Skip to content

Reviews and outputs

You review the work, not the log. Workers attach what they made to the item, and when an item needs your eye, the outputs come to you together in one review.

An output is something work produced, attached to the item that produced it: a screenshot, a mockup, a report, a video, a pull request, a link.

  • Files are stored with the project, never public. Up to 25 MB each: PNG, JPEG, GIF, WebP and SVG images, PDF, HTML, MP4, WebM and QuickTime video, Markdown, plain text, CSV and JSON.
  • Previews: images show inline, Markdown is rendered (that covers most reports), and PDF, HTML and video open through a short-lived link. HTML previews are static: no scripts, no forms.
  • Versions: a new version of an output replaces the old one on the item (“B → B2”), and the review shows both.
  • Who attaches: workers attach files from their worktree; Lea can attach a file you dropped into the chat; you can attach files on the item page.

Plan › Outputs lists every output and pull request in the project, newest first, by milestone.

A file from outside the project (a tester’s screenshot, say) is attached as untrusted and stays marked that way.

When a worker says an item is done, the item goes into review. Depending on its checks, it passes up to three stages:

  1. Agent: the item’s owner agent (usually Lea) reviews the result against the checks and approves or requests changes.
  2. Validator: a separate run (V-3) checks the rubric and commands, on the other model family by default.
  3. Person: a review for a person in Needs you. An item gets this stage when one of its checks asks for a person’s approval. Items that publish something always end here.

A review is titled with its question (“Pick the welcome screen”). Each output shows large, with the worker’s one-line pitch under it.

Control What it does
Pick this (under each output) Makes that output your pick. It then reads Your pick; Undo takes it back.
Not this (under each output) Sets one output aside. It is greyed and never passed on. Undo brings it back.
The note One box for the whole review. It goes with your pick or with your request (“Bigger mic button on B”).
Approve B Appears once you have picked, named after your pick. Approves the item with that output. The pick becomes final; the other outputs and older versions are set aside.
Request changes Sends the item back with your note. The next round starts from it.
More › Approve the round Approves without one pick. Outputs you set aside stay set aside.
More › Cancel the item Cancels the item, after you confirm.

With one output, there is nothing to pick: the main button simply reads Approve. With more than five outputs, or for a code review with a pull request, the review also has one plain Approve. Open wide shows the outputs side by side.

The review waits in Needs you and on the item’s page. In the chat it shows as one line that opens it. It always matches the item’s current outputs: when a new version arrives, the review updates.

  • Approve: the item is done. Lea hears about it in her next turn and plans what it unlocks. An item that this one informs gets only your pick as its input file.
  • Request changes: a new worker starts on the item with your note. It posts a new version, and the next review shows the old and the new side by side.

Your rounds are free. Rounds you send back never count toward the item’s review limit. Rounds an agent sends back do count; when they run out, the item goes back to its owner agent as blocked, and the agent rewrites, splits or asks you.

Your picks stand. Picks and Not this are yours: no agent can change an output you approved or set aside. An agent can only post a new version.

Workers in the repo preset open a draft pull request from their item’s branch when their checks pass, and attach it to the item. They are instructed that a person reviews and merges it.

Planned Merge as the review button: runners watch open pull requests with your gh, and the review’s main button reads Merge and merges the reviewed commit, so “done” means merged. Approve without merging moves under More.

Not every output has a review stage, for example a research report. On the item page, each output has Comment. A comment wakes the item’s owner; on a finished item Lea usually creates a follow-up item that gets the report as input.

Claude Code is a trademark of Anthropic; Codex of OpenAI. Minga is not affiliated.

PrivacyImprintTerms