Plan a bounded text-inference workflow.

Your decision

Build a safe text-request path without relying on unimplemented features.

• 1. Choose an executable local model from GET /v1/models using a synthetic project key with models:read; public reference IDs are not automatically executable.
• 2. On a trusted backend, submit one bounded text request with inference:write and an explicit output limit. Never put the key in browser code.
• 3. Check HTTP status, consume bounded SSE frames and terminal usage/[DONE], then close the reader. EOF, disconnect and timeout are not success or proof that compute stopped.
• 4. Record only sanitized request IDs for diagnosis. Do not automatically retry an ambiguous POST or log prompts, cookies or credentials.

Continue your decision