fix(anthropic): strip uniqueItems + other unsupported array/object constraints from output_format schema

[dkindlund]

Relevant issues

No existing issue tracks the uniqueItems gap specifically. Related prior work on this same filter: #19444 (the filter's own tracking issue), #21045, #22447, #22500, #29203 and #29624 (stripping numeric/count/length constraints for Anthropic and Bedrock structured outputs), and #18625 / #19201 (why vertex_ai Claude avoids this path; it is forced onto tool-use, which ignores these keywords)

Linear ticket

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have added meaningful tests
• My PR passes all CI/CD checks (e.g., lint, format, unit tests)
• 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 before requesting a maintainer review (Greptile reviews automatically once the PR is opened; only comment @greptileai to re-request a review after pushing changes)

Screenshots / Proof of Fix

Anthropic's structured outputs compile the output_format JSON schema into a constrained-decoding grammar and reject keywords the grammar cannot enforce with a 400 invalid_request_error, e.g. output_format.schema: For 'array' type, property 'uniqueItems' is not supported. filter_anthropic_output_schema already strips minimum/maximum/exclusiveMinimum/exclusiveMaximum/minLength/maxLength/minItems/maxItems into description notes, but not uniqueItems, contains/minContains/maxContains, or minProperties/maxProperties, so those reached the API and 400'd. vertex_ai Claude is unaffected because it is routed through tool-use emulation (input_schema), which silently ignores these keywords; that asymmetry is what made this easy to miss

All runs below are fully e2e with zero mocks: a live LiteLLM proxy in front of a real Azure AI Foundry Claude deployment (azure_ai/claude-sonnet-4-6), costing real $. The before run is at fcd236097e (current litellm_internal_staging tip, pre-fix); the after run is at this PR's head dba83b71ba. Same five requests both times, each a response_format: json_schema call whose schema carries one group of unsupported keywords:

Schema keyword(s)	Before fcd236097e	After dba83b71ba
none (control)	200	200
uniqueItems	400	200
contains+minContains+maxContains	400	200
minProperties+maxProperties	400	200
multipleOf (not in this PR)	400	400

Example request (the uniqueItems case; the other payloads differ only in which keyword they set):

curl -sS "http://localhost:$PORT/v1/chat/completions" \
  -H "Authorization: Bearer sk-1234" -H 'Content-Type: application/json' -d '{
    "model": "claude-sonnet-4-6",
    "messages": [{"role": "user", "content": "List three distinct colors."}],
    "response_format": {"type": "json_schema", "json_schema": {"name": "colors", "schema": {
      "type": "object",
      "properties": {"colors": {"type": "array", "items": {"type": "string"}, "uniqueItems": true}},
      "required": ["colors"], "additionalProperties": false}}}
  }'

Raw results, before (fcd236097e):

clean            | HTTP 200 | {"colors":["red","blue","green"]}
uniqueItems      | HTTP 400 | output_format.schema: For 'array' type, property 'uniqueItems' is not supported
contains         | HTTP 400 | output_format.schema: For 'array' type, property 'contains' is not supported
minMaxProperties | HTTP 400 | output_format.schema: For 'object' type, property 'minProperties' is not supported
multipleOf       | HTTP 400 | output_format.schema: For 'integer' type, property 'multipleOf' is not supported

Raw results, after (dba83b71ba):

clean            | HTTP 200 | {"colors":["red","blue","green"]}
uniqueItems      | HTTP 200 | {"colors":["red","green","blue"]}
contains         | HTTP 200 | {"nums":[1,2,3]}
minMaxProperties | HTTP 200 | {"city":"Paris"}
multipleOf       | HTTP 400 | output_format.schema: For 'integer' type, property 'multipleOf' is not supported

The multipleOf row shows the strip list is still not exhaustive after this PR; it fails identically before and after and is left for a follow-up, since this PR only targets the array/object keywords that were reported failing. These keywords are also confirmed rejected by the Anthropic API directly (not an Azure quirk): the same schemas 400 against api.anthropic.com with the structured-outputs-2025-11-13 beta header

Type

🐛 Bug Fix

Changes

litellm/llms/anthropic/chat/transformation.py: add uniqueItems, contains, minContains, maxContains, minProperties, maxProperties to filter_anthropic_output_schema's unsupported_fields set and constraint_labels map, so they are stripped from the output_format schema and their intent is moved into the field description, exactly as already done for the numeric/length constraints and as the Anthropic SDK does. A disabled boolean constraint (uniqueItems: false) is stripped without adding a misleading advisory note, and sub-schema constraints (contains) are JSON-serialized into the note so the requirement is preserved

Tests: tests/litellm/llms/anthropic/test_anthropic_schema_filter.py covers uniqueItems, contains/minContains/maxContains, minProperties/maxProperties, and the uniqueItems: false edge case; tests/test_litellm/llms/anthropic/test_anthropic_output_format_filter.py mirrors them so the coverage-uploading CI job exercises the new branches

Final Attestation

• The tests check the right things, including the edge cases, and regressions in the respective real-world customer use-cases are not possible after this PR

[greptile-apps]

Greptile Summary

This PR extends filter_anthropic_output_schema to strip the remaining JSON Schema keywords that Anthropic's output_format rejects with a 400 — uniqueItems, contains/minContains/maxContains, and minProperties/maxProperties — and moves their intent into the field description, matching the existing behaviour for numeric/string/array count constraints.

• transformation.py: Adds six new keywords to unsupported_fields and constraint_labels; adds logic to skip advisory notes for falsy boolean values (uniqueItems: false) and to JSON-serialize sub-schema values (contains) so the description carries the full constraint detail.
• Test files: Four new unit tests covering each new keyword group and the uniqueItems: false edge case; the new tests/test_litellm/ mirror file ensures the CI coverage job also runs these cases. Existing test changes are purely cosmetic reformatting.

Confidence Score: 5/5

Safe to merge — the change is narrowly scoped to a schema-filtering utility that removes unsupported keywords before they reach Anthropic's API, with no effect on other providers or code paths.

The fix is additive (expanding an existing set/map), handles edge cases correctly (falsy booleans, sub-schema serialization), and is covered by both existing and new unit tests. No production logic outside the filter function is touched.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/chat/transformation.py	Adds uniqueItems, contains/minContains/maxContains, and minProperties/maxProperties to filter_anthropic_output_schema's unsupported-field set; correctly skips advisory notes for falsy boolean constraints and JSON-serializes sub-schema values.
tests/litellm/llms/anthropic/test_anthropic_schema_filter.py	Adds four new tests covering uniqueItems, contains/minContains/maxContains, minProperties/maxProperties, and the uniqueItems: false edge case; existing test reformatting is cosmetic-only with no assertion changes.
tests/test_litellm/llms/anthropic/test_anthropic_output_format_filter.py	New file mirroring the schema-filter tests under tests/test_litellm/ (mock-only, no network calls) to ensure the CI coverage job exercises the new keyword handling.

_{Reviews (3): Last reviewed commit: "test(anthropic): cover output_format arr..." | Re-trigger Greptile}

[dkindlund]

@greptileai I've addressed the contains advisory-note feedback in ce142f8 (sub-schema now serialized as JSON). 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-array-object-constraints (1ea0ef1) with litellm_oss_daily_2026_07_17 (1ebf2a7)}

[mateo-berri]

@greptileai

[mateo-berri]

LGTM; thanks!