Reasoning settings and interpretation of context metrics.
Some providers return reasoning content or a summary that ZeroChat displays in a collapsible block. This is what the API exposes, not a complete view of the model's internal process.
The composer's reasoning control lets you choose from None to Maximum. ZeroChat maps the level to parameters supported by each adapter. None asks to turn it off with each protocol's value (reasoning_effort: "none" or thinking disabled). If the server rejects it, because the model has a minimum reasoning level like some Gemini models, ZeroChat repeats the request without that parameter and the model reasons with its default; in a turn with tools, the remaining steps are sent without it. With Gemini, Maximum asks for its highest level, high. Some servers accept the value and reason anyway. Availability and effect depend on the model.
While the agent chains tools within one answer, ZeroChat sends each step's reasoning back to the model so it keeps its plan from one call to the next. It only does so when the server sends reasoning in its own field (reasoning_content or reasoning, as llama.cpp, vLLM, Ollama or OpenRouter do) or in Messages API thinking blocks, which are sent back with their signature, and drops it when the answer ends: the next message does not include it and it is not saved in the conversation.
The same panel has the "Project mode" switch, and the dot on the button shows that a project is in use: see Project management.
The indicator, for example ~1.4k / 128k*, compares used tokens with the known or assumed capacity. ~ marks estimated usage and * marks assumed capacity.
These are UI guidance thresholds. Context includes instructions, history, tools and results, not just the last message. If it grows too large, reduce attachments or start a conversation with a checked summary. Branching while retaining all history does not free that context.
When the history reaches 70% of the input budget, ZeroChat automatically compacts it into a summary before the next request and shows the line “Compacting context.....” in the chat. The model itself writes the summary, which may leave out details.
Click the indicator to see usage, capacity and available space. Provider data is used when available, with estimates otherwise. Manually setting capacity changes ZeroChat's reference; it does not enlarge the server's actual window. Sent context may exclude older messages to fit the budget even while they remain visible in history.
Cache read and write tokens appear when reported by the provider. They do not measure financial savings: rates and conditions depend on the service.
The panel displays input and output tokens, generation speed and time to first token when data is available. Metrics may be missing or vary between providers; they are neither an invoice nor a controlled performance comparison.