Conversations

A conversation is a single chat thread. This page covers solo conversations, per-conversation settings, slash commands, attachments and keyboard shortcuts. For multi-agent group conversations, see Multi-agent teams.

Two kinds of conversations

Both kinds have the same conversation settings. Group chats also have per-member settings.

Creating a solo conversation

1. On the home screen, click New Chat (or press Ctrl+N).
2. The conversation opens with the default provider and model from your global settings.
3. Type your message. Press Enter to send, Shift+Enter for a newline.

Per-conversation settings

Each conversation has its own settings, separate from the global defaults. Open Chat Settings (the gear button in the chat header, or Ctrl+Shift+S) to change them.

Provider and model

Override the global default provider and model for this chat only, in the Model section. Useful when you want one conversation on cloud and another on local.

Pro-tip: When you switch providers mid-conversation, the new provider is used for the next message. Earlier replies are not re-run.

System prompt

A free-text instruction prefixed to every request sent on this conversation. The agent receives it as the system role.

Note: The system prompt is conversation-scoped, not global. Other conversations are unaffected.

Temperature

A value from 0.0 to 2.0 that controls randomness. New conversations use the model's own default until you move the slider.

• 0.0: most predictable.
• 0.7: a common value for chat.
• 1.0 and above: more varied and less predictable.

Max tokens

Upper bound on the assistant's response length. Auto (no cap), the default for new conversations, lets the model decide its own stopping point. This is usually what you want, especially with thinking models whose reasoning shares the same budget. Explicit values are passed to providers that support the parameter. The built-in llama.cpp engine is the exception: on Auto it caps a reply at 1,024 tokens.

Streaming

When enabled (default), tokens stream into the bubble as they arrive. When disabled, the full reply appears at once. Disable only if your provider does not support streaming (rare).

Thinking mode

For models that support a separate reasoning pass before answering (qwen3, DeepSeek-R1, Claude with extended thinking, Gemini 2.5, OpenAI o1 / o3, gpt-oss-thinking, etc.), the Thinking toggle (in the message bar, or Thinking Mode in Chat Settings) tells the model whether to reason before it produces the visible reply. Off by default.

• On: the model thinks before answering. Replies are usually slower but tend to be sharper on planning, multi-step problems, careful tool-call selection, and anything that benefits from internal deliberation.
• Off: the model skips the reasoning pass (where the provider supports an explicit switch) and answers directly. Faster and lighter, well suited to small talk and obvious questions.

The reasoning the model produces is rendered as a collapsed Reasoning disclosure inside the assistant bubble: a single muted row labelled ▸ Reasoning (N chars) above the reply. Click the row to expand and see the full thinking text in a muted scrollable panel; click again to collapse. The disclosure appears only when the model produced reasoning; replies from other models look the same as before.

When the model produces reasoning but no visible reply (some reasoning models go silent after a long thinking pass on a hard follow-up to a tool result), the bubble carries an italic muted note ("The model produced reasoning only. Expand below to see it.") and the disclosure starts expanded so you can read the reasoning.

Reasoning is shown to you only. It is never re-sent to the model on later turns (so it doesn't double your per-turn context cost), never indexed for search, and not included in exports. The toggle asks the model to reason before answering, and any reasoning it produces is shown in the collapsible disclosure.

Verzeta sends the toggle to every provider that exposes a per-request thinking parameter, choosing the right shape per provider:

Use app-recommended sampling

Some models ship with factory settings that misbehave in long multi-agent chats (the best-known case: Qwen 3.5/3.6 on Ollama cutting replies off mid-sentence). For those models Verzeta carries a tested sampling recipe and applies it automatically. This checkbox (visible only when a recipe exists for your current model) controls that.

• Checked (default): the recipe overrides the advanced sliders below at send time. Recommended.
• Unchecked: your own slider values win. The recipe only fills in sliders you left on Auto.

Beneath it sit the advanced sampling sliders (Top-K, Top-P, Repeat / Presence / Frequency penalty), each with an Auto position that means "use the model's own default". Leave them on Auto unless you know what you're tuning. See Troubleshooting for the Qwen story.

Context window

The model context size Verzeta budgets requests against (and passes to Ollama as its window size). Bigger windows fit more conversation before compaction is needed. Group chats should use 16k or more when the model and hardware allow.

Tools

When on (the default), the model can call tools (files, shell, web search, MCP tools). When off, tools are hidden from the model for this conversation. You can also switch this with the Tools toggle in the message bar.

Tool steps per turn

How many tool calls an agent may chain in one turn before it pauses and asks you to continue. Default 25; 0 means unlimited. This setting applies to all conversations. A separate check always stops a turn after 5 tool calls fail in a row.

Describe tools in system prompt

Default off. Models receive their tool definitions through a structured channel either way; this switch also adds a written list of every tool to the agent's instructions. That written list is redundant for modern models and costs a meaningful slice of every request (roughly 3,000 tokens on a full tool roster) that is better spent on conversation history. Turn it on for a conversation only if a smaller model keeps "forgetting" its tools; some behave better with the written list. Conversations created before this default changed keep whatever setting they had; flip the switch in Chat Settings to update them.

Dynamic compaction

Default on. Verzeta summarises the older part of a long conversation in the background (before the model's context window runs out) so decisions and assignments stay in the model's view. Your visible transcript is never altered. A ~Dynamic Compact Performed~ entry appears in the chat each time.

Beneath the toggle, "Refresh summary every N agent replies" (default 20, 0 = off) also refreshes the summary on a fixed rhythm, whether or not the context is full. This helps because long team exchanges drift even when everything still fits.

Full explanation, the context gauge, and the manual /compact and /flashmemory commands: see Conversation memory.

Auto-complete tasks when the model goes quiet

Default off. When on, an open task is marked done automatically if the model stops right after a tool call. Leave it off for multi-part work, so the task stays open until someone completes it. See Canvas and tasks.

Agent pattern

Affects how the conversation orchestrates the assistant. The Pattern list in the Agent section shows these values:

Note: Most people should keep direct. The other patterns are experimental.

Require confirmation

When enabled, you are asked to confirm every tool call before it runs. Useful when an agent has shell or file write access and you want to check each step yourself.

Team autonomy: rounds before pausing

In group chats, how many autonomous rounds the team runs (agents taking turns and @mentioning each other) before pausing to check in with you. Default 6; raise it (up to 50) for more independence, or set it to Unlimited to let the team keep going until it stops making progress on its own. Lower it to steer the team every couple of exchanges. A round ends and a new one begins only when the team made real progress (a tool call, a correction, or a new contribution); a round that only repeats earlier messages pauses. See Multi-agent teams → Turn limits, rounds, and pauses.

Preferred skills

Choose which approved skills agents in this chat are pointed to, with Preferred Skills…. See Skills.

Heartbeat

Controls whether scheduled heartbeat reports may be posted into this chat. See Heartbeats.

RAG

When Enable RAG is on, Verzeta searches this conversation's indexed messages and your knowledge-base documents for relevant context, and includes the most relevant passages in the system prompt. Off by default. You can also switch it with the RAG toggle in the message bar.

Clear RAG memory deletes the retrieval index built for this conversation. Your messages are not deleted. See Conversation memory.

Agent memory and team memory

Whether agents recall facts they saved, and whether this chat shares a project's team memory. See Conversation memory.

Slash commands

Type a slash command in the input field instead of a message. Known commands run locally and are not sent to the model. Text that starts with / but is not a known command is sent as a normal message.

Pro-tip: Use /showtools to check which tools an agent in this chat can use.

Attachments

You can attach files to a message. The supported behaviour depends on the type:

• Images (PNG, JPEG, WebP, ...): sent to vision-capable providers (OpenAI GPT-4o, Claude, Gemini) as base64-encoded image content alongside your text.
• Text-like files (plain text, source code, JSON, Markdown): the content is read inline and appended to your message. The model sees it as text in the user message.

Use the Attach files button in the message bar.

Warning: Attachments count toward the conversation's context window. Large files may push earlier messages out of context.

Conversation actions

Message actions

• Copy: select text in any message and press Ctrl+C.
• Retry: the Retry last response button in the message bar sends your last message again for a new reply.
• Stop: while a reply is streaming, the Send button becomes Stop, which cancels the reply.

Sent messages cannot be edited.

Keyboard shortcuts

Storage

All conversations and messages are stored locally in ~/.local/share/Verzeta/verzeta-studio/verzeta-studio.db on Linux, or %APPDATA%\Verzeta\verzeta-studio\verzeta-studio.db on Windows. The database is a regular SQLite file; you can back it up by copying the folder while the app is closed.

Note: No conversation content is ever sent anywhere except the provider you address the request to, and the services you set up yourself (see question 2 in the FAQ). There is no telemetry, no usage reporting, and no cloud sync.

What's next

• Multi-agent teams: group chat with multiple agents
• Projects: organising conversations into projects
• Activity timeline: viewing the audit trail of a conversation