@camstack/addon-osd-manager 0.1.179 → 0.1.181

package/dist/index.js

@@ -10614,7 +10614,53 @@ var ModelFormatEntrySchema = object({
 	/** Multi-file payload (directory members or sibling files). */
 	files: array(string()).readonly().optional(),
 	/** Runtime(s) that can use this format. If omitted, inferred from ModelFormat key */
-	runtimes: array(_enum(["python"])).readonly().optional()
+	runtimes: array(_enum(["python"])).readonly().optional(),
+	/**
+	* Content revision of the artefact, for one hosted OFF the CamStack mirror.
+	* A format on the mirror leaves it unset: its revision (the sha256 of its
+	* weight file) comes from the downloader's generated table, kept in step
+	* with the mirror by `scripts/check-hosted-model-revisions.ts` (D744). A
+	* node records the revision it downloaded next to the file
+	* (`<file>.revision`); a copy with no record or a different one is stale
+	* unless its weight file hashes to the revision, and a stale copy is
+	* downloaded again, the old one serving until the new one is complete. No
+	* revision at all ⇒ a file on disk is current, whatever it holds (D743).
+	*/
+	revision: string().min(1).optional()
+});
+/**
+* What a node holds for one (model, format) (D746). `current` is a complete
+* copy at its catalog revision — exactly what `downloaded: true` means.
+* `stale` is bytes on disk that are not (yet) proven current under D743/D744:
+* no or a different revision record, or a missing sibling. `missing` is no
+* copy. `downloaded` alone folds `stale` into "not downloaded", which hid
+* reclaimable bytes and a copy about to be replaced.
+*/
+var ModelOnDiskStateSchema = _enum([
+	"current",
+	"stale",
+	"missing"
+]);
+/**
+* The numeric precision of one format's BUILD (D746) — reported only when the
+* build states it (file-name token, an INT8 variant, an EdgeTPU compile, a
+* CamStack Core ML export). Absent means "not stated", never "fp32".
+*/
+var ModelFormatPrecisionSchema = _enum([
+	"fp32",
+	"fp16",
+	"int8"
+]);
+/**
+* One format of one model as a node's `pipelineExecutor.getSchema` reports it.
+* `onDisk` / `precision` are optional for version skew: a runner that predates
+* them omits them, and a reader falls back to `downloaded`.
+*/
+var PipelineModelFormatStatusSchema = object({
+	downloaded: boolean(),
+	sizeMB: number(),
+	onDisk: ModelOnDiskStateSchema.optional(),
+	precision: ModelFormatPrecisionSchema.optional()
 });
 /**
 * Extra file that must be downloaded alongside the model (e.g., labels JSON, dict.txt).
@@ -27377,7 +27423,7 @@ var FrameRefSchema = object({
 	timestamp: number(),
 	capturedAt: number().optional()
 });
-var ModelFormatSchema$1 = _enum([
+var ModelFormatSchema$2 = _enum([
 	"onnx",
 	"coreml",
 	"openvino",
@@ -27395,7 +27441,7 @@ var PipelineSlotSchema = _enum([
 var PipelineEngineChoiceSchema = object({
 	runtime: literal("python"),
 	backend: string(),
-	format: ModelFormatSchema$1,
+	format: ModelFormatSchema$2,
 	device: string().optional()
 });
 /**
@@ -27452,10 +27498,7 @@ var PipelineTemplateSchema$1 = object({
 var PipelineModelOptionSchema = object({
 	id: string(),
 	name: string(),
-	formats: record(string(), object({
-		downloaded: boolean(),
-		sizeMB: number()
-	})),
+	formats: record(string(), PipelineModelFormatStatusSchema),
 	group: ModelVariantGroupSchema.optional(),
 	legacy: boolean().optional(),
 	provider: ModelProviderIdSchema.optional(),
@@ -27670,6 +27713,32 @@ var pipelineExecutorCapability = {
 			auth: "admin"
 		}),
 		/**
+		* Best settings (D745): the MODEL-BOUND thresholds (`bindsToModel` keys of
+		* the object-detection root, and their binding keys) stored in THIS node's
+		* per-camera detector settings — and, unless `dryRun`, remove them so the
+		* running model's own defaults apply. Every other key of a camera's store
+		* is kept. `nodeId` is the ROUTING key. `deviceIds` absent ⇒ every camera
+		* with a store. A camera whose store could not be read is `refused`, never
+		* reported as clean.
+		*/
+		clearModelBoundDeviceOverrides: method(object({
+			nodeId: string(),
+			deviceIds: array(number()).readonly().optional(),
+			dryRun: boolean()
+		}), object({
+			entries: array(object({
+				deviceId: number(),
+				keys: array(string()).readonly()
+			})).readonly(),
+			refused: array(object({
+				deviceId: number(),
+				reason: string()
+			})).readonly()
+		}), {
+			kind: "mutation",
+			auth: "admin"
+		}),
+		/**
 		* Which of THIS node's inference devices the executor currently refuses,
 		* and why. `nodeId` is the ROUTING key (stripped by the generated router).
 		*
@@ -27764,12 +27833,12 @@ var pipelineExecutorCapability = {
 		downloadModel: method(object({
 			addonId: string(),
 			modelId: string(),
-			format: ModelFormatSchema$1
+			format: ModelFormatSchema$2
 		}), DownloadModelResultSchema, { kind: "mutation" }),
 		deleteModel: method(object({
 			addonId: string(),
 			modelId: string(),
-			format: ModelFormatSchema$1
+			format: ModelFormatSchema$2
 		}), object({ success: literal(true) }), { kind: "mutation" }),
 		/**
 		* Stateless single-frame execution. Callers (runner, benchmark) pass
@@ -28015,6 +28084,307 @@ var pipelineExecutorCapability = {
 		getDetectionConfigSchema: method(_void(), ConfigUISchemaNullableBridge)
 	}
 };
+/**
+* Wire shapes of the Load balancing surface (D748): the session placement
+* policy, the atomic share write, the observed placement statistics, the
+* hardware share suggestion and the stale device pins.
+*/
+/**
+* How a new detection session picks its node and device.
+*  - `least-loaded` — the historical rule: lowest `attached / weight`, a
+*    constant node order at a tie (D37). An idle node always ties at 0, so a
+*    weight only bites once a node already holds a session.
+*  - `share` — the share of PARALLEL sessions is held to the weights, and a
+*    full or drowning node yields to a free one (D748).
+*/
+var SessionPlacementPolicySchema = _enum(["least-loaded", "share"]);
+var LoadBalancingSharesInputSchema = object({
+	/** nodeId → share (> 0). A node left out keeps its weight. */
+	nodes: record(string(), number().positive()).default({}),
+	/** nodeId → deviceKey → share (> 0). A device left out keeps its weight. */
+	devices: record(string(), record(string(), number().positive())).default({})
+});
+var PlacementStatsInputSchema = object({ 
+/** Window the observed shares are averaged over (default 1 h, max 24 h). */
+windowMs: number().int().min(6e4).max(1440 * 60 * 1e3).default(3600 * 1e3) });
+var PlacementDeviceStatsSchema = object({
+	deviceKey: string(),
+	/** Sessions running on this device now. */
+	sessionsNow: number(),
+	/** Sessions started on this device inside the window. */
+	sessionsStarted: number(),
+	/** Time-weighted share of the node's parallel sessions over the window (0–1); null = no session ran. */
+	observedShare: number().nullable()
+});
+var PlacementNodeStatsSchema = object({
+	nodeId: string(),
+	sessionsNow: number(),
+	sessionsStarted: number(),
+	/** Time-weighted share of the cluster's parallel sessions over the window (0–1); null = none ran. */
+	observedShare: number().nullable(),
+	/** Most sessions this node ran at once inside the window. */
+	peakSessions: number(),
+	/** Sessions a share election handed elsewhere because this node was full or drowning. */
+	yieldsAway: number(),
+	devices: array(PlacementDeviceStatsSchema).readonly()
+});
+/** A camera's device pin that the session path cannot honour. */
+var StaleDevicePinSchema = object({
+	deviceId: number(),
+	deviceKey: string(),
+	/** `disabled` — the key exists on some node but is not enabled anywhere; `absent` — no node has it. */
+	reason: _enum(["disabled", "absent"])
+});
+var PlacementStatsSchema = object({
+	policy: SessionPlacementPolicySchema,
+	windowMs: number(),
+	/** How much of the window was actually observed (the hub keeps it in memory). */
+	observedForMs: number(),
+	sessionsNow: number(),
+	/** Most parallel sessions across the cluster inside the window. */
+	clusterPeakSessions: number(),
+	nodes: array(PlacementNodeStatsSchema).readonly(),
+	stalePins: array(StaleDevicePinSchema).readonly()
+});
+var SuggestSharesInputSchema = object({ headroom: number().min(.3).max(.5).default(.4) });
+var SuggestedDeviceShareSchema = object({
+	deviceKey: string(),
+	/** Share of the node (0–100, rounded). */
+	share: number(),
+	modelId: string().nullable(),
+	/** Inferences per second at that model, from the measured table. */
+	fps: number().nullable(),
+	measured: boolean(),
+	/** Advisory session cap: ⌊fps × (1 − headroom) / camera fps⌋. Never enforced. */
+	suggestedMaxSessions: number().nullable()
+});
+var SuggestedNodeShareSchema = object({
+	nodeId: string(),
+	/** Share of the cluster (0–100, rounded); null = node skipped (offline, unknown hardware). */
+	share: number().nullable(),
+	/** Parallel sessions the node sustains at the suggested models with the headroom kept free. */
+	capacitySessions: number(),
+	devices: array(SuggestedDeviceShareSchema).readonly(),
+	skippedReason: string().nullable()
+});
+var SuggestSharesResultSchema = object({
+	headroom: number(),
+	/** Camera detection rate the capacity is counted in (the fastest camera's fps). */
+	cameraFps: number(),
+	/** How many share → model → capacity rounds ran before the shares stopped moving (≤ 3). */
+	iterations: number(),
+	nodes: array(SuggestedNodeShareSchema).readonly()
+});
+/**
+* The wire shapes of the Best settings plan and apply (D745), used by the
+* `pipeline-orchestrator` cap. The recommender's own types
+* (`recommend.ts`, `demand.ts`) are the TS face of the same data; these
+* schemas are what crosses the cap boundary.
+*/
+var ModelFormatSchema$1 = _enum(MODEL_FORMATS);
+var HardwareClassIdSchema = _enum([
+	"intel-core-ultra",
+	"intel-n-series",
+	"apple-m-base",
+	"apple-m-pro",
+	"unknown"
+]);
+var AcceleratorKindSchema = _enum([
+	"intel-npu",
+	"intel-igpu",
+	"x86-cpu",
+	"arm-cpu",
+	"apple-ane",
+	"coral"
+]);
+var RecommendationReasonSchema = discriminatedUnion("kind", [
+	object({
+		kind: literal("fits"),
+		utilisation: number()
+	}),
+	object({
+		kind: literal("tie-faster"),
+		over: string(),
+		recallDelta: number()
+	}),
+	object({
+		kind: literal("stepped-down"),
+		from: string(),
+		to: string(),
+		limitedBy: string()
+	}),
+	object({
+		kind: literal("no-fit-lightest"),
+		utilisation: number()
+	}),
+	object({
+		kind: literal("derived-latency"),
+		modelId: string()
+	}),
+	object({
+		kind: literal("estimated-load"),
+		observedForMs: number().nullable()
+	}),
+	object({ kind: literal("cpu-device") }),
+	object({ kind: literal("coral-untouched") }),
+	object({ kind: literal("device-disabled") }),
+	object({ kind: literal("device-unavailable") }),
+	object({ kind: literal("node-offline") }),
+	object({ kind: literal("no-hardware-facts") }),
+	object({
+		kind: literal("no-measured-data"),
+		hardwareClass: HardwareClassIdSchema
+	}),
+	object({
+		kind: literal("no-family-build"),
+		format: ModelFormatSchema$1
+	})
+]);
+var DeviceDemandSchema = object({
+	nodeId: string(),
+	deviceKey: string(),
+	demandFps: number(),
+	maxCameraFps: number(),
+	alwaysOnCameras: number(),
+	onMotionSessions: number(),
+	onMotionSource: _enum([
+		"observed-24h",
+		"estimate",
+		"none"
+	]),
+	observedForMs: number().nullable(),
+	/** How on-motion sessions were charged: per-device observation, or split by share (D748). */
+	placement: SessionPlacementPolicySchema
+});
+var DeviceModelChoiceSchema = object({
+	modelId: string(),
+	p50Ms: number(),
+	fps: number(),
+	measured: boolean(),
+	utilisation: number(),
+	personRecall: number()
+});
+var DeviceRecommendationSchema = object({
+	deviceKey: string(),
+	accelerator: AcceleratorKindSchema.nullable(),
+	enabled: boolean(),
+	available: boolean(),
+	demand: DeviceDemandSchema.nullable(),
+	chosen: DeviceModelChoiceSchema.nullable(),
+	reasons: array(RecommendationReasonSchema).readonly()
+});
+var NodeRecommendationSchema = object({
+	nodeId: string(),
+	reachable: boolean(),
+	hardwareClass: HardwareClassIdSchema.nullable(),
+	cpuModel: string().nullable(),
+	skipped: boolean(),
+	acceleratorModelId: string().nullable(),
+	devices: array(DeviceRecommendationSchema).readonly(),
+	reasons: array(RecommendationReasonSchema).readonly()
+});
+var ModelPinChangeSchema = object({
+	stepId: string(),
+	before: string().nullable(),
+	after: string().nullable()
+});
+var NodeDeviceChangeSchema = object({
+	nodeId: string(),
+	deviceKey: string(),
+	madeExplicit: boolean(),
+	modelPins: array(ModelPinChangeSchema).readonly(),
+	thresholds: array(object({
+		stepId: string(),
+		keys: array(string()).readonly()
+	})).readonly()
+});
+var CameraOverrideClearSchema = object({
+	deviceId: number(),
+	source: _enum(["step-override", "wholesale"]),
+	nodeId: string(),
+	deviceKey: string().nullable(),
+	stepId: string(),
+	modelIdBefore: string().nullable(),
+	keys: array(string()).readonly()
+});
+var DetectorStoreClearSchema = object({
+	deviceId: number(),
+	keys: array(string()).readonly()
+});
+var BestSettingsWarningSchema = discriminatedUnion("kind", [
+	object({
+		kind: literal("detector-store-cluster-wide"),
+		skippedNodes: array(string()).readonly()
+	}),
+	object({
+		kind: literal("detector-store-unreadable"),
+		error: string()
+	}),
+	object({
+		kind: literal("coral-enabled"),
+		nodeId: string(),
+		deviceKey: string()
+	})
+]);
+var BestSettingsOptionsSchema = object({
+	/** Share of each accelerator's throughput kept free (0.3–0.5). */
+	headroom: number().min(.3).max(.5).default(.4),
+	/**
+	* What-if shares (D748): plan as if these node/device shares were stored
+	* and the policy were `share`. Absent = the stored shares under the stored
+	* policy. A what-if plan is for display; apply always re-plans from what
+	* is stored.
+	*/
+	shares: LoadBalancingSharesInputSchema.optional()
+});
+var BestSettingsPlanSchema = object({
+	/** Hash of everything the apply will write — the apply refuses a different one. */
+	planHash: string(),
+	family: literal("yolov9"),
+	headroom: number(),
+	/** Planned under what-if shares (D748): for display — apply refuses it as stale. */
+	whatIf: boolean(),
+	generatedAt: number(),
+	nodes: array(NodeRecommendationSchema).readonly(),
+	nodeChanges: array(NodeDeviceChangeSchema).readonly(),
+	cameraClears: array(CameraOverrideClearSchema).readonly(),
+	detectorStoreClears: array(DetectorStoreClearSchema).readonly(),
+	warnings: array(BestSettingsWarningSchema).readonly()
+});
+var BestSettingsApplyInputSchema = object({
+	planHash: string(),
+	headroom: number().min(.3).max(.5).default(.4)
+});
+var BestSettingsPhaseResultSchema = object({
+	phase: _enum([
+		"download",
+		"camera-overrides",
+		"detector-thresholds",
+		"node-models",
+		"redispatch"
+	]),
+	ok: boolean(),
+	detail: string()
+});
+var BestSettingsApplyResultSchema = discriminatedUnion("kind", [
+	object({
+		kind: literal("applied"),
+		planHash: string(),
+		phases: array(BestSettingsPhaseResultSchema).readonly()
+	}),
+	object({
+		kind: literal("aborted"),
+		phases: array(BestSettingsPhaseResultSchema).readonly()
+	}),
+	object({
+		kind: literal("stale"),
+		plan: BestSettingsPlanSchema
+	}),
+	object({
+		kind: literal("refused"),
+		reason: _enum(["apply-in-progress"])
+	})
+]);
 object({
 	activeCameras: number(),
 	throttledCameras: number(),
@@ -29427,7 +29797,6 @@ var CameraStepOverridePatchSchema = object({
 	settings: record(string(), unknown()).readonly().optional()
 });
 var CameraPipelineSettingsSchema = object({
-	pinnedAgentNodeId: string().optional(),
 	stepToggles: record(string(), boolean()).optional(),
 	stepOverridesByDevice: record(string(), record(string(), record(string(), CameraStepOverridePatchSchema))).optional(),
 	pipelineByAgent: record(string(), CameraPipelineForAgentSchema).optional()
@@ -30075,18 +30444,6 @@ var pipelineOrchestratorCapability = {
 			auth: "admin"
 		}),
 		/**
-		* Set a node's detection WEIGHT (relative share for the quota balancer).
-		* `null` clears it (back to the implicit weight 1). Unpinned cameras
-		* distribute proportionally to weight; `maxCameras` still clamps on top.
-		*/
-		setAgentDetectWeight: method(object({
-			agentNodeId: string(),
-			detectWeight: number().positive().nullable()
-		}), object({ success: literal(true) }), {
-			kind: "mutation",
-			auth: "admin"
-		}),
-		/**
 		* Set one node's placement CAPABILITIES — the per-node record that
 		* replaced the four flat orchestrator lists (`enabledNodes`,
 		* `enabledDecoderNodes`, `enabledAudioNodes`, `remoteSourcingNodes`).
@@ -30205,6 +30562,54 @@ var pipelineOrchestratorCapability = {
 			kind: "mutation",
 			auth: "admin"
 		}),
+		/**
+		* Best settings (D745) — DRY RUN. Recommends ONE detector family (YOLOv9)
+		* for the whole cluster from each node's static hardware class, the
+		* measured latency table and the detection demand, and lists every write
+		* the apply would make: the node model pins set, the enrichment pins and
+		* model-bound thresholds removed (node level and per camera), and the
+		* per-camera detector thresholds cleared. Writes nothing. Step
+		* enablement, device enablement, pins, recording, notifications, zones and
+		* rules are never part of it. `planHash` names exactly this diff.
+		*/
+		planBestSettings: method(BestSettingsOptionsSchema, BestSettingsPlanSchema, { auth: "admin" }),
+		/**
+		* Best settings (D745) — APPLY the plan the operator confirmed. Recomputes
+		* the plan and refuses (`stale`, with the fresh plan) when its hash is not
+		* `planHash`, so what is written is exactly the confirmed diff. Offline
+		* nodes are skipped; there is no undo.
+		*/
+		applyBestSettings: method(BestSettingsApplyInputSchema, BestSettingsApplyResultSchema, {
+			kind: "mutation",
+			auth: "admin"
+		}),
+		/**
+		* Load balancing (D748) — write node shares (`detectWeight`) and per-node
+		* device shares (inference-device `weight`) in ONE atomic read-modify-write
+		* of the agent settings — the only writer of node shares (the per-node
+		* `setAgentDetectWeight` it replaced raced its siblings, one RMW each).
+		* A node or device left out keeps its weight. Never changes the
+		* `sessionPlacement` policy.
+		*/
+		setLoadBalancingShares: method(LoadBalancingSharesInputSchema, object({ success: literal(true) }), {
+			kind: "mutation",
+			auth: "admin"
+		}),
+		/**
+		* Load balancing (D748) — what placement actually did: per node and
+		* device, sessions now, sessions started and the time-weighted share of
+		* PARALLEL sessions over `windowMs`; the cluster's parallel peak; yields
+		* a share election made; and every device pin the session path cannot
+		* honour. In memory on the hub — a restart starts a new window.
+		*/
+		getPlacementStats: method(PlacementStatsInputSchema, PlacementStatsSchema, { auth: "admin" }),
+		/**
+		* Load balancing (D748) — shares proportional to each node's measured
+		* capacity at the models Best settings would pick, iterated (≤ 3 rounds)
+		* until the shares stop moving; per device, an ADVISORY session cap.
+		* Writes nothing.
+		*/
+		suggestShares: method(SuggestSharesInputSchema, SuggestSharesResultSchema, { auth: "admin" }),
 		/** Read one camera's settings. Null when never touched (inherits agent defaults fully). */
 		getCameraSettings: method(object({ deviceId: number() }), CameraPipelineSettingsSchema.nullable()),
 		/** Set or clear the 3-state toggle for one (camera, addonId). Pass `enabled: null` to clear and revert to agent default. */
@@ -51058,6 +51463,12 @@ Object.freeze({
 		addonId: null,
 		access: "delete"
 	},
+	"pipelineExecutor.clearModelBoundDeviceOverrides": {
+		capName: "pipeline-executor",
+		capScope: "system",
+		addonId: null,
+		access: "delete"
+	},
 	"pipelineExecutor.deleteModel": {
 		capName: "pipeline-executor",
 		capScope: "system",
@@ -51274,6 +51685,12 @@ Object.freeze({
 		addonId: null,
 		access: "view"
 	},
+	"pipelineOrchestrator.applyBestSettings": {
+		capName: "pipeline-orchestrator",
+		capScope: "system",
+		addonId: null,
+		access: "create"
+	},
 	"pipelineOrchestrator.applyDeviceSettingsPatch": {
 		capName: "pipeline-orchestrator",
 		capScope: "system",
@@ -51424,6 +51841,12 @@ Object.freeze({
 		addonId: null,
 		access: "view"
 	},
+	"pipelineOrchestrator.getPlacementStats": {
+		capName: "pipeline-orchestrator",
+		capScope: "system",
+		addonId: null,
+		access: "view"
+	},
 	"pipelineOrchestrator.getReconcileFromDiskStatus": {
 		capName: "pipeline-orchestrator",
 		capScope: "system",
@@ -51448,6 +51871,12 @@ Object.freeze({
 		addonId: null,
 		access: "create"
 	},
+	"pipelineOrchestrator.planBestSettings": {
+		capName: "pipeline-orchestrator",
+		capScope: "system",
+		addonId: null,
+		access: "view"
+	},
 	"pipelineOrchestrator.rebalance": {
 		capName: "pipeline-orchestrator",
 		capScope: "system",
@@ -51508,12 +51937,6 @@ Object.freeze({
 		addonId: null,
 		access: "create"
 	},
-	"pipelineOrchestrator.setAgentDetectWeight": {
-		capName: "pipeline-orchestrator",
-		capScope: "system",
-		addonId: null,
-		access: "create"
-	},
 	"pipelineOrchestrator.setAgentInferenceDevices": {
 		capName: "pipeline-orchestrator",
 		capScope: "system",
@@ -51568,12 +51991,24 @@ Object.freeze({
 		addonId: null,
 		access: "create"
 	},
+	"pipelineOrchestrator.setLoadBalancingShares": {
+		capName: "pipeline-orchestrator",
+		capScope: "system",
+		addonId: null,
+		access: "create"
+	},
 	"pipelineOrchestrator.setPipelineDevicePin": {
 		capName: "pipeline-orchestrator",
 		capScope: "system",
 		addonId: null,
 		access: "create"
 	},
+	"pipelineOrchestrator.suggestShares": {
+		capName: "pipeline-orchestrator",
+		capScope: "system",
+		addonId: null,
+		access: "view"
+	},
 	"pipelineOrchestrator.unassignAudio": {
 		capName: "pipeline-orchestrator",
 		capScope: "system",
@@ -57046,6 +57481,11 @@ Object.freeze({
 		form: "single",
 		optional: false
 	}],
+	"pipelineExecutor.clearModelBoundDeviceOverrides": [{
+		name: "deviceIds",
+		form: "array",
+		optional: true
+	}],
 	"pipelineExecutor.runPipeline": [{
 		name: "deviceId",
 		form: "single",