chore(typing): clear 2.7k basedpyright Any errors across 15 hotspot files

[mateo-berri]

TLDR

Problem this solves:

• Backend Any counts keep drifting up toward their basedpyright ceilings
• 15 hotspot files carried 3,105 reportAny/reportExplicitAny errors

How it solves it:

• Real types at each Any source; zero casts, ignores, or suppressions
• Prisma models, TypedDicts, and Protocols replace Any-typed dicts
• Ratchets budgets down: basedpyright -2,869, ruff-strict -86, type-discipline -90
• Raises the lint job's node heap so basedpyright survives the typed tree
• Adds transformation tests for the retyped volcengine and evals modules

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

Whole-tree basedpyright before and after this branch, measured with the same command the CI gate runs (uv run basedpyright --outputjson | python scripts/type_check_gate.py), counting severity=error diagnostics in-tree:

rule                 before   after   delta
reportAny            27,005   24,427  -2,578
reportExplicitAny     7,439    7,280    -159
all rules combined  157,700  154,832  -2,868

No basedpyright rule increased, repo-wide or in any individual file; every one of the 48 budgeted rules is at or below its baseline count. make lint-budget-update output confirming the fixes are real and the ceilings now hold them:

Ratcheted basedpyright limits down by 2869 errors this branch fixed across 48 rules
Ratcheted strict-rule limits down by 86 violations this branch fixed
Ratcheted LIT-rule limits down by 90 violations this branch fixed

from litellm import * import safety and the circular import check both pass, and make pre-commit is green

Type

🧹 Refactoring

Changes

Typing-only changes across the 15 files with the highest reportAny/reportExplicitAny density, plus the small blast radius those fixes surfaced in 7 neighboring files. No runtime behavior changes: annotations, TYPE_CHECKING imports, TypedDicts, Protocols, and pydantic model_validate swaps that are runtime-equivalent to the Model(**dict) constructors they replace

The recurring root causes and their fixes: untyped Prisma reads now go through typed helpers that pay the Any-to-typed crossing once per boundary (mcp_server/db.py, verification_token_repository.py); dict[str, Any] payloads became local TypedDicts (OAuth credential payloads, aggregated spend rows, guardrail usage responses); duck-typed spend records got a DailySpendRecord Protocol shared by the daily-activity endpoints; response transformation modules (volcengine, openai evals, azure batches, azure_ai count_tokens, ocr) now use the concrete request/response types they already imported

Forbidden constructs were not used anywhere in the diff: no cast(), no # type: ignore, no # noqa, no suppression comments, no new Any annotations. Diagnostics that could not be fixed without one of those were left in place rather than hidden, which is why roughly 380 target errors remain in the touched files (mostly the irreducible one-flag-per-seam residue where upstream sources like PrismaWrapper.__getattr__ still return Any)

Budget files are ratcheted by make lint-budget-update so the cleared headroom cannot silently grow back. The existing mapped suites for every touched module pass (1,140 tests, 0 failures)

Two follow-up commits round the PR out. 26ab846 raises the lint job's node heap to 12GB for the basedpyright budget step: with the hotspots now carrying real types, basedpyright's inference load exceeds node's ~4GB default cap on ubuntu-latest, so the process died with a JS heap OOM and the gate correctly refused the vacuous run (deterministic; a rerun at fbfb63c OOM'd the same way). 8b08c31 adds tests for the volcengine responses and openai evals transformation modules, covering the streaming field-fill heuristics, the model_construct fallbacks, and the get/cancel/delete/list request and response transforms that previously had no tests; these are retyped surfaces of this PR, and covering them also lifts the diff's patch coverage past the codecov target

QA runbook

1. make lint-basedpyright lint-type-discipline lint-ruff-budget all green on this branch
2. git diff litellm_internal_staging --stat shows only typing edits, budget JSONs, and no test deletions
3. Boot the proxy (python litellm/proxy/proxy_cli.py --config litellm/proxy/dev_config.yaml) and confirm clean startup, then exercise a touched surface end to end, e.g. curl http://localhost:4000/user/daily/activity?start_date=2026-07-01&end_date=2026-07-25 -H "Authorization: Bearer sk-1234" and the MCP credential list at http://localhost:4000/ui/?page=mcp-servers; responses are byte-identical to litellm_internal_staging because no runtime code path changed

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

---

Note

Low Risk
Refactor is typing-only with equivalent Pydantic validation patterns; main operational risk is CI memory for basedpyright, which this PR explicitly addresses.

Overview
This PR lowers basedpyright error ceilings in basedpyright-code-budget.json (notably reportAny and related rules) and sets NODE_OPTIONS=--max-old-space-size=12288 on the lint workflow’s basedpyright step so the heavier typed tree does not OOM in CI.

Across the code changes, the work is annotation and boundary typing, not new product behavior: modern union/list syntax, model_validate instead of **dict construction, and TYPE_CHECKING + Prisma model types with small DB accessor helpers (especially in MCP db.py and enterprise project endpoints). OAuth credential payloads get explicit TypedDicts; MCP REST/outbound OAuth stores use Mapping instead of loose dicts. LLM provider modules (Azure batches, OpenAI evals, Volcengine responses, Azure AI count_tokens, OCR) get stricter request/response typing and shared JSON-parse helpers where needed.

^{Reviewed by Cursor Bugbot for commit 8b08c31. Bugbot is set up for automated code reviews on this repo. Configure here.}

[mateo-berri]

bugbot run

[greptile-apps]

Greptile Summary

This PR strengthens static typing across provider transformations, MCP credential handling, management endpoints, usage aggregation, registries, and Prisma boundaries

• Replaces loosely typed payloads with concrete models, TypedDicts, Protocols, and typed helper boundaries
• Adds focused OpenAI evals and Volcengine response transformation tests
• Ratchets static-analysis budgets downward and increases the basedpyright CI heap allocation

Confidence Score: 5/5

The PR appears safe to merge

No blocking failure remains

Important Files Changed

Filename	Overview
litellm/proxy/_experimental/mcp_server/db.py	Adds typed Prisma and OAuth credential boundaries while preserving the existing stored credential fields and refresh flow
litellm/proxy/_experimental/mcp_server/server.py	Narrows MCP server and prefetched OAuth credential types without changing admission or dispatch behavior
litellm/proxy/management_endpoints/common_daily_activity.py	Introduces typed spend-record and aggregate-row representations while retaining nullable SQL aggregate handling
litellm/proxy/guardrails/usage_endpoints.py	Retypes guardrail usage aggregation and log transformation helpers
enterprise/litellm_enterprise/proxy/management_endpoints/project_endpoints.py	Adds typed Prisma table access and Pydantic validation at project-management data boundaries
litellm/llms/openai/evals/transformation.py	Uses concrete eval request and response models and adds transformation coverage
litellm/llms/volcengine/responses/transformation.py	Retypes response parsing and field-fill transformations with expanded streaming and fallback tests
.github/workflows/test-linting.yml	Raises the Node.js heap limit for the basedpyright budget check
basedpyright-code-budget.json	Ratchets type-checking diagnostic ceilings downward to retain the reductions from this refactor

_{Reviews (2): Last reviewed commit: "test: cover volcengine responses and ope..." | Re-trigger Greptile}

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit fbfb63c. Configure here.}

[codecov]

Codecov Report

❌ Patch coverage is 74.11765% with 286 lines in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...oxy/management_endpoints/organization_endpoints.py	33.96%	70 Missing ⚠️
...roxy/management_endpoints/common_daily_activity.py	63.30%	40 Missing ⚠️
...y/management_endpoints/tag_management_endpoints.py	61.84%	29 Missing ⚠️
litellm/proxy/_experimental/mcp_server/db.py	81.56%	26 Missing ⚠️
...ellm/repositories/verification_token_repository.py	62.12%	25 Missing ⚠️
...xy/management_endpoints/internal_user_endpoints.py	55.55%	24 Missing ⚠️
litellm/proxy/policy_engine/policy_registry.py	77.57%	24 Missing ⚠️
litellm/proxy/guardrails/usage_endpoints.py	82.50%	21 Missing ⚠️
litellm/proxy/agent_endpoints/agent_registry.py	76.47%	12 Missing ⚠️
...y/management_endpoints/mcp_management_endpoints.py	82.14%	5 Missing ⚠️
... and 3 more

📢 Thoughts on this report? Let us know!

[codspeed-hq]

Merging this PR will not alter performance

✅ 31 untouched benchmarks

---

_{Comparing litellm_decrease_anys_fable (8b08c31) with litellm_internal_staging (bb6bb66)}

[ryan-crabbe-berri]

LGTM; thanks!

[mateo-berri]

@greptileai

[mateo-berri]

bugbot run

[cursor]

✅ Bugbot reviewed your changes and found no new issues!

Comment @cursor review or bugbot run to trigger another review on this PR

^{Reviewed by Cursor Bugbot for commit 8b08c31. Configure here.}