Sizes, safety and errors

The rest of the OpenAI-compatible API: how big a request may be, what is checked before the model sees it, the errors, and what isn't offered.

Prompt size and reply size

Nothing is ever trimmed for you. A request whose messages and tool definitions come to more than your plan’s prompt limit is refused, with OpenAI’s wording so tools that shorten their own history when they see it (Cline, Roo Code) do:

HTTP 400
{ "error": {
  "message": "This API's maximum context length for your plan is 12,000 tokens. However, your messages resulted in about 13,750 tokens. Please reduce the length of the messages.",
  "type": "invalid_request_error",
  "code": "context_length_exceeded",
  "param": null
} }

Plan	Largest prompt	Longest reply	Largest request body
Free	12,000 tokens	2,000 tokens	about 200 KB
Pro, Standard	64,000 tokens	8,000 tokens	about 500 KB
Pro, other	Whatever your plan grants: GET /models and the dashboard show it

For this check a token is estimated as 4 characters, so it can be a little off either way. The real counts are in usage. max_tokens can only lower your reply limit. A coding agent’s system prompt alone is often 10,000 tokens or more, so the free plan is for chat tools; agents need a Pro API plan.

Safety

Requests are checked on the server before the model sees them, with the same detectors as the web chat. A refusal is a normal 200 whose only choice has finish_reason: "content_filter", a fixed message, and zero usage; it is not an error.

yahoobot-1: only the most serious class of the child-safety check is applied, the wording that asks for the material and has no innocent reading. Code and shell commands are full of words the broader rules misread, so those are not applied here. A match is refused and the account is suspended for review, as everywhere else on the site.
israelbot-1: the full check, as on the web chat. A broader match gets a polite refusal and a flag for review.
A message that reads as someone in crisis, when it is the newest message and comes from the user, is answered with support resources instead of the model: finish_reason: "stop" and the resources as the content. Tool results aren’t treated as a person speaking.
A very long request is checked newest text first, up to 48,000 characters, with your system message last. Older text was checked when it was new; an agent resends its history every turn.
The model’s replies aren’t checked, only what is sent.

See the native API for the personas’ own endpoint, and Safety & Crisis Resources.

Errors

Errors are OpenAI’s shape plus our code: {"error":{"message","type","code","param":null}}, with retry_after_seconds on rate limits.

Status	code	type	Meaning
400	invalid_request	invalid_request_error	The body or a field is malformed. `message` says which.
400	invalid_json	invalid_request_error	The body is not valid JSON.
400	context_length_exceeded	invalid_request_error	The messages (and tools) are longer than the plan's context limit. Shorten them; nothing is ever trimmed for you.
400	unsupported_parameter	invalid_request_error	A field with a value this API refuses (n above 1, functions, function_call).
400	unsupported_content	invalid_request_error	A message has image or audio content. Text only.
400	tools_not_supported	invalid_request_error	tools, tool_choice or tool messages sent to israelbot-1.
400	invalid_tools	invalid_request_error	A tool definition or tool-call message is malformed or over a limit.
401	missing_api_key	authentication_error	No "Authorization: Bearer igpt_..." header.
401	invalid_api_key	authentication_error	Not a key, revoked, or from before the rebuild.
403	suspended	authentication_error	The account is suspended; the message says how to appeal.
404	model_not_found	invalid_request_error	The model is not "yahoobot-1" or "israelbot-1".
413	payload_too_large	invalid_request_error	The body is over the plan's size limit.
415	unsupported_media_type	invalid_request_error	Send Content-Type: application/json.
429	rate_limited	rate_limit_error	Over the requests-per-minute or per-day allowance. Retry-After says when.
429	token_limit	rate_limit_error	The account used its tokens for today (UTC). They return at 00:00 UTC.
502	upstream_unavailable	server_error	The model did not answer. Try again.
503	capacity_exceeded	rate_limit_error	Everyone's shared capacity is used up right now.
503	maintenance	server_error	The API is switched off.

See Errors for the rest, Rate limits for the headers, and honour Retry-After.

What isn't here, and what is kept

Only /chat/completions and /models: no /completions, /embeddings, /images, /audio, /responses, assistants, files, fine-tuning or /models/{id}.
No image or audio input, no n above 1, no log probabilities, no structured-output mode.
Nothing is stored for you to fetch later, and there are no memories or About Me.
Model names are ours: the OpenAI names (gpt-4o...) get 404 model_not_found.

Each request leaves a row in the key’s log on the dashboard: when, the result, the model, token counts, how long it took, and what the caller said it is (never the text, never the address). The newest user message and the reply, clipped to 2,000 and 4,000 characters, are kept in the chat log, encrypted, and read only if something needs looking into. See the Privacy Policy.