You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Changes
test(mcp): the missing-agent dry run names an id the sequence already burned (#1715)
fix(retention): pace the pass by the WAL it writes, and let it survive a failing thread (#1710)
feat(mcp): alert channels take their destination URL write-only (#1712)
fix(setup): the setup token is not a password field for password managers (#1711)
fix(graph): a replayed thinking block bound to an older prefix is dropped instead of failing the turn (#1709)
🇺🇸 English
⚡ Improvements
Configure alert channels entirely through the MCP integration. You can now provide the destination URL directly when configuring alerts for instance and agent failures, without having to finish setup in the console. The URL stays securely stored and is never shown back.
🐛 Fixes
Automatic cleanup finishes on large databases. The cleanup introduced in the previous release now runs at a pace the database can handle and continues past problematic conversations, instead of stopping early and putting the database under heavy load for a few minutes. The first complete cleanup on a large database may take a few hours in the background.
Agents using Anthropic Claude 5.5 keep responding in existing conversations. Editing an agent’s prompt or tools, or installing an update that changes them, no longer causes the agent to stop responding in conversations with prior history.
Password managers stop offering to save the setup token as a password. On the instance’s initial setup page, they no longer mistake the setup token for the administrator’s new password.
🇧🇷 Português
⚡ Melhorias
Configure canais de alerta por inteiro pela integração MCP. Agora você pode informar a URL de destino diretamente ao configurar os alertas de falhas da instância e dos agentes, sem precisar concluir a configuração no console. A URL continua guardada de forma protegida e nunca é exibida de volta.
🐛 Correções
A limpeza automática termina em bancos grandes. A limpeza que entrou na versão anterior agora avança num ritmo que o banco aguenta e continua mesmo se encontrar uma conversa problemática, sem parar logo no começo nem sobrecarregar o banco por alguns minutos. A primeira limpeza completa numa base grande pode levar algumas horas em segundo plano.
Agentes com Claude 5.5 da Anthropic continuam respondendo em conversas existentes. Editar o prompt ou as ferramentas do agente, ou instalar uma atualização que os altere, não faz mais o agente parar de responder em conversas que já têm histórico.
Gerenciadores de senha deixam de oferecer salvar o token de configuração como senha. Na página de configuração inicial da instância, eles não confundem mais o token de configuração com a nova senha do administrador.