fix(anthropic): coerce explicit additionalProperties to false in output_format schema

[dkindlund]

Relevant issues

Fixes #35808

Pre-Submission checklist

• I have added meaningful tests
• My PR passes all CI/CD checks (lint, format, unit tests) — ruff format clean, target suite green
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have received a Greptile Confidence Score of at least 4/5

Screenshots / Proof of Fix

Root cause

Anthropic's structured outputs reject any additionalProperties value other than false. Their docs list under Supported features: "required and additionalProperties (must be set to false for objects)", and under Not supported: "additionalProperties set to anything other than false"

filter_anthropic_output_schema only added the key when it was absent:

if result.get("type") == "object" and "additionalProperties" not in result:
    result["additionalProperties"] = False

additionalProperties isn't in the unsupported/strip set, so an explicit value is copied verbatim by the catch-all else: result[key] = value, and the not in result guard never fires. The explicit value reaches output_format.schema and 400s

This PR coerces the value for object schemas instead. That is what the official SDKs do — anthropic-sdk-python _parse/_transform.py (pop(...) then strict_schema["additionalProperties"] = False) and anthropic-sdk-typescript transform-json-schema.ts

The permissive tool-use path (map_response_format_to_anthropic_tool, which sets _input_schema["additionalProperties"] = True and is deliberately forced for vertex_ai per #18625 / #19201) is not touched

Before (on litellm_internal_staging, v1.96.0)

from litellm.llms.anthropic.chat.transformation import AnthropicConfig
AnthropicConfig.filter_anthropic_output_schema(
    {"type": "object", "additionalProperties": True, "properties": {"a": {"type": "string"}}}
)
# -> {'type': 'object', 'additionalProperties': True, 'properties': {...}}   <-- 400s at Anthropic

Every explicit form leaked through — top level, nested under properties, inside array items, and the dict sub-schema form:

top-level explicit true    -> leaked non-false additionalProperties: [True]
nested under properties    -> leaked non-false additionalProperties: [True]
inside array items         -> leaked non-false additionalProperties: [True]
dict sub-schema form       -> leaked non-false additionalProperties: [{'type': 'string'}]
control: key absent        -> leaked non-false additionalProperties: NONE (ok)

End-to-end against a real azure_ai/claude-sonnet-4-6 deployment (un-mocked, real API call), sending a schema with "additionalProperties": true:

HTTP 400
litellm.BadRequestError: Azure_aiException - {"type":"error","error":{"type":"invalid_request_error",
"message":"output_format.schema: For 'object' type, 'additionalProperties: true' is not supported.
Please set 'additionalProperties' to false"},"request_id":"req_011CdiGH686teY8qpkxwe2mP"}

After (commit 46751ad)

All five cases clean:

top-level true   -> NONE (ok)
nested           -> NONE (ok)
array items      -> NONE (ok)
sub-schema       -> NONE (ok)
control absent   -> NONE (ok)

$ uv run pytest tests/litellm/llms/anthropic/test_anthropic_schema_filter.py -q
21 passed

Regression check on the surrounding suites (tests/litellm/llms/anthropic/ + tests/test_litellm/llms/anthropic/): 41 failed / 1218 passed with this change vs 41 failed / 1214 passed on the unmodified branch — identical pre-existing failure set, +4 new passing tests. (The pre-existing failures are unrelated experimental_pass_through/context_management tests; a tests/test_litellm/llms/vertex_ai/realtime collection error is a missing local websockets dep and also reproduces unmodified)

Note on the live "after": reproducing the 400 requires addressing the Azure deployment directly — if the model group has a fallback to a vertex_ai deployment, LiteLLM transparently retries there and returns 200, masking the rejection

Type

🐛 Bug Fix

Changes

• litellm/llms/anthropic/chat/transformation.py — in filter_anthropic_output_schema, drop the and "additionalProperties" not in result guard so object schemas always get additionalProperties: False (applies at every recursion depth and to the dict sub-schema form)
• tests/litellm/llms/anthropic/test_anthropic_schema_filter.py — 4 tests: explicit true at top level, nested (under properties and inside array items), the dict sub-schema form, and an explicit false preserved

Semantic caveat (disclosed)

Coercing narrows what the schema permits. For a bare free-form map ({"type": "object", "additionalProperties": true} with no properties) this means "object with no allowed keys", so the model can only emit {} — genuine semantic loss. Anthropic has no native encoding for that shape and the official SDKs degrade it identically; the alternative on this path is a guaranteed 400. A more conservative future option would be detecting that shape and routing it to the permissive tool-use path (the one already forced for vertex_ai), but that is a larger change and out of scope here

Prior art

• fix(anthropic): filter unsupported JSON Schema constraints for structured outputs #22447 and fix(anthropic): strip unsupported schema constraints in tool-based response_format path #22500 (both closed, unmerged) contained this same coercion — fix(anthropic): filter unsupported JSON Schema constraints for structured outputs #22447 even had a test named "additionalProperties: true should be overridden to false for objects" — but were bundled with unrelated changes. This PR is deliberately minimal: one condition + its tests
• fix anthropic structured output 400 error for unsupported schema fields. #21045 (open) introduced the current add-only guard; its test only covers the missing-key path
• fix(anthropic): strip uniqueItems + other unsupported array/object constraints from output_format schema #33981 / fix(anthropic): strip uniqueItems + other unsupported array/object constraints from output_format schema #34313 / fix(anthropic): strip all remaining output_format schema keywords rejected by Anthropic #34319 (merged, shipped v1.95.0) extended the same function for uniqueItems/contains/prefixItems/etc. but contain zero additionalProperties changes

[greptile-apps]

Greptile Summary

The PR updates Anthropic structured-output schema transformation to force additionalProperties to false for every object schema.

• Handles explicit true and dictionary-valued forms at all recursion depths.
• Adds regression coverage for top-level, nested, array-item, sub-schema, and already-false cases.

Confidence Score: 5/5

The PR appears safe to merge.

No blocking failure remains, and the previously reported comment-convention issue is fixed on the current HEAD.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/chat/transformation.py	Always normalizes object schemas to Anthropic's required additionalProperties: false; the previously flagged production comment expansion has been fully reverted.
tests/litellm/llms/anthropic/test_anthropic_schema_filter.py	Adds focused regression tests covering explicit and recursively nested additionalProperties values.

_{Reviews (2): Last reviewed commit: "fix(anthropic): coerce explicit addition..." | Re-trigger Greptile}

[dkindlund]

@greptileai addressed in 46751ad — comment reverted to the original two lines, plus the other CLAUDE.md violations I'd missed (commit trailer, PR-body attribution, trailing periods). Please re-review

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing dkindlund:fix/anthropic-output-format-additional-properties (46751ad) with litellm_internal_staging (355ae99)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (dcb4e50) during the generation of this report, so 355ae99 was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[mateo-berri]

LGTM. Thanks for the contribution!