[upstream] Bedrock Runtime: chat GPT com tools + reasoning via Responses e perfis Claude Fable 5/5.1 #497

O que mudou no upstream

• fix: serve Bedrock Runtime GPT chat with function tools and reasoning through Responses mnfst/llm-gateway#3110 (issue [Bug] HTTP 400 error on global.openai.gpt-6-luna (Bedrock) when using function tools with max reasoning_effort in /v1/chat/completions mnfst/llm-gateway#3107): perfis GPT do Bedrock Runtime (e GPT-6 Astra com tools) rejeitam Chat Completions com function tools + reasoning; as requisições passam pela Responses API. Flag BEDROCK_RUNTIME_CHAT_VIA_RESPONSES (desliga só com false). Arquivos: bedrock-runtime-capabilities.ts, function-tools.ts, chatgpt-adapter.ts, provider-client*.ts, provider-endpoints.ts. Substitui o PR aberto citado na [upstream] Bedrock: prefill de assistente no Sonnet 5.5 e perfis GPT (Luna) com tools + reasoning_effort #136.
• List Claude Fable 5 and 5.1 Bedrock profiles and document the data retention requirement mnfst/llm-gateway#3111 (issue aberta AWS Bedrock request fails for global.anthropic.claude-fable-5-1 with "data retention mode 'default' is not available for this model" (HTTP 400) mnfst/llm-gateway#3114): lista os perfis Claude Fable 5 e 5.1 e documenta que exigem data retention mode diferente de default.
• PR aberto List the GPT-6.1 Sol Bedrock Runtime profiles in model discovery mnfst/llm-gateway#3115 (perfis GPT-6.1 Sol) ainda não mesclado; acompanhar.

Relevância para o Thinker

Thinker tem packages/backend/src/routing/bedrock-runtime-capabilities.ts, sem chatViaResponses. Sem chave Bedrock real o teste é difícil; validar com o agente providers.

Classificação

adaptar: trocar o nome da env var para THINKER_* (ex.: THINKER_BEDROCK_RUNTIME_CHAT_VIA_RESPONSES) e conferir o dossiê em docs/externos/.

Esforço estimado

M

Generated by Claude Code