Reasoning (Thinking) & Telemetry

Learn how to control model deliberation effort and supervise context window consumption and token metrics in real time.

1. Reasoning-Capable Models

Next-generation models (such as DeepSeek-R1, OpenAI o1/o3-mini, Claude 3.7 Sonnet with thinking, and Qwen QwQ) generate internal thought chains prior to answering.

ZeroChat streams these thinking chunks in real time, rendering them inside collapsible thought blocks with timers.

The Reasoning Effort Selector

The composer toolbar includes a speedometer indicator. Its needle reflects the selected intensity; click it to open a compact panel with an intensity slider and an X close button.

None: No reasoning configuration is sent; it is the right choice for fast, concise answers.
Low to Maximum: Use the slider to adjust reasoning effort to the task's complexity.
Compatibility: ZeroChat adapts the level to the format understood by the provider. If the model does not support reasoning, it continues with a normal response.

Agent Checkpoint (agent_checkpoint): Enable this tool from the reasoning menu to let models store intermediate conclusions before complex tool loops or RAG searches.

2. Unified Context Hub

Next to the profile selector, the telemetry chip (e.g., 1.4k / 128k) reflects context health via color-coded status:

Green (Optimal): Under 75% of context limit used.
Amber (Warning): Between 75% and 90% consumed; consider pruning or branching.
Red (Critical): Over 90% capacity used; truncation risk is imminent.

3. Context Hub Popover Diagnostics

Click the token pill to open the comprehensive telemetry diagnostic view:

Section 1: Global Context Window

Progress Bar: Visual percentage of context window used.
Used, Limit, and Available Tokens: Exact token metrics synthesized from server usage data and client tokenizer estimates.
Assumed Capacity Override: Manually input custom limits (e.g., 8k or 32k) if your local server does not publish its capacity.

Section 2: Prompt Cache Savings (KV Cache)

Cached Read Tokens (Savings): Tokens served from cache with reduced pricing and latency.
Cached Write Tokens: New prefix tokens cached for subsequent turns.

Section 3: Last Turn Performance

Prompt & Completion Tokens: Token breakdown for the latest message.
Generation Speed: Measured in tokens per second (t/s).
Latency (TTFT): Time to First Token in milliseconds.