AI CodeOps

Workflows

Last updated

The 16 built-in AI CodeOps workflows and their steps, plus how to start one, schedule it, re-run it, cancel it, and build on a finished session with follow-ups.

A workflow is the recipe a session follows: a fixed list of steps, each run by one agent. You pick the workflow when you start a session. Whichever you pick, a session that writes code ends the same way: your tests run, the AI review records findings, and a person approves before a pull request opens.

The built-in workflows

Browse them under Develop → Workflows, in the Workflow Catalog.

Building and changing code

WorkflowSteps (agent)
Plan, Build, Review & Test (default)Plan (Architect) → Build (Builder) → Review (Reviewer) → Test (Tester)
Quality-Gated ChangeBuild (Builder) → Test (Tester) → Review (Reviewer)
Security-Sensitive ChangeScope change (Architect) → Risk analysis (Security Auditor) → Implement (Builder) → Audit diff (Security Auditor) → Validate (Tester)
UI ImplementationUI plan (UI Designer) → Implement (Builder) → Code review (Reviewer) → Behavior test (Tester)
DevOps ChangeDevOps plan (DevOps Engineer) → Risk review (Security Auditor) → Apply (DevOps Engineer) → Validate (Tester)
Feature SDLCRequirements (Product Analyst) → Delivery plan (Project Planner) → Technical design (Architect) → Backend → Frontend → E2E tests → Security review → User docs (Technical Writer)

Review and readiness

WorkflowSteps (agent)
Parallel ReviewCorrectness review (Reviewer) → Security review (Security Auditor) → Test coverage review (Tester) → Merge findings (Architect)
Swarm InspectionReviewer slice → Security slice → Tester slice → Merge summary (Architect)
Release ReadinessCorrectness review → Security review → E2E sanity → Dependency audit → Go / no-go (Architect)

Discovery and design

WorkflowSteps (agent)
Architecture DebateArchitect A → Architect B → Critic review (Reviewer) → Final synthesis (Architect)
Discovery & SpecPrior art (Researcher) → Requirements (Product Analyst) → Delivery plan (Project Planner)
Threat ModelScope & data-flow (Architect) → STRIDE analysis (Security Auditor) → Record ADR (ADR Writer)

These three save their decision as a proposed architecture decision record under Plan → ADRs.

Operations and maintenance

WorkflowSteps (agent)
Incident ResponseTriage (Incident Responder) → Root cause (Architect) → Mitigation (Builder) → Verify (Tester) → Postmortem (Site Reliability)
Dependency RefreshAudit (Dependency Auditor) → Upgrade (Migration Specialist) → Test (Tester) → Verify secure (Security Auditor)
Performance TuningProfile (Performance Tester) → Optimization plan (Architect) → Apply (Builder) → Re-measure (Performance Tester)
Data MigrationSchema (Database Engineer) → Migration (Migration Specialist) → Validate (Tester) → Security → Final review (Reviewer)

Steps always run one after another, and each step sees what the earlier steps wrote. The catalog's Parallel, Debate and Swarm labels describe how the steps divide the work, not that they run at the same time. A step whose name contains "plan" also adds its items to the Plan board.

When the code is written

The code is written after the step that implements the change, not after the last step:

  1. The steps up to and including the one that implements the change run first. That step is Build, Implement, Apply, Mitigation, Upgrade or Migration, and in Feature SDLC it is Frontend. In Plan, Build, Review & Test, Plan and Build run first.
  2. The change is written, and your project's tests run on it.
  3. The steps after the implementing step run next, such as Review and Test, Audit diff and Validate, or Re-measure. In Feature SDLC they are E2E tests, Security review and User docs. Each is shown the change as written and the recorded test result, so it checks the code, not a plan. If one of them fails or runs out of time, the session keeps its change and the step is marked failed.
  4. The AI review of the diff records findings.

Parallel Review, Swarm Inspection and Release Readiness have no implementing step, so all their steps run before the code is written. The three discovery and design workflows write documents and no code.

The session's summary is the last step's text, so in a workflow with steps after the code it is written after the tests ran. The session page shows a warning if what the summary says about tests contradicts the recorded result.

Start a workflow

You can start one from several places:

  • Develop → AI Sessions → New session. Pick any workflow in the Workflow list. The default is Plan, Build, Review & Test.
  • Develop → Workflows. Choose Launch on a workflow, describe the task, pick a Project, and choose Start session.
  • A project's page. Run workflow starts the default workflow on that project.
  • Fix with AI, on an incident, the Quality page or the Security page, and Fix on a finding. These start the default workflow on the problem. A fix for a finding is asked to include a test that reproduces it.
  • A schedule (below).

Every workflow run needs a project. Your workspace policy can require approval to start a workflow, or deny it. See Approvals and policies.

Schedule a workflow

Open Release → Schedules and choose New schedule.

  1. Enter a Name, pick a Workflow and a Project.
  2. Enter a 5-field Cron expression and a Timezone. Choose Preview next 5 to check the times.
  3. Optionally describe the Task. Without one, the session runs "Run the scheduled workflow" with the schedule's name.
  4. Choose what happens if a scheduled tick is missed: catch up once, skip missed, or exact.
  5. Choose Create schedule.

Each schedule row has Run now, History, Edit, Clone, Disable and Delete. Deleted schedules can be restored from Recently deleted.

Cancel a session

On a running session's page, choose Cancel session. Enter a reason, then confirm Cancel session. The session stops within a few seconds, including any model call in flight, and its change is marked abandoned. You need a role that can manage sessions; viewers can't cancel.

Re-run a session

On a finished session's page, choose Re-run and confirm. It starts a new session with the same task, workflow, project and model. The new session isn't linked to the original. If the original workflow can't be launched any more, the re-run uses the default workflow and the dialog says so.

Follow up on a session

A follow-up builds on an earlier session instead of starting over.

  1. On the session page, write in the Follow-up prompt box, for example "Also handle the empty-input case."
  2. Pick the workflow for the follow-up. It defaults to the same one.
  3. Choose Send follow-up.

The follow-up is a new session, linked to its parent: its page shows Parent session under Linkage. It uses the parent's project and model, and gets the parent's task and summary as context. Its page shows the original Task once and, under it, Latest request (follow-up N) with what you asked this time. If the parent's change hasn't been merged yet, the follow-up starts from it, so the new diff contains both.

While a session is still running, the same box says Mid-session handoff / follow-up and the button Spawn handoff. It starts a new linked session. The running session carries on. Choosing a step in the session's step strip starts the new session from that step.

Next steps

Frequently asked questions

Can I build my own workflow?

Not yet. Every session runs one of the 16 built-in workflows, each with a fixed list of steps.

Do the steps of Parallel Review run at the same time?

No. Every workflow runs its steps one after another. Parallel, debate and swarm describe how the steps divide the work, not timing.

Is Re-run the same as a follow-up?

No. Re-run starts the same task again from scratch, with no link to the original. A follow-up builds on the earlier session and is linked to it.

Was this page helpful?