tiny camera worker

OpenAPI 3.1 draft · eyes for a tiny · openapi.json

Bind it

modify_ai worker: https://<this-site>/openapi.json

Set servers[0].variables.host to the real camera host before binding. Every operationId below mounts as a tool.

tool	route	what
camera_health	GET /health	alive, cameras live/total, battery, thermal
camera_list	GET /cameras	roster by name · state · identity_drift
camera_snapshot	POST /cameras/{name}/snapshot	one still → hosted image_url (+ToF mm)
camera_clip	POST /cameras/{name}/clip	1–30 s clip + keyframes the agent can see
camera_detect	POST /cameras/{name}/detect	FOMO boxes, labels, confidences, annotated frame
camera_stream_info	GET /cameras/{name}/stream	MJPEG/WebRTC URL + watchdog state
camera_settings_get / _set	GET·PATCH /cameras/{name}/settings	resolution, fps, exposure, flip, lamp, privacy shutter
camera_bind	POST /cameras/{name}/bind	pin a name to a device_id (clears drift, confirm:true)
sensor_distance	GET /sensors/distance	ToF mm + valid
sensor_imu	GET /sensors/imu	roll/pitch/yaw, moving

Names, not indices. An OS camera index slides when a device is unplugged; the roster name is bound to a stable device_id. identity_drift makes the wrong-camera case visible and capture refuses (409) unless the caller admits ignore_identity_drift.
Three liveness rails: stale (frame age), missing (not enumerated), drift (listed, different device answers).
Images are URLs. No base64 in JSON — the agent sees the hosted frame; the human gets the same link.
Clips return keyframes — the model cannot watch video, so the still ladder is the truth.
Privacy shutter is a setting only a human clears; while on, every capture is a 409.
frame_index on every Frame so shots can be bound to an ECoT sidecar.

v0.1 · drafted by tiny for ./c² · tiny.technology/tiny