central_ai_llm_master/docs/hub-api.md

The Hub Inbound API (v1)

Purpose: one frozen interface from every way in to the Hub. Applies to: phone app, ESP32 voice, dsh chat, admin app, console portal. Status: frozen. Build the Hub server and every client to this document. Decided: 2026-10-10.

hub_brief.md is the source of truth for the design. This document freezes the wire format. If this document and the brief disagree, the brief wins.

---

1. Service language (decided)

Go 1.25.

Reason (one line): The Hub is a small, always-on gateway on the fast path, so a single static Go binary gives hard timeouts, no runtime dependency drift, instant start, and a tiny container, and nothing in the Hub needs the Python data or ML stack because every model call goes through OmniRoute over HTTP.

Build notes:

• Standard library net/http with context deadlines. Small dependency list.
• Multi-stage Docker build. Run on a static or distroless base image.
• Deploy as one container on 192.168.20.13 (the agreed host).
• One binary, one port, one folder. Do not merge a destination into the Hub.

Go 1.25 is available from nixpkgs (a derivation is already in the local store). Docker 29 runs on the build host.

2. Two different /ask endpoints

The name /ask appears twice. They are different services on different ports. Do not confuse them.

Endpoint	Caller	Server	Defined in
POST /ask	A way in (phone, voice, chat, admin)	The Hub	This document
POST /ask	The Hub	A data project (email, school, internet control)	docs/ask-contract.md

This document freezes only the Hub endpoint. The Hub sends the project endpoint with the shape in docs/ask-contract.md.

3. Transport and authentication

Base

• Base URL: http://192.168.20.13:8788 (the allocated Hub port, HUB_PORT; recorded in docs/outline/02-config-and-env.md).
• Plain HTTP on the private LAN for v1. Put a TLS reverse proxy in front later.
• Request and response body: JSON, application/json; charset=utf-8.
• Maximum request body: 64 KiB. Larger bodies get 413.

Authentication

Every request carries a bearer token, except GET /health:

Authorization: Bearer <token>

• One token per client, not per person. Store each token in ~/.config/environment.d/10-secrets.conf.
• The Hub holds the accepted tokens. It compares them in constant time.
• The token names the device or channel. The token does not name the person.

Env key	Client
HUB_TOKEN_PHONE	Phone app.
HUB_TOKEN_ESP32	House voice (ESP32, "Hey Jervis").
HUB_TOKEN_DSH	dsh chat.
HUB_TOKEN_CONSOLE	Console portal.
HUB_TOKEN_ADMIN	Admin app. The only trusted door.

Auth matrix

Route	Accepted tokens
POST /ask	Any client token.
POST /approve	HUB_TOKEN_ADMIN only.
GET /health	None.

Voice is shared, not trusted

Voice is open to everyone in the house. The ESP32 token proves the device, not the speaker. The Hub never treats a voice request as proof of identity. A restricted action always needs the admin app.

One identity

There is one identity: Sam. The asker field must be "sam". Any other value gets 422 unsupported_asker. Keep the field so a second person can be added later.

4. Common response rules

Success and partial success

A well-formed, authorized, processed request returns HTTP 200, even when the Hub cannot route the request. The top-level status field says what happened.

status	Meaning
ok	A reply is ready.
needs_clarification	The Hub needs one short answer. answer holds the question.
needs_approval	A Hub-owned approval is required. approval holds the id.
no_match	The Hub cannot route the request.
refused	The request is a high-consequence control. The Hub does not do it.
unavailable	A destination or dependency is down.

Error envelope

An error uses a non-200 status and this shape:

{
  "status": "error",
  "error": {
    "code": "bad_request",
    "message": "Field 'text' is required.",
    "request_id": "01J8Z3K4M5N6P7Q8R9S0T1V2W3"
  }
}

message is one short sentence. It is safe to speak aloud. It holds no secret, no internal path, and no stack trace.

HTTP status and error codes

HTTP	code	Use
400	bad_request	Bad JSON, missing field, or bad enum value.
401	unauthorized	Missing or invalid bearer token.
403	forbidden	Valid token, wrong route. For example, a non-admin calls /approve.
403	source_mismatch	The source field does not match the token.
404	not_found	Unknown path, or unknown approval_id.
409	not_hub_owned	The approval belongs to another service.
409	already_decided	The approval was already decided.
409	approval_expired	The approval passed its expires time.
413	payload_too_large	Body over 64 KiB.
415	unsupported_media_type	Body is not application/json.
422	unprocessable	Valid JSON, bad content. For example, empty text.
422	unsupported_asker	asker is not "sam".
429	rate_limited	Too many requests. See Retry-After.
500	internal	Unexpected Hub fault.
502	destination_error	A destination failed to answer.
503	unavailable	A dependency is down (OmniRoute, MQTT).
504	timeout	A deadline passed.

Rate limit

Cap each token at 30 requests per minute. On breach, return 429 with a Retry-After header in seconds. This protects the Hub. It is not a security control.

Versioning

Add fields. Do not remove or rename a field. Clients must ignore unknown fields. A breaking change needs a new path, for example /ask/v2, /approve/v2, or /health/v2.

---

5. POST /ask

A way in sends words to the Hub. The Hub runs the five steps: hear, understand, check, route, answer. A way in needs only this route to ask.

Request

{
  "text": "what does Harry have on today?",
  "asker": "sam",
  "source": "esp32",
  "person": "harry",
  "want": "speech",
  "device": "kitchen-esp32",
  "request_id": "01J8Z3K4M5N6P7Q8R9S0T1V2W3"
}

Field	Required	Type	Meaning
text	yes	string	The words the user said, or typed. 1 to 1000 characters.
asker	yes	string	Who asked. Must be "sam" in v1.
source	yes	enum	Where it came from. See below.
person	no	string	Who the request is about. Defaults to asker.
want	no	enum	Reply shape the caller prefers: text, speech, both. Default comes from source.
device	no	string	Device name, for speech routing.
request_id	no	string	Client id. The Hub makes one when absent.

source values: esp32, phone, chat, console, admin.

The token must belong to the source. For example, HUB_TOKEN_DSH cannot claim source: "admin". A mismatch gets 403 source_mismatch.

Response — a normal reply

{
  "status": "ok",
  "request_id": "01J8Z3K4M5N6P7Q8R9S0T1V2W3",
  "answer": "Harry has maths and a music lesson today.",
  "shape": "speech",
  "speech": { "text": "Harry has maths and a music lesson today.", "target": "house" },
  "route": { "kind": "project", "project": "compass_nhs" },
  "approval": null,
  "clarify": null,
  "expires": "2026-10-10T09:00:00Z"
}

Field	Required	Meaning
status	yes	See the table in section 4.
request_id	yes	Echo of the request id, or a new one.
answer	yes	One short sentence. Safe to speak aloud.
shape	yes	text, speech, notification, or none.
speech	no	TTS detail. Present when shape is speech.
route	no	What the Hub did. For logging and debug.
approval	no	Present when status is needs_approval.
clarify	no	Present when status is needs_clarification.
expires	no	When the answer stops being true.

shape values:

• text — text only. Chat and console.
• speech — the Hub also speaks the reply.
• notification — a phone or desktop notification.
• none — no reply body. The action ran.

speech.target values: house (house speakers), phone (the asking phone), none.

route.kind values: direct (Home Assistant, speakers), project (a data project through /ask), fanout (several projects), paperclip, clarify, no_match, refused.

Response — the Hub needs a short answer

{
  "status": "needs_clarification",
  "request_id": "01J8Z3K4M5N6P7Q8R9S0T1V2W3",
  "answer": "School or email?",
  "shape": "speech",
  "speech": { "text": "School or email?", "target": "house" },
  "route": { "kind": "clarify" },
  "clarify": { "question": "School or email?", "options": ["school", "email"] }
}

Response — a Hub-owned approval is required

A slow, recorded action (a Paperclip job) needs a yes. The Hub does not run it yet. It returns an approval id.

{
  "status": "needs_approval",
  "request_id": "01J8Z3K4M5N6P7Q8R9S0T1V2W3",
  "answer": "I need a yes on the admin app to file that job.",
  "shape": "speech",
  "speech": { "text": "I need a yes on the admin app to file that job.", "target": "house" },
  "route": { "kind": "paperclip" },
  "approval": {
    "approval_id": "ap_01J8Z3K4M5N6P7Q8R9S0T1V2W3",
    "action": "paperclip_job",
    "summary": "File a job: plan the kitchen lights.",
    "reason": "A slow, multi-step job needs a yes.",
    "expires": "2026-10-10T09:10:00Z"
  }
}

approval.action values in v1: paperclip_job. Add values later. Never remove one.

Response — a high-consequence control

The Hub does not route a control action. It refuses and points at the admin app.

{
  "status": "refused",
  "request_id": "01J8Z3K4M5N6P7Q8R9S0T1V2W3",
  "answer": "Use the admin app to change internet limits.",
  "shape": "speech",
  "speech": { "text": "Use the admin app to change internet limits.", "target": "house" },
  "route": { "kind": "refused", "reason": "high_consequence_control" }
}

Response — no match

{
  "status": "no_match",
  "request_id": "01J8Z3K4M5N6P7Q8R9S0T1V2W3",
  "answer": "I do not know where to send that.",
  "shape": "text",
  "route": { "kind": "no_match" }
}

Behaviour rules

1. The Hub never parses a project's data. It forwards the question.
2. The Hub never changes the meaning of an answer. It only changes the shape.
3. The Hub never places an LLM in the path of a fast or consequential action (brief rule 4). Routing uses an LLM. The action call does not.
4. One project down is not an outage. Return status: "unavailable" and say so.
5. Every request is logged. See section 10.

6. POST /approve

The admin app answers "Did you ask for this?". This route records one yes or no for a Hub-owned approval and then runs the approved action.

In v1 the only Hub-owned approval is a Paperclip job. The Hub files the job once, then returns the result.

Request

{
  "approval_id": "ap_01J8Z3K4M5N6P7Q8R9S0T1V2W3",
  "decision": "approve",
  "asker": "sam",
  "device": "sam-phone"
}

Field	Required	Type	Meaning
approval_id	yes	string	The id from the /ask response.
decision	yes	enum	approve or deny.
asker	yes	string	Who decided. Must be "sam" in v1.
device	no	string	The admin device.

Response

{
  "status": "ok",
  "approval_id": "ap_01J8Z3K4M5N6P7Q8R9S0T1V2W3",
  "decision": "approve",
  "result": {
    "kind": "paperclip_job",
    "status": "filed",
    "detail": "Issue filed."
  },
  "answer": "Filed the job."
}

result.status values: filed, denied, failed.

Rules

1. An approval is single use. A second decision gets 409 already_decided.
2. An unknown id gets 404 not_found.
3. An id owned by another service gets 409 not_hub_owned. The admin app must call that service directly.
4. An expired approval gets 409 approval_expired. The user must ask again.
5. Only HUB_TOKEN_ADMIN may call this route.

Parental controls are not here

A parental control change is a high-consequence control. The Hub does not execute it (brief rule 11). The admin app shows the confirmation and calls the internet control service directly. This route is not in that path.

7. GET /health

Liveness only. No authentication.

Response — healthy

{
  "ok": true,
  "service": "hub",
  "version": "1",
  "uptime_s": 4321,
  "checks": {
    "omniroute": "ok",
    "mqtt": "ok",
    "log": "ok"
  }
}

Response — degraded

{
  "ok": false,
  "service": "hub",
  "version": "1",
  "uptime_s": 4321,
  "checks": {
    "omniroute": "ok",
    "mqtt": "down",
    "log": "ok"
  }
}

HTTP 200 when ok is true. HTTP 503 when ok is false.

Each checks value is ok, down, or unknown. A check never blocks the health reply for longer than 1 second. The health reply holds no secret and no personal data.

8. Timeouts

Every timeout is a hard cap. The Hub cancels the work when the cap passes.

Step	Cap
Read the request body	1 s
Routing LLM call (OmniRoute)	3 s
One project /ask call	3 s
Fanout (all projects, in parallel)	4 s
One destination command (Home Assistant, MQTT)	2 s
POST /ask total	8 s
POST /approve total	5 s
GET /health total	1 s

House TTS is out of band. The Hub starts it and does not wait for the speakers. Phone TTS is part of the reply.

Data project rules stay as written in docs/ask-contract.md: a project answers in under 3 seconds, or it returns answer: null.

9. Reply shape and TTS target

The want field is a preference. The Hub picks the final shape. It uses source and want:

source	Default shape	Default speech.target
esp32	speech	house
phone	speech	phone
chat	text	none
console	text	none
admin	text	none

want: "both" returns text and adds speech. want: "text" never speaks. A spoken reply is always one short sentence.

This settles the open question "when does phone TTS get used?": the Hub speaks on the asking phone when the source is the phone, and on the house speakers when the source is the house voice. The caller may override with want. Artemis transport (stream or file) is still open. See hub_brief.md.

10. Logging

The Hub writes one append-only log (brief rule 13).

One line per request. One line holds:

Field	Example
time	2026-10-10T08:55:03Z
request_id	01J8Z3K4M5N6P7Q8R9S0T1V2W3
source	esp32
asker	sam
text	what does Harry have on today?
route	project:compass_nhs
status	ok
latency_ms	412
shape	speech

Rules:

• Append only. Never edit a line.
• Cap the size. Rotate at a fixed size. Keep a fixed number of days.
• No secrets. Never log a token.
• No full personal data. Log the question and the route, not a project's data.
• Never log the Authorization header.

11. Safety confirmation

The three frozen routes are not high-consequence control actions.

Route	High-consequence control?	Why
POST /ask	No	Routes a query or a fast command. Refuses a control action with status: "refused".
POST /approve	No	Decides a Hub-owned approval only. Files a Paperclip job. Runs no parental control.
GET /health	No	Read only.

A parental control change goes: admin app → internet control service, directly. The Hub is not in that path. See brief rules 7, 8, and 11.

12. Not in v1

POST /notify (a project tells the Hub) is part of the design. Its shape is defined in docs/ask-contract.md. It is not one of the three frozen Hub-inbound routes in this document. Freeze it in a later task.

routing.md (the human-edited routing rules) is still open. POST /ask depends on it. See hub_brief.md, "What is still open", item 5.

The Artemis transport (stream or file) is still open. See hub_brief.md, "What is still open", item 3.