The built-in flows.

01How a flow runs

A flow is a series of phases, and each phase is a series of steps. Some of those steps are human gates: the run parks there and waits for your answer. When the run has filed findings that are still untriaged, the gate’s body ends with a count of them.

Who walks the steps is a setting: Workflow Orchestration, at Settings → AI → Session settings, with two options: Programmatic (labeled Default) and Orchestrated. Per-agent runtime pins, rewinding a single sprint lane, and Retry failed step exist on programmatic runs. Pause and Resume follow the runtime instead: they work on a Chat (SDK) run and are disabled on the interactive CLI.

• Chatting with a programmatic run at its final gate converts it into a full orchestrated agent that adopts the conversation, one way only. Approving or rejecting without chatting leaves the run as it was.
• While a Planner or Ship run is parked at a gate, you can highlight text on its Idea spec or Architecture design tab and attach comments. Send them to a revision agent that rewrites the document.
• Steering a live run is chat, not buttons. See the run monitor and run controls.

02Tuning levels

Every built-in flow carries a four-way dial: Efficient, Standard (the default), Thorough, and Custom. It sits on the first page of the flow editor, and the launch wizard can override it for one run under Workflow configuration.

• Efficient collapses Sprint’s per-task lane to Implement → Verify, folding tests and self-review into the implement turn, and drops Planner’s optional design track and its separate task-detail step.
• Standard pins a model and an effort level per agent, and makes no structural edits.
• Thorough changes no Sprint steps. It pins heavier models and makes Planner’s design steps and its Approve design gate always on.
• Custom is selectable only once the flow has a saved custom definition.

The level is a stamp resolved when the flow is read, so switching back and forth loses nothing. Launch also stamps a per-project thoroughness that sets the dial’s starting point: prototype gives Efficient, v1 Standard, production Thorough.

03Runtime mix

A second dial picks which provider runs each step: Claude only (the default), Claude primary, Codex primary, and Codex only. Claude primary keeps execution on Claude and sends the review and verify steps to Codex. Codex primary is the reverse.

• The mix is independent of the wizard’s Runtime row.
• It’s frozen into the run’s spec at launch, so a mid-run edit to the flow changes nothing about the run in front of you.
• The two mixed options are greyed out on Compound and Verify Setup, which have no verification steps to cross. The whole row is grayed out on an OMP launch, and while an A/B variant is pinned.
• An explicit per-agent runtime pin beats the mix.

04Launch

Launch bootstraps a brand-new project. Picking it opens a modal headed What are you trying to build?, with a Start interview button. Your answer seeds the interview and can run to 4,000 characters.

The interview asks one question at a time with no cap on rounds, and every fourth question it stops to offer Clarify further or Draft the brief. It then decomposes the brief into an ordered idea set: four to eight ideas, capped at 10.

Launch never materializes a sprint. Run Sprint or Ship afterwards.

05Planner

Planner turns an idea into a reviewed backlog.

• The first gate, headed Approve idea in the run, reviews a short stub: at most five bullets of problem and five of proposal, before the full spec is expanded.
• Both the stub and the plan offer Approve, Revise, or Reject. At the plan gate, Revise reconciles the existing drafts in place and re-presents the gate, and Reject deletes every draft the run created.
• Only a single-task idea is epic-free. Anything larger gets an epic.
• A batch can seed one to four ideas, metered as N of 4 selected. Any seed that sizes large is guarded out with a blocking idea-size-guard decision and dropped, so plan a large idea on its own run. A single-idea run has no size guard.
• In a batch the per-idea approvals move onto an Approve ideas artifact tab, with Approve or Deny per row and one Submit. A batch mints one combined hub tab instead of one per idea, labeled Idea summaries followed by the idea count, and it appears once at least one idea’s ledger has something on it.
• When the context step flags the idea for design, the plain Approve is replaced by Approve → design mode and Approve → keep planning.

Approving the plan retires the originating idea from the board, as long as it got at least one task out of the run. An idea with no task stays.

06Sprint

Sprint executes a batch of ready tasks inside the session’s shared worktree. Each integrated task lands as its own commit.

Dispatch is dependency-aware: independent lanes start as capacity frees, at most five task lanes at once, and lanes that would touch the same files run one after another. A lane starts queued, shows running while its subagent works, and ends integrated, failed, or blocked. Blocked means a prerequisite lane failed.

The Visual check step is optional. It runs only where visual verification is turned on, which is off by default.

A failing stage loops the lane back to Implement at most three times. On a programmatic run the lane then gets one rescue pass. The run’s monitor takes it over, and may rewind it to an earlier step or edit the task body where the acceptance criteria conflict with the repo.

One rescue per lane, four per run, logged as a non-blocking finding. Set CYBOFLOW_DISABLE_LANE_TRIAGE=1 to turn it off.

• If every lane integrated, the run finishes with a full verification pass and a review summary. Then an Address review findings step fixes the findings worth fixing, re-verifies, commits once, and only then resolves them as fixed or triaged. Anything it defers stays open on the queue.
• If any lane failed or was blocked, all three closing steps are skipped. The run goes straight to the gate with a per-lane summary naming each failed task and the step it died on.

Either way, you decide what merges. On the queue the final gate reads Human gate: Human review.

07Ship

Ship is Planner and Sprint in one run, and the longest built-in.

• There’s no Archive idea gate. Approving the plan is what retires the idea.
• Approve plan doubles as the pre-execution gate: you approve the plan and pick which tasks run now. Only Reject tears the drafts down.
• If the subset you approve exceeds the batch cap, materialization is refused and the run stops. It never truncates your selection.
• Ship has no design-mode fork at its idea gate, even when the context agent suggests one.

08Compound

A Compound run mines recently merged work for durable learnings. It writes the approved ones back in three buckets: a quick fix applied in place in the worktree, an edit to your project’s guidance docs (CLAUDE.md or CODE-PATTERNS.md), or a follow-up backlog task.

A finding is Compound’s input, not its output.

• Two gates: Approve over the recommendations artifact, with Approve all, Pick subset, or Reject. Then a terminal Approve compound over the committed diff.
• A run seeded from the Insights triage tray skips the first gate, acts on exactly your selection, and still ends on the merge gate. Each seeded finding is resolved as fixed, triaged, or promoted.
• Compound is exempt from the code review eval jury that grades Sprint and Ship diffs.

09Verify Setup

Verify Setup is the per-project flow that derives a run-book from evidence. It waits for your approval before anything is written to the repo, then commits the run-book and proves it by running it. The flow ends on a human-review merge gate: five steps, two gates, both headed Approve setup.

It’s kept out of the launch wizard’s flow list on purpose. The usual way in is the Verify Queue health block, where each project row carries Set up or Re-run setup. The Workflows gallery and a project overview’s Set up verification recommendation start it too, and each opens the wizard with the flow already picked.

The flow can’t mark its own run-book proven: only a passing proof run does that. The full account is on the verification page.

10The design pass

Planner, Ship, and Launch can each build a prototype (Launch calls it a Concept prototype) and an architecture design, run an adversarial reviewer over them, and stop at an Approve design gate.

Each design step is optional on its own trigger: a prototype when the idea has real UI surface, an architecture design when the change is structurally complex. The adversarial review runs only when one of them did, and the gate is skipped when neither ran. Thorough makes all four always on in Planner and Ship. Launch’s Thorough preset pins models only, so its design steps stay conditional there.

The critique lands as its own Adversarial review artifact tab. The gate is a blocking decision card in the review queue, with two buttons.

Continue, log as findings files every reviewer entry as a non-blocking accepted-risk finding and carries on. Rerun planning with findings re-runs the design steps with the critique threaded in.

Revisions are bounded at five. A sixth ends the run as rejected. Design mode covers what the prototype itself can do.