deepseek-harness/packages/workflow/workflow

@deepseek-ai/dsh-workflow

The workflow seam (ctx.workflows) executes a model-written orchestration script that can fan out subagents. The seam defines the script, run, result, error, and event contracts; an engine decides how to isolate and execute the script.

@deepseek-ai/dsh-workflow-workerthread is the current engine and @deepseek-ai/dsh-tool-workflow is the model-facing consumer. A future process or sandbox engine can replace the implementation without changing the tool.

Service and run contract

WorkflowService.start(request): WorkflowRun validates enough synchronously to reject a malformed meta block or unparseable script before a run exists. Once returned, WorkflowRun.result never rejects: execution failures resolve with stopReason: 'error', and cancellation resolves with cancelled within the engine's bounded grace.

A run is holder-owned. Engine-plugin unload prevents new starts but does not revoke accepted runs. The holder must call dispose() on every path; disposal cancels remaining work and reaches or abandons quiescence within the documented bound.

WorkflowStartRequest contains { meta, script, args?, parent, signal? }. parent attributes every child agent to the invoking agent. meta and args are plain data, not script fragments.

WorkflowRun exposes { id, meta, result, cancel(reason?), dispose() }. WorkflowResult contains { value, stopReason, error?, agentsStarted }; value is plain JSON data or null.

Events

Workflow events are observe-only. They carry WorkflowRunInfo (id plus meta) rather than the live run, so listeners cannot acquire cancellation or disposal authority.

• workflow/start / workflow/end pair the run.
• workflow/phase and workflow/log expose script narration.
• workflow/agent-start / workflow/agent-end pair each child call by seq; a child whose async provider start rejects emits neither.

Same-process event payloads are borrowed immutable values. Every listener is independently contained: a synchronous throw or rejected returned promise is logged without starving peers or changing execution.

Failure discipline

WorkflowError carries a code and a fatal flag. Fatal errors always escape parallel() and pipeline() instead of becoming an ordinary per-item null:

• SCRIPT_PARSE / META_INVALID — the workflow cannot start.
• INVALID_ARGUMENT / UNSUPPORTED_OPTION / UNSUPPORTED_SCHEMA — a hook call violates the engine contract.
• AGENT_CAP / ITEM_CAP — configured safety limits were exceeded.
• AGENT_START — the provider's async start rejected.
• AGENT_RESULT — a ready child's result rejected with an infrastructure fault.
• RESULT_UNSERIALIZABLE — a script/worker value is not plain JSON data.
• CANCELLED — cancellation owns the run and pending/future hooks reject.

A child that resolves normally with a non-completed stop reason is not an infrastructure exception: agent() returns null, allowing the script to handle an ordinary child failure.

Model Experience

Indirectly, through dsh-tool-workflow and a workflow engine, which create child-agent requests and return a retained parent tool result.

KV Cache effect

No direct invalidation; the named consumer owns any request-prefix changes.

Known Limitations and Deferred Work

• Foreground collection only — the caller owns one live run and awaits it; background start/poll, spill handles, and detached collection are deferred.
• No journaling or resume — scripts, child progress, and intermediate values are not checkpointed, so a process restart cannot continue a run.
• No saved or nested workflows — the seam starts caller-supplied scripts only, and a workflow script receives no workflow() hook for recursive orchestration.
• No token-budget vocabulary — engines cap concurrency, items, and children, but neither the request nor result accounts for model tokens across children.
• Runs are holder-owned, not service-tracked — unloading the engine does not discover independent live handles; every consumer must dispose the run it started.

See the dynamic-workflows Agent Note for the deferred workflow surface.