Reference

Settings

Generated page

Generated from the source by scripts/docs/build-reference.js; edit the code, not this page. It lists the shipped client defaults and the control that edits each field. A saved account or a request may hold other values; the precedence below says which one a turn uses.

Shipped client defaults

The values a new profile starts with, from DEFAULT_SETTINGS in public/app.js. Control is the element in Settings that edits the field, read from readSettingsFromForm; a field without one is kept from the saved settings or written by the page's own logic (setup cards, the decision-model cards, tests of a connection). Credential marks a field the secret broker holds in encrypted account storage: saved secrets are returned to the page as references. The Settings save endpoint accepts newly entered credentials; ordinary turn requests cannot override the account credentials.

Setting	Shipped default	Control	Credential
permissionMode	auto	#defaultPermissionModeSelect
permissionsEnabled	true	#permissionsEnabledToggle
completionSoundEnabled	true	#completionSoundToggle
mainProvider	custom	#mainProviderSelect
baseUrl	empty	#baseUrlInput
apiKey	empty	#apiKeyInput	yes
braveApiKey	empty	#braveApiKeyInput	yes
model	empty	—
decisionsEnabled	false	#decisionsEnabledToggle
decisionsProvider	typesafe	#decisionsProviderSelect
decisionsBaseUrl	empty	—
decisionsApiKey	empty	#decisionsApiKeyInput	yes
decisionsModel	empty	—
decisionsActive	jev	—
decisionsLayaBaseUrl	empty	#layaBaseUrlInput
decisionsLayaApiKey	empty	#layaApiKeyInput	yes
decisionsLayaModel	empty	—
imageGeneration	{"enabled":false,"activeServerId":"","servers":[]}	—
temperature	null (no override; the provider's default)	#temperatureInput
reasoningControl	auto	#reasoningControlSelect
reasoningEffort	high	#reasoningEffortSelect
codingProvider	main	#codingProviderSelect
codingBaseUrl	empty	#codingBaseUrlInput
codingApiKey	empty	#codingApiKeyInput	yes
codingModel	empty	—
codingTemperature	null (no override; the provider's default)	#codingTemperatureInput
codingEnableThinking	true	—
codingReasoningControl	auto	—
codingReasoningEffort	high	—
classifierProvider	main	#classifierProviderSelect
classifierBaseUrl	empty	#classifierBaseUrlInput
classifierApiKey	empty	#classifierApiKeyInput	yes
classifierModel	empty	—
classifierTemperature	0	—
themeId	nawa-dark	#themeSelect
transcriptView	normal	#transcriptViewSelect
llmPresets	empty list	—
activeLlmPresetId	empty	—
providerConnectionMemory	{}	—
skillMatcherSettings	{"profile":"classifier","baseUrl":"","apiKey":"","model":"","temperature":0,"maxTokens":512,"enableThinking":false,"reasoningControl":"auto","reasoningEffort":"off"}	—
enableThinking	true	—
webSearchEnabled	true	—
disabledConnectors	empty list	—
maxParallelAgents	4	#maxParallelAgentsInput
stageRepairAttempts	3	#stageRepairAttemptsInput
chatDirectory	./data/chats	#chatDirectoryInput
sidebarWidth	290	—
maxAttachmentChars	100000	#maxAttachmentCharsInput
docAgentThreshold	15000	#docAgentThresholdInput
executionTimeoutMinutes	40	—
contextTokenLimit	32768	#contextTokenLimitInput
maxOutputTokens	null (no override; the provider's default)	#maxOutputTokensInput
compactionThresholdPercent	80	#compactionThresholdPercentInput
docEditAnimation	morph	—

Precedence

How a turn's settings are resolved, as the owners apply them:

1. System defaults. config/system-settings.json holds clientDefaults for a profile; sanitizeClientDefaults (server/client-settings.js) drops retired keys from whatever it reads.
2. The account replaces them. effectiveClientDefaultsForRequest (server.js) loads the signed-in user's encrypted settings; a non-empty sanitized account object replaces the system defaults, it is not merged with them. Saving from Settings writes the whole client settings object to the account (POST /api/system-settings/client-defaults).
3. The request lies on top. runtimeRoutingSettings (server.js) lays the request body's fields over the account's, then pins every credential to the account's value, so a body can change a model or a URL for one turn but never present a key. Chat instructions guide model behavior; they are not another runtime-settings layer. Asking for a different temperature, provider or permission mode does not itself change those settings.

Thinking preferences

The Main and Coding thinking controls save the user's toggle, control and effort. Their change handlers update saved preferences directly; capability discovery only changes availability and display. A temporarily unavailable preference is retained across model switches. Send does not wait for the browser's capability check: the server resolves automatic controls from the selected model's capabilities and applies the supported provider fields. The old derived reasoning override and client capability copies are no longer settings inputs.

Sanitizer examples from synthetic fixtures

Values below come from named fixtures run through sanitizeClientDefaults and credential-key classification, never from a live profile. These test sanitization only, not account/request resolution or encryption.

Fixture	Input	After sanitizeClientDefaults	Fields classified as credentials
A — a new profile	the shipped defaults	unchanged (54 fields, changed: false)	apiKey, braveApiKey, decisionsApiKey, decisionsLayaApiKey, codingApiKey, classifierApiKey
B — a saved account	the defaults with model, apiKey, decisionsActive: "laya", decisionsLayaBaseUrl and the retired plannerModel set	the retired key is dropped (changed: true); decisionsActive = laya, decisionsLayaBaseUrl = http://127.0.0.1:8080	apiKey is classified as a credential; sanitization does not encrypt it

Retired keys the sanitizer removes: plannerProvider, plannerBaseUrl, plannerApiKey, plannerModel, plannerSettings, maxPlannerPasses, presencePenalty, presencePenaltyModels, codingPresencePenalty, reasoning, reasoningCapabilities, codingReasoningCapabilities.

Not covered here

The server section of config/system-settings.json (ports, database, skills, memory); per-chat state such as a chat's own permission mode; the LLM setup cards' saved connections beyond the flat fields they project.