Write down the backend <-> ai-worker contract

`stream:camera_frames` and `channel:print_alerts` are the only interface
between `continuum-backend` and `continuum-ai-worker`, and there was
nothing specifying either. They had drifted completely: the backend
wrote `farmId`, `printerId`, `r2ObjectKey` and `capturedAt`; the worker
read a single field called `printer_id` and ignored the rest. Every task
was delivered, acknowledged, and attributed to a printer called
"unknown" - and nothing errored, because a stream entry is a flat string
map and asking it for a field that is not there is not a failure. Same
shape as the /ws/edge/v1 drift, found the same way.

`schemas/continuum/v1/ai/` is now canonical, with a TypeBox mirror in
ts-types and a Rust mirror in the worker.

Two things the schemas have to say that JSON Schema does not naturally
express, both of which caused real bugs:

- A Valkey stream entry is a flat field/value map of *strings*. No
  nesting, no numbers, no null. So FrameTask has no numeric or nested
  properties, and a field with no value is omitted from the XADD rather
  than written as `""` - an empty string is a value, and the producer
  was writing one for `print_job_id`, which a consumer would carry into
  a lookup or a uuid column.
- PrintAlert, by contrast, *is* a JSON document, because PUBLISH carries
  one opaque string. It is also fire-and-forget with no persistence, so
  it can only ever be a live-console feed and never the record of a
  detection.

The README's layout section was stale (it still listed four DTOs that
were removed in 2026-09) and did not mention `edge/` or `ai/` at all. It
now names both, says why the inter-service protocols live here, and
notes that their Rust mirrors sit in the single repo that consumes each
rather than in `packages/rust-types`.

README.md

@@ -13,19 +13,39 @@ JSON Schema, compiled into two consumer packages:
 ## Layout
 
 ```
-schemas/continuum/v1/     Continuum's own domain model (PrintJob, Template, ...)
+schemas/continuum/v1/     Continuum's own model
+  ai/                     continuum-backend <-> continuum-ai-worker: the
+                          stream:camera_frames work queue and the
+                          channel:print_alerts notification channel
+  edge/                   continuum-backend <-> continuum-proxy: the
+                          /ws/edge/v1 gateway uplink protocol
+  material/  printer/     domain objects
+  vendor.schema.json
 schemas/bambulab/         Bambu Lab MQTT wire schemas (v1/v2, report + request)
 packages/ts-types/        TypeBox schemas (hand-written, mirrors schemas/)
 packages/rust-types/      serde structs (hand-written, mirrors schemas/)
 ```
 
+The two inter-service protocols - `edge/` and `ai/` - are here for a
+specific reason: each has two independent implementations, one per repo, and
+when there was nothing written down between them they drifted completely
+while every test passed. `/ws/edge/v1` had an `auth` handshake the gateway
+never sent; `stream:camera_frames` had a producer writing `printerId` and a
+consumer reading `printer_id`, so every frame was attributed to a printer
+called "unknown". Neither failure was visible from either side alone, which
+is what a canonical schema plus a contract test on each end is for.
+
 `schemas/continuum/v1/*.schema.json` is the canonical shape for each
 domain object; `packages/ts-types/src/schemas/` and
-`packages/rust-types/src/lib.rs` both mirror it by hand. There are only
+`packages/rust-types/src/lib.rs` both mirror it by hand. A mirror is added
-four small, stable DTOs here (`PrintJob`, `MaterialProfile`,
+per schema only once something actually consumes it, so not every file here
-`PrintParameters`, `Template`), so keeping two hand-synced copies in sync
+has one, and the two inter-service protocols keep their Rust mirror in the
-is less overhead than running a codegen pipeline (`protoc`/`ts-proto`/
+single repo that consumes it rather than in `packages/rust-types`:
-`prost-build`) nothing actually consumed - see git history for the
+`edge/` in `continuum-proxy/src/uplink/protocol.rs`, `ai/` in
+`continuum-ai-worker/src/task.rs`. The set is small and changes slowly,
+which is why keeping hand-synced copies in sync is less overhead than a
+codegen pipeline (`protoc`/`ts-proto`/`prost-build`) nothing actually
+consumed - see git history for the
 protobuf version this replaced, and why it was dropped: every service in
 this system talks JSON (JSON Schema for the wire, TypeBox/serde for
 runtime validation), so protobuf's actual value - a single schema compiled

packages/ts-types/src/schemas/ai.ts

@@ -0,0 +1,96 @@
+import { Type, type Static } from "@sinclair/typebox";
+
+/**
+ * The `stream:camera_frames` work queue and the `channel:print_alerts`
+ * notification channel, mirroring `schemas/continuum/v1/ai/*.schema.json`
+ * by hand - the JSON Schema files are canonical, and the Rust mirror lives
+ * in `continuum-ai-worker/src/task.rs`.
+ *
+ * These two are the only interface between `continuum-backend` and
+ * `continuum-ai-worker`. They are written down here because the previous
+ * arrangement - a producer and a consumer each naming their own fields,
+ * with nothing in between - produced exactly the drift it sounds like: the
+ * backend wrote `printerId` and the worker read `printer_id`, so every
+ * frame the worker ever handled was attributed to a printer called
+ * "unknown", and nothing failed.
+ */
+
+/**
+ * One entry on `stream:camera_frames`.
+ *
+ * ## Every value is a string
+ *
+ * A Valkey stream entry is a flat field/value map of strings. There is no
+ * nesting and no numeric type, which is why `FrameTaskSchema` has no
+ * numbers and no nested objects - a producer that needs one has to encode
+ * it, and then both sides have to agree on the encoding, which is the thing
+ * this file exists to avoid.
+ *
+ * ## An absent field is absent
+ *
+ * `print_job_id` is omitted from the XADD when the frame belongs to no
+ * tracked job - not written as `""`. An empty string is a value: a consumer
+ * that passes it along ends up looking up job id `""`, and one that writes
+ * it to Postgres violates a uuid column. So the producer omits the field and
+ * the consumer treats absent and empty alike.
+ */
+export const FrameTaskSchema = Type.Object({
+  /** `farms.id`. */
+  farm_id: Type.String({ format: "uuid" }),
+  /**
+   * `printers.id` - the backend's UUID, already resolved from the gateway's
+   * own printer id, which means nothing outside its farm.
+   */
+  printer_id: Type.String({ format: "uuid" }),
+  /** `print_jobs.id`, or omitted. Never `""`. */
+  print_job_id: Type.Optional(Type.String({ format: "uuid" })),
+  /** Where the frame is in R2. The bytes are never in the stream. */
+  r2_object_key: Type.String({ minLength: 1 }),
+  /**
+   * The gateway's clock, at capture - not enqueue time. A gateway that
+   * buffered frames through an uplink outage produces tasks whose
+   * `captured_at` is minutes old.
+   */
+  captured_at: Type.String({ format: "date-time" }),
+});
+export type FrameTask = Static<typeof FrameTaskSchema>;
+
+/**
+ * One message on `channel:print_alerts`, as JSON.
+ *
+ * Unlike a stream entry this one is a document, because PUBLISH carries a
+ * single opaque string. It is also fire-and-forget: an alert published with
+ * no subscriber is gone, so this channel is a live-console feed and not the
+ * record of a detection.
+ */
+export const PrintAlertSchema = Type.Object({
+  printer_id: Type.String({ format: "uuid" }),
+  print_job_id: Type.Optional(Type.String({ format: "uuid" })),
+  /**
+   * Free string, not an enum: the label set belongs to the model file and
+   * changes without this schema changing.
+   */
+  label: Type.String(),
+  /**
+   * The model's own score. Thresholding is the consumer's policy - an
+   * operator may well want to see near-misses.
+   */
+  confidence: Type.Number({ minimum: 0, maximum: 1 }),
+  /** From the frame, so a console shows when the printer looked like this. */
+  captured_at: Type.String({ format: "date-time" }),
+  /** So a console can show the picture, not just the label. */
+  r2_object_key: Type.String(),
+});
+export type PrintAlert = Static<typeof PrintAlertSchema>;
+
+/** The stream the tasks go on, and the group the worker reads as. */
+export const CAMERA_FRAMES_STREAM = "stream:camera_frames";
+export const AI_WORKER_CONSUMER_GROUP = "ai-worker";
+/**
+ * Where entries go that cannot be processed and must not be retried - a
+ * malformed task, or one that has been redelivered too many times. Separate
+ * from the main stream so a poison message cannot block the queue and
+ * cannot be lost silently either.
+ */
+export const CAMERA_FRAMES_DEAD_LETTER_STREAM = "stream:camera_frames:dead";
+export const PRINT_ALERTS_CHANNEL = "channel:print_alerts";

packages/ts-types/src/schemas/index.ts

@@ -9,3 +9,8 @@
 // a TypeBox mirror here, consumed by continuum-backend's edge WebSocket
 // route. Mirrors schemas/continuum/v1/edge/.
 export * from "./edge";
+
+// The continuum-backend -> continuum-ai-worker interface: the
+// stream:camera_frames work queue and the channel:print_alerts notification
+// channel. Mirrors schemas/continuum/v1/ai/.
+export * from "./ai";

schemas/continuum/v1/ai/frame_task.schema.json

@@ -0,0 +1,43 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://git.octoturge.com/Continuum/continuum-schemas/raw/branch/main/schemas/continuum/v1/ai/frame_task.schema.json",
+  "title": "FrameTask",
+  "description": "One camera still, handed from continuum-backend to continuum-ai-worker for print-failure scoring. Produced by XADD onto the stream named below; consumed by XREADGROUP.\n\nThis is a Valkey stream entry, not a JSON document: the wire form is a flat field/value map and EVERY value is a string. That is why there are no numeric or boolean properties here and why a field with no value is omitted from the XADD rather than written as an empty string - an empty string is a value, and a consumer that treats it as one ends up looking up the object key \"\".\n\nField names are snake_case, matching every other schema in this repository. The backend's HTTP routes are camelCase; this wire is not an HTTP route, and the consumer on the other end is serde-derived Rust.",
+  "x-stream": "stream:camera_frames",
+  "x-consumer-group": "ai-worker",
+  "type": "object",
+  "properties": {
+    "farm_id": {
+      "type": "string",
+      "format": "uuid",
+      "description": "`farms.id`. Carried so a consumer can attribute load and alerts per farm without a database round trip."
+    },
+    "printer_id": {
+      "type": "string",
+      "format": "uuid",
+      "description": "`printers.id` - the backend's own UUID, already resolved from the gateway's printer id. NOT the gateway's `printers.toml` id, which means nothing outside its own farm."
+    },
+    "print_job_id": {
+      "type": "string",
+      "format": "uuid",
+      "description": "`print_jobs.id`, when the frame belongs to a tracked job. Omitted entirely when it does not - a printer may be streaming while idle, or running something queued outside Continuum. Consumers must treat an absent key and an empty value alike, and must not fabricate a job id."
+    },
+    "r2_object_key": {
+      "type": "string",
+      "minLength": 1,
+      "description": "Where the frame actually is, in the R2 bucket. The frame bytes are deliberately NOT in the stream: a stream entry is held in memory by the server and replayed to every consumer group, and a JPEG per printer per second would make the stream the largest thing in the deployment."
+    },
+    "captured_at": {
+      "type": "string",
+      "format": "date-time",
+      "description": "When the camera took the frame, from the gateway's clock - not when the backend enqueued it. These can differ by a lot: a gateway buffers frames while its uplink is down, so a task may arrive minutes after the moment it describes."
+    }
+  },
+  "required": [
+    "farm_id",
+    "printer_id",
+    "r2_object_key",
+    "captured_at"
+  ],
+  "additionalProperties": false
+}

schemas/continuum/v1/ai/print_alert.schema.json

@@ -0,0 +1,47 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$id": "https://git.octoturge.com/Continuum/continuum-schemas/raw/branch/main/schemas/continuum/v1/ai/print_alert.schema.json",
+  "title": "PrintAlert",
+  "description": "A print-failure detection, published by continuum-ai-worker for live operator consoles.\n\nUnlike FrameTask this one IS a JSON document: it travels as the payload of a Valkey PUBLISH, which carries a single opaque string, so the structure has to be inside it.\n\nPUBLISH is fire-and-forget with no persistence and no delivery guarantee: an alert published while nothing is subscribed is gone. That is acceptable for the live-console path it feeds and is NOT acceptable as the record of a detection, which belongs in Postgres. As of 2026-10-11 nothing subscribes to this channel and nothing persists it - see the README section on the AI path.",
+  "x-channel": "channel:print_alerts",
+  "type": "object",
+  "properties": {
+    "printer_id": {
+      "type": "string",
+      "format": "uuid",
+      "description": "`printers.id`, taken from the frame task that produced the detection."
+    },
+    "print_job_id": {
+      "type": "string",
+      "format": "uuid",
+      "description": "Omitted when the frame was not attributed to a tracked job."
+    },
+    "label": {
+      "type": "string",
+      "description": "What the model thinks it saw, e.g. \"spaghetti\", \"layer_shift\", \"detachment\". A free string rather than an enum on purpose: the label set is a property of the model file, which changes without this schema changing."
+    },
+    "confidence": {
+      "type": "number",
+      "minimum": 0,
+      "maximum": 1,
+      "description": "The model's own score. Whether it is high enough to act on is the consumer's policy decision, not the worker's - the worker publishes what it found and does not threshold on behalf of an operator who may want to see near-misses."
+    },
+    "captured_at": {
+      "type": "string",
+      "format": "date-time",
+      "description": "From the frame, not from the detection. An operator needs to know when the printer looked like this, not when a GPU got round to looking."
+    },
+    "r2_object_key": {
+      "type": "string",
+      "description": "The frame the detection came from, so a console can show the operator the picture rather than only a label."
+    }
+  },
+  "required": [
+    "printer_id",
+    "label",
+    "confidence",
+    "captured_at",
+    "r2_object_key"
+  ],
+  "additionalProperties": false
+}