feat(proxy): resolve Cursor thinking/fast model-name suffixes on /cursor/chat/completions

[mateo-berri]

TLDR

Problem this solves:

• Cursor appends -thinking-high / -thinking-xhigh-fast to custom model names
• Those suffixed names miss wildcard routes and fail with no healthy deployments
• Users must hand-maintain one config alias per model x level x speed combo

How it solves it:

• /cursor/chat/completions parses the suffixes instead of requiring aliases
• Routes the base name and maps the level to reasoning_effort
• Only rewrites when the raw name is unroutable, so explicit aliases still win

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

Stacks on #34029; the diff collapses to one commit once that merges

Tested end-to-end with the real Cursor app (no curls, no mocks) against a live proxy at f25f1d2 exposed through a cloudflared tunnel, with only wildcard routes configured and no per-variant aliases:

model_list:
  - model_name: openai/*
    litellm_params:
      model: openai/*
      api_key: os.environ/OPENAI_API_KEY
  - model_name: anthropic/*
    litellm_params:
      model: anthropic/*
      api_key: os.environ/ANTHROPIC_API_KEY

Cursor is configured with the proxy master key in the OpenAI API key field and Override OpenAI Base URL pointing at <tunnel>/cursor. Selecting the custom model claude-opus-5 with a thinking level makes Cursor send claude-opus-5-thinking-high and claude-opus-5-thinking-xhigh-fast, which previously returned "no healthy deployments" and now route to anthropic/claude-opus-5 with reasoning_effort high and xhigh respectively. Built-in model names like gpt-5.6-sol continue to route untouched through the openai wildcard. App screenshots of both models answering in chat, plus the proxy spend logs showing the rewritten model and real provider cost, are attached in the PR comments

Type

🆕 New Feature

Changes

Cursor encodes the thinking level and fast toggle for BYOK custom models in the model name itself, so a model the user adds as claude-opus-5 reaches the proxy as claude-opus-5-thinking-high or claude-opus-5-thinking-xhigh-fast. Wildcard routes like anthropic/* cannot match those names because provider inference fails for names absent from the model cost map, so every model x level x speed combination needed its own config alias

This PR teaches the /cursor/chat/completions endpoint added in #34029 to parse the suffixes instead. When the requested name is not servable by the router but the name with -fast and -thinking-<level> stripped is, the body is rewritten to the base model and the level is carried into reasoning_effort for chat completions bodies or reasoning.effort for Responses bodies. Levels are validated against litellm's REASONING_EFFORT literal (none, minimal, low, medium, high, xhigh)

The rewrite is deliberately conservative. It only fires when the raw name would otherwise fail routing, so an explicitly configured alias such as claude-opus-5-thinking-high keeps winning, provider-inferable bare names like gpt-5.6-sol are never touched, and an unknown base name leaves the body unchanged so the original routing error still names the model the client sent. An effort the client already set is never clobbered; only the model name is rewritten in that case

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

The PR adds Cursor-specific model suffix parsing and conditionally rewrites routable base models while preserving explicit aliases and existing reasoning settings

• Adds parsing and router-servability checks for -thinking-<level> and -fast variants
• Applies the resolved model and reasoning effort to chat-completions and Responses-shaped Cursor requests
• Adds unit and endpoint coverage for parsing, conservative routing, body propagation, and both request shapes

Confidence Score: 4/5

The restricted-key failure should be fixed before merging because permitted base-model requests can still be rejected when Cursor appends a thinking or fast suffix

Authentication validates the original suffixed model before the handler resolves it, so exact base-model permissions prevent the newly supported Cursor variants from reaching routing

Files Needing Attention: litellm/proxy/response_api_endpoints/endpoints.py

Important Files Changed

Filename	Overview
litellm/proxy/response_api_endpoints/endpoints.py	Adds Cursor variant resolution, but performing it after authentication prevents restricted keys from using permitted base-model variants
tests/test_litellm/proxy/response_api_endpoints/test_endpoints.py	Adds broad parser and routing coverage, but does not exercise exact allowed-model restrictions through the authentication dependency

_{Reviews (1): Last reviewed commit: "feat(proxy): resolve Cursor thinking/fas..." | Re-trigger Greptile}

[greptile-apps]

Suffix resolution follows authorization

When a key or team permits an exact base model such as claude-opus-5, authentication checks the original Cursor-suffixed name before _resolve_cursor_model_variant runs, causing otherwise permitted thinking and fast variants to be rejected with a model-access error.

Context Used: CLAUDE.md (source)

Knowledge Base Used: Proxy Server Request Flow

[veria-ai]

High: Per-model budget bypass

Authentication checks model_max_budget against the original suffixed name, but this line subsequently routes the request to the base model. A user can call restricted-model-thinking-high after the budget for restricted-model is exhausted because no budget configuration is found for the suffixed name. Resolve the model before authorization, or rerun the resolved-model authorization and budget checks before routing.

[veria-ai]

PR overview

This pull request updates the /cursor/chat/completions proxy endpoint to resolve Cursor thinking and fast model-name suffixes and route requests to the corresponding base model.

One security issue remains open: model suffix resolution occurs after authorization, allowing callers to bypass per-model budget limits by requesting a suffixed name that routes to the restricted base model. No issues have yet been addressed, so the authorization and routing mismatch remains unresolved.

Open issues (1)

• High: Per-model budget bypass — litellm/proxy/response_api_endpoints/endpoints.py:499

Fixed/addressed: 0 · PR risk: 6/10

[codecov]

Codecov Report

❌ Patch coverage is 97.77778% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
litellm/proxy/response_api_endpoints/endpoints.py	97.77%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!