Not every step needs a model.

Tuckflow draws the whole conversation, then lets you pick what runs each step: a line you wrote, a model held to a schema, or an agent with a goal and a budget. The consent disclosure stays a line you wrote. The hint ladder gets to improvise.

Deterministic

Fixed lines, slots, API calls, branches. Same input, same output, every run. Anything you would have to defend in an audit belongs here.

Constrained generative

A model picks between named exits, fills a schema, or rewords a line you already approved. It chooses. It does not author.

Agentic

A goal, a budget, the skills it may call, the exits it has to leave through. It improvises inside the box, and the box is drawn on the canvas.

Nodes that carry a legal obligation are pinned to the first row. A model never gets to rewrite a disclosure.

WHAT THE SAMPLE FLOW GETS WRONG

The flow loaded here is a system design interview, and it has a real bug in it. In staging session 812, the candidate said they were probably more L5. The agent believed them and wrote candidate.level from that sentence. Nothing verified it, and the scorer reads it. Then the candidate asked to stop, and the candidate_quit exit had no target, so the session fell through to the default end.

Three screens are looking at that one bug from different angles. The trace shows it happening, turn by turn. The release review blocks on it. Production stays on v13 until someone fixes it or Compliance signs off on the risk. That is the part worth judging here, not the node shapes.

THE SCREENS

• Canvas · same flow as Reviewer and Viewer
  Mid-edit, with the inspector open on the agent that runs the interview. Reviewers and viewers get the same picture with the editing taken away.
• Skills
  Seven capabilities, five certified, each one versioned and owned by a named team. Locked rows stay visible so an author asks for access instead of rebuilding the thing badly.
• Release review
  One blocking check, two warnings, four passed. The prompt diff from v6 to v7 sits next to the policy result, because that edit is what caused it.
• Roles and access
  Who edits, who approves, who sees personal data. Authors cannot approve their own release, and that rule is not a role setting anyone here can turn off.
• Simulation trace · and under a JIT grant
  The transcript beside the hops, each one labeled with what held the turn and for how long. Personal data arrives redacted. Reading it costs a two-hour grant, a written reason, and a line in the log.

Sample data throughout, and nothing is wired up. Buttons do not submit, nothing you type survives a reload, and you change role by opening a different page rather than by signing in. It is here to be argued about, not signed into.