Webui/server config defaults

[ServeurpersoCom]

Make sure to read the contributing guidelines before submitting a PR

server/webui: add CLI support for configuring WebUI defaults

Summary

Implements issue #17940: adds '--webui-config-file' CLI argument and 'LLAMA_WEBUI_CONFIG' environment variable to configure WebUI default settings from the server side.

Problem

WebUI settings are hardcoded in the frontend with no server-side configuration. This causes problems:

• Server operators can't set sensible defaults for their deployment
• Users must reconfigure settings every time localStorage is cleared
• Multi-user servers can't enforce defaults (like disabling auto-file-paste)

Example: Issue #17940 requests ability to set 'pasteLongTextToFileLen: 0' by default, but currently every user must manually configure this in every new browser session.

The TODO that guided this implementation

While exploring the codebase, I found this intentional placeholder in 'server-models.cpp' (lines 6-16):

// TODO: add support for this on web UI
{"role", "router"},
// this is a dummy response to make sure webui doesn't break
{"default_generation_settings", {
    {"params", json{}},  // <-- Empty JSON waiting for implementation
    {"n_ctx", 0},
}},

This empty JSON was left by the original developers as a clear signal for where webui settings should go. This PR completes that TODO.

Backend Changes (C++)

What was added

1. CLI argument ('common/arg.cpp')

• New flag: '--webui-config-file '
• Loads JSON file with WebUI default settings

2. Environment variable ('server.cpp')

• New env var: 'LLAMA_WEBUI_CONFIG'
• Accepts JSON string, overrides file config
• Example: 'LLAMA_WEBUI_CONFIG='{"pasteLongTextToFileLen":0}''

3. JSON parsing logic ('server.cpp')

• Parses config at startup, before any router/child fork
• Simple override: env var OR file (env var takes precedence)
• Error handling: warn on missing file, error on invalid JSON

4. Expose in /props ('server-models.cpp' + 'server-context.cpp')

• Added 'webui_settings' field to '/props' response
• Works in both router mode and child mode
• Fills in the TODO placeholder mentioned above

How it works

Server startup
  |
  v
Parse LLAMA_WEBUI_CONFIG env var (if set) OR config.json (if provided)
  |
  v
Store result in params.webui_config
  |
  v
Return in /props as webui_settings

The C++ backend is completely agnostic: it doesn't know what settings exist. It just:

• Reads JSON
• Passes through to frontend

This means adding new WebUI settings never requires C++ changes.

Frontend Changes (TypeScript)

What was added

1. API type ('types/api.d.ts')

• Added optional 'webui_settings' field to props response
• Type: 'Record<string, string | number | boolean>'

2. Syncable parameters ('services/parameter-sync.ts')

• Added 14 WebUI settings to 'SYNCABLE_PARAMETERS':
  • pasteLongTextToFileLen
  • pdfAsImage
  • showThoughtInProgress
  • showToolCalls
  • disableReasoningFormat
  • keepStatsVisible
  • showMessageStats
  • askForTitleConfirmation
  • disableAutoScroll
  • renderUserContentAsMarkdown
  • autoMicOnEmpty
  • pyInterpreterEnabled
  • enableContinueGeneration
  • showSystemMessage
• Updated 'extractServerDefaults()' to merge webui_settings

3. Store integration ('stores/server.svelte.ts' + 'stores/settings.svelte.ts')

• Added getter for 'webuiSettings'
• Plumbed through to settings store

4. Tests ('parameter-sync.spec.ts')

• Test for webui_settings extraction
• Validates user-specific settings are filtered out
• Uses 'pasteLongTextToFileLen: 0' example from issue

Settings intentionally excluded

These are user-specific and should NOT be server-configurable:

• apiKey (personal credentials)
• systemMessage (personal prompt)
• theme (UI preference)
• mcpServers (personal MCP config)
• custom (user's custom JSON)

Priority Order

Server operators set defaults, but users keep control:

1. WebUI hardcoded defaults (settings-config.ts)
2. Server config: LLAMA_WEBUI_CONFIG env var OR --webui-config-file
3. User's localStorage (always wins)

This means:

• Server can provide sensible defaults
• Users can still override anything
• No breaking changes

Usage Examples

Config file

# webui-defaults.json
{
  "pasteLongTextToFileLen": 0,
  "pdfAsImage": true,
  "showThoughtInProgress": false
}

./llama-server --webui-config-file webui-defaults.json

Environment variable

LLAMA_WEBUI_CONFIG='{"pasteLongTextToFileLen":0}' ./llama-server

Both (env var overrides file)

LLAMA_WEBUI_CONFIG='{"pdfAsImage":true}' \
  ./llama-server --webui-config-file webui-defaults.json

Check it works

curl http://localhost:8080/props | jq .webui_settings

Testing

Backend

# Create test config
echo '{"pasteLongTextToFileLen":0}' > test.json

# Test file loading
./llama-server --webui-config-file test.json

# Test env var
LLAMA_WEBUI_CONFIG='{"pdfAsImage":true}' ./llama-server

# Check /props
curl http://localhost:8080/props | jq .webui_settings

Frontend

cd tools/server/webui
npm test -- parameter-sync.spec.ts

Backward Compatibility

No breaking changes:

• 'webui_settings' is optional in /props
• Old servers work fine (webui_settings: null or missing)
• WebUI falls back to hardcoded defaults
• No recompilation needed for existing deployments

What this enables

Now server operators can:

• Set deployment-wide defaults without forking
• Disable problematic features (like auto-file-paste)
• Provide better defaults for their use case
• Support multiple configurations via env vars

And adding new configurable settings only requires TypeScript changes: no C++ recompilation needed.

Fixes #17940

[ServeurpersoCom]

Testing...

cd /var/www/ia/models
/root/llama.cpp.pascal/build/bin/llama-server \
 -ngl 999 -ctk q8_0 -ctv q8_0 -fa on --mlock -np 4 -kvu --port 8082 \
 --models-max 1 --models-preset backend.ini --webui-config-file frontend.json \
 > llama-server.log 2>&1 &

frontend.json

{
 "showSystemMessage": true,
 "showThoughtInProgress": true,
 "showToolCalls": true,
 "disableReasoningFormat": false,
 "keepStatsVisible": true,
 "showMessageStats": true,
 "askForTitleConfirmation": false,
 "pasteLongTextToFileLen": 0,
 "pdfAsImage": false,
 "disableAutoScroll": false,
 "renderUserContentAsMarkdown": false,
 "autoMicOnEmpty": false,
 "samplers": "top_k;typ_p;top_p;min_p;temperature",
 "temperature": 0.8,
 "dynatemp_range": 0.0,
 "dynatemp_exponent": 1.0,
 "top_k": 40,
 "top_p": 0.95,
 "min_p": 0.05,
 "xtc_probability": 0.0,
 "xtc_threshold": 0.1,
 "typ_p": 1.0,
 "repeat_last_n": 64,
 "repeat_penalty": 1.0,
 "presence_penalty": 0.0,
 "frequency_penalty": 0.0,
 "dry_multiplier": 0.0,
 "dry_base": 1.75,
 "dry_allowed_length": 2,
 "dry_penalty_last_n": -1,
 "max_tokens": -1,
 "pyInterpreterEnabled": false,
 "enableContinueGeneration": true
}

[ServeurpersoCom]

It work

[allozaur]

Looking good on my end