Title: File Context is Not Shared Sequentially Across All Models in Multi-Agent Mode ​Body: I’ve noticed a critical issue with context/file sharing when running multiple agents simultaneously. ​When I uploaded a PDF manuscript (ASCOM-D-26-00177.pdf) and asked for an analysis, only a few models (like ChatGPT 5.5 and Gemini 3.1 Pro) successfully recognized and summarized the document. ​However, at the exact same time, other models (like Claude Sonnet 4.6 and Claude Opus 4.7) completely failed to read the uploaded file. Instead of analyzing it, they just outputted their default greeting templates (e.g., "What would you like me to analyze? Please paste your text..."). ​The core strength of the Multi-Agent mode should be "cross-validation"—allowing users to compare how different models analyze the exact same data to spot overlapping issues. Currently, because the file context isn't being properly fed to all selected models simultaneously, this collective intelligence feature is broken. ​Please review the multi-agent token/file-routing logic so that all active agents can access the uploaded attachment seamlessly.