fix(proxy): don't enforce budgets on model-discovery / info routes (#27923)

[Ar-maan05]

Relevant issues

Fixes #27923

Pre-Submission checklist

• 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

Type

🐛 Bug Fix

Changes

When a team, virtual key, organization, or user budget is exhausted, the proxy returned 429 Budget Exceeded on the model-discovery endpoints (GET /v1/models, /models, and the /model/info family). Every OpenAI-compatible client (Open WebUI, Cursor, Aider, Continue, LibreChat, and others) calls GET /v1/models at startup to populate its model list, so an exhausted budget made all models invisible, including free or self-hosted ones, and broke the whole integration rather than just blocking paid inference.

The cause was an inconsistency in litellm/proxy/auth/auth_checks.py. _global_proxy_budget_check already exempted /v1/models and /models, but the entity-level checks in common_checks (team, virtual key, organization, user) ran inside a blanket if not skip_budget_checks: block with no route awareness, so they raised BudgetExceededError on discovery routes. Model listing runs no inference and incurs no spend, so it should never be budget-gated.

The fix sets skip_budget_checks = True in common_checks when the route is a read-only model-discovery route. The exempt set is a small explicit frozenset (MODEL_DISCOVERY_ROUTES) rather than RouteChecks.is_info_route(). This is deliberate: info_routes also contains side-effectful routes such as /health/services, whose handler can send Slack, email, and webhook test messages, and exempting those would let an out-of-budget key keep triggering them. Keeping the set narrow closes that hole while still fixing discovery. Only the budget-skip flag is touched, so authentication, route allow-listing, model-access, and RBAC checks are unchanged, and inference routes stay fully budget-enforced.

The configurable general_settings behavior the issue also proposes (return-all vs return-free-only vs block) is intentionally left out so this PR stays isolated to the bug; it can follow separately.

Regression tests live in tests/test_litellm/proxy/auth/test_auth_checks.py: the six discovery routes are allowed under an exhausted team and user budget, /health/services and /v1/chat/completions still raise BudgetExceededError. The /health/services case is the mutation guard; widening the exemption back to is_info_route() fails it.

Screenshots / Proof of Fix

Run a local proxy against a config that has at least one real model and a master key:

python litellm/proxy/proxy_cli.py --config litellm/proxy/dev_config.yaml --detailed_debug --reload --use_v2_migration_resolver 2>&1 | tee litellm.log

Create a virtual key with a tiny budget so the first real call exhausts it, then drive one real completion to push spend over budget (this hits a live provider and costs a few cents):

KEY=$(curl -s http://localhost:4000/key/generate \
  -H "Authorization: Bearer sk-1234" -H "Content-Type: application/json" \
  -d '{"max_budget": 0.00001, "models": ["<a-model-in-your-config>"]}' | jq -r .key)

curl -s http://localhost:4000/v1/chat/completions \
  -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -d '{"model": "<a-model-in-your-config>", "messages": [{"role": "user", "content": "hi"}]}' > /dev/null

Now the key is over budget. Model discovery succeeds and inference is still blocked:

# returns 200 with the model list (was 429 Budget Exceeded before this fix)
curl -i http://localhost:4000/v1/models -H "Authorization: Bearer $KEY"

# still returns 429 Budget Exceeded
curl -i http://localhost:4000/v1/chat/completions \
  -H "Authorization: Bearer $KEY" -H "Content-Type: application/json" \
  -d '{"model": "<a-model-in-your-config>", "messages": [{"role": "user", "content": "hi"}]}'

To see the old behavior for contrast, check out litellm_internal_staging, restart the proxy, and repeat the two curls above: GET /v1/models returns 429 there.

[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

This PR fixes a bug where an exhausted team, user, or organization budget caused 429 Budget Exceeded on read-only model-discovery endpoints (/v1/models, /model/info, etc.), breaking OpenAI-compatible clients that call these routes at startup. The fix adds a narrow MODEL_DISCOVERY_ROUTES frozenset in common_checks and sets skip_budget_checks = True before entity budget gates when the request targets one of those routes.

• litellm/proxy/auth/auth_checks.py: Introduces MODEL_DISCOVERY_ROUTES frozenset and a two-line early assignment of skip_budget_checks = True, leaving all authentication, RBAC, and model-access checks intact; only budget enforcement is bypassed for the listed routes.
• tests/test_litellm/proxy/auth/test_auth_checks.py: Adds four focused regression tests (team-budget bypass, user-budget bypass, /health/services still blocked, inference routes still blocked), though MODEL_DISCOVERY_ROUTES is re-declared locally instead of imported from the production module.

Confidence Score: 5/5

Safe to merge — the change is a small, well-scoped addition to common_checks that only skips budget enforcement for six read-only, no-spend model-discovery routes.

The production change is minimal: a frozenset constant and a two-line guard that sets an existing boolean flag. Authentication, RBAC, model-access, and route allow-listing are all unaffected. The narrowness of the frozenset (deliberately kept tighter than is_info_route) avoids unintended bypass of side-effectful endpoints. Regression tests cover the key scenarios, and no logic errors or data-path issues were found.

No files require special attention beyond the minor test-maintainability note on test_auth_checks.py.

Important Files Changed

Filename	Overview
litellm/proxy/auth/auth_checks.py	Adds a narrow MODEL_DISCOVERY_ROUTES frozenset and sets skip_budget_checks=True before entity budget gates when the route is a model-discovery endpoint, fixing #27923 without widening the exemption to side-effectful info routes.
tests/test_litellm/proxy/auth/test_auth_checks.py	Adds four regression tests for #27923 (team budget bypass, user budget bypass, /health/services still blocked, inference still blocked) but duplicates MODEL_DISCOVERY_ROUTES as a local list instead of importing the source-of-truth frozenset from auth_checks.

_{Reviews (2): Last reviewed commit: "fix(proxy): narrow model-discovery budge..." | Re-trigger Greptile}

[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

[Ar-maan05]

@greptile-apps

[Sameerlite]

LGTM, thanks!