feat(naming): an admin section that shows what Claude costs

- Every question asked of Claude is recorded: when, the kind of list, the model, the tokens read and written (also for an answer that was no name but was billed) and how it ended: a name made, a name taken from the kept ones, or a failure and why. Nothing of the facts that were sent is recorded. Records older than a year are removed (migration 0022).
- GET /providers/anthropic/usage (admin) gives the totals of the last days (1 to 365, 30 by default), a row for each day with calls and an estimate of the cost in dollars, worked out from the tokens and the price of the model. It says it is an estimate and flags calls to a model whose price is not known.
- A Costs section in the admin area shows the last 7, 30 or 90 days.
This commit is contained in:
tgiachi 2026-11-13 11:14:45 +01:00
commit 1b79a7370d

View file

@ -258,7 +258,7 @@ The address can have a path if Lidarr runs under a URL base, and it cannot carry
The key is stored in the database, in clear text, and is never returned by the API or written to the log: the list only says `has_api_key`. Because the database holds it, treat database backups like secrets. Removing the key turns the provider off, and each provider describes in the list exactly what it sends.
**Claude (Anthropic)** is the one provider that names things: with it on, a playlist, a mix or a smart playlist can get a name and a line of description that fit its songs. An admin turns it on with an Anthropic API key (`PUT /providers/anthropic`, the key is write-only, and `POST /providers/anthropic/test` reads one model, which costs nothing and sends no text). `POST /playlists/{id}/suggest-name`, `POST /mixes/{id}/suggest-name` and `POST /smart-playlists/{id}/suggest-name` answer with a name, a description and a `source`. They are a pipeline: the facts about the list (up to 5 genres, up to 8 artist names, how many songs, for a smart playlist a few words about its rules, and the language) go to a namer. With the provider off, or when Claude fails or declines, the local namer builds a name from the genres and the artists on this server, and nothing leaves it. When Claude answers, the name is kept (under a hash of the facts, which holds no artist name), so the same facts are not asked, or paid, twice. A suggestion changes nothing: the client decides what to do with it, and the web app offers it where a playlist is renamed. With the provider on, names are also made by themselves: a mix, a smart playlist or a suggested smart playlist that is kept as a playlist with `auto_name` (the web app always sends it) gets Claude's name and line of description, with `name` as the fallback when Claude is off, fails or declines; a smart playlist can be made with no name at all and Claude names it from the songs its rules give (without Claude that is `invalid_input`); and `GET /naming` says whether names are automatic, so that the editor does not ask for one. The model is `claude-opus-5-5`, at low effort; the key is never taken from the environment of the server, and a redirect is never followed.
**Claude (Anthropic)** is the one provider that names things: with it on, a playlist, a mix or a smart playlist can get a name and a line of description that fit its songs. An admin turns it on with an Anthropic API key (`PUT /providers/anthropic`, the key is write-only, and `POST /providers/anthropic/test` reads one model, which costs nothing and sends no text). `POST /playlists/{id}/suggest-name`, `POST /mixes/{id}/suggest-name` and `POST /smart-playlists/{id}/suggest-name` answer with a name, a description and a `source`. They are a pipeline: the facts about the list (up to 5 genres, up to 8 artist names, how many songs, for a smart playlist a few words about its rules, and the language) go to a namer. With the provider off, or when Claude fails or declines, the local namer builds a name from the genres and the artists on this server, and nothing leaves it. When Claude answers, the name is kept (under a hash of the facts, which holds no artist name), so the same facts are not asked, or paid, twice. A suggestion changes nothing: the client decides what to do with it, and the web app offers it where a playlist is renamed. With the provider on, names are also made by themselves: a mix, a smart playlist or a suggested smart playlist that is kept as a playlist with `auto_name` (the web app always sends it) gets Claude's name and line of description, with `name` as the fallback when Claude is off, fails or declines; a smart playlist can be made with no name at all and Claude names it from the songs its rules give (without Claude that is `invalid_input`); and `GET /naming` says whether names are automatic, so that the editor does not ask for one. Every question to Claude is counted (a made name, a name taken from the kept ones, a failure with its reason, and the tokens that were billed, also for a failure that was billed) and nothing of what was sent is kept; `GET /providers/anthropic/usage?days=30` (admin) gives the totals, a row for each day and an **estimate** of the cost in dollars, worked out from the tokens and the price of the model (the invoice of Anthropic is the truth), and the Costs section of the admin area shows it for 7, 30 or 90 days; records older than a year are removed. The model is `claude-opus-5-5`, at low effort; the key is never taken from the environment of the server, and a redirect is never followed.
**What leaves your server, when a provider is on:** to Anthropic, only when someone asks for a name, up to 5 genres and 8 artist names of that list, how many songs it has, a few words about the rules of a smart playlist, the language, your API key, your IP address and a User-Agent (never your users, your listens, track titles or file paths); to Last.fm, the artist name (or MusicBrainz id) being looked up, your API key, your IP address and the `User-Agent`; to MusicBrainz, the MusicBrainz id of the artist (and, only for the covers of albums that have no id, the artist's name and the album's title: see **Covers**), your IP address and a `User-Agent` that carries your contact. Never your users, your listens or file paths. Requests only happen when someone asks for them, never as a bulk pass over your library (the exceptions are the three jobs that an admin starts on purpose, each of them one request that goes through the whole library, one item after the other, spaced out, and each stops when its providers keep failing: `POST /library/covers`, with the Cover Art Archive on, sends the release group id of every album that has no picture in its files (and, for the albums whose files have no id, with MusicBrainz on, the artist and the title, one request a second); `POST /library/artist-images`, with Commons and MusicBrainz on, sends the MusicBrainz id of every artist that has one, to MusicBrainz, Wikidata and Commons; `POST /library/descriptions`, with Wikipedia and MusicBrainz on, sends it to MusicBrainz, Wikidata and Wikipedia), and they are spaced out, retried with backoff, and stopped by a circuit breaker if a provider is down. When MusicBrainz answers that an id was merged into another, Sonda follows that one redirect, and only to an artist or a release group of musicbrainz.org over https; the new id is used for that request and is not stored, and any other redirect is refused. **Descriptions** (the Wikipedia provider, which also needs MusicBrainz on) send the MusicBrainz id to MusicBrainz, the Wikidata item to Wikidata and the title of the page to Wikipedia, with your IP address and a `User-Agent` that carries a contact (the one set for each provider; Wikimedia refuses requests without one), and never a name. **Photos** (the Commons provider, which also needs MusicBrainz on) send the MusicBrainz id to MusicBrainz, the Wikidata item to Wikidata, the name of the picture to Commons and the address of a copy to Wikimedia, with your IP address and a `User-Agent` with a contact, and never a name. **Covers** (the Cover Art Archive provider) send the MusicBrainz release group id of an album, read from the tags of its own files, and only when the files have no picture. When the files have no id either and the MusicBrainz provider is on too, the name of the album's artist and its title go to MusicBrainz, which is asked for the release group: a result is taken only when MusicBrainz scores it 95 or more and its title and artist are the ones asked for (compared without case, accents or punctuation), and anything doubtful is no cover. A title that ends with a note about the disc or the edition (`(disc 2)`, `(bonus disc)`, `(Deluxe Edition)`) is asked as it is first and, if MusicBrainz is not sure, once more without the note, because a release group is named after the album and not after a disc. The id that comes back is used for that request and is not stored: it is not what your files say, and it is never used to read an artist or to send anything to Lidarr; the redirects of the archive are followed only to archive.org, over https. Its covers have no licence of their own: Sonda keeps a copy for your library and never treats it as free. **Discover** asks Last.fm for the similar artists of each seed (the seed's name or MusicBrainz id) and, for the best suggestions that have no tags yet, for the tags of that suggestion (its name); never MusicBrainz. Seeds you name are sent when you ask. Seeds taken from your listening, which are the names of the artists you play most, are sent only if you turned on `discover_from_listening` yourself; otherwise Discover refuses to run without a `seed`. **Lidarr** (your own server, at the address an admin set) gets the MusicBrainz id of an artist an admin sends, or of an artist someone opens `GET /artists/{id}/lidarr` for, and the API key, nothing else; asking it what it offers and testing the key send only the key. Never a name, your users, your listens or your files. What Lidarr then does is up to its own indexers and download clients. Besides the three jobs, the other exception to "only when someone asks" is `POST /providers/{name}/test`, which an admin sends to check a key (or, for Lidarr, the address and the key); it sends nothing but the key.
@ -316,6 +316,7 @@ Everything lives under `/api/v1`. Bodies are JSON. Errors always look like `{"er
| admin | `POST /library/scan`, `POST /libraries/{id}/scan` | Scan everything, or one library. `?full=true` reads every file again. |
| admin | `GET /library/duplicates`; `POST`, `GET`, `DELETE /library/duplicates/check`; `POST /library/duplicates/clean` | Find the songs that are in more than one file, read their files to tell the identical ones, and delete the extra copies. See **Duplicates**. |
| admin | `GET /providers`, `PUT /providers/{name}`, `POST /providers/{name}/test`, `DELETE /providers/{name}/data` | See, set up and check the providers (Last.fm, MusicBrainz, Wikipedia, Commons, the Cover Art Archive, Lidarr), and remove what they gave. The API key is write-only. |
| admin | `GET /providers/anthropic/usage` | What the Claude provider costs over the last days (query `days`, 1 to 365): totals, a row for each day with calls and an estimate in dollars. |
| admin | `GET /providers/lidarr/options` | The profiles and root folders that Lidarr offers, to choose the ones artists are sent with. |
| admin | `POST /artists/{id}/lidarr` | Send an artist to your own Lidarr, by its MusicBrainz id, which then looks for the music. Never by name, never by itself. |
| signed in | `GET /artists/{id}/lidarr` | Whether Lidarr has the artist and monitors it. |