v1.7.16 v0.5.6

Name: vllm
Author: vllm-project

medium risk

vllm-project/vllmAug 19, 2026 → Aug 20, 20261 day between releases

What's New

Switch from Eino tool pairing fixes to Agent Sandbox reliability and MCP filesystem tools

Breaking Changes

Strict Sandbox-to-Pod mapping prevents duplicate pod creation, potentially breaking environments relying on this behavior.

New Features

Added `list_files` and `file_exists` tools to the MCP server.
Opt-in Prometheus Operator monitoring resources.
Pod scheduling status mirroring to backing pods.
Warm pool reliability improvements.
New examples for Pi coding agent and E2B envd daemon.

Improvements

Enhanced controller reliability and warm pool lifecycle management.
Improved race condition handling in sandbox adoption.
Added bounded wait logic for transient pod networking.

Bug Fixes

Fixed race condition where `SandboxClaim` adopts stale template pods.
Added resilience to informers cache lag.
Added bounded wait and requeue logic for rotating warm-pool candidates.

Upgrade Notes

Update Helm charts to include Prometheus Operator resources. Verify `ownerReference` UIDs are correctly set for Sandbox-to-Pod mapping to avoid pod creation conflicts.

The Two Releases

v1.7.16

Aug 19, 2026· 2 months ago

A bug fix for Eino addressing 400 errors caused by broken tool_call/tool_result pairing in long conversations by adding defense-in-depth middleware.

Full release notes

v0.5.6

Aug 20, 2026· 2 months ago

This release focuses on controller reliability and warm pool lifecycle management, introduces filesystem tools to the MCP server, adds Prometheus Operator monitoring resources, and provides new examples.

Full release notes