escapeboy/agent-fleet-o/fleetq-manage-workflow

Manage a FleetQ workflow

A workflow is a reusable directed graph of work. Nodes are typed: start, end, agent, conditional, human_task, switch, dynamic_fork, do_while. Workflows are templates; running one materialises an experiment.

Requires a working connection — see the fleetq-connect skill first.

Fastest path: generate from a goal

workflow_generate decomposes a natural-language goal into a graph. Use it as a starting point, then refine — it produces a draft, not a final answer.

Build by hand

REST equivalent:

POST {{APP_URL}}/api/v1/workflows
PUT  {{APP_URL}}/api/v1/workflows/{id}/graph

Always validate before activating

workflow_validate   →  structural check (unreachable nodes, missing start/end, cycles)
workflow_estimate_cost  →  projected credit cost per run
workflow_simulate   →  dry run without spending credits or sending outbound
workflow_activate   →  make it runnable

workflow_activate on an invalid graph fails. Run workflow_validate first and fix what it reports — the errors name the offending node.

Execute and observe

• workflow_execution_chain — trace what ran, in order.
• workflow_snapshot_list — historical versions of the graph.
• workflow_compensation_runs — compensating actions triggered on failure.

Portability

workflow_export_yaml / workflow_import_yaml move a workflow between teams or environments as plain YAML. workflow_duplicate copies within a team.

Rules that will bite you

• `human_task` nodes block. They create an approval request and wait.

An unattended agent run will stall there until a human completes it, or until the SLA check escalates or expires the task.

• `switch` nodes route on `case_value` on the outgoing edges. An edge

with no matching case is never taken; a switch with no matching edge dead-ends the branch.

• Cost estimates are projections, not caps. Budget enforcement happens

at execution time against the team's real balance.