[wozhenbang2004]

Tools returning mutually recursive Person.pet / Pet.owner models can send correct structured_content while the client's .data silently uses the wrong class several levels down. Without schema titles, distinct generated classes share the display name Root, and recursive references bind to that shared name.

Bind recursive references to each generated class's identity, resolve the complete graph, and only then share it through the class cache. Public class names remain unchanged. For the four-level result in #6135, the final pet retains its owner field:

result = await client.call_tool("get_person", {})
assert result.data.pet.owner.pet.name == "Max"
assert result.data.pet.owner.pet.owner is None

Fixes #6135.

🤖 Generated with OpenAI Codex.

Co-authored-by: Alperen Keleş 19610315+alpaylan@users.noreply.github.com

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	⚠️ Failed 2026-10-28T14:27:01.823207Z We couldn’t complete this review. Comment @codex review to try again.	93961c2	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[github-actions]

This pull request is awaiting maintainer assignment. Linked PRs stay open while maintainers triage the issue; this check remains failing until assignment.

Per CONTRIBUTING.md, an external PR must reference an issue that's assigned to its author. To get there:

1. A maintainer will review the issue and decide whether to assign you. Please don't comment just to request assignment.

Once you're assigned and the link is present, this check is re-run automatically. Previously gate-closed PRs also reopen — no further action needed.

Maintainers: reopen this PR or remove the missing-issue-link label to bypass this check.

[himanshuchawla0807-gif]

Additional regression coverage for your fix at 93961c29742218b2e929ec145f49bf4b0f7398c8: a three-type cycle through an array (A → list[B] → C → A), covering dataclass/Pydantic generation and absent/shared titles. All four cases fail on current main d07877e062337842cfd46afef4efe85fc87bab28 because the returning A is validated as C; all four pass on your head. They also check distinct class identities, serialization round-trip, repeated cached conversion and input immutability.

With these cases added temporarily, uv run pytest tests/utilities/json_schema_type -n 0 -q gives 427 passed, 1 skipped on your head (Python 3.12.14, Pydantic 2.14.1). The test passes Ruff check/format. A separate Codex agent independently repeated the four base/head cases. This is scoped converter validation, not current-main composition, a full-suite/platform result or merge readiness; the assignment and fork-workflow approval gates remain separate.

Runnable coverage if useful alongside your two-type/client cases:

from copy import deepcopy
from typing import Any

import pytest
from pydantic import TypeAdapter

from fastmcp.utilities.json_schema_type import json_schema_to_type


@pytest.mark.parametrize("extra", [False, True], ids=["dataclass", "pydantic"])
@pytest.mark.parametrize("title", [None, "Record"], ids=["untitled", "same-title"])
def test_three_type_cycle_through_array(extra: bool, title: str | None) -> None:
    definitions: dict[str, Any] = {
        "A": {
            "type": "object",
            "properties": {
                "a": {"type": "string"},
                "children": {"type": "array", "items": {"$ref": "#/$defs/B"}},
            },
            "required": ["a", "children"],
        },
        "B": {
            "type": "object",
            "properties": {
                "b": {"type": "integer"},
                "next": {"$ref": "#/$defs/C"},
            },
            "required": ["b", "next"],
        },
        "C": {
            "type": "object",
            "properties": {
                "c": {"type": "boolean"},
                "next": {"$ref": "#/$defs/A"},
            },
            "required": ["c", "next"],
        },
    }
    for definition in definitions.values():
        if extra:
            definition["additionalProperties"] = True
        if title:
            definition["title"] = title
    schema = {
        "type": "object",
        "properties": {"node": {"$ref": "#/$defs/A"}},
        "required": ["node"],
        "$defs": definitions,
    }
    original = deepcopy(schema)
    value = {
        "node": {
            "a": "first",
            "children": [
                {"b": 7, "next": {"c": True, "next": {"a": "last", "children": []}}}
            ],
        }
    }
    generated = json_schema_to_type(schema)
    for _ in range(2):
        result = TypeAdapter(generated).validate_python(value)
        first = result.node
        middle = first.children[0]
        last = middle.next.next
        assert last.a == "last"
        assert last.children == []
        assert type(last) is type(first)
        assert len({type(first), type(middle), type(middle.next)}) == 3
        assert TypeAdapter(generated).dump_python(result) == value
        assert json_schema_to_type(schema) is generated
    assert schema == original

Generated and executed with OpenAI Codex. No human review before posting.

[wozhenbang2004]

Thanks for checking this. The three-type cycle through an array is useful additional regression coverage.

[fastmcp-triage-reviewer]

Automated review by the FastMCP triage reviewer bot.

The identity-bound recursive references fix the reported silent hydration error at the schema-conversion layer. Nested Person/Pet values retain their correct fields and types, and the additional three-type array cycle also preserves distinct classes and serialization. No consequential defect was found in this pass.

Checks

All seven regression cases fail on the base and pass with this revision. Four array-cycle probes and 93 neighboring recursion/default/cache tests also pass.

[zzstoatzz]

Independent adversarial review passed for this exact revision. Mutually recursive schema identity and complete-graph caching passed independent mixed dataclass/Pydantic probes. Applicable local static checks and the combined recovery tests passed. This is code approval; current CI, contribution gate and repository merge requirements remain separate prerequisites.

[fastmcp-triage-reviewer]

Automated review by the FastMCP triage reviewer bot.

The converter fix preserves distinct recursive model shapes without changing public class names. Independent three-type array cycles also retain class identity and serialize correctly across mixed dataclass/Pydantic graphs. No actionable defect found; this is not a merge-readiness determination.

Checks

At this revision, seven hydration regressions and 15 cache tests pass. The seven regressions fail on the pinned baseline. All 16 independent mixed-model cycle probes pass; the workflow change was inspected but its corpus run was not executed.

[zzstoatzz]

thanks — the fix and tests landed as #6394 with you as co-author, without the schema-crash workflow commit so that change can be reviewed on its own. closing this in favor of that one.