Main and auxiliary models

After the first model is configured, it becomes the main model automatically. This is the model used for normal coding conversations.

When adding later models with /model add, choose one of:

Use as main model: make it the current main model.
Add to fallback chain: use it as a backup candidate.
Save model only: store it without assigning it yet.

Auxiliary models are used for smaller tasks such as titles, prompt suggestions, summaries, and goal judging. They do not always need the strongest main model. There is no need to configure all of them at first; a practical starting point is to assign a cheap stable model to high-frequency tasks such as sessionTitle, promptSuggestion, and goalJudge.

Use /model aux list to view auxiliary tasks. When assigning a model, replace <task> with an auxiliary task name from the first column of the reference table below, such as goalJudge:

/model aux set goalJudge <model-id>

Common auxiliary task reference

Auxiliary task	What it does	Recommended model type
`sessionTitle`	Generates session titles.	Cheap, fast, stable text model.
`promptSuggestion`	Generates the faint next-step suggestion in the input box.	Cheap fast small model, usually with a short output limit.
`tokenCounting`	Helps estimate or count tokens.	Cheap model; failure should not block the main chat.
`sessionSearchSummary`	Summarizes `/resume` or SessionSearch matches.	Cheap summarization model.
`memdirRelevance`	Picks relevant memory files from frontmatter.	Mid-quality model with stable JSON output.
`goalJudge`	Checks whether `/goal` is complete after each turn.	JSON-strict model, ideally cheaper than the main model.
`hookPrompt` / `hookAgent`	Runs prompt hooks or full agent hooks.	`hookAgent` usually needs a higher-quality model.
`skillImprovement`	Suggests skill improvements.	Cheap model that follows formatting instructions.
`mcpDateTimeParse`	Parses natural-language date/time fields in MCP forms.	Fast text model.
`shellPrefix`	Extracts shell command prefixes for permissions or display.	Fast text model.
`permissionExplainer`	Generates permission request explanations.	Fast text model.
`compact`	Summarizes long conversations and results so later turns can continue with less context.	Quality-first; usually stronger than title or counting models.
`sideQuestion`	Answers temporary questions outside the main task, such as explaining code or adding background context.	Quality-first; usually stronger than title or counting models.
`forkedAgent`	Provides a model for separate agent tasks, such as parallel research, checks, or subtasks.	Quality-first; usually stronger than title or counting models.