dsh-work-continuity

You log off mid-task at 1 a.m. The next morning you open a fresh session and the agent draws a blank. You end up scrolling through chat history to re-explain the goal, the decisions you made, and what was left on the table. dsh-work-continuity turns that hand-off into an automatic, auditable trail. Goal changes, todo writes, and conversation-level milestones are captured into a dedicated work-state layer, so a new session picks up exactly where the last one left off.

The plugin layers three complementary capture mechanisms that back each other up. Event-driven capture listens to authoritative host events (goal/change and todo/write) for deterministic, LLM-free recording. LLM-autonomous capture injects a compact ~150-token summary every round so the model can decide on its own when to log a thought, a decision, or a milestone via the work_state tool. And the explicit /checkpoint command keeps full manual control in the user's hands—set goals, record decisions, add next steps, track artifacts and unresolved issues, all viewable, editable, and clearable at a glance.

Two design principles hold firm: human-verifiability and data ownership. Every entry written to work.db can be inspected and audited at any time; there is no hidden per-turn summarization. Uninstalling the plugin never deletes your data, and reinstalling restores it instantly. All failure paths are fail-open—plugin errors are logged but never block the conversation. This plugin is built for engineers who drive agent-assisted development and need reliable, cross-session continuity across goals, decisions, and next steps.