Sections
Harness entities
The configuration, run, work, council, approval, and Flownix records that form an executable Harness.
Configuration
01Definition → Version
Role · Workflow · Policy
Run
02Run → Step → Work item
Agent · Assignment · Event
Decision
03Proposal · Claim · Evidence
Critique · Vote · Decision
Definition and version
- HarnessDefinition names a reusable process.
typeis a label; execution comes from the graph. - HarnessVersion owns an input schema, an enforced final output schema, a context policy, a default budget, roles, workflow nodes, edges, and policies. A published version is immutable.
- AgentRole carries the objective, instructions, responsibilities, prohibited actions, an output schema enforced on
work.complete, and instance bounds. - WorkflowNode selects runtime behavior and node-specific
config; WorkflowEdge carries an optional condition over the source output.
Run and work
- HarnessRun is one execution with input, Flownix source nodes, result nodes, a budget snapshot, usage, and status.
- StepRun is one node iteration and stores its output and outcome.
- WorkItem is a leased unit for an external agent.
work.get_contextjoins it with the run, node, role, predecessors, and ballot. - ApprovalRequest represents a decision only a human can answer. RunEvent is the audit trail; artifacts and node IDs point to durable results.
Council
A Deliberation groups Proposal, Claim, Evidence, Critique, responses, Ballot, Vote, and Decision records. Multiple nodes share it only through the same config.deliberation_key. council.get_deliberation returns the whole aggregate in one call.
Creating tracked work
There is no separate task-plan generator. A preceding agent can return a structured flow_nodes list; after a human_gate, create_flow_nodes creates ordinary Flownix epic, feature, plan, or task nodes under the initiating user's permissions and project rules.