dsh-workflow
Upgrade one-shot multi-agent scheduling into a generatable, savable, governable, recoverable workflow layer
Install
dsh plugin --profile web add @dsh-external/workflow dsh-workflow
Turn multi-agent scheduling from a one-off trick into a generatable, savable, governable, recoverable engineering asset — DSH doesn’t just “call agents” anymore; it can host a long-lived workflow library.
What problem it solves
DSH’s built-in foreground workflow tool is great for “run several things in parallel this once”, but how to decompose, parallelize, verify, and synthesize has to be re-described every time, making strategies hard to reuse or govern. Parallel results scatter across the session, interrupted runs usually start over, and provider / model / concurrency / budget can only be constrained through prompts.
This plugin (npm: @dsh-external/workflow, official bundle form, zero core patches) independently implements a KodaX-style workflow design on top of DSH’s Cordis, ctx.subagents, sessions, background jobs, approval, commands, and tools, adding the higher-level product capabilities: naming, discovery, generation, reuse, pause/resume, re-run/continue, persistent evidence, cost records, and governance. Run graphs, events, artifacts, summaries, and costs are persisted; interrupted runs can restart from a run snapshot or continue with an effect cache.
Key features
- Reusable workflow capsules: versioned
dsh.workflowv1 with manifest / source / intent / inputs / requires / provenance - Full WorkflowApi:
phase,spawnAgent,runAgent,parallel,pipeline,synthesize, single-level nested workflows, and more - Two built-in workflows: the parameterizable
parallel-investigationandscoped-reviewwith a full review chain - Six standard patterns: classify-and-act, fan-out-and-synthesize, adversarial-verification, generate-and-filter, tournament, loop-until-done
- Persistence and continuation: run graphs / events / artifacts / costs are persisted; restart from a run snapshot or continue with an effect cache
- Governance and hard limits: provider / model / concurrency / budget enforced by manifest + preflight + runtime hard limits, with graded approvals
- Safe sandbox: generated scripts run in a QuickJS WebAssembly isolated heap, capability-only VM, JSON boundaries, and deterministic guards
Quick start
Install and restart the profile, then use /workflow commands in the session: /workflow list to discover workflows, /workflow create <request> to generate one, /workflow <name> [JSON args] to run by name, /workflow review to trigger a review, /workflow runs / show / pause / resume / stop to manage runs, and /workflow save to reuse. Models can also call the three tools workflow_list, run_workflow, and workflow_manage.
Who it is for
- Teams that want to turn reusable multi-agent flows into engineering assets
- Advanced users who need observability, recoverability, and governance (cost / permission / audit) for their flows
- Developers who want KodaX-style workflow capabilities while staying in the DSH ecosystem
Compiled from the project README · All rights belong to the original author
Report issue / Request removal