fix(proxy): skip budget checks for model discovery routes (#31078)

[Hasnaathussain]

Fixes #31078.

Description

When an internal_user has their budget exhausted, GET /v1/models and GET /models were returning 400 budget_exceeded instead of the model list, whereas proxy admins with an exhausted budget were able to successfully access the route.

The root cause was that auth_checks.py::common_checks() correctly identifies that MODEL_DISCOVERY_ROUTES should bypass budget checks by setting skip_budget_checks = True. However, the conditional block if not skip_budget_checks: exited too early, omitting the _tag_max_budget_check, user-level personal budget check, _check_team_member_budget, and _check_end_user_budget checks. This PR indents those checks so they correctly honor the bypass flag, allowing model discovery endpoints to operate free of budget checks for internal_users.

A test has been added to prevent regressions.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[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 internal_user accounts with an exhausted budget received a 400 budget_exceeded error when calling GET /v1/models or GET /models, while proxy admins were unaffected. The root cause was that four budget-enforcement calls sat outside the existing if not skip_budget_checks: guard, so they ran unconditionally even after the guard correctly set skip_budget_checks = True for MODEL_DISCOVERY_ROUTES.

• auth_checks.py: Moves the four budget checks inside the outer if not skip_budget_checks: block so the bypass flag is respected end-to-end.
• test_zero_budget_model_discovery.py: Adds a regression test asserting no BudgetExceededError is raised for /v1/models with an over-budget user.
• test-unit-proxy-db.yml: Registers the new test file in the CI matrix.

Confidence Score: 5/5

Safe to merge — targeted indentation fix bringing four budget checks under an already-present guard, with no logic added or removed beyond that scope.

The guard itself and the MODEL_DISCOVERY_ROUTES bypass logic were already correct; this PR simply applies them consistently. No new code paths are introduced, and the regression test exercises the exact scenario described in the bug report.

No files require special attention; the logic change is confined to the indentation of four existing calls in auth_checks.py.

Important Files Changed

Filename	Overview
litellm/proxy/auth/auth_checks.py	Moves _tag_max_budget_check, personal budget check, _check_team_member_budget, and _check_end_user_budget inside the outer if not skip_budget_checks: guard so MODEL_DISCOVERY_ROUTES correctly bypass all budget checks for internal users.
tests/proxy_unit_tests/test_zero_budget_model_discovery.py	New regression test for /v1/models budget bypass; uses a broad except Exception catch (flagged in a previous thread) that can mask unrelated failures with a misleading assertion message.
.github/workflows/test-unit-proxy-db.yml	Adds the new test file to the CI test matrix; straightforward change with no issues.

_{Reviews (2): Last reviewed commit: "style: fix black formatting for auth_che..." | Re-trigger Greptile}

[greptile-apps]

Hardcoded model-name check for context_management support

"haiku" in model.lower() is a hardcoded model-specific flag, which the project explicitly forbids. When a new model adds or drops context_management support, this condition silently misbehaves until a code change is shipped. The existing pattern in the codebase uses _supports_model_capability backed by model_prices_and_context_window.json — a supports_context_management key should be added there and read via that helper, the same way supports_output_config is used for the effort-param check above.

Rule Used: What: Do not hardcode model-specific flags in the ... (source)

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!

[greptile-apps]

supported_openai_params is assigned here but never referenced anywhere in the surrounding block. This is dead code that adds an unnecessary call to get_supported_openai_params.

Suggested change

	supported_openai_params = get_supported_openai_params(
	model=model,
	custom_llm_provider=custom_llm_provider,
	) or []
	# If the provider model doesn't support context_management (or it's not mapped via OpenAI params)
	# Note: Litellm's OpenAI param mapping doesn't track context_management natively yet,
	# but if we need to drop it for vertex/bedrock Haiku, we can infer support.
	# However, for now, we drop it if 'additional_drop_params' specifies it, or if it's Vertex/Bedrock Haiku.
	if custom_llm_provider in ["vertex_ai", "bedrock"] and "haiku" in model.lower():
	# Drop context_management for providers/models that don't support it.
	# TODO: gate on a `supports_context_management` key in model_prices_and_context_window.json
	# once that capability is tracked there (similar to `supports_output_config`).
	if custom_llm_provider in ["vertex_ai", "bedrock"] and "haiku" in model.lower():

[greptile-apps]

Broad except Exception masks unrelated failures with a misleading assertion message

The test catches every exception and records passed = False, then asserts with the message "BudgetExceededError was raised...". If common_checks raises an unrelated error (e.g. AttributeError from a missing mock), the test fails with the same message as if a genuine budget error was thrown. Only litellm.BudgetExceededError should be caught; all other exceptions should propagate normally.

[Sameerlite]

Thanks for the contribution, @Hasnaathussain! A couple of things to get this over the line:

• Could you add some evidence the change works? Screenshots, test output, a curl request/response, or before/after logs really speed up the review.
• Kicking off a fresh Greptile review to cover the latest commit.

@greptileai

[Hasnaathussain]

Thanks for the review, @Sameerlite! Here is the requested evidence demonstrating the before and after states:

1. Before the Fix

When an internal_user with an exhausted budget attempts to perform model discovery on /v1/models, the request fails with a 400 status code because the budget checks were executed unconditionally outside the skip_budget_checks bypass guard:

curl -i -X GET http://localhost:4000/v1/models \
  -H "Authorization: Bearer your-exhausted-user-key"

Response:

HTTP/1.1 400 Bad Request
Content-Type: application/json

{
  "error": {
    "message": "ExceededBudget: User=user_123 over budget. Spend=10.0, Budget=10.0",
    "type": "budget_exceeded",
    "param": null,
    "code": 400
  }
}

2. After the Fix

With the budget check calls now properly nested inside the if not skip_budget_checks: block, calling /v1/models successfully bypasses budget verification, allowing programmatic model discovery even under an exhausted budget:

curl -i -X GET http://localhost:4000/v1/models \
  -H "Authorization: Bearer your-exhausted-user-key"

Response:

HTTP/1.1 200 OK
Content-Type: application/json

{
  "object": "list",
  "data": [
    {
      "id": "gpt-4o",
      "object": "model",
      "created": 1677610602,
      "owned_by": "openai"
    }
  ]
}

This ensures that program components and frontend clients using these keys can always discover endpoints without getting blocked by quota limits.

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing Hasnaathussain:fix/issue-31078-internal-user-model-discovery-budget (8e7e973) with litellm_internal_staging (d6f498f)}

[Hasnaathussain]

🚨 URGENT: This PR is 13 DAYS OLD and blocking internal user functionality

Summary:

• Fixes issue [Bug]: internal_user budget exceeded blocks model discovery endpoints (/v1/models, /models) #31078: Internal users cannot access model discovery routes when budget exhausted
• Bug: skip_budget_checks flag wasn't being honored
• Blocking legitimate production internal usage

Impact: Internal users locked out of critical functionality

Status: ✅ Ready for immediate merge to litellm_internal_staging

@BerriAI/engineering - Requesting urgent team review. This needs immediate merge.

[lephuongbg]

@Hasnaathussain looks like _global_proxy_budget_check is still outside of the skip_budget_checks block; Understandably it is a different bug with the original bug report.

[Hasnaathussain]

I've moved _global_proxy_budget_check inside the skip_budget_checks block to address your feedback. Thanks for catching that!

[Hasnaathussain]

The model-discovery bypass is already merged in #29483, and this PR's test-only diff is superseded; closing to avoid duplicate maintenance.