Settings¶

The Settings section configures project-level infrastructure and observability integrations. Changes are written to YAML config files and take effect after a reload.

Settings page

The left panel groups settings into four categories:

Category	Item	Purpose
Infrastructure	Redis	Shared state for multi-worker deployments
Observability	Telemetry	OpenTelemetry trace and metric export
Testing	LLM Judge	Automated quality evaluation for CI
Security	API Access	Control which HTTP API surfaces are mounted

Redis¶

Redis is required when running tuvl with more than one uvicorn worker. Without Redis, OAuth2 CSRF state tokens and Biscuit token revocation lists are stored only in each worker's memory — causing random failures when a request lands on a different worker than the one that started the session.

Redis YAML format¶

kind: RedisConfig
version: v1
enabled: true
metadata:
  name: redis-primary
spec:
  type: redis
  connection:
    host: "${REDIS_HOST:localhost}"
    port: "${REDIS_PORT:6379}"
    db: 0
    password: "${REDIS_PASSWORD:}"

Redis fields¶

Field	Description	Default
`host`	Redis server hostname	`localhost`
`port`	Redis port	`6379`
`db`	Database index (0–15)	`0`
`password`	AUTH password (leave blank for no auth)	(empty)

env var syntax

All fields support ${ENV_VAR:default} substitution. Set REDIS_HOST, REDIS_PORT, and REDIS_PASSWORD in your .env file or environment.

Telemetry¶

tuvl ships with built-in OpenTelemetry support. Traces and metrics can be exported to any OTLP-compatible backend (Jaeger, Tempo, Honeycomb, Datadog, …).

Telemetry YAML format¶

kind: TelemetryConfig
version: v1
enabled: true
metadata:
  name: otel
spec:
  exporter: otlp-grpc              # otlp-grpc | otlp-http | jaeger | console
  endpoint: "${OTEL_EXPORTER_OTLP_ENDPOINT:http://localhost:4317}"
  service_name: "${OTEL_SERVICE_NAME:tuvl}"
  sample_rate: 1.0                 # 0.0–1.0  (1.0 = trace everything)

Exporter options¶

Exporter	Description
`otlp-grpc`	OpenTelemetry Protocol over gRPC (default)
`otlp-http`	OpenTelemetry Protocol over HTTP/JSON
`jaeger`	Jaeger native UDP exporter (legacy)
`console`	Print spans to stdout — useful for debugging

See the Telemetry configuration guide for a full guide including Grafana Tempo setup.

LLM Judge¶

The LLM Judge powers tuvl's tuvl test command. It evaluates workflow outputs against natural-language quality criteria using a secondary LLM.

LLM Judge YAML format¶

kind: LLMJudgeConfig
version: v1
enabled: true
metadata:
  name: judge
spec:
  model: default          # references an AgentModel metadata.name
  temperature: 0.0        # deterministic for consistent scoring
  pass_threshold: 0.8     # 0.0–1.0; runs below this score are flagged

Judge fields¶

Field	Description
`model`	AgentModel to use for evaluation (see AI Models)
`temperature`	Set to `0.0` for reproducible judgements
`pass_threshold`	Minimum pass score (0.0–1.0). Runs below this fail the test.

See the LLM Judge guide and the Testing Workflows reference for how to write test cases and run them in CI.

API Access¶

Control which HTTP API surfaces are mounted. This edits .tuvl/system.yaml (kind: SystemConfig) — see Configuration Overview for the full knob reference.

Expose model CRUD APIs¶

A toggle for spec.api.expose_model_crud. When off, only workflow APIs are mounted — /models/* CRUD endpoints return 404. This applies in production too; the TUVL_EXPOSE_MODEL_CRUD env var overrides whatever is saved here.

Restart required to apply changes

The API surface is mounted once at startup. Save your config here, then restart the engine with tuvl run (or tuvl dev) for changes to take effect.