Auto Mode and Decision Model Configuration

Keys of model auto mode and of the shared decision model. For the idea, read Model Auto Mode; for the setup in the Web UI, the guide Let Pando pick the right model.

Decision model

One small model answers quick routing and relevance questions for auto mode, persona auto-select and the context relevance filter. Web UI: Settings > Decision model.

[DecisionModel]
TimeoutMs = 0            # 0 = 1500 ms for Ollama, 3000 ms for remote providers

[DecisionModel.Router]
Provider  = 'ollama'     # 'ollama' (default), 'typesafe' or 'custom'
BaseURL   = ''           # empty = provider default
APIKey    = ''           # stored encrypted; '$ENV_VAR' references are allowed
Model     = 'tev1:0.8b'
KeepAlive = '30m'        # how long Ollama keeps the model loaded
# Headers = { 'X-Team' = 'docs' }   # extra headers for gateways

Key	Default	Description
`Router.Provider`	`ollama`	`ollama` (local), `typesafe` (TypeSafe Jev) or `custom` (a Jev-compatible gateway)
`Router.BaseURL`	provider default	API root. For TypeSafe the default is `https://api.typesafe.ai`
`Router.APIKey`	empty	For TypeSafe, the `TYPESAFE_API_KEY` environment variable is used when empty
`Router.Model`	none	For example `tev1:0.8b`
`Router.KeepAlive`	`30m`	Ollama only
`Router.Headers`	none	Extra HTTP headers
`TimeoutMs`	`0`	Upper bound of one decision call

Ollama must be 0.35 or later; Pando does not install it. Install the model with ollama pull tev1:0.8b, or with the Pull button the settings page shows next to the suggested models (tev1:0.8b, tev1, nimble) that are not installed yet.

Test connection reports: Reachable, Authorized, Version ≥ 0.35, Model present, Decision-capable model, Latency.

Who uses the decision model:

Feature	Where it is switched on
Auto mode routing	Settings > Auto mode (below)
Persona auto-select	Settings > Agents > Persona Selector > Use decision model (`useDecisionModel` on the `persona-selector` agent). The agent’s own model becomes the fallback
Context relevance filter	Settings > Remembrances > Decision model relevance filter, see the Remembrances reference

The idea is explained in Decision model; the setup, in the guide Give Pando quick reflexes.

Auto mode

Web UI: Settings > Auto mode.

[ModelAutoMode]
Enabled        = true
DefaultAuto    = true    # new sessions start with Auto selected
Threshold      = 0.60    # minimum probability of the chosen route (0.05–1)
MinConfidence  = 0       # extra confidence check; 0 disables it
HistoryPrompts = 0       # previous user prompts added as context for the decision

[[ModelAutoMode.Routes]]
ID          = 'quick'
Description = 'Short question or explanation about code, a concept, an error message or a command; no code changes needed.'
Model       = 'ollama.qwen2.5-coder:7b'

[[ModelAutoMode.Routes]]
ID          = 'implementation'
Description = 'Write, modify, refactor or fix code across one or more files, including adding tests.'
Model       = 'anthropic.claude-sonnet-4'
Fallbacks   = ['copilot.gpt-5.4']

[[ModelAutoMode.Routes]]
ID          = 'planning'
Description = 'Design, architecture, trade-off analysis or planning a feature before implementing it.'
Model       = 'anthropic.claude-opus-4'

Key	Default	Web UI label	Description
`Enabled`	`false`	Enable Auto mode	Adds “Auto” as the first entry of every model selector
`DefaultAuto`	`true`	Use Auto by default	New sessions start with Auto selected
`Threshold`	`0.60`	Threshold	Below it, the turn runs on the coder model
`MinConfidence`	`0`	Minimum confidence	0 disables the extra check
`HistoryPrompts`	`0`	History prompts	Up to 20

Routes

Key	Description
`ID`	Stable short name. `none` is reserved for “no route matches”
`Description`	The kind of prompt, in natural language. Up to 500 characters
`Model`	Primary model
`Fallbacks`	Up to 2 models tried in order when the primary fails on a rate limit, a server error or a network problem
`Disabled`	`true` removes the route from the decision without deleting it

Up to 25 routes. Order only matters for ties.

Behaviour

The choice is made once per prompt; the model does not change while the agent chains tool calls.
If no route reaches the threshold, or the decision model does not answer, the turn runs on the coder model. Auto mode never blocks a prompt.
Picking a concrete model turns Auto off for that session until it is selected again.
Delegated sub-agents keep their own configured model.
Switching models between prompts reduces provider prompt-cache reuse for that conversation.
The chat reports each choice:

Auto: implementation → anthropic.claude-sonnet-4 (p=0.93, 38 ms via ollama/tev1:0.8b)

pando doctor checks that the decision model answers and that every route points to a model that exists.

Where Auto appears

Surface	Where
Web UI and desktop	First entry of the model switcher; shows `Auto · <model picked>` during a turn
TUI	First entry of the model dialog
Editors over ACP (Zed, Xcode and others)	First model in the list