dsh_workflow

If you are already building multi-agent collaboration on DeepSeek Harness, you know the pain: model routing, sub-agent providers, tool permissions, approvals, Session logs, background jobs — all powerful, but every task feels like a one-off improvisation. Teams have to repeatedly describe how to decompose, parallelize, verify, and synthesize from scratch in each session. Strategies cannot be reused, parallel results are scattered across conversations, interruptions usually mean starting over, and complex flows are only understood by their authors. DSH Workflow puts an end to that: it upgrades one-shot scheduling into a workflow layer that is generatable, saveable, governable, observable, and resumable — turning multi-agent from a one-time trick into an auditable, shareable, evolvable engineering asset.

On the capability side, it plugs into DSH as an official bundle with zero core patches. It provides versioned capsules (manifest, source, intent, inputs, requires, provenance), a unified async run model, a full WorkflowApi (phase, spawnAgent, runAgent, parallel, pipeline, synthesize, artifact, log, budget, etc.), two built-ins (parallel-investigation and scoped-review), and six standard patterns (classify-and-act, fan-out-and-synthesize, adversarial-verification, generate-and-filter, tournament, loop-until-done). Every run has a stable id and a complete lifecycle, with support for rerunning from an immutable run snapshot, resuming unfinished parts via effect cache, and persisting run graphs, events, artifacts, result summaries, and costs to disk. Discovery and reuse follow a deterministic order: plugin built-ins → project .dsh/workflows → personal directory. Generated scripts run in an isolated QuickJS WebAssembly heap, exposing only JSON capability RPC, with static policy denying file, shell, network, timer, and other non-deterministic APIs.

This plugin is for two kinds of people. First, teams doing complex multi-agent orchestration on DeepSeek Harness who are tired of starting from zero every time — they can save mature strategies as project or personal workflows, run them by name, and pause/resume on demand. Second, platform teams that want to turn agent capabilities into long-lived assets with cost tracking and governance boundaries — using preflight hard limits, approval tiers, read-only allow-lists, and verification adapters to safely hand the workflow library to more users and maintainers.

In short, DSH Workflow makes DeepSeek Harness not just about “calling agents” but about hosting a long-term, maintained agent workflow library. For teams that already have strong execution primitives and are only missing a process layer on top, this is the missing piece that closes the loop.