AI as a native plane
LLM routing
Capability-aware model selection. Tier requirements. Deprecation fallback policy.
This page is the documentation contract for the LLM routing surface in Vadyl's final form. It is not a marketing summary: it names the authorities, projections, runtime behavior, examples, limits, errors, and observability expectations that every product implementation must honor.
What this surface owns
- LLM routing owns the canonical product-facing contract described here; provider-specific machinery stays behind capability declarations.
- Capability-aware model selection. Tier requirements. Deprecation fallback policy.
- The final docs treat this as complete: REST, GraphQL, gRPC, SDK, CLI, MCP, dashboard, observability, limits, errors, and explainability are all covered as projections of one authority.
Canonical authorities
| Authority | Role |
|---|---|
AgentManifest | Canonical owner for LLM routing; downstream surfaces derive from this rather than inventing their own truth. |
AutomationDefinition | Canonical owner for LLM routing; downstream surfaces derive from this rather than inventing their own truth. |
PlaneCapabilityGraph | Places LLM routing inside the UCSA taxonomy and enforces binding, grants, conformance, and consumption. |
ILlmModelRouter | Canonical owner for LLM routing; downstream surfaces derive from this rather than inventing their own truth. |
TokenAccountingPreflight | Canonical owner for LLM routing; downstream surfaces derive from this rather than inventing their own truth. |
Projection coverage
| Surface kinds | AgentSkillSurface, AutomationActionSurface, ToolSurface, ModelInferenceAdapter, AnalyticsSurface |
| Projection facets | Agents, Automations, PlaneCapabilityGraph, Measures, ExposureBindings |
| Protocols | Rest, OpenApi, Sdk, Cli, Mcp, Dashboard |
| Public projections | agent APIs; automation APIs; MCP tools; SDK agents/automation/mcp; CLI agent/automation; dashboard run views |
Project-scope parity
Publish, install, consume
Consumption evidence
Runtime behavior
- agent plan validator
- LLM capability router
- MCP projection
- immutable memory supersession
- automation compiler
REST and controller surface
Code-backed controllers are listed here so the docs menu does not hide the real endpoint surface. The complete route-by-route table remains in the REST controller atlas.
| Controller | Base route | Endpoint count | Examples |
|---|---|---|---|
| Agent | /api/Agent | 1 | POST {agentId}/run |
| AgentRun | /api/AgentRun | 6 | GET /GET {runId}POST {runId}/cancelGET {runId}/steps |
| AgentPlan | /api/AgentPlan | 10 | GET /GET {planId}POST /POST {planId}/validate |
| AgentMemory | /api/AgentMemory | 7 | GET namespacesPOST namespacesGET namespaces/{namespaceId}POST apply |
| AgentSkill | /api/AgentSkill | 5 | GET /GET {id}POST /PUT {id} |
| Automation | /api/Automation | 18 | POST compilePOST runs/startGET runs/{runId}GET runs/{runId}/approvals |
| KnowledgeCorpus | /api/KnowledgeCorpus | 5 | GET /GET {id}POST /PUT {id} |
| AgentModelBinding | /api/AgentModelBinding | 5 | GET /GET {id}POST /PUT {id} |
| Connector | /api/Connector | 9 | GET contractsGET contracts/{kind}/{majorVersion:int}GET implementationsGET implementations/{implementationId} |
SDK and CLI surface
| Projection | Namespace / group | Coverage |
|---|---|---|
| SDK | agents | definitions, runs, plans, memory, skills, model bindings, token accounting, MCP exposure. Rendered methods: 6. |
| SDK | automation | compile, definitions, runs, approvals, attempts, signals, compensation. Rendered methods: 6. |
| SDK | mcp | token issuance, exposure descriptors, tool invocation, resources, prompts. Rendered methods: 6. |
| CLI | vadyl agent | List, inspect, run agents. Rendered commands: 7. |
| CLI | vadyl automation | Compile definitions, manage runs, approvals, attempts, signals, compensation. Rendered commands: 10. |
| CLI | vadyl connector | Contracts, implementations, bindings, builds, conformance, invocation. Rendered commands: 6. |
Input request and output
POST /api/Agent/{agentId}/run HTTP/1.1
Host: api.vadyl.app
Authorization: Bearer $VADYL_TOKEN
X-Vadyl-Tenant: acme
X-Vadyl-Project: billing
Content-Type: application/json
{
"surface": "llm-routing",
"publicationVersion": 412,
"explain": true
}Limits and quotas
Error model
| Error | Meaning |
|---|