dsh-shared-handoff

When an agent session drags on and context starts to drift, or you are juggling the same project across dsh, Codex, and Claude with task state scattered everywhere and handoffs relying on manual copy-paste, dsh-shared-handoff steps in. It packages the shared-handoff-kit workflow into two zero-dependency, build-free dsh skills that activate conversationally after install, requiring no extra commands to memorize.

The two skills complement each other. handoff handles session-level handoff: it exports your workspace, branch, completed items, and next steps into a paste-ready prompt for a fresh thread; in the new session, saying resume lets the model rebuild context from state files instead of chat history. task-id-bootstrap manages repo-level task state under .agents/state/tasks/, binding the current dsh session to the task and persisting progress, auto-log, and compaction-guard metadata. Crucially, the state layout is identical to the Codex and Claude editions, all three sharing the same session-tasks.json, so a handoff exported from dsh can be resumed in Codex or Claude and vice versa.

The host-side automation, equivalent to the original kit hooks, runs automatically via the harness event system: session-start baseline injection, per-turn snapshot refresh, pre and post-compaction guard markers, and a single-writer rule that keeps subagents read-only. Everything uses Node builtins with zero external dependencies, adapts to macOS, Linux, and Windows 10/11, and never installs Python silently, reporting the gap and asking for consent first. Ideal for developers who frequently switch between multiple agent runtimes and need continuous semantic state across threads.