fix(bedrock_mantle): stop dropping web search tools on /responses

[devin-ai-integration]

TLDR

Problem this solves:

• Mantle /responses accepts web_search, LiteLLM silently dropped it
• Users got a plain answer with no search, no error
• Cost map claimed no web search on Mantle models

How it solves it:

• Keep every web_search* tool variant in the request
• Still drop what Mantle rejects: file_search, image_generation, code_interpreter
• Mark supports_web_search on the 5 Mantle models that accept it

User Flow

Before: a developer asking a Mantle model to search the web gets an answer from model memory, with no sign the tool was ignored

1. They send POST https://litellm-domain/v1/responses with "model": "bedrock_mantle/openai.gpt-5.6-sol" and "tools": [{"type": "web_search"}]
2. A 200 comes back, but the output has no web_search_call item, so the model answered without searching
3. Nothing in the response says the tool was removed, so stale answers look like normal answers

After: the same request runs a real search server side

1. They send the same POST https://litellm-domain/v1/responses with "tools": [{"type": "web_search"}]
2. The 200 response now carries web_search_call items alongside the message, showing the queries the model ran
3. bedrock_mantle/openai.gpt-5.6-sol reports supports_web_search: true on https://litellm-domain/model/info

Relevant issues

Linear ticket

Pre-Submission checklist

• 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

Live LiteLLM proxy on localhost:4000 against real Bedrock Mantle (https://bedrock-mantle.us-east-1.api.aws, us-east-1), no mocks. Same config and same curl in both runs, only the checked out commit differs

Proxy config used in both runs (mantle_web_search_config.yaml):

model_list:
  - model_name: mantle-sol
    litellm_params:
      model: bedrock_mantle/openai.gpt-5.6-sol
      aws_region_name: us-east-1
      aws_access_key_id: os.environ/AWS_BEDROCK_MANTLE_ACCESS_KEY_ID
      aws_secret_access_key: os.environ/AWS_BEDROCK_MANTLE_SECRET_ACCESS_KEY

general_settings:
  master_key: sk-1234

Note on the search backend: our AWS test account's search egress returns 403, so each web_search_call item comes back with "status": "failed". The point of the before/after is that the tool now reaches Mantle and gets executed server-side at all, instead of being stripped by LiteLLM before the request leaves

Before (merge base 9c38d6d002cf663c41985be62b51bf3ebc5a9279)

1. Start the proxy on the merge base commit

git worktree add /home/ubuntu/mantle_before 9c38d6d002cf663c41985be62b51bf3ebc5a9279
cd /home/ubuntu/mantle_before
uv run --no-sync litellm --config mantle_web_search_config.yaml --detailed_debug --port 4000

2. Send a /v1/responses request with a web_search tool

curl -s -X POST http://localhost:4000/v1/responses \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer sk-1234' \
  -d '{"model":"mantle-sol","input":"Search the web: what is the newest LiteLLM release version?","tools":[{"type":"web_search"}]}'

HTTP 200, and the tool is gone. output has no web_search_call item and the echoed tools is empty:

{
  "id": "resp_sNzjXjEvdtXftPDKLPq8fFyVryJGZADM...",
  "model": "mantle-sol",
  "status": "completed",
  "tools": [],
  "output": [
    { "type": "reasoning", "id": "rs_6781c756d1e55ec69e58cc95b233663b" },
    { "type": "reasoning", "id": "rs_..." },
    { "type": "message", "role": "assistant", "content": [{ "type": "output_text",
      "text": "I can\u2019t access the live web from this environment, so I can\u2019t reliably identify the newest LiteLLM release.\n\nCheck the current version here:\n\n- PyPI: https://pypi.org/project/litellm/\n- GitHub releases: https://github.com/BerriAI/litellm/releases" }] }
  ]
}

output item types: ['reasoning', 'reasoning', 'message']

3. The proxy log shows LiteLLM dropping it before the call

22:39:32 - LiteLLM:WARNING: transformation.py:119 - Bedrock Mantle Responses API: dropping unsupported tool type(s) ['web_search'] (supported: ['custom', 'function', 'mcp', 'namespace', 'tool_search']).

After (branch tip 4859f53bb0be6f30c96b051d1c2031c108320afe)

1. Start the proxy on the PR branch, same config

cd /home/ubuntu/repos/litellm   # litellm_1787092214_mantle_web_search @ 4859f53bb0
LITELLM_LOCAL_MODEL_COST_MAP=True uv run --no-sync litellm --config mantle_web_search_config.yaml --detailed_debug --port 4000

2. Same curl as before

curl -s -X POST http://localhost:4000/v1/responses \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer sk-1234' \
  -d '{"model":"mantle-sol","input":"Search the web: what is the newest LiteLLM release version?","tools":[{"type":"web_search"}]}'

HTTP 200, and Mantle now runs the search server-side: web_search_call items are present, with real action payloads, and the tool is echoed back in tools:

{
  "id": "resp_Gbx1Dc44LL0iHdJ5oLyJzjHc4tc...",
  "model": "mantle-sol",
  "status": "completed",
  "tools": [{ "type": "web_search", "filters": null, "search_context_size": null, "user_location": null }],
  "output": [
    { "type": "web_search_call", "id": "ws_19d7fe0b31be53d8871f475bb627ab8e", "status": "failed",
      "action": { "type": "search",
                  "query": "site:github.com/BerriAI/litellm releases latest LiteLLM",
                  "queries": ["site:github.com/BerriAI/litellm releases latest LiteLLM", "PyPI LiteLLM latest version"] } },
    { "type": "reasoning", "id": "rs_..." },
    { "type": "web_search_call", "id": "ws_6543216809535dc0bc1fae6cfb9b9e74", "status": "failed",
      "action": { "type": "open_page", "url": "https://api.github.com/repos/BerriAI/litellm/releases/latest" } },
    { "type": "reasoning", "id": "rs_..." },
    { "type": "message", "role": "assistant", "content": [{ "type": "output_text",
      "text": "I couldn\u2019t verify the newest LiteLLM release because web access returned a 403 error.\n\nYou can check the current version with:\n\n```bash\npip index versions litellm\n```" }] }
  ]
}

output item types: ['web_search_call', 'reasoning', 'web_search_call', 'reasoning', 'message'], and the proxy log has zero dropping unsupported tool type(s) lines for these requests. The "status": "failed" on each call is our account's search egress returning 403, not LiteLLM: the model itself says "web access returned a 403 error". An earlier identical run on the same tip produced three web_search_call items (search, then two open_page fetches of the GitHub releases URLs), same shape

3. Unsupported tool types are still filtered instead of blowing up. image_generation still gets dropped, request still succeeds

curl -s -X POST http://localhost:4000/v1/responses \
  -H 'Content-Type: application/json' \
  -H 'Authorization: Bearer sk-1234' \
  -d '{"model":"mantle-sol","input":"Say hi in 3 words.","tools":[{"type":"image_generation"}]}'

HTTP 200, tools: [], output item types ['reasoning', 'message'] with text Hi there, friend!, and the log line:

22:56:37 - LiteLLM:WARNING: transformation.py:123 - Bedrock Mantle Responses API: dropping unsupported tool type(s) ['image_generation'] (supported: ['custom', 'function', 'mcp', 'namespace', 'tool_search']).

file_search also still returns HTTP 200 ({"type":"file_search","vector_store_ids":["vs_fake"]} -> status: completed, output item types ['file_search_call', 'message']), served by LiteLLM's emulated file-search handler rather than being sent to Mantle, so no Mantle 400

4. The cost map change is visible on the proxy

curl -s http://localhost:4000/model/info -H 'Authorization: Bearer sk-1234'

model_name: mantle-sol | model: bedrock_mantle/openai.gpt-5.6-sol
supports_web_search: True

This needs LITELLM_LOCAL_MODEL_COST_MAP=True, otherwise the proxy pulls the published cost map from the network and reports supports_web_search: null until this PR's model_prices_and_context_window.json is released

Type

🐛 Bug Fix

Caveats (if any)

• Chat Completions unchanged: Mantle rejects web_search there
• Mantle's own error message omits web search from its supported list
• Capability flags were set from live per-model probes, not AWS docs
• Test account's search egress 403s, so calls run but fail

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

Link to Devin session: https://app.devin.ai/sessions/f751987e5dab49f7b560e50acd9927ee
Requested by: @mateo-berri

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR. Add '(aside)' to your comment to have me ignore it.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[greptile-apps]

Greptile Summary

This PR preserves web-search tools in Bedrock Mantle Responses requests and advertises that capability for five Mantle Responses models

• Allows web_search-prefixed tools through Mantle Responses parameter filtering
• Keeps unsupported file, image-generation, and code-interpreter tools filtered out
• Synchronizes supports_web_search metadata across the primary and backup model maps
• Adds direct parameter-mapping and request-transformation regression tests

Confidence Score: 4/5

The PR appears safe to merge, with one non-blocking regression-test gap for web-search tools supplied through additional_tools

The production change consistently preserves web-search tools across both filtering entry points, while tests cover the main request path but not the separate hoisting path

Files Needing Attention: tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py

Important Files Changed

Filename	Overview
litellm/llms/bedrock_mantle/responses/transformation.py	Preserves web_search-prefixed tools in both ordinary and hoisted Mantle Responses tool filtering
model_prices_and_context_window.json	Advertises web-search support for the five registered Bedrock Mantle Responses models
litellm/model_prices_and_context_window_backup.json	Keeps backup model capability metadata synchronized with the canonical model map
tests/test_litellm/llms/bedrock_mantle/test_bedrock_mantle_responses_transformation.py	Covers top-level web-search preservation and unsupported-tool removal but omits preservation through additional_tools hoisting

_{Reviews (1): Last reviewed commit: "feat(bedrock): support Mantle web search..." | Re-trigger Greptile}

[greptile-apps]

Cover hoisted web-search preservation

The changed tests cover unsupported hoisted tools, but not web-search preservation, leaving this distinct transformation path without regression protection

Context Used: CLAUDE.md (source)

Knowledge Base Used: LLM Provider Adapters

Note: If this suggestion doesn't match your team's coding style, reply to this and let me know. I'll remember it for next time!

[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 litellm_1787092214_mantle_web_search (4859f53) with litellm_internal_staging (852368d)¹}

Footnotes

1. No successful run was found on litellm_internal_staging (e75b4b1) during the generation of this report, so 852368d was used instead as the comparison base. There might be some changes unrelated to this pull request in this report. ↩