Two API surfaces: Runtime APIs are stable contracts that ship VAOS as infrastructure. Cognition APIs are an evolving surface where reasoning, prediction, and trace endpoints live.
Scenes · Actions · Memory · Sensors · Plugins. Versioned contracts, semver-compatible across minor releases. Build production systems against these.
Reason · Predict · Summarise · Score · Trace. These surfaces evolve as reasoning improves. Expect minor breaks across versions — pin your client.
Default base URL is http://localhost:8800. All responses are JSON. Authentication via API key header (X-VAOS-Key) when configured.
Return the most recent Scene Contract from a sensor.
curl http://localhost:8800/scenes/latest?camera=0
Response: Scene Contract object.
Query historical scenes from local memory.
| Param | Type | Description |
|---|---|---|
type | string | Filter by event type |
since | ISO 8601 | Earliest timestamp |
limit | int | Default 50, max 500 |
Emit a validated action. Validation gate is on by default.
{
"type": "notify",
"target": "safety-team",
"channel": "slack://ops-safety",
"validation": true
}Return per-sensor reliability scores and ISP metadata.
Liveness and readiness probe. Returns runtime status, plugin versions, hardware backend.
Stream Scene Contracts as they're published. JSON-encoded frames.
const ws = new WebSocket("ws://localhost:8800/ws/scenes");
Stream emitted actions and their validation results.
Per-camera topic. Payload is the Scene Contract.
Emitted actions with validation result.
Sensor reliability scores and warnings.
Query the local Memory Engine — temporal summaries, priors, expectation models.
List installed plugins with versions, capabilities, and runtime status.
The reasoning surface. Endpoints below evolve as the Reasoning Runtime and SceneLM improve. Minor breaks may land in minor versions — production systems should pin a client version.
Run the orchestrator over a scene + memory + policy snapshot. Returns candidate actions with confidence breakdown.
{
"scene": { ... },
"engines": ["scenelm", "rules"],
"include_trace": true
}Forward-model the next-state from the current scene and memory. Used for proactive intervention. Backed by SceneLM-predictive variant when available.
Compact natural-language summary of a scene window. Suitable for logs, dashboards, and human handoff.
Compute the validation envelope C_v for a candidate action without emitting it. Useful for shadow-mode testing.
{
"C_v": 0.84,
"components": {
"C_s": 0.92,
"C_m": 0.78,
"C_t": 0.81,
"C_h": 0.88
},
"threshold": 0.70,
"would_pass": true
}Replay the full reasoning trace for a past decision — scene, engine outputs, scoring, validation envelope, emitted action. Audit-ready.