PyAutoEyes

The Eyes of the PyAuto organism: where the human checks in on what every library's figures look like. Dashboard: dashboard.md, or with a library → dataset → figure browser and copyable review prompts on Pages at https://pyautolabs.github.io/PyAutoEyes/.

Two layers

• Project repos <lib>_visualization (today: autolens_visualization, autogalaxy_visualization, autofit_visualization and autocti_visualization) make, store and track the figures: producers, datasets, the tracked PNGs, GALLERY.md, and a tracked figure manifest. They re-render on every library release.
• This organ reads each project repo's manifest and builds one dashboard across all of them. It links to the PNGs where they live. It renders nothing, copies no figures and never judges them (the Brain's Eyes conductor does). It never edits library plot code, because accepted critiques route through intake → start_dev.

It is the same layering as autolens_profiling / autolens_inference under the Brain board.

Use

bin/pyauto-eyes board            # re-render dashboard.md + dashboard.html + badge.json + state.json
bin/pyauto-eyes check            # registry, manifests, every PNG URL, dashboard current
bin/pyauto-eyes survey lens      # pyauto-brain eyes survey on the lens project repo
bin/pyauto-eyes board --offline --from lens=../../lens/autolens_visualization   # no network

To ask for a figure to be improved, use its Suggest an improvement link on the dashboard. It opens a pre-filled issue on the project repo (title figure: <domain>/<file>, the raw PNG link and a Suggested improvement: stub, label eyes-critique). Or copy its Use the eyes skill. review <instance> <figure> line into an AI assistant session. The dashboard files nothing itself, and an accepted critique becomes a PyAutoMind intake prompt that goes through start_dev.

Expand a library, then a dataset, and choose one figure to inspect. Click the image to enlarge it without leaving the dashboard. Open critiques are tucked into a separate disclosure. Local survey evidence stays in the context markers and state feed rather than adding text to the gallery.

Beside the page on Pages sit badge.json (the one-line headline) and state.json, the organ-cockpit feed: one status, one headline and the rows that ask something of a human, in the shared contract PyAutoBrain/board/state_schema.json (validated on publish).

What is here

Path	What
registry.yaml	one row per instance (lens, galaxy, fit and cti)
REFERENCE.md	the manifest contract (schema 1), the registry fields and the refresh chain
eyes/	registry.py, manifest.py, context.py (survey + critiques), board.py, cli.py
bin/pyauto-eyes	board, check, survey
dashboard.md / dashboard.html / badge.json / state.json	generated, never hand-edited
tests/	hermetic pytest tests and an optional real-browser check
.github/workflows/	lint.yml, pages_dashboard.yml, dashboard_refresh.yml

Requires Python 3.11+ and PyYAML. There is nothing to install: bin/pyauto-eyes puts the repo on sys.path.

AI policy

See AI_POLICY.md.

Shared board presentation

Rendering the HTML board requires a current PyAutoBrain checkout beside this repo, at _brain/, or selected with PYAUTO_BRAIN. The banner, logo, responsive layout and section navigation come from its board/_theme.py; CI and dashboard refresh workflows check it out automatically. Counts and section links remain owned by this board.

Browser validation

The optional interaction check uses Playwright with Chromium and loads real project PNGs. It covers five viewport widths in light/dark, keyboard navigation, selection and critique matching, in-page enlargement, retry and response races. From an environment with Playwright available, run:

node tests/browser.cjs dashboard.html .scratch/browser

NODE_PATH may point to an existing Playwright installation's node_modules. Screenshots and a JSON result summary go to the requested output directory.