Flow-Driven — Run

Outcome

One move executed well: the right item advanced through the right step under its contract, its exit evidence produced and independently verified, its state recorded, and the next decision made visible to the human who owns it. One cycle, one move — the power is in repetition, not in scope.

Modes

• mode:pair (default) — ask when the run card is ambiguous; a human is present.
• mode:loop — headless. No blocking questions. When a cycle reaches a human decision, write the brief, set next_decision, log it, and end reporting DECISION-PENDING <slug>.
• mode:brief <slug> — skip selection; prepare the decision brief for this item.

Workflow

1. Validate. flow_lint.py <ws>. GATE: violations mean the definition of workflow is lying to you — fixing them is this cycle's move, and nothing else is.
2. Project. flow_board.py <ws> — regenerate before reading. Never steer from memory, a stale board, or a summary of a board.
3. Get the run card. flow_next.py <ws> prints the rule that fired, why, the item, the step's full run contract, the exit evidence to produce, the candidate next edges, and what to record. Exit code 4 means the top move is a human decision.
4. Sanity-check the selection, do not shop for a better one. The rule ordering encodes the system's beliefs. If the card is genuinely wrong, that is a finding for flow-driven-evolve, not a licence to pick your favourite item.
5. Run the step under its contract. Load the named context packs. Use only the tools the contract allows. Respect the budget and the stop rule; if there is an inner graph, follow it and honour max_iterations. When the escalation condition fires, stop and escalate — that is the contract being satisfied, not a failure.
6. Verify. Run verify_with in a context independent of the run, with a sceptic's brief. Record its findings — including "none, and here is what I examined". If the step declares no verification and sits at rung 3+, verify anyway and open an improvement card.
7. Decide the edge. More than one edge out of the step means a branch condition to evaluate against evidence; record which edge you took and why in the item's ## Notes. Entering a decision point without a dated Decision-log entry is forbidden — prepare the brief instead and set next_decision with a date and a name.
8. Record. Update step, step_entered, holder, evidence_exits_met, and add a dated row to the item's ## Exit evidence log. Re-run flow_board.py (it appends the transition). Append one line to reviews/loop-log.md. Capture at most one learning.
9. Report. CYCLE-COMPLETE <rule> <slug> or DECISION-PENDING <slug>, plus the one thing the next cycle should know.

Decision-brief mode

When the move is a human decision, the deliverable is a brief on the item (or in reviews/): what you can rely on (evidence, with its receipts) · what is still open (and what it would cost to close) · evidence vs opinion, labelled · options with trade-offs, including doing nothing · a recommendation with its reasoning · who decides and by when. Then stop. Agents prepare decisions; they do not make them, and a brief that argues one option without the alternatives is advocacy wearing a brief's clothes.

Rules

• One move. Everything else discovered goes to improvements/, learnings/, or the loop log — never into this cycle's scope.
• Never fabricate exit evidence. If the evidence cannot be produced, the honest move is the back-edge or an escalation. A recorded "we could not" beats an invented "we did".
• Never advance across a decision point without a dated entry naming a human.
• Never edit workflow.md or a step contract during a run cycle, however obvious the fix.
• Escalate rather than improvise when the contract does not cover the situation. Then log the escalation — repeated escalations are the meta-loop's best input.
• Respect WIP limits: over-limit means finish something, not start something.
• If the item is blocked, the move is to clear or escalate the block — not to work around it by pulling something easier.

Quality Gates

• Lint clean before the move (or the move was fixing it); board regenerated after.
• Exactly one move executed.
• Exit evidence exists, is recorded on the item, and was verified by something other than what produced it.
• Any decision-point entry has its dated human decision.
• Loop-log line appended; at most one learning captured.

References

• skills/flow-driven/references/loop-engineering.md — the six beats and why the leverage table is ordered the way it is.
• skills/flow-driven/references/step-contracts.md — running a step, escalation, verification.
• skills/flow-driven/references/workflow-definition.md — what to record and where.
• flow-driven-evolve — where a wrong run card, a missing check, or a recurring escalation goes.