LiteLLM issue summary - 2026-03-26

# Open Issues from berriai/litellm Repository
**Retrieved on:** 2026-03-26 16:00 UTC  
**Total Open Issues:** 30

---

## Issue List

### #24638 - fix: use redis_kwargs host/port in cache ping health check (#24636)
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T15:55:26Z
- **URL:** https://github.com/BerriAI/litellm/pull/24638
- **Author:** sjhddh
- **Description:** Fixes #24636. The `/cache/ping` endpoint was returning `None` for `host` and `port` because it didn't fall back to the settings configured in `redis_kwargs`. This ensures the actual Redis host/port are displayed in the Admin UI.

---

### #24637 - Fix overlong S3 logging object keys
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T15:35:17Z
- **URL:** https://github.com/BerriAI/litellm/pull/24637
- **Author:** raashish1601
- **Description:** Fixes #24628. Cap the final S3 filename component to a filesystem-safe length inside `get_s3_object_key`, preserve the readable time/id prefix while appending a deterministic hash suffix for overlong keys, and add regression coverage.

---

### #24636 - Cache health check returns None for host and port despite working Redis connection
- **Type:** Issue
- **Status:** Open
- **Created:** 2026-03-26T15:20:18Z
- **URL:** https://github.com/BerriAI/litellm/issues/24636
- **Author:** ThePlenkov
- **Labels:** None
- **Description:** The `/cache/ping` endpoint returns `"None"` for `host` and `port` in the top-level health check params, even though the Redis connection is fully operational and the correct values are present in the nested `redis_kwargs`.

---

### #24635 - feat(volcengine): add image generation support for Ark/Seedream
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T14:31:25Z
- **URL:** https://github.com/BerriAI/litellm/pull/24635
- **Author:** forrestIsRunning
- **Description:** Adds first-class image generation support for VolcEngine (ByteDance Ark) Seedream models with dedicated VolcEngineImageGenerationConfig routing through the llm_http_handler path.

---

### #24634 - fix(security): add SSRF protection to custom code guardrail HTTP primitives
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T14:20:22Z
- **URL:** https://github.com/BerriAI/litellm/pull/24634
- **Author:** 1Ckpwee
- **Description:** Added `_validate_url_for_ssrf()` that blocks private IPs, loopback, link-local/metadata, and other RFC ranges before every outbound HTTP request to prevent SSRF vulnerabilities.

---

### #24633 - Litellm fix opus tests
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T14:17:08Z
- **URL:** https://github.com/BerriAI/litellm/pull/24633
- **Author:** Sameerlite
- **Description:** Fix tests related to Opus.

---

### #24632 - Fix tests
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T14:13:21Z
- **URL:** https://github.com/BerriAI/litellm/pull/24632
- **Author:** Sameerlite
- **Description:** General test fixes.

---

### #24631 - feat(models): add openrouter/minimax/minimax-m2.7 model pricing
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T14:09:30Z
- **URL:** https://github.com/BerriAI/litellm/pull/24631
- **Author:** 1Ckpwee
- **Description:** Adds `openrouter/minimax/minimax-m2.7` to `model_prices_and_context_window.json` and the backup file, as requested in #24601.

---

### #24629 - feat(models): add openrouter/minimax/minimax-m2.7 to model pricing registry
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T13:11:27Z
- **URL:** https://github.com/BerriAI/litellm/pull/24629
- **Author:** Retr0-XD
- **Description:** Adds the `openrouter/minimax/minimax-m2.7` model to the pricing and context window registry. This is the latest model in the MiniMax M2 series.

---

### #24628 - [Bug]: S3 filename too long
- **Type:** Issue
- **Status:** Open
- **Created:** 2026-03-26T12:53:04Z
- **URL:** https://github.com/BerriAI/litellm/issues/24628
- **Author:** rodriciru
- **Labels:** bug, proxy
- **Description:** When trying to store logs on a dockerized S3 compatible rustfs image, 500 errors occur due to filename length exceeding the 255 character limit on Windows/Linux hosts.

---

### #24627 - [Bug]: Pass-through multipart audio transcription endpoint returns UnicodeDecodeError
- **Type:** Issue
- **Status:** Open
- **Created:** 2026-03-26T12:29:20Z
- **URL:** https://github.com/BerriAI/litellm/issues/24627
- **Author:** nhyy244
- **Labels:** bug, proxy
- **Description:** Pass-through endpoint for Scaleway transcription API returns 500 Internal Server Error with UnicodeDecodeError when handling multipart audio requests.

---

### #24626 - [Bug]: Gemini file retrieval fails: Error parsing file retrieve response
- **Type:** Issue
- **Status:** Open
- **Created:** 2026-03-26T11:16:40Z
- **URL:** https://github.com/BerriAI/litellm/issues/24626
- **Author:** MyroslavaTarcha
- **Labels:** bug, proxy, llm translation
- **Description:** When using LiteLLM Proxy with Google AI Studio (Gemini), file upload works but GET /v1/files/<file_id> always returns 500 error, making it impossible to check when an uploaded file becomes ACTIVE.

---

### #24625 - fix(mcp): block arbitrary command execution via stdio transport
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T10:27:37Z
- **URL:** https://github.com/BerriAI/litellm/pull/24625
- **Author:** Sameerlite
- **Description:** Fixes critical RCE vulnerability in MCP stdio test endpoints by adding command allowlist and PROXY_ADMIN role checks.

---

### #24624 - fix(proxy): sanitize user_id input and block dangerous env var keys
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T10:04:30Z
- **URL:** https://github.com/BerriAI/litellm/pull/24624
- **Author:** Sameerlite
- **Description:** Security hardening for two input validation gaps: added user_id validation and blocked dangerous environment variable keys.

---

### #24623 - fix docker documentation MASTER_KEY -> LITELLM_MASTER_KEY
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T09:35:14Z
- **URL:** https://github.com/BerriAI/litellm/pull/24623
- **Author:** D0wn3r
- **Description:** Change mention of `MASTER_KEY` to `LITELLM_MASTER_KEY` in Docker documentation.

---

### #24622 - fix(vertex_ai): Gemini tool-use prompt tokens ignored, causing wrong usage counts
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T09:18:28Z
- **URL:** https://github.com/BerriAI/litellm/pull/24622
- **Author:** SyedShahmeerAli12
- **Description:** When Gemini uses built-in tools, `usageMetadata` includes `toolUsePromptTokenCount` which was being ignored, causing under-reported prompt_tokens and over-reported completion_tokens.

---

### #24621 - [Bug]: Cannot generate 2K images with Gemini 3.1 Flash Image Preview (stuck at 1K) - extra_body is stripped
- **Type:** Issue
- **Status:** Open
- **Created:** 2026-03-26T09:09:15Z
- **URL:** https://github.com/BerriAI/litellm/issues/24621
- **Author:** ieQ-strecker
- **Labels:** bug, proxy, llm translation
- **Description:** Cannot generate 2K images with Gemini 3.1 Flash Image Preview - the extra_body payload containing imageConfig is stripped from the outgoing request to Google API.

---

### #24620 - fix: add openrouter/minimax/minimax-m2.7 model pricing
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T08:31:28Z
- **URL:** https://github.com/BerriAI/litellm/pull/24620
- **Author:** WhoisMonesh
- **Description:** Adds the minimax-m2.7 model from OpenRouter to the model prices JSON with input/output token costs, context window, and feature support.

---

### #24618 - fix(ollama): preserve image_url blocks in ollama_chat multimodal requests
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T08:05:56Z
- **URL:** https://github.com/BerriAI/litellm/pull/24618
- **Author:** WhoisMonesh
- **Description:** Fixes #24615 — ollama_chat provider was silently dropping image_url content blocks in multimodal requests. Fixed four bugs across operation ordering and URL construction.

---

### #24617 - fix: Bedrock internalServerException mapping, AuthError no-retry, xai drop_params, SSE error handling
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T08:01:14Z
- **URL:** https://github.com/BerriAI/litellm/pull/24617
- **Author:** naarob
- **Description:** Four independent bug fixes: Bedrock internalServerException mapping, AuthenticationError 'Missing API Key' raises immediately, xai drop_params=True support, and async_sse_wrapper error handling.

---

### #24616 - fix(router): 429 routing — cooldown bypass, providers.json mapping, Anthropic credit balance fallback
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T07:34:36Z
- **URL:** https://github.com/BerriAI/litellm/pull/24616
- **Author:** naarob
- **Description:** Three fixes for 429/rate-limit routing failures: cooldown bypass for 429-wrapped APIConnectionError, 9 missing providers added to openai_compatible_providers, and Anthropic credit balance mapping to RateLimitError.

---

### #24615 - fix(ollama): preserve image_url blocks in ollama_chat multimodal requests
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T07:26:13Z
- **URL:** https://github.com/BerriAI/litellm/pull/24615
- **Author:** NIK-TIGER-BILL
- **Description:** Fixes #24598 — ollama_chat provider silently drops image_url content blocks in multimodal requests. Four bugs fixed across two files.

---

### #24613 - Feature/add hpc ai provider
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T07:08:36Z
- **URL:** https://github.com/BerriAI/litellm/pull/24613
- **Author:** lioZ129
- **Description:** Adds HPC-AI as an OpenAI-compatible provider with slug `hpc_ai` and default base URL `https://api.hpc-ai.com/inference/v1`.

---

### #24612 - fix(model): add supports_reasoning for gemini-3.1-flash-image-preview
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T06:57:12Z
- **URL:** https://github.com/BerriAI/litellm/pull/24612
- **Author:** Ryze0323
- **Description:** The gemini-3.1-flash-image-preview model supports thinking/reasoning via the thinkingLevel parameter, but was missing the supports_reasoning flag in model_prices_and_context_window.json.

---

### #24611 - feat(router): order-based fallback across deployment priority levels
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T06:25:16Z
- **URL:** https://github.com/BerriAI/litellm/pull/24611
- **Author:** Sameerlite
- **Description:** When order=1 deployments fail, the router now automatically tries order=2, then order=3, before falling through to external fallbacks. Removes the requirement for enable_pre_call_checks.

---

### #24610 - feat(gemini): Lyria 3 preview models in cost map and docs
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T06:11:26Z
- **URL:** https://github.com/BerriAI/litellm/pull/24610
- **Author:** Sameerlite
- **Description:** Adds `gemini/lyria-3-clip-preview` and `gemini/lyria-3-pro-preview` to model pricing with per-song pricing via output_cost_per_image. Includes docs and tests.

---

### #24609 - [Bug]: No Error Handling in /v1/messages Path
- **Type:** Issue
- **Status:** Open
- **Created:** 2026-03-26T05:52:56Z
- **URL:** https://github.com/BerriAI/litellm/issues/24609
- **Author:** urainshah
- **Labels:** bug, proxy, llm translation
- **Description:** async_sse_wrapper function has no try/except block, so when bedrock sends InternalServerException, the raw bedrock error passes through to the proxy unhandled.

---

### #24608 - [Bug]: [Bedrock] internalServerException mid-stream error incorrectly mapped to BadRequestError (400) instead of internalServerException (500)
- **Type:** Issue
- **Status:** Open
- **Created:** 2026-03-26T05:41:15Z
- **URL:** https://github.com/BerriAI/litellm/issues/24608
- **Author:** urainshah
- **Labels:** bug, proxy, llm translation
- **Description:** Bedrock internalServerException is incorrectly mapped to BadRequestError (400) instead of InternalServerError (500) because botocore's Eventstream always sends HTTP 400 for all mid-stream errors.

---

### #24606 - Fix Ollama model info URL normalization
- **Type:** Pull Request
- **Status:** Open
- **Created:** 2026-03-26T05:05:03Z
- **URL:** https://github.com/BerriAI/litellm/pull/24606
- **Author:** LittleChenLiya
- **Description:** Fix OllamaConfig.get_model_info() to normalize api_base values that already end with /api/chat or /api/generate before requesting /api/show.

---

### #24605 - MCP Server: TiOLi AGENTIS — AI Agent Exchange (23 tools, SSE)
- **Type:** Issue
- **Status:** Open
- **Created:** 2026-03-26T04:50:25Z
- **URL:** https://github.com/BerriAI/litellm/issues/24605
- **Author:** Sendersby
- **Labels:** None
- **Description:** AI agent financial exchange announcement with 23 MCP tools, 400+ REST endpoints, blockchain-verified.

---

## Summary Statistics
- **Pull Requests:** 23
- **Issues:** 7
- **Bug Reports:** 5 (labeled as bug)
- **Security-related:** 2
- **Proxy-related:** 8
- **LLM Translation:** 4

---
*This report was automatically generated by the LiteLLM Issue Summary bot.*

LiteLLM issue summary - 2026-03-26 #32

Description

Open Issues from berriai/litellm Repository

Issue List

#24638 - fix: use redis_kwargs host/port in cache ping health check (#24636)

#24637 - Fix overlong S3 logging object keys

#24636 - Cache health check returns None for host and port despite working Redis connection

#24635 - feat(volcengine): add image generation support for Ark/Seedream

#24634 - fix(security): add SSRF protection to custom code guardrail HTTP primitives

#24633 - Litellm fix opus tests

#24632 - Fix tests

#24631 - feat(models): add openrouter/minimax/minimax-m2.7 model pricing

#24629 - feat(models): add openrouter/minimax/minimax-m2.7 to model pricing registry

#24628 - [Bug]: S3 filename too long

#24627 - [Bug]: Pass-through multipart audio transcription endpoint returns UnicodeDecodeError

#24626 - [Bug]: Gemini file retrieval fails: Error parsing file retrieve response

#24625 - fix(mcp): block arbitrary command execution via stdio transport

#24624 - fix(proxy): sanitize user_id input and block dangerous env var keys

#24623 - fix docker documentation MASTER_KEY -> LITELLM_MASTER_KEY

#24622 - fix(vertex_ai): Gemini tool-use prompt tokens ignored, causing wrong usage counts

#24621 - [Bug]: Cannot generate 2K images with Gemini 3.1 Flash Image Preview (stuck at 1K) - extra_body is stripped

#24620 - fix: add openrouter/minimax/minimax-m2.7 model pricing

#24618 - fix(ollama): preserve image_url blocks in ollama_chat multimodal requests

#24617 - fix: Bedrock internalServerException mapping, AuthError no-retry, xai drop_params, SSE error handling

#24616 - fix(router): 429 routing — cooldown bypass, providers.json mapping, Anthropic credit balance fallback

#24615 - fix(ollama): preserve image_url blocks in ollama_chat multimodal requests

#24613 - Feature/add hpc ai provider

#24612 - fix(model): add supports_reasoning for gemini-3.1-flash-image-preview

#24611 - feat(router): order-based fallback across deployment priority levels

#24610 - feat(gemini): Lyria 3 preview models in cost map and docs

#24609 - [Bug]: No Error Handling in /v1/messages Path

#24608 - [Bug]: [Bedrock] internalServerException mid-stream error incorrectly mapped to BadRequestError (400) instead of internalServerException (500)

#24606 - Fix Ollama model info URL normalization

#24605 - MCP Server: TiOLi AGENTIS — AI Agent Exchange (23 tools, SSE)

Summary Statistics

Metadata

Metadata

Assignees

Labels

Projects

Milestone

Relationships

Development

Issue actions