Workspace and Timeline

The central idea in Agent Mode is the workspace.

In ordinary generation, the model usually returns text directly. Agent Mode works differently: the Agent first handles files inside a constrained workspace, then submits a chosen output to the chat.

This lets TauriTavern turn generation into a more visible process. What was read, what was written, what was changed, and when the final reply was submitted can all be traced.

Why a Workspace Exists

The workspace solves four practical problems:

Problem	What the workspace changes
Intermediate drafts have nowhere to live	Agent can draft first, then revise into the final reply
Tool results pollute chat history	Tool results stay in the run process instead of becoming chat messages
Model behavior is hard to understand	The timeline shows key reads, searches, edits, and commits
Output is more than body text	Status blocks, summaries, notes, and side material can have stable locations

This is one of the most important differences between TauriTavern Agent and traditional tool calling: a tool call is a run event, not a chat message.

What Is in the Workspace

The current Agent Profile can decide which workspace roots are visible and writable. Common roots are:

Root	Purpose
output/	Final output. The default message body is output/main.md
scratch/	Temporary notes, drafts, and working material
plan/	Planning files. The directory is reserved, though full Plan Mode runtime is not yet open
summaries/	Summaries and reusable stage notes
persist/	Durable information that later runs in the same chat may continue to use

These paths are logical paths inside the Agent workspace, not system file paths. Agent cannot use ../, absolute paths, or drive prefixes to reach outside the workspace.

WARNING

persist/ is best for short, stable information that remains useful later, such as unresolved plot threads, relationship state, or user writing preferences. It is not a good place for complete chat history, model reasoning, or large temporary tool results.

How Output Enters the Chat

The default output file is:

output/main.md

When Agent has finished writing, it calls the commit tool to submit this file as the chat message. The chat shows the final content, while the workspace and timeline retain the process for this run.

This gives two benefits:

• The chat stays clean and only contains the reply the user is meant to read.
• The Agent process remains inspectable for understanding and troubleshooting.

What the Timeline Shows

The timeline is a user-facing projection of the run journal. It does not expose every internal detail, but it shows the steps that matter to users.

Common timeline items include:

Item	Meaning
Read	Reads chat, world info, SKILLS, or workspace files
Search	Searches chat or files for relevant snippets
Write	Creates or rewrites a workspace file
Patch	Makes a precise replacement in an existing file
Commit	Submits an output file to the chat
Persist	Saves durable workspace changes for later runs in the same chat
Done	The Agent run finished normally
Failed	A profile, model, tool, or save operation failed

The timeline is not meant to make you read every low-level detail. It is there so Agent behavior is not a black box.

What Can Appear in Details

Some events can be expanded. For example:

• Search result snippets.
• Workspace files that were read.
• Tool call results.
• Output content after a write.
• Model responses or error summaries.

Details are read from files associated with the current run. They still follow workspace path restrictions and do not expose system files.

Cancelled, Failed, and Completed

The timeline distinguishes several states:

State	Meaning
Running	Agent is still in a model call or tool loop
Completed	Agent submitted successfully and ended
Cancelled	The user or system requested cancellation
Failed	A non-recoverable error occurred

Cancellation is not failure. Failure usually means a required condition was not met: the model did not return the required tool-call shape, the profile was invalid, a path was illegal, or the chat commit failed.

TauriTavern tries to show these problems clearly instead of turning them into a reply that only looks normal.