Capture procedure for the product site

The site’s figures must be real captures of the real surfaces, taken over a temporary fixture repository driven by the simulated agent — never over a real project and never under a real agent account (sitio-web-producto design D5). No image may leak a personal path, an identity, a third-party product name or any credential material.

The captures are made here and written to docs/media/. The product site has lived in meltemi-web since 2026-10-21 and copies them from there (release-con-el-push D8).

Every figure carries two things the site’s lint, in meltemi-web, enforces: descriptive alternative text and its provenance (surface, capture source and product state). What is inside the pixels is the human’s judgement; the lint cannot inspect a raster.

Terminal surface — automated, reproducible

The terminal capture is generated from the renderer itself, so it can never drift into a mockup:

cargo run -p meltemi --example capture_svg > docs/media/tui-sessions.svg

The example builds a fixture session set (two generic project roots, the simulated agent, two launch profiles) and renders it through the same code path the shell uses. Regenerate it whenever the sessions view changes; the diff is reviewable because the output is text.

Desktop surface — scripted, run by a maintainer

The desktop capture needs a running window, so it is a maintainer action rather than a CI step. It is still scripted, so the provenance in the caption is a fact anybody can re-derive:

cargo build --release -p mock-agent -p meltemi -p meltemid
cd desktop && ui/node_modules/.bin/tauri build --no-bundle && cd ..
pwsh -NoProfile -File scripts/capture-desktop.ps1

The script builds a throwaway fixture repository with two launch profiles over the simulated agent, dispatches four tasks and one proposal so the views have content, seeds the window geometry for the machine’s display scale, captures the window with PrintWindow/PW_RENDERFULLCONTENT (a WebView2 surface comes out blank without it), downscales to 1600 px wide and writes docs/media/desktop-sessions.png. Its endpoint, data directory and config directory are its own: it never touches the maintainer’s daemon, projects or agents. No real agent, no network.

Only the platform differs — it is PowerShell because the surface it captures is a Windows window, and Windows is a first-class platform here. On macOS and Linux the equivalent is a manual capture of the same fixture; the caption must then name that platform.

Before committing the image, check it: no user name in a path, no real project name, no third-party product name, no e-mail, no token. Then commit it under docs/media/; meltemi-web copies it into both language trees of the site, with alt text and a caption declaring surface, fixture origin, product version and platform, and its lint requires both.

Refreshing

Captures age against the interface. Refresh them as part of a release when the surface they show has changed; a caption that names a version the product left behind is worse than no caption.