when triggered manually

Snapshots are stored as JSON files (atomic writes, .tmp + rename to prevent corruption) and can be used to recover an interrupted session via resumeSessionId.

Local Conversations and Remote Conversations

In multi-node scenarios, conversations in the session list display different source icons:

Icon	Meaning
Computer icon	Conversations initiated or executed by a local APP node
Server icon	Conversations executed by a remote node
Platform icon	Conversations from IM channels (Telegram, WeChat, etc.)

This helps users quickly distinguish the execution source of a conversation.

Agent Loop Complex Task Capability

The Agent Loop enables Work to handle complex tasks that require multiple steps:

Example scenario: The user requests "Analyze the project directory structure and generate architecture documentation"

Round 1: Agent calls List Directory to browse the project structure Round 2: Agent calls Glob Files to search for key configuration files Round 3: Agent calls Read File to read the content of core files Round 4: Agent calls Search Files to find module dependencies Round 5: Agent synthesizes the analysis and generates the Markdown architecture documentation Round 6: Agent calls Write File to save the document to the specified path

                      
                      Round 1: Agent calls List Directory to browse the project structure
Round 2: Agent calls Glob Files to search for key configuration files
Round 3: Agent calls Read File to read the content of core files
Round 4: Agent calls Search Files to find module dependencies
Round 5: Agent synthesizes the analysis and generates the Markdown architecture documentation
Round 6: Agent calls Write File to save the document to the specified path

This code block in the floating window

Throughout the process, the Agent autonomously decides which tools to call and in what order to execute them, without the user needing to provide step-by-step guidance. The default maximum is 25 loop rounds, and a circuit breaker mechanism ensures automatic stopping in case of anomalies.

For technical details, refer to Agent Loop Engine.

Tips: How to Communicate with the Agent More Effectively

Tip	Example	Effect
Be clear about the goal	"Help me refactor the formatDate function in utils.ts to use date-fns"	The Agent accurately locates the target and approach
Request step by step	"First list all .tsx files, then tell me which ones use class components"	The Agent executes step by step, with more controllable results
Specify the tool	"Please use Bash to run `npm test`"	The Agent directly uses the specified tool
Ask for confirmation	"Tell me your plan first, and execute only after I confirm"	Prevents the Agent from modifying files directly
Provide context	"This is a React + TypeScript project built with Vite"	The Agent gives more relevant suggestions

How to Abort a Task in Progress

When the Agent is performing an operation you did not intend:

Click the Stop button at the bottom of the message flow (or press Esc)
The Agent will stop after the current tool execution is complete
Operations that have already been executed (such as files already created) will not be automatically undone
You can continue the conversation to instruct the Agent to undo the previous operations

Note: A Bash command in progress will be terminated (SIGTERM is sent), but the effects of commands that have already completed cannot be automatically undone.

What to Do When a Tool Call Fails

Failure Reason	Solution
Insufficient permissions	The Agent will explain the required permission; you can allow it in the popup, or adjust the tool's security level
File does not exist	Check whether the path is correct; you can have the Agent first use Glob / List Directory to confirm the file location
Command execution error	The Agent usually analyzes the error and retries automatically. If it keeps failing, manually provide more context
Network timeout	Network tools such as Web Fetch may fail due to network issues; just retry later

Multimodal Input and Output — File input and preview capabilities
Subagent — Task decomposition and delegation
Agent Loop Engine — The underlying loop mechanism
5-Layer Context Compression — Long conversation context management