Build workflows

The workflow editor is where you describe “what happens when a task is created” without writing code. Open Workflows in the sidebar.

Built-in workflows ship with Sybra. Fork them to customize. Reset them anytime from the list view.

Edit a workflow

Click any workflow for the editor.

Left side: step list. Center: canvas with the DAG. Right: inspector panel for the selected step.

Configure the trigger

Every workflow starts with one trigger. Click the trigger node to open the inspector.

Pick an event:

• task.created — fires when a new task lands on the board
• task.status_changed — fires on any status transition
• pr.event — fires on GitHub PR events (opened, updated, closed, labeled)

Add conditions to scope the trigger. Conditions are AND-combined and reference one of the fields below:

Field	Use for
task.id	single-task targeting
task.title	title pattern
task.status	tasks in a specific status
task.tags	”only backend tasks”
task.agent_mode	”only headless”
task.project_id	”only this repo”
task.branch	specific branch
task.pr_number	tasks with a linked PR
task.reviewed	already-reviewed tasks
pr.issue_kind	PR failure class (ci_failure, merge_conflict, …)

Operators: equals, not_equals, contains, not_contains, exists.

Add steps

From the canvas, click the + on any node to add a downstream step. Pick the step type:

Type	Use when
run_agent	Dispatch a Claude Code agent
wait_human	Pause until you approve/reject
set_status	Unconditionally set the task’s status
condition	Branch the DAG based on state
shell	Run a shell command (tests, linters)
ensure_pr_closes_issue	Require Closes #N in PR body
verify_commits	Fail if branch has no commits
link_pr_and_review	Open PR from branch, attach review workflow
evaluate	LLM judge step (for grading agent output)

Configure a run_agent step

The most common step. Inspector fields:

Field	Default	Notes
Role	""	triage / plan / eval / pr-fix / "" (implementation)
Mode	headless	or interactive
Provider	—	claude / codex
Model	—	Override the default
Prompt	—	Template; {{ task.body }}, {{ task.title }} interpolate
Allowed tools	[]	Empty = all with skip-permissions
Needs worktree	false	Create a worktree before the run
Max retries	0	Retry on failure
Reuse agent	false	Resume the prior session ID

Prompt templates

Prompts support {{ }} interpolation:

Implement the task described below.

Task title: {{ task.title }}
Task body:  {{ task.body }}
Tags:       {{ task.tags | join: ", " }}
Project:    {{ task.project_id }}

Missing fields render as empty strings. No exceptions.

Human steps

wait_human pauses the workflow and surfaces buttons on the task detail page.

type: wait_human
config:
  status: plan-review
  human_actions: [approve, reject]

status sets the task’s status while waiting. human_actions renders as buttons. Approving completes the step; rejecting fails it.

Branches with condition

- id: check-tag
  type: condition
  config:
    branches:
      - when: { field: task.tags, operator: contains, value: backend }
        next: backend-agent
      - when: { field: task.tags, operator: contains, value: frontend }
        next: frontend-agent
    else: generic-agent

Each branch has a list of conditions (AND). The first matching branch wins. else is the fallback.

Test a workflow

Click Start workflow on any task that matches the trigger conditions. Sybra runs the workflow with live step-by-step tracing. The task’s frontmatter gets a workflow: block:

workflow:
  execution_id: exec-abc123
  status: running
  current_step_id: plan

Reset a built-in

If you forked a built-in workflow and want to revert: open the workflow list, click Reset to built-in on the row. Your version is discarded.

Built-in workflows

Ship and maintained:

• simple-task — default general-purpose workflow (triage → plan → implement → PR)
• testing-task — drives a task through manual testing
• pr-review — attaches to review-tagged tasks
• pr-fix — triggered when a linked PR has CI red or a merge conflict

Treat these as examples. Real teams fork them.

Debugging a workflow

When a step fails, the task surfaces the error. Inspect:

• The agent log (if run_agent)
• The shell command output (if shell)
• The condition evaluation (if condition — Sybra shows which branch was chosen and why)

Fix the step and retry from the task detail page.

Keep workflows small

Every step is a place where something goes wrong. A 12-step workflow has 12 failure modes. Ship 4-step workflows and let the orchestrator handle the rest by conversation.