Workflow Profiles

WorkflowProfile is a v1alpha2 serving-intent resource. It exposes a stable client-facing name such as fast-chat, long-context-review, code-agent, or image-edit, then selects existing Model resources that can satisfy that intent.

The first implementation wave is deliberately lightweight:

• Profiles select and explain existing Model resources.
• Profiles do not create or mutate underlying models.
• Direct concrete model names continue to work.
• Proxy routing and CLI explain commands are separate follow-up slices.

Minimal shape

apiVersion: ai.flexinfer/v1alpha2
kind: WorkflowProfile
metadata:
  name: fast-chat
spec:
  intent: Low-latency chat with tool-calling support
  servedName: fast-chat
  aliases:
    - chat-fast
  selector:
    matchServiceLabels:
      - textgen
      - fast
    capabilities:
      toolCalling: true
  routes:
    - name: ready-fast
      objective: PreferReady
      maxPromptTokens: 12000
      requiredCapabilities:
        toolCalling: true
    - name: fallback-long
      objective: LongestContext
      maxPromptTokens: 32768
  gates:
    requireReadyOrColdStart: true
    requirePromotionEvidence: true
    allowedPhases:
      - Idle
      - Loading
    minSupportLevel: Validated

Selector

spec.selector narrows candidate models. At least one selector field must be set:

• modelRefs: same-namespace Model names.
• matchLabels: Kubernetes labels on the Model.
• matchServiceLabels: semantic labels from Model.spec.serviceLabels.
• backends: allowed backend names, such as vllm, ollama, or diffusers.
• capabilities: explicit Model.spec.capabilities requirements.

The selector is intentionally about existing resources. To add a new model to a profile, create or update the Model separately and make it match the selector.

Routes

Routes are ordered. A request matches the first route whose constraints fit. Supported objectives are:

Objective	Intent
PreferReady	Prefer warm/ready models before cold starts
LowestLatency	Prefer the lowest observed latency candidate
LongestContext	Prefer the largest sufficient context window
CheapestSufficient	Prefer the smallest validated lane that satisfies the request
Pinned	Route to a configured model until policy changes

Gates

Gates fail closed. A candidate blocked by gates should appear in status as a rejected candidate rather than receiving traffic.

MVP gate fields:

• requireReadyOrColdStart: permit Ready plus explicitly allowed cold-start phases.
• allowedPhases: extra phases that may be routed, commonly Idle for scale-to-zero.
• requirePromotionEvidence: require validation or promotion evidence.
• minSupportLevel: require Experimental, Canary, Validated, or Production.
• backendAllowlist: narrow routable backends after broad selector discovery.

Promotion evidence initially maps to existing model promotion conditions and annotations. See .loom/60-validation-matrix.md for the operator evidence matrix used by current runtime-promotion decisions.

Status

The controller status slice records:

• resolved served name and aliases
• active route and concrete model
• aggregate gate state
• eligible candidates
• rejected candidates with reasons and evidence references
• conditions: Resolved, CandidatesAvailable, GatesSatisfied, Routable

Examples

• examples/v1alpha2/workflow-profiles/fast-chat.yaml
• examples/v1alpha2/workflow-profiles/long-context-review.yaml
• examples/v1alpha2/workflow-profiles/code-agent.yaml
• examples/v1alpha2/workflow-profiles/image-edit.yaml