fix(proxy): restore spend logging for /cursor/chat/completions (#30126)

[discobot]

Relevant issues

Fixes #30126

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 unit tests on make test-unit
• My PR's scope is as isolated as possible; it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

Screenshots / Proof of Fix

Re-ran the A/B against a live proxy with a real provider and a Postgres-backed SpendLogs table, so the row actually lands instead of hitting a mock. The OpenAI key I had was rate-limited, so the model here is claude-haiku-4-5; the fix is proxy-side metadata routing so it is provider-agnostic, and the cursor response-format conversion is OpenAI-specific and separate from spend tracking, so the check is the SpendLogs write itself (which is what #30126 reports missing)

Config is one model claude-haiku-4-5 -> anthropic/claude-haiku-4-5-20251001, master_key: sk-1234, DATABASE_URL set. Generate a key, hit the cursor route, and read SpendLogs (truncating it before each run so the diff is clean):

KEY=$(curl -s -X POST localhost:4000/key/generate -H "Authorization: Bearer sk-1234" \
  -H "Content-Type: application/json" -d '{"models":["claude-haiku-4-5"],"key_alias":"cursor-demo"}' | jq -r .key)

curl -s -X POST localhost:4000/cursor/chat/completions -H "Authorization: Bearer $KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"claude-haiku-4-5","input":[{"role":"user","content":"ping"}]}' > /dev/null

# after the spend-log batch flush:
psql "$DATABASE_URL" -c 'select call_type, model, spend from "LiteLLM_SpendLogs" order by "startTime" desc'

Without the fix (current litellm_internal_staging) the cursor call's cost is never recorded; only the proxy-level row lands, at spend 0:

 call_type | model            | spend
-----------+------------------+-------
           | claude-haiku-4-5 |     0

With the fix the same call writes the LLM spend row with the real cost:

 call_type  | model                               | spend
------------+-------------------------------------+---------
            | claude-haiku-4-5                    |       0
 aresponses | anthropic/claude-haiku-4-5-20251001 | 3.3e-05

So proxy auth metadata now reaches the spend callback and the SpendLogs write for /cursor/chat/completions fires the same way it does for /v1/chat/completions

Type

🐛 Bug Fix

Changes

/cursor/chat/completions is served through the Responses API pipeline, which keeps proxy-internal metadata under litellm_metadata because metadata is the user-facing Responses API param. The route was missing from LITELLM_METADATA_ROUTES, so add_litellm_data_to_request wrote user_api_key and friends into the user-facing metadata param, the spend tracking callback saw user_api_key=None, and the SpendLogs write was silently skipped; hence 200 + provider billed + nothing in the Logs UI

The fix adds the route to LITELLM_METADATA_ROUTES, matching /v1/responses which this endpoint bridges to. Two regression tests cover it (the _get_metadata_variable_name mapping and an endpoint test asserting the kwargs reaching llm_router.aresponses carry litellm_metadata.user_api_key); both fail without the one line change

[CLAassistant]

All committers have signed the CLA.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[greptile-apps]

Greptile Summary

Restores spend logging for /cursor/chat/completions by adding the route to LITELLM_METADATA_ROUTES. Without this, add_litellm_data_to_request wrote auth metadata into the user-facing metadata param instead of litellm_metadata, causing _should_track_cost_callback to see user_api_key=None and silently skip the SpendLogs write.

• litellm/proxy/litellm_pre_call_utils.py: One-line addition of /cursor/chat/completions to LITELLM_METADATA_ROUTES, matching the existing pattern used by /v1/responses and related routes.
• Tests: A unit test for _get_metadata_variable_name and an integration test that mocks llm_router.aresponses and asserts litellm_metadata.user_api_key is populated correctly; both tests fail without the fix.

Confidence Score: 4/5

The production change is a single-entry addition to a routing tuple with no behavioral impact outside the cursor endpoint; safe to merge.

The core fix is minimal and correct. The integration test covers the regression path end-to-end but accesses call_args.kwargs without first asserting the mock was actually invoked, which would produce a confusing AttributeError rather than a clear failure if the endpoint ever returned 200 without reaching aresponses.

tests/test_litellm/proxy/response_api_endpoints/test_endpoints.py — the new integration test should assert mock_router.aresponses.assert_called_once() before accessing call_args.

Important Files Changed

Filename	Overview
litellm/proxy/litellm_pre_call_utils.py	Adds /cursor/chat/completions to LITELLM_METADATA_ROUTES so proxy auth fields are written to litellm_metadata instead of the user-facing metadata param; minimal and correct one-line change.
tests/test_litellm/proxy/response_api_endpoints/test_endpoints.py	Adds TestCursorChatCompletionsSpendTracking integration test; correctly mocks llm_router.aresponses and verifies metadata routing, but accesses call_args.kwargs without first asserting the mock was called.
tests/test_litellm/proxy/test_litellm_pre_call_utils.py	Adds a clean unit test for _get_metadata_variable_name with the new /cursor/chat/completions route; no issues.

_{Reviews (1): Last reviewed commit: "fix(proxy): restore spend logging for /c..." | Re-trigger Greptile}

[greptile-apps]

Missing guard before call_args access

call_args is None when the mock was never invoked, so call_args.kwargs would raise AttributeError rather than producing a useful failure message. If the handler returns 200 through an alternative code path without ever reaching llm_router.aresponses (e.g., a middleware short-circuit or an exception swallowed upstream), the test will blow up with an opaque AttributeError instead of a clear assertion failure. Adding mock_router.aresponses.assert_called_once() immediately after the status check pins the failure message to the actual absence of the call.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: efcb0f8b02

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Preserve proxy metadata for callbacks with client metadata

When a /cursor/chat/completions request includes the Responses API metadata field, this route now puts proxy auth fields only in litellm_metadata while leaving the client metadata in metadata. litellm.utils.function_setup only copies litellm_metadata into litellm_params["metadata"] when metadata is absent, and callbacks such as the Lago integration read user_api_key_user_id / user_api_key_team_id only from litellm_params["metadata"], so Cursor calls with user metadata lose callback billing/attribution even though the spend log path is fixed. Please also mirror the proxy identity metadata for callback consumers or update those consumers to read litellm_metadata.

Useful? React with 👍 / 👎.

[discobot]

The responses pipeline overrides this in update_from_kwargs, which sets the logging metadata from kwargs["litellm_metadata"] whenever it is present (litellm/responses/main.py), so a client supplied metadata field lands in optional_params for the provider request body while litellm_params["metadata"] keeps the proxy auth fields. I verified with a spy CustomLogger passing both metadata and litellm_metadata through aresponses; the callback saw user_api_key, user_api_key_user_id and user_api_key_team_id in both litellm_params["metadata"] and get_litellm_metadata_from_kwargs, matching the existing /v1/responses route which has been in LITELLM_METADATA_ROUTES all along

[krrish-berri-2]

@discobot — could you add a screenshot or short video showing that this change works as expected? It really helps reviewers verify the fix quickly. Thanks!

[discobot]

Added a real-provider run to the PR description under Proof of Fix. Short version: on a live proxy with a Postgres SpendLogs table, a /cursor/chat/completions call records nothing for the request on current litellm_internal_staging (only a proxy-level row at spend 0), and with this change the same call writes the aresponses spend row with the real cost. My OpenAI key was rate-limited so I ran it against claude-haiku-4-5; the fix is proxy-side metadata routing so it is provider-agnostic, and the SpendLogs write is exactly what #30126 reported missing.