Content use

Some vendors say that what you send them, and what their models answer, may be used to train or improve their models. On those routes crucible asks you once before anything is sent, and remembers your answer.

What is asked, and when

The first time a message would go on such a route, a panel stands in place of the prompt box. It names the route, says what the vendor's own terms say in English, and names the page and the day it was read. Nothing has been sent at that point: not the message, not a web search, not a sign-in, not the renewal of a token.

• Use it anyway sends the message and writes the route into your own configuration file, so the question is not asked again, in this run or a later one.
• Go back, or Escape, sends nothing and leaves your message in the prompt box.

The same question is asked when you choose such a route in /login, or a model of one in /model, before the choice is taken. A yes given in /login is written down once the credential is stored; a sign-in that fails or is left writes nothing down.

A window too short for the whole panel says so and sends nothing:

! this route needs an answer first; make the window taller and send again

A run with no terminal to ask on ends before sending, with what the vendor's terms say and Nothing was sent; answer it once in a terminal. on standard error, and exit status 1. A client of the application with no terminal is put the same question as a pending warning, answered with an accepted or a declined decision naming it.

Behind the question, every request crucible sends to a host of such a route is held until the route has its yes, however the route was reached: /login, /model, a key in the environment, the configuration file, --model, or a session picked up with --continue or --resume.

Where the answer is kept

Each yes is a route name in contentUse.accepted, in the configuration file in your home directory:

{ "contentUse": { "accepted": ["key:google"] } }

Only that file is read for it. Either project file that sets contentUse.accepted stops crucible before it draws anything, naming the file and the key, as every key a checkout may not set does.

A yes is taken out of the file before a credential of its route is removed or replaced by one of another row: /logout, a sign-in or key on another row of the same provider, and a second credential taken out at a start. With it go the yes of every model of that provider and the yes of the route its baseUrl answers for. A new account on the route has agreed to nothing, so it is asked again. A key changed in the environment moves nothing. To take a yes back yourself, delete its name from the list.

The routes asked about

A route is a /login row, named by its list and the name its credential is stored under; a model that is itself asked about, named model:, its provider, / and the model; or a Kimi open platform address, named by its host. The sentence is what the panel says; the caution is what the row says in /login. Each quote is the vendor's own, in the language it was written in.

The OpenAI help pages refuse automated readers, so that row rests on the Internet Archive's copies of 28 and 14 September 2026, the date shown being the later.

The two Meta contributor models are asked about when one is chosen in /model, and at the first send on either, whichever way it was reached; the yes is kept for each model on its own. The Meta key row and the standard models, muse-spark-1.3 and muse-spark-1.2, are not asked about: Meta's terms say it does not use what is sent to them to train its models. A web search names the model the session is asking now, so it is held while that model is a contributor model without its yes, and goes out on the standard model once the session has moved to one.

The Qwen rows of aliyun.com rest on the plan pages of that site, read in Chinese; Alibaba Cloud's service agreement says the same of both plans. The Token Plan Team edition is not trained on, but its key cannot be told from a Personal one, so that row states the edition first.

These are not asked about, because their vendors' terms say what is sent is not used for training unless you agree:

• the OpenAI and Anthropic key rows;
• xAI: "SpaceXAI never trains on your API inputs or outputs without your explicit permission";
• MiMo: Xiaomi's privacy policies for the platform say it does not use what you send for model training: the one outside mainland China outright, the mainland one without your prior consent;
• Z.ai · z.ai: "We will not use End User Content to develop or improve Services, unless you explicitly agree to such use";
• Qwen · alibabacloud.com, Qwen · aliyun.com, and the Qwen Coding Plan and Token Plan rows of alibabacloud.com: Alibaba Cloud's terms on both sites say it does not use what is sent to train its models without your consent, and the alibabacloud.com plan pages add nothing to that.

DeepSeek is not asked about either: its open platform terms say nothing of training on what is sent through the API, either way. Its general terms of use say it may use inputs and outputs to improve its services, and do not say whether they reach the API.

A key from a provider's variable belongs to one row of it: for MoonshotAI, the kimi.com key row; for MiniMax, the minimax.io key row, which is asked about; for Qwen, the alibabacloud.com key row; for Z.ai, the z.ai key row.

A baseUrl crucible recognises

A baseUrl is warned only at an address crucible documents. Where two rows whose credential is a key send to one address, as a MiniMax plan and a MiniMax key do, the address answers for the API key row. It is recognised when its scheme is the address's, its host is the address's (case, a trailing dot and the default port written out do not matter), and its path is the address's path or lies under it by whole segments, under any of six readings: with adjacent slashes merged or kept, and for each, . and .. resolved before percent-encoding is decoded, or after it with an encoded / kept inside its segment, or after it with an encoded / taken as a separator. Any one of those readings being documented is enough to be asked. A path written with \ for /, or with ; parameters, is not read as the documented path. Every other baseUrl, including https://chatgpt.com/backend-api/codex/responses, where no key row stands, is not asked about.

Rolling back to 0.43.3

0.43.3 does not know contentUse, and stops before drawing anything on a file that has it, with a line that begins:

crucible: <home>/config.json: contentUse is not a setting crucible has at line <n>, column <m>

Delete the contentUse block from the configuration file in your home directory before running 0.43.3.