[xuhaihan]

Summary

• detect consecutive deterministic failures for the same Responses tool in the current user turn
• temporarily remove only the failing tool after two failures and add a model instruction to continue without it
• support top-level, flattened, and Responses Lite namespaced tools across HTTP, WebSocket, and the WS-to-HTTP bridge
• reconstruct state from request history so the guard works across reconnects, failover, and gateway replicas

This is intentionally separate from the declaration/translation fixes in #6062: it protects the runtime when a compatible provider still keeps selecting a tool with an invalid handle or another deterministic failure.

Validation

• go test ./internal/service -run TestOpenAIResponsesToolLoopGuard -count=1
• go test ./internal/service -count=1
• go test ./...
• git diff --check
• sensitive-content scan

[xuhaihan]

@Wei-Shaw 您好，已将此 PR 更新到最新 main（当前 head：f6473765f1），GitHub 显示可合并，当前测试、前端、lint、安全及 CLA 检查均已通过。方便时烦请 review 并合并，谢谢。

[xuhaihan]

@Wei-Shaw 您好，再次提醒您检查这个 PR。当前 head 为 f249e3dc76，测试、前端、lint、安全及 CLA 检查均已通过（cla-lock 按流程跳过），GitHub 当前显示可合并。方便时烦请 review，如无其他问题请帮忙合并；如还需要调整，我会继续跟进。谢谢！