Community App Operator

Small, local utilities for Codex-assisted review of the Pollinations community app catalog.

Codex performs visual inspection with the in-app Browser. This repository does not install or control Chromium, manage browser profiles, authenticate to apps, or call a visual model. It keeps only the deterministic parts of the workflow: catalog validation, target queues, an auditable decision ledger, isolated data outputs, and draft PR creation.

Review a catalog change

Fetch or otherwise save the base and candidate catalogs locally, then prepare a review run:

npm run review -- \
  --base=workspace/base.catalog.json \
  --candidate=workspace/candidate.catalog.json \
  --limit=100

When --candidate is supplied, the queue contains only rows removed by the candidate. Duplicate rows that resolve to the same target are grouped into one review. Without --candidate, the command reviews the base catalog itself.

The command creates:

runs/<timestamp>/
  base.catalog.json
  candidate.catalog.json
  report.json
  review.json

Ask Codex to inspect each report.json target with the in-app Browser. Record a decision without approving it. Use the browser review policy for the verdict criteria:

npm run record -- runs/<timestamp>/review.json \
  --id=<target-id> \
  --outcome=keep \
  --reason="The live product matches the catalog entry." \
  --evidence="HTTP 200 and visible matching product UI"

Valid outcomes are keep, remove, and retry. Add --apply=true only after a human approves that exact decision. Review evidence is ordinary text or a safe URL; never store browser credentials or session state in a run.

Metadata corrections have their own approval ledger:

npm run record -- runs/<timestamp>/review.json \
  --kind=metadata \
  --id=<metadata-id> \
  --outcome=approve \
  --reason="The canonical product name is visible on the live site." \
  --apply=true

Use --outcome=reject without --apply=true when the proposed correction is not supported. Unapproved metadata never appears in the split catalog.

Split the benchmark into three changes

After review, generate independent catalog artifacts:

npm run apply -- runs/<timestamp>/report.json

This writes:

runs/<timestamp>/split/
  metadata.catalog.json
  removals.catalog.json
  screenshots.catalog.json
  manifest.json

• metadata.catalog.json applies only non-screenshot field corrections.
• removals.catalog.json deletes only approved remove decisions.
• screenshots.catalog.json applies only screenshotUrl changes.

The original benchmark PR and its candidate catalog remain unchanged.

Create draft PRs

Each output can be proposed separately:

npm run propose -- runs/<timestamp>/split/manifest.json --kind=metadata
npm run propose -- runs/<timestamp>/split/manifest.json --kind=removals
npm run propose -- runs/<timestamp>/split/manifest.json --kind=screenshots

propose clones pollinations/pollinations into a temporary directory, regenerates and validates catalog outputs, pushes one data-only branch, and opens one draft PR. It never changes or closes the benchmark PR.