The workflow builder

The builder is the canvas on the Workflows page. You drag steps from the palette, connect them, and edit each one in the inspector on the right. Changes are saved to the draft. See Workflow graphs for saving, publishing and dry runs.

The palette

Group	Holds
Triggers	Webhook events, label, finding, kanban, cron and manual
GitHub reads	Where a step reads GitHub from: posture (the default), raw files in storage, the webhook payload, or a live API read that costs 1 to 3 calls
System One	Classify, gate, route and score decisions
Crews	Your library’s crews, and the built-in crews with their members
Agents	Every agent you can dispatch, workers first
Gates	Quality gate, release jail, tests, budget check, human approval
State	S3 object, Postgres table, kanban card
Outputs	Draft PR, issue comment, TEA publish, notify, VEX, PR review, approve when ready

Your own crews and agents come from your library. Built-in agents are listed by their canonical names, such as belai:scout. A graph may also name a built-in by its bare name (scout), which the engine accepts.

Which agents a step can dispatch

A crew step dispatches a worker, and hosts advertise only worker-mode agents. Of Belai’s built-in agents, the workers include belai:builder, belai:designer, belai:patcher, belai:planner, belai:repro, belai:reviewer, belai:scout, belai:verifier, belai:vuln-scout, belai:test-fixer, belai:lint-fixer, belai:pr-reviewer and the event workflow agents. See Agents for the full list.

The others run as background agents (mode single), for example the deps-* agents, triage, debug and vulnetix-scanner. The palette shows them greyed out with the reason: “Runs as a background agent, not a worker, so a workflow cannot dispatch it.” If a graph names one anyway, the validator warns with agent_not_dispatchable. It is a warning, not an error, but such a step would find no host.

The decision inspector

Select a decision step of a published workflow to see what it decided over the last 30 days:

• how many decisions it made
• the share of each outcome
• how often confidence was under the threshold
• escalations, and how many a person overrode

The stats route also takes 7 or 90 days, and reports which models answered.

Handoffs

The edges inside a crew come from your agents’ handoff rules: which list and labels each agent claims, what it does on success and on failure, and who it hands to. The map of them appears in the agent builder as “Who hands work to whom”. Labels that an agent adds but no agent claims are flagged, because work with those labels would sit unclaimed.

API

Method	Path
GET	/v1/factory/workflows/palette
GET	/v1/factory/workflows/{id}/nodes/{nodeId}/stats?range=30d
GET	/v1/factory/handoffs