fix(anthropic): stop tool translation from mutating caller input_schema

[cognis-digital]

TLDR

Problem this solves:

• translate_anthropic_tools_to_openai assigned the caller's input_schema dict to the translated parameters by reference, then the "pass additional computer kwargs" loop merged extra top-level tool keys (e.g. a computer tool's display_width_px) into that shared dict in place
• Callers reuse the same in-memory tool list (a pre-call guardrail translates the tools, then the real call translates them again), so the second pass forwarded a schema polluted with non-schema keys, and providers that validate the tool parameters JSON schema can reject the request. The corruption is silent

How it solves it:

• Deep-copy input_schema before assigning it to function_chunk["parameters"], so translation is a pure read and later kwarg merges land only on the copy
• This mirrors the sibling translate_anthropic_output_format_to_openai in the same file, which already copy.deepcopys its schema for the same reason; copy is already imported

Relevant issues

Fixes #34510

Linear ticket

Type

🐛 Bug Fix

Changes

litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py: deep-copy input_schema before it becomes the translated OpenAI parameters

Added a regression test test_translate_anthropic_tools_to_openai_does_not_mutate_caller_input_schema that asserts the caller's input_schema is untouched after translation and that the vendor kwargs still reach the translated parameters. It fails on the pre-fix code and passes after

QA runbook

Pure translation logic, so it reproduces with no provider credentials

from litellm.llms.anthropic.experimental_pass_through.adapters.transformation import LiteLLMAnthropicMessagesAdapter
import copy
adapter = LiteLLMAnthropicMessagesAdapter()
tool = {
    "name": "computer",
    "type": "computer_20241022",
    "input_schema": {"type": "object", "properties": {"action": {"type": "string"}}},
    "display_width_px": 1024,
    "display_height_px": 768,
}
before = copy.deepcopy(tool["input_schema"])
adapter.translate_anthropic_tools_to_openai(tools=[tool], model="claude-3-5-sonnet")
print("mutated:", tool["input_schema"] != before)

Before this PR prints mutated: True (the caller's schema now contains display_width_px / display_height_px); after, it prints mutated: False

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Greptile Summary

Fixes Anthropic tool translation so caller-owned input schemas remain unchanged.

• Deep-copies input_schema before merging provider-specific tool fields into translated OpenAI parameters.
• Adds a regression test verifying input immutability, vendor-field propagation, nested schema preservation, and distinct object identity.

Confidence Score: 5/5

The PR appears safe to merge with no actionable defects identified.

The copied value is a request JSON-schema dictionary, downstream mutations now apply only to the translated object, and the regression test directly covers both immutability and expected translated output.

Important Files Changed

Filename	Overview
litellm/llms/anthropic/experimental_pass_through/adapters/transformation.py	Deep-copies the caller-provided schema before downstream parameter mutation, preventing translation from corrupting reused request data.
tests/test_litellm/llms/anthropic/experimental_pass_through/adapters/test_anthropic_experimental_pass_through_adapters_transformation.py	Adds focused regression coverage demonstrating that translation preserves the source schema while retaining provider-specific fields in the translated result.

_{Reviews (1): Last reviewed commit: "fix(anthropic): stop tool translation fr..." | Re-trigger Greptile}

[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 cognis-digital:litellm_fix_tool_translation_schema_mutation (7fa14bd) with litellm_internal_staging (b9b27c2)}