Skip to content

RFC 0006 - Frigate Config Health Report

Status: Ready for Review 2026-11-13, not yet built

Summary

A periodic check of Frigate's config and logs, and of how Camera AutoNotify is set up against them, written up by an AI Task entity the user chooses, and sent by e-mail through Supernotify, like the periodic GenAI report. Off to begin with. The user can choose to hear nothing when there's nothing wrong.

Motivation

Spiking RFC 0004 on the live install turned up config problems that had been there, silently, for days or longer, none of which anything reported:

  • an old model.labelmap override, from before a Frigate+ model, had turned the model's package class into car, so packages were reported as cars
  • cameras' own objects.track replaced the global one, leaving out the logo attributes the model offers, and on one camera person
  • a zone drawn in pixels for one camera's resolution made Frigate's profile manager fail to reset another camera's detect section on every profile change, logged as an error and nothing more
  • the Frigate integration's copy of the config stayed stale until it was reloaded, so anything reading it, Camera AutoNotify's settings included, saw the old config

Each was found by reading the config and the logs together and asking whether they made sense, which is what a model is good at, given the right material.

Design

Settings

A Frigate Config Health Report section in Global Settings, beside the periodic GenAI report, and on the same pattern:

Setting Default What it does
Send a report Off Whether the check runs at all
AI Task None The ai_task entity that writes the report, needed to switch it on
How often Once a week Daily or weekly, from a time of day
Starting at When it runs
Review config On Send Frigate's config and Camera AutoNotify's settings to the model
Review logs 1 day Send a digest of the logs to the model: Off, 1 day, 1 week or All, everything Frigate still holds since it last started
Only when there's an issue On Nothing is sent when the check finds nothing worth saying
Deliveries Supernotify's e-mail As the GenAI report, too long for a phone

It needs Supernotify, as the GenAI report does, and says so where it's not installed.

Review config and Review logs are there to keep the tokens sent, and so the cost, down: config alone is the cheaper review, and finds most of what Motivation lists, while logs find what only shows when Frigate runs, such as the profile reset error. With both off, only the checks in code run, and the AI Task isn't called at all. The checks in code that read the logs, such as the same error logged again and again, look back as far as Review logs says, or a day when it's off.

What's checked

Two layers, so the model's work is the part only a model can do.

Checks in code, each a plain finding with a fix, given to the model as facts it needn't rediscover:

  1. model.labelmap overriding the labels of a Frigate+ model, model.plus.labelMap
  2. a camera's objects.track leaving out attributes the model offers for objects it does track, such as logos for car, or face for person
  3. a label chosen in a camera's Objects setting here that the camera doesn't track, so it can never match
  4. a zone chosen in a camera's Zones setting here that's no longer in Frigate's config
  5. the Frigate integration's copy of the config differing from Frigate's own, so it needs a reload
  6. the notification proxy switched off or expiring, as the settings screens already warn
  7. a camera without video that has no Powered by to say it's off on purpose
  8. GenAI on for a camera, but no review summaries seen from it since the last report
  9. the same error logged again and again, such as a config validation error on each profile change

The model's review, given:

  • with Review config, Frigate's config, from /api/config, which Frigate already redacts: MQTT, ONVIF and go2rtc passwords, GenAI keys, and the user and password in camera paths. environment_vars, which an admin is given, is dropped here too
  • with Review logs, Frigate's warnings and errors from /api/logs/frigate and /api/logs/go2rtc, as a digest: each distinct message once, with how often and when first and last, over the period chosen
  • Frigate's stats, from the integration's coordinator
  • with Review logs, Home Assistant's own warnings and errors from the Frigate, MQTT and Camera AutoNotify integrations, from system_log
  • with Review config, Camera AutoNotify's settings for each camera and zone, with targets left out
  • the findings of the checks in code

It's asked for, by the structure of ai_task.generate_data:

  • a short summary, and whether anything needs attention
  • a list of issues, each with a severity (critical, warning, info), the camera or area, what's wrong, and the fix, as a few lines of config where that helps

Nothing is changed in Frigate or Home Assistant. The report only suggests.

Sending

The issues are formatted as e-mail by Camera AutoNotify, not by the model, so its layout is stable: critical first, each with its camera and fix. With Only when there's an issue on, a report with no issues, or only info, isn't sent. Rules apply to it like any other notification.

Without an AI Task entity

The checks in code need no model. They could run on their own, or raise repair issues as each is found, which would also meet the repair-issues rule of the quality scale. See the open questions.

Feasibility

  • AI Task: ai_task.generate_data takes task_name, instructions, entity_id, structure and attachments, and returns the data in its response. It's in every Home Assistant release this supports
  • Frigate's config: async_get_config() on the integration's client, already used here
  • Frigate's logs: /api/logs/{frigate,nginx,go2rtc}, read through the client's general request method, as async_get_path does. The logs live in memory since Frigate last started, so 1 week may see less than a week, and All is only as far back as that. Each line has its time, so a period is cut from it. They need the admin role, so with a login that isn't admin, the logs are left out and the report says so
  • Size: the config of a few cameras is tens of kilobytes, the log digest smaller once each message is counted rather than repeated, and growing with the period chosen. Within the limits of cloud models, but maybe not of a small local one, which is what Review config and Review logs are for. The config may also need cutting to what's set rather than every default
  • Privacy: the config names cameras, zones and the house's layout, and goes to whichever provider the AI Task entity uses, which may be in the cloud. Frigate's own redaction covers credentials. The settings should say what's sent, and the user chooses the provider
  • Cost: one call per report, so small, but subject to the provider's charges, as the GenAI report is

Open questions

  • Checks in code without AI: run them anyway when the report is on with no AI Task entity, as with both reviews off, and as repair issues whether the report is on or not
  • How often: weekly by default, since config changes rarely, with a report straight after Frigate's config changes as an option
  • Home Assistant's logs: system_log only keeps the most recent entries, so whether that's enough, or the recorder is needed
  • Other servers: one report per Frigate server, or one for all

Spikes, against the live install

  1. The size, in tokens, of the redacted config, and of the log digest for a day, a week and all, to set the defaults
  2. ai_task.generate_data with the install's AI Task entity, on today's config and logs, to see whether it finds the problems above, and what it adds
  3. Whether the Frigate integration's login has the admin role, so the logs can be read