Activities & Reasoning

Activities

Every message you send creates an activity. Think of it as one request-response cycle. When you ask Hefty to do something, you'll see the activity progress through several stages in the UI:

1. Preparing - Hefty searches its knowledge base for relevant skills, entities, and conversation history.
2. Reasoning - Hefty streams its thinking in real-time. You'll see the response text appear as it's generated.
3. Acting - if Hefty decides it needs to take action (read a file, run a command, etc.), you'll see each action appear with its status - what tool was used, why, and whether it succeeded.
4. Done - Hefty delivers its final response. Behind the scenes, it also learns from the interaction for next time.

You can click on any activity to see its full details: which knowledge Hefty recalled, what it planned, and the result of every action it took.

What You See During Reasoning

For simple questions, Hefty responds directly - just like a normal chat. For tasks that require action, Hefty works in loops:

Extended Thinking

For models that support extended thinking (e.g., DeepSeek-R1), a collapsible "Thinking" section appears showing the model's internal chain-of-thought reasoning before it produces its visible plan and actions. This is separate from Hefty's own reasoning — it's the raw model-level thinking that precedes the structured output.

In the UI, each loop shows:

• What Hefty is thinking - the streamed reasoning text
• What it plans to do - listed actions with the tool name and a brief rationale
• What happened - each action's result: success or failure, output, and how long it took

If an action fails, Hefty sees the error and can try a different approach in the next loop. This continues until Hefty has a final answer (or reaches the loop limit).

Reasoning Loop Indicators

When the agent goes through multiple reasoning cycles, each loop displays a "Reasoning (loop N)..." indicator. Each loop can produce its own plan update, action sequence, and reflection — you see the full progression of the agent's thinking in real-time.