fix(bedrock_mantle): hoist Codex additional_tools input items to top-level tools

[lyb0307]

Relevant issues

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)

Delays in PR merge?

If you're seeing a delay in your PR being merged, ping the LiteLLM Team on Slack (#pr-review).

Screenshots / Proof of Fix

Setup: proxy on localhost:4000 hitting the real Bedrock Mantle endpoint in us-east-2 (SigV4 from the ambient AWS credentials, real $). gpt-5.6 is not in the price map yet, so mode and wire path are declared via model_info, the same way a proxy admin sets them in the Admin UI

model_list:
  - model_name: gpt-5.6-sol
    litellm_params:
      model: bedrock_mantle/openai.gpt-5.6-sol
      aws_region_name: us-east-2
    model_info:
      mode: responses
      use_openai_responses_path: true

general_settings:
  master_key: sk-1234

Before, at the PR base 10d5804, a Responses request shaped the way Codex CLI 0.144.x sends it (tool definitions inside input as an additional_tools item) fails:

curl -sS http://localhost:4000/v1/responses \
  -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-5.6-sol",
    "input": [
      {"type": "additional_tools", "role": "developer", "tools": [
        {"type": "function", "name": "wait", "description": "Waits on a yielded exec cell.", "strict": false,
         "parameters": {"type": "object", "properties": {"cell_id": {"type": "string"}}, "required": ["cell_id"], "additionalProperties": false}}
      ]},
      {"type": "message", "role": "user", "content": [{"type": "input_text", "text": "Say hi in one word."}]}
    ]
  }' -w "\nHTTP %{http_code}\n"

{"error":{"message":"litellm.APIConnectionError: Bedrock_mantleException - {\"error\":{\"code\":\"validation_error\",\"message\":\"invalid request body: Invalid 'input': value did not match any expected variant\",\"param\":null,\"type\":\"invalid_request_error\"}}. Received Model Group=gpt-5.6-sol\nAvailable Model Group Fallbacks=None","type":null,"param":null,"code":"500"}}
HTTP 500

After, at 20398d9, the exact same curl succeeds:

HTTP 200
status: completed | error: None
assistant said: ['Hi!']
usage: 56

End to end with the real client at 20398d9: Codex CLI 0.144.4 pointed at the proxy (wire_api = "responses", base_url = "http://127.0.0.1:4000/v1", model gpt-5.6-sol) completes a full agentic task. The run made 5 /v1/responses calls (the first turn plus tool-call turns that replay reasoning, custom_tool_call and function_call_output items) and every one returned 200

$ codex exec --skip-git-repo-check --sandbox workspace-write "Create a file named proof.txt containing exactly 'mantle fix works', then run 'wc -c proof.txt' and report the byte count."
...
Created `proof.txt` exactly as requested. Byte count: **16**.
diff --git a/proof.txt b/proof.txt
new file mode 100644
--- /dev/null
+++ b/proof.txt
@@ -0,0 +1 @@
+mantle fix works

tokens used
10,003

$ cat proof.txt
mantle fix works

$ grep -o 'POST /v1/responses HTTP/1.1" [0-9]*' litellm.log | sort | uniq -c
   5 POST /v1/responses HTTP/1.1" 200

Before the fix, the same Codex CLI setup failed on the first request of every session with the HTTP 500 above and retried forever

Type

🐛 Bug Fix

Changes

Codex CLI 0.144.x uses a "responses lite" wire format for the gpt-5.5 and gpt-5.6 model families: instructions move into developer messages and the tool definitions move out of the top-level tools param into input as a {"type": "additional_tools", "role": "developer", "tools": [...]} item. api.openai.com accepts that item type; Bedrock Mantle's request parser rejects the whole request with 400 invalid request body: Invalid 'input': value did not match any expected variant, which litellm surfaces as a 500 APIConnectionError. The result is that Codex CLI cannot talk to any bedrock_mantle/ Responses model through the proxy

Testing against the live endpoint (bedrock-mantle.us-east-2.api.aws, openai.gpt-5.6-sol) shows Mantle accepts the very same tool definitions when they are sent in the top-level tools param. BedrockMantleResponsesAPIConfig.transform_responses_api_request therefore now strips additional_tools items from input and merges their tools into the top-level tools param, running them through the existing unsupported-tool-type filter. Requests that carry no such item are passed through unchanged, and the item is stripped even when none of its tools survive the filter, since Mantle would otherwise 400 on the item itself

Unit tests cover the hoist (item stripped, tools merged after any existing tools, multiple items merged in order), the filter interaction, the malformed-item and string-input edge cases, and a pass-through case built from the item types Codex replays on later turns

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

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ mateo-berri
❌ BingBingL
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[greptile-apps]

Greptile Summary

This PR fixes a Bedrock Mantle Responses API incompatibility where Codex CLI's "responses lite" wire format places tool definitions inside input as {"type": "additional_tools", ...} items rather than the top-level tools param — a format OpenAI accepts but Mantle rejects with a 400.

• Adds transform_responses_api_request to BedrockMantleResponsesAPIConfig that strips additional_tools items from input, merges their tools (after applying the existing unsupported-type filter) into the top-level tools param, then delegates to the parent implementation.
• Adds _hoist_codex_additional_tools, _is_codex_additional_tools_item, and _tools_of_additional_tools_item helpers; all edge cases (string input, no matching items, all tools filtered, multiple items, malformed item) are handled and covered by unit tests.

Confidence Score: 5/5

Safe to merge — the change is additive, scoped entirely to the Bedrock Mantle responses path, and does not touch any shared or critical infrastructure.

The override intercepts only requests destined for Bedrock Mantle's Responses API and is fully isolated from all other providers. Requests that contain no additional_tools items are passed through unchanged (confirmed by the passthrough test). All edge cases — empty filter results, malformed items, string inputs, multiple items — are explicitly tested with unit tests that make no real network calls. The tool-filter logic reuses the existing _filter_unsupported_tools helper, keeping behavior consistent with the existing map_openai_params path.

No files require special attention.

Important Files Changed

Filename	Overview
litellm/llms/bedrock_mantle/responses/transformation.py	Adds transform_responses_api_request override that hoists additional_tools input items to the top-level tools param before forwarding to the parent implementation; logic is clean and handles all edge cases correctly
tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py	Adds 9 new unit tests covering the hoist, merge ordering, filter interaction, empty-filter strip, malformed item, string input passthrough, agentic replay passthrough, and debug logging; all tests use mocked/local calls with no real network requests

_{Reviews (2): Last reviewed commit: "fix(bedrock_mantle): log additional_tool..." | Re-trigger Greptile}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[veria-ai]

PR overview

All previously flagged issues have been addressed. No open security concerns remain on this pull request.

Security review

No open security issues remain on this pull request.

Fixed/addressed: 1 · PR risk: 0/10

[lzy7071]

Thank you! I am experiencing the same issue as well. Although I am not entirely certain if this is the "best solution" because it seems to be a deliberate design choice codex has made.

For additional evidence that this is not on the bedrock mantle side - using amazon-bed rock directly, I am able to use codex with amazon-bedrock set as provider directly.

[mateo-berri]

Thanks for the fix and the thorough live proof. To get this mergeable I retargeted the base to litellm_internal_staging (the default base for OSS contributions), rebased your branch onto its current tip, and pushed one commit on top that logs the hoist at debug level so the request rewrite is visible when diagnosing proxy traffic. Please also sign the CLA (see the cla-assistant comment above) so we can merge

One heads up: a separate change will gate the unsupported service_tier param that Codex sends when a speed tier is selected in its config, which is the other half of what breaks Codex against Mantle. Your PR stays scoped to the input item hoist

[codspeed]

Merging this PR will improve performance by 31.47%

⚠️ Different runtime environments detected

Some benchmarks with significant performance changes were compared across different runtime environments,
which may affect the accuracy of the results.

Open the report in CodSpeed to investigate

⚡ 4 improved benchmarks
✅ 27 untouched benchmarks

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	test_completion_streaming	58.8 ms	34.6 ms	+70%
⚡	test_response_to_model_response_object	543.5 µs	410.7 µs	+32.34%
⚡	test_cost_per_token_openai	597.7 µs	517.7 µs	+15.46%
⚡	test_cost_per_token_anthropic	602.6 µs	524 µs	+15%

Tip

Curious why this is faster? Comment @codspeedbot explain why this is faster on this PR, or directly use the CodSpeed MCP with your agent.

---

_{Comparing lyb0307:litellm_bedrock_mantle_codex_additional_tools (8745f35) with litellm_internal_staging (214945a)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (9ad8698) during the generation of this report, so 214945a was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩

[mateo-berri]

@greptileai

[mateo-berri]

LGTM; thanks!