Models, favorites, and defaults

Vera can use models from any connected provider. You choose the model for the conversation you’re working in. Favorites make models easier to find; defaults choose models for jobs such as compaction and delegated work.

Switch model changes what runs next and writes nothing that outlives the conversation, except a favorite. /models is the opposite: it browses, favorites, assigns, and connects, and never changes what runs next.

These choices are independent. Selecting a model does not favorite it, and removing a favorite does not clear a default assignment.

Choose a model

1. Enter /model to open Switch model.
2. Type a model name. Search includes connected models outside your favorites.
3. Use Up/Down to highlight a model, then press Enter.
4. If the model offers reasoning effort levels, choose one to finish.

The list opens as Favorites, then Recent from this conversation, then every connected model grouped by provider. Typing drops the grouping and ranks one list by how well each name matches. Ctrl+F adds or removes the highlighted favorite; that is the only thing this dialog writes to disk.

Switch model has no filters, sort, scores, prices, or details panel. Those are in /models.

In a conversation, the selection applies to its next request. From Home, it opens a conversation. The composer footer shows the current model and effort.

You can also open Switch model from Ctrl+P, or press Ctrl+X, release Ctrl, and press M. Escape cancels that shortcut without changing your draft.

Browse every model

/models opens the browse page: a searchable list of models with Filter and sort, Connect provider, and Manage models under it. Nothing here changes the model for your next request.

Filter and sort

Browse starts with Favorites unless you saved another collection. To see every connected model, open Filter and sort and choose All connected. The header lists Favorites, Recommended, and All connected, with a filled dot on the one you are looking at. Ctrl+G moves to the next and Shift+Ctrl+G to the previous. The header only shows where you are; it is not a set of tabs.

Filters narrow results by provider, availability, known price, image support, and intelligence score. They also apply to searches. Clearing the search returns to the selected collection; it does not clear filters.

Enter on a toggle, such as view, an on/off filter, variants, or Clear filters, changes that row and leaves the menu open. Choosing a value from Show, Sort, or Provider returns you to Filter and sort on the row you changed.

The Intelligence cutoff slider runs from Any to Smarter. Left and Right change it. Enter returns to the model list and keeps the cutoff; Escape returns and keeps it too. A positive cutoff also hides models without scores, including favorites. If a model is missing, check the cutoff and other filters before reconnecting its provider.

Older and extra variants are normally hidden unless favorited or found by search. Turn on variants in Filter and sort to browse them. Clear filters keeps your search, view, sort order, and favorites.

Compare prices and capabilities

Choose Detailed view in Filter and sort to add WA Score, input price, output price, and status columns when the terminal is wide enough. When there is room for both, a details panel sits beside those columns. If the columns need the width, the panel is omitted. A narrower terminal keeps facts in the list, including image support, verification, and the exact model ID.

Prices are dollars per million tokens. Unknown scores and prices show ?. An unknown price is not zero. WA Score, when known, is a rating from blind comparisons of responses. Cheapest first sorts within each provider using three input tokens for every output token, with unknown prices last. Other orders are A to Z and, within Favorites, your saved favorite order.

Vera saves the collection, view, and sort for the next time you open the picker.

Save favorites

Open Favorites from /models, highlight a model, and press Enter to add or remove it. Ctrl+S does the same on a highlighted model anywhere in the model page, and Ctrl+F does it in Switch model.

Connected providers start with a few favorites already saved, so the list is not empty before you curate it.

Favoriting saves a shortcut. It does not switch the conversation or make a verification request. Removing a favorite keeps its verification results and any default assignments.

Rename or verify a favorite

In Favorites, Ctrl+R changes the highlighted model’s display name. Leave the name empty to restore its catalog label. Ctrl+Y tests whether the model can answer; it also works on connected models you have not favorited.

For bulk checks, choose Verify favorites from the command palette. Tab switches between unverified favorites and all favorites. Choose all of them or one provider, then press Enter to start.

Verification makes real requests and can incur provider charges. Results show whether each check passed or failed. Escape returns to the previous screen while those checks continue.

Assign defaults

Open Defaults from /models to assign connected models to snappy, eco, extra, classifier, compaction, and subagents. A model does not have to be a favorite to appear here. Assigning one that is not a favorite adds it to your favorites, because a subagent can run on any of these assignments and subagents only run favorites.

A new assignment needs permission to use the model and takes effect right away. To verify each model before it is assigned, set "verify_model_assignments": true in config.json. With that on, Vera warns that the verification request may cost money, a failed check or canceled assignment leaves the previous value in place, and Escape cancels the assignment, though an already-started check may still finish.

After you pick a model for one default, you are back on the Defaults page with the new value showing. To give every default the same model, choose Assign one model to all at the bottom of the page and pick the model. Subagents get it first and keep the rest of their list, and a classifier failsafe stays in place. A toast confirms how many defaults changed.

Assigning a default leaves your conversation model unchanged. Subagents use an ordered list of allowed models and an optional fallback to the spawning conversation’s model. See Agents and delegated work.

Adjust agent and access together

Press Shift+Tab from the composer, or choose Dial strip from Ctrl+P, to open the quick controls.

Access offers readonly, ask, and auto. Unavailable choices are skipped.

Model and effort are not here. Use /model and /effort. The composer footer shows all four values whether or not the quick controls are open.

Set reasoning effort

/effort lists the levels the current model supports, each with a line saying what it is for. /effort <level> sets one directly. Levels differ by model, and not every model has an off setting.

Switching models keeps your effort level when the new model also supports it. You are asked to choose again only when it does not.

Choosing a level saves it as the default for new conversations, except max. max applies to the current conversation only, and new conversations keep the level they had before. Vera says which of the two happened when you choose.

Refresh or recover a connection

Choose Refresh model catalog from Manage models (the row under the list on the browse page) to reload connected catalogs. For a ChatGPT subscription, Vera asks OpenAI for the current model list using your sign-in. A brief overlay types catalogs refreshed, new models, and any failures, then a tick. Press Escape to close it when no dialog is open; otherwise it clears after a few seconds. Another overlay waits its turn. A model that has disappeared from its provider is marked unavailable and cannot be selected. To change credentials or endpoints, use Configure providers.

Vera retries rate-limit, server, and dropped-connection failures up to six times, waiting 1, 2, 4, 8, 15, and 30 seconds. Other transient request failures get up to two retries. A connection that drops while the model is still reasoning is retried too; once answer text has appeared, Vera keeps it and shows the error instead. If a model sends nothing for 5 minutes, Vera treats that as a timeout and retries. Retry activity shows the next attempt and total attempt count. Press Escape to cancel at any point, including while a request is stuck. Missing credentials or provider credit need your attention and stop immediately. Malformed tool arguments can also be retried when no tool could have run. If retries fail, Vera keeps the last partial response and shows the error.

Configure an overload fallback

A fallback entry in the home configuration can name a backup on the same provider:

{
    "fallback": {
        "model": "your-backup-model",
        "after_failures": 3
    }
}

after_failures accepts 1, 2, or 3 consecutive rate-limit or server failures. Once selected, the backup stays active for the rest of the turn. This setting is separate from the Subagents default list.

Use a local model

Connect a running local server, then choose its model in Switch model. Vera does not install or download models.

A successful connection does not establish the model’s context window or reasoning controls. If the server omits those facts, context occupancy and effort can remain unknown until configured or supplied by a supported mapping. Thinking controls vary by server; do not assume every local model has the same effort levels.

Prompt caching

Providers keep the start of a request they have already processed and charge less to reuse it. Vera keeps that start the same within a conversation and sends each provider what it needs to find it again. There is nothing to set up.

• The ChatGPT subscription gets the conversation’s id as a cache key.
• Anthropic-style endpoints get cache markers on the system prompt, the tools, and the latest message.
• anthropic/* models through OpenRouter get cache markers too. Every OpenRouter request carries the conversation’s id, so OpenRouter keeps the conversation on the same upstream provider.
• Other OpenAI-style providers cache on their own.

The usage page counts cached input for every provider that reports it, and cache writes where the provider reports those.

To change how long providers are asked to keep the cache, set prompt_cache.retention in config.json to short (the default), long, or none to send no cache hints at all:

"prompt_cache": { "retention": "long" }

A provider you declare under providers caches the way its protocol suggests. If your endpoint caches differently, give it its own cache block:

"providers": {
  "crows-nest": {
    "protocol": "openai-chat",
    "base_url": "https://crows-nest.example/v1",
    "cache": { "mode": "key", "usage": ["read", "write"] }
  }
}

mode is key, automatic or none for openai-chat, and breakpoints or none for anthropic-messages. usage lists what the endpoint reports back: read, write, both, or neither. affinity_header names a header that carries the conversation’s id, for a gateway that pins a conversation to one server by it.

Keyboard reference

The footer shows the action for the focused control. Typing or pasting from any picker control moves into Search.

Switch model has no controls to move between: type to search, Up/Down to move, Enter to switch. In the model page’s Search, Left/Right move the caret and Space enters a space. Arrows that have no action in the current control move focus to another control.

Ctrl+K on a model opens actions for that model only: add or remove the favorite, verify it, and request options where the provider has them. The Manage models row under the list holds actions for the whole list: refresh, Edit favorites, verify favorites, defaults, and provider configuration. Opening Edit favorites does not change membership until you make a choice there.

Command-line reference

vera library list
vera library add "<provider>/<model>"
vera library add "<provider>/<model>" --verify
vera library remove "<provider>/<model>"
vera models refresh

Adding without --verify saves the reference without contacting the provider. With --verify, Vera saves it only after the required tool-calling check passes. Refreshing updates a running host’s catalog, or the cache for the next host. It covers a signed-in ChatGPT subscription along with your other providers and reports each one. No TUI restart is needed.

The older vera shortlist spelling remains an alias for vera library.

The model page’s collection, view, and sort persist under model_picker in the home’s tui.json.