approvals
set who checks each step of a pipeline, and approve staged results right on the card.
Put a person where a bad result would be expensive, and nobody where it would be noise. Each card in a pipeline has its own approval setting, and staged results wait on the card until you click.
the policy chain
A card's chip, at the bottom of the card, says how its runs deliver: auto, human, self×3 or self×3+human. The setting is looked up in this order, and the first layer that has one wins:
- the card itself
- the pipeline's default
- the process it points at (its chip in the Runs header)
- the canvas that hosts the process
- the built-in default, human review
A card with nothing of its own reads inherit · human (or whatever it inherits). Pick a value to override it, pick inherit to go back.
| option | on a card |
|---|---|
| auto approve | results publish and flow downstream at once |
| human review | results park on the card until you approve |
| self review | the pixie critiques and retries up to the self-review retries setting (1 to 10, default 3), then publishes |
| self + human | self review first, then parks for you |
Whatever the chain says is fixed for a run at the moment it spawns. Changing a chip affects runs started afterwards.
approving on the card
When a run stages results, its card shows a fuchsia review box with thumbnails of the actual outputs and a note such as "poster v2 ready". Hover a thumbnail to enlarge it; click a text result to read it in full.
- Approve publishes everything staged. Downstream cards start the moment it lands.
- The small check on a thumbnail rejects just that output (it needs two or more). The button then reads Approve 2/3, and the dropped output is not published.
- Redo opens what should change? (optional); send it and the pixie reworks. The box hides until the new staging arrives.
- The stop button rejects and ends this run. Nothing publishes and downstream steps skip it.
- The speech bubble opens the run's chat.
Only a click on Approve publishes. The same box appears in the session chat, and both act on the same run, so it doesn't matter where you click.
the review badge
A card with staged work shows a badge, the chat tab shows a count, and the session's row in the panel reads needs approval (or 3 to approve). The Deliverables panel marks the step needs you. The badge on a card clears only when none of its runs are still waiting. If a fan-out staged four runs, approving three leaves it lit.
limits & good to know
- An unanswered review times out after 48 hours and the run is marked timed out; it stops waiting.
- Fan-out multiplies reviews: 5 results into a human-review card means 5 boxes. Use combine upstream, or auto approve on steps you trust.
- Runs reached by email or the API have nobody to click, so they publish directly.
related
checked against the product · 2026-10-23