How Vera works

Vera is an agent runtime. The model generates responses and requests tool calls; Vera supplies context, checks permissions, runs tools, and saves the conversation. You can inspect changes, stop work, or continue with another message.

Conversations and tools

A conversation holds your messages, responses, and tool results. The selected agent definition supplies instructions and a list of available tools. Your permission setting limits what those tools may do.

What the model sees

>

1

Ask for a change.

2

The model reads your message and asks to read src/map.ts.

3

Reading is allowed in ask mode, so Vera runs the tool. The file contents go back to the model.

4

It asks to edit the file. Edits inside the workspace are allowed too, so the loop runs again without stopping.

5

Next it asks to run a command. No rule allows that in ask mode, so Vera asks you first.

6

Press 1 to allow it once.

7

Vera runs bun test. The group header says Running until it finishes.

8

The tests pass. The first line of output shows here, and the whole output goes back to the model.

9

The model needs nothing else, so it answers.

Wide view1 of 9

Tool use can repeat while Vera works on the task. You can review workspace changes with /diff and stop the active turn with Ctrl+C. Messages you submit during a turn are queued, and can join the turn after its next tool step. See Queue messages for the controls that send them.

Each conversation has a scratch folder for Vera’s own working notes. When Vera keeps a task list there, the transcript shows it as a checklist with the file’s full path and a done count instead of a file diff. Other notes in that folder show as one Updated notes line with the note’s full path. Tool rows show scratch files by their full path too, so you can open them.

Models and definitions

The model generates responses and tool requests. The definition sets the instructions and available tools for a role, such as planning or investigation. Changing one does not necessarily change the other.

Use /model to choose a connected model and /agent to choose a definition. See Models and Agents and delegated work.

Permissions

The quick controls offer readonly, ask, and auto access. Open them with Shift+Tab. A definition can restrict the conversation’s permissions further; it cannot grant broader access.

Automatic approval checks rules first and uses a classifier when needed. See Automatic approval for configuration and failure behavior.

What to expect

• No harness can guarantee safety, Vera included. A model can misread a request or follow instructions hidden in a file it reads.
• Ask and auto are for normal work. Both can still go wrong, so keep an eye on what Vera does.
• Vera does not ask you to trust a new folder. The folder’s instruction files, skills, and rules load straight away. Extensions load only from your Vera home, never from the folder.

Context and customization

Each request includes the conversation’s working context. Instructions, tool results, and other loaded sources use space in the model’s context window. Open /context to inspect that usage and its sources.

You can provide project instructions, reusable skills, standing instructions, and extensions. /customize lets you browse their sources. Start with Customize Vera and Vera’s files and settings.

As the conversation grows, compaction reduces earlier context. Its limits and related controls are under /settings, then Overrides. See Context limits and compaction.

Saved and running work

Vera’s resident host owns live conversations and tool execution. The terminal and other clients display that state and send your actions to the host. Closing a client does not necessarily stop the work it was displaying.

Use /resume or Ctrl+E to switch, choosing whether the current work stops or keeps running. See Saved conversations and Switch conversations.

Delegated work runs in child conversations. /subagents opens their list; /parent returns to the parent conversation.