Models and providers

#Supported protocols

#Configuring in the app

Control → Models. The configuration is edited as a draft and saved as a whole:

1. Add a provider: from the catalog or custom (name, API URL including the version path such as /v1, protocol).
2. Add keys: several keys per provider are fine; they rotate.
3. Choose models: tick from the public catalog (models.dev, with context lengths and prices), fetch from the provider's models endpoint, or add a custom model by hand.
4. Test: shows latency and the result.
5. Global model order: all enabled models from all providers sit in one list you can drag.

Before changing an API URL you must remove its old keys, and saving carries a version number: if the configuration was changed elsewhere first (for instance from a Feishu card) the save is rejected and you are asked to reload. Both rules exist so a key is never sent to the wrong address.

#Call order and failover

Each call walks the global order of enabled models; within a provider, keys rotate with at most two attempts:

Loading diagram…

• Rate limits, timeouts, 5xx: try another key;
• 400/401/403/404 ("this model or configuration is wrong"): move to the next model;
• A streaming response with no data chunk for 90 seconds counts as failed (with a 15-minute absolute cap).

#Introspection model

You can nominate a quick (introspection) model. When she wakes she uses it first for a lightweight judgement: does she feel like doing anything, and what. A cheap, fast model here saves most of the cost of "woke up and went back to sleep".

#Models that can see images

When you send her pictures, routing only picks models that accept images. The order of evidence: your manual vision flag → the public catalog's input modalities → the model name. If none qualifies the image is dropped and the message says so.

#Special providers

A few providers have extra conventions handled by separate compatibility modules that activate automatically. Currently: OpenCode Go, activated when the catalog ID is opencode-go or the URL is opencode.ai/zen/go; it picks the right API per model and sends a session identifier. Nothing to configure.

#How keys are stored

• Keys are encrypted with AES-256-GCM on the phone; the master key lives in WINDLER_HOME/secrets/;
• Every interface returns only the last four characters;
• Keys never enter the soul repository: they do not travel when she changes bodies; each body is configured separately.

#Usage and budget

Tokens and estimated cost per call are recorded locally and shown in Now and Control → Budget. When the budget is spent her wake rate drops sharply (see Permissions and safety).