fix(nim): dynamically clamp max_tokens to resolve context window exhaustion (400 Bad Request)

fix(nim): classify expected context length / max context 400/500 errors as context_window_exceeded for automatic client compaction

fix(provider): remove invalid flush_task_arg_buffers call causing HTTP 500 error on streaming responses

feat: sync upstream commits up to f17c92bc

feat: apply upstream commits f077c358, 8696e132, cb5b4305

Revert "fix(nim): handle progressive payload downgrades on NIM 500 internal server errors"

fix(nim): handle progressive payload downgrades on NIM 500 internal server errors

fix(nim): pass body argument to _normalize_stream in KeyRotator loop to fix 502 error

feat: sync upstream 25th July commits (MiniMax-M3 native tool recovery, HTTP 403 permission denials)

feat: sync proxy with upstream commits up to a4a0fb8 (model catalog refactor, desktop launcher, telemetry feature flags)

fix(streaming): remove Task tool buffering and fix thinking bypass

fix(compatibility): resolve python 3.12 compatibility NameErrors and SyntaxErrors

feat: synchronize proxy with upstream commits up to 5305bd7 and integrate NIM key rotator

fix(nim): limit rate-limit retries to number of keys to prevent hanging

Fix: Bypass think_parser to emit raw thinking tags as plain text when reasoning isn't explicitly requested by client

Fix HF Space timeouts and infinite rotation loops

Fix tight RateLimitError loop causing timeouts by sleeping when keys are cooling down

Deploy latest upstream with key rotation, vision adapter, and HF configs

fix: next_client typo and dataclass super() issue in ExecutionFailure

fix: implement missing _get_client methods for key rotation

fix: change timeout_s to http_read_timeout in Nim Provider initialization

fix: add missing inject_vision_into_request function

fix: update legacy core import in vision adapter

Deploy to Hugging Face Space