dsh-taskswarm

The native subagents and conversational workflows in DSH excel at ad-hoc delegation, but when you face a batch of multi-task, dependency-ordered projects that need to run in parallel, there is no isolation, no quality gate, and no crash recovery—kill the process and all in-flight work is lost. TaskSwarm adds a project-level orchestration layer on top: it topologically layers tasks into waves by their dependency DAG, runs tasks within each wave concurrently in isolated git worktree lanes, routes output through an independent reviewer (PASS or REVISE), resolves merge conflicts semantically via a dedicated merger agent, and integrates everything into a single branch.

The whole pipeline is built for production resilience: every lane commits checkpoints at step boundaries so a crash never loses committed work; durable on-disk state supports pause, resume, and single-lane re-runs; a file mailbox lets the supervisor and workers communicate asynchronously without sharing context; and a zero-dependency Web Dashboard auto-starts with each batch, giving you real-time visibility into wave progress and lane status. Notifications are generated once by the engine in fully formed bilingual messages, so the supervisor never burns tokens re-explaining them.

If you regularly run multiple AI agents in parallel inside DSH and want them to stay isolated, produce reviewable output, survive process crashes, and remain observable end-to-end, TaskSwarm is the orchestration layer built for that scenario.