dsh-atlas

Long DSH sessions in a flat text log are hard to review, organize, and resume. Multiple rounds of reasoning, tool calls, and chart artifacts blend into hundreds of lines of output, burying the conclusions you actually need. Atlas treats the DSH Session Log as the single source of truth and projects every turn—user message, model response, and tool activity—into a card on a visual canvas so context, branches, and key findings are visible at a glance.

Core capabilities revolve around three dimensions: viewing, organizing, and continuing. Viewing: cards display preparation, reasoning, tool calls, and generation in real time; once a turn completes, the persisted history locks the final state and eliminates stale “Thinking” indicators. The detail panel preserves original tool-event order and embeds ECharts, Mermaid, and sandboxed render_html canvases. Organizing: the canvas supports drag, 50-to-400-percent zoom, auto-arrangement, and collapsible downstream cards. After searching sessions and cards, you can tag nodes as Important, Key Conclusion, or Needs Verification to track long-running work. Continuing: branching from any answer creates a real DSH Fork with full context preserved. Every card also surfaces aggregated metrics—LLM time, time to first token, token rate, and cache hit rate—without exposing individual model-call breakdowns.

Atlas does not duplicate the chat system or rewrite original conversations. Model selection, reasoning levels, permission presets (Read Only, Workspace Write, Full Access), Slash commands, and the at-sign context menu (local files, conversation history, Skills) all run through the active DSH session API. Atlas stores only auxiliary data such as layout, status, and versioned summaries. Local files are parsed in the browser and injected as turn-level context; original binaries are never uploaded or retained.

Built for developers who rely on DSH for multi-round deep reasoning, tool-chain orchestration, and code analysis, and for anyone who needs to quickly locate conclusions, compare branch paths, and take structured notes from long sessions.