Advisor

In the advisor section of "Settings → Models → Default models", pick a source, then a model, and turn the advisor on. The model row lists only that source's models; turning the advisor on without choosing a model is the same as leaving it off. All chats share this choice. The advisor can come from a different model source than the main model, does not have to be stronger than the main model, and does not consume the high/medium/low effort levels.

When it consults

Once enabled, the main model receives an advisor tool, along with guidance to consult it before starting substantive work, when it runs into difficulty, when it is considering a change of approach, and before finishing. The main model decides exactly when to call it — it does not make an extra request on every reply.

The main model hands the question to the tool, which sends the currently effective conversation context to the selected advisor model. The context includes the conversation and the tool calls and results so far; compacted history is represented by its current summary, and the full history logs are not re-read. The advisor only returns advice — it does not execute tools, modify files, or call the advisor again. After receiving the advice, the main model continues working on the task.

Models and settings

• The advisor model is saved separately; switching chats or main models does not change it.
• Changing the advisor model saves immediately and takes effect on the next turn; it does not alter the turn currently running.
• Turning the advisor off keeps the selected model, making it easy to turn back on later.
• If the selected source is not signed in, the model is unavailable, or the request fails, an error is shown — the advisor is never silently swapped for another model.
• If the advisor cannot read images or documents in the context, it reports the error explicitly. When the context exceeds the advisor's capacity, the main model used for the current chat request first prepares a summary for the originally selected advisor. The main conversation history remains intact.
• The summary preserves relevant objectives, constraints, decisions, evidence references, and open questions. A failed or cancelled summary, or one that still does not fit, produces an explicit error without silent truncation or model substitution.

Context, quota, and caching

An advisor consultation sends the current context or its summary to the source you chose and uses that source's quota or incurs its charges. Summarization also uses the current main model's quota. The summary and advisor requests are billed separately and respect the session spend limit. Before selecting a different source, confirm that it is appropriate to receive this conversation and the work it contains.

The advisor uses its own request and cache identity: the fixed instructions stay stable, and the current question is appended after the context. Do not assume the main model and the advisor share a cache; whether a cache hit occurs and how it is billed depends on the corresponding model source. Summaries have a separate cache identity and are not written back into the main conversation. Stopping the current turn cancels any summary or advisor request in flight; both stages share one consultation timeout.