Auto Model Routing
When you select Auto in the model dropdown (labelled Smart Auto in Chat mode), Starfish takes care of the model choice. You don't need to think about which model to use -- Starfish reads each request and picks the model and the way to answer it. In both Chat and Developer mode, Auto runs GPT-5.6 Luna at extra-high reasoning; a level or a named model you choose overrides it.
How It Works
- You send a message with the model set to Auto.
- A lightweight LLM classifier reads your message, the conversation context, the number of active agents, available MCP tools, and attachment types. It runs on GPT-5.6 Luna and has a short timeout so it never holds up a reply.
- The classifier works out what the turn owes: an answer in chat, a saved file, an external change, or a fresh read, plus which methodology skills to load and whether the request needs research.
- In Chat and Developer mode, Auto resolves to GPT-5.6 Luna at extra-high reasoning for the turn, so a request that needs judgement gets the deeper pass rather than the cheapest one.
- In both Chat and Developer mode, a brief "Preparing your request…" indicator appears while routing completes.
If the classifier fails or times out, the system falls back to heuristic rules based on message length, attachment presence, and tool count. A turn that requests image or video generation is routed to Deep for reliability.
Picking a Model Yourself
You are never forced onto Auto:
- Chat mode offers outcome-based levels -- Light, Balanced, and Deep -- plus up to five named models you choose under Settings → Models. Light is the cheapest, on DeepSeek V4.1 Flash. Balanced runs GPT-5.6 Luna at high reasoning, and Deep runs the same Luna at its maximum reasoning -- a step up in effort, not a different model.
- Developer mode lists the full catalog, grouped by provider, with a Thinking effort control (low / medium / high) for models that support it.
A model you pick always wins over Auto.
Model Stickiness
Auto does not consider or continue a previously resolved model. In both Chat and Developer mode it always runs GPT-5.6 Luna at extra-high reasoning. What is remembered across chats is your own choice: a level (Light, Balanced, Deep) or a named model, which overrides Auto.
Cost
The model you chose handles the whole turn. Starfish deliberately does not swap to a cheaper or smaller model partway through a long, multi-tool task -- capability matters most there -- so cost is bounded by a step and tool budget rather than by downgrading.
Context Compaction
When a conversation grows large, the system automatically compacts older messages into a summary to stay within model context limits. Compaction triggers at roughly 75% of the model's context window. You'll see an amber "Context Compacted" banner in the chat when this happens.
Token Usage
The header displays a token usage badge that shows how full the model's context window is (for example, 120k / 200k ctx), not a lifetime spend total:
| Color | Context fill | Meaning |
|---|---|---|
| Neutral | Under 60% | Normal usage |
| Amber | 60% -- 80% | Elevated usage, consider starting a new session |
| Red | Over 80% | Near context limit, compaction likely |
Hover the badge for the exact counts and the total tokens processed this session.
Model Tiers and Prompt Optimization
Models are grouped into tiers: Standard, Fast & Cheap, and Ultra Cheap (the model metadata also reserves a Premium tier). Each model family -- OpenAI, Anthropic, Google, Moonshot, and everything else -- gets a short behavioural prompt tuned to the habits that family tends to need, rather than one generic instruction set for every model.