Historical record

This page records what happened on a date, with its observed failures and limits. It is not a requirement for the current implementation. Current behaviour is in Status.

Layout checks in a real browser

Updated 2026-11-04.

Question

Eight untracked probe scripts sat in the checkout root. Each opened fixed local pages in Chrome and printed what it found: overflow, overlapping boxes, page errors, screenshots. Each was a one-off, with hard-coded paths and one Chrome location. The question was whether those checks could be a capability anyone can run on any page: a developer, or an Agent that has the Rho Skill installed.

Setup

• rho-screen-check <url or page.html> (scripts/screen-check.mjs) opens the page at 1440×900, 1024×768, 768×1024 and 390×844, waits 800 ms, writes one screenshot per width and report.json, and prints the same report as JSON. The measurements are pageFindings in scripts/browser-checks.mjs.
• The browser is Google Chrome through playwright-core 1.64.1, then Microsoft Edge, then Playwright's Chromium, unless --browser names one.
• The Skill rho.screen-check serves the practice text (src/screen-check-skill.ts). skills/rho/references/screen-check.md carries the same text for an Agent that has no connection.
• The project was a copy of the shipped example/pa14-qs, opened with project.open in a scratch service with its own state folder. The Open folder button was not used: on this machine it opens a native dialog.
• The checks ran from a verification copy built from the checkout's tracked and new files, not from the checkout's own dist/.

The eight probes

• findover.mjs and overlap.mjs found overlapping boxes at one width. That is --overlap SELECTOR.
• shot2.mjs, shot3.mjs and shot-artifact.mjs took 1440 px screenshots of named files, and shot3 also collected page errors. That is one screenshot per width, and pageErrors in the report.
• narrow.mjs took 390 px screenshots at 2× scale. The tool captures at 1×.
• shot3all.mjs and final4.mjs checked fixed design pages. Their pages are passed to the tool one at a time. final4 is a subset of shot3all.

Several probes blocked outside requests before loading the page. The tool records those requests and blocks nothing, so the page loads as it does for a person.

Observed

The home page (/) at four widths: no horizontal overflow, no elements past the edge, no unnamed buttons, no page or console errors, and no outside requests. The 390 and 1440 screenshots show the project list and its actions whole.

The workspace (/#project=<id>) was run twice: first with the copy named pa14-qs, then with the install name PA14 quorum sensing. Both runs gave the same numbers.

Width	Horizontal overflow	Past the edge	Not kept in by a box	Failing responses
1440 × 900	no	89	0	1
1024 × 768	no	291	0	1
768 × 1024	no	325	0	1
390 × 844	no	317	0	1

• The past-edge elements are span.tl-band, span.tl-fold and i.graph-tick.tl-block. Each sits inside the scrolling main pane, so none is unkept. The totals count boxes, not defects.
• The failing response is POST /api/v2/agent.activity.capture. It answers 404 not_found with "This file is not in the project any more." The request path is PA14 quorum sensing/analysis/summary_figure.R. The file is at analysis/summary_figure.R inside the project.
• The sidebar study title "Co-culture rescue: does mixing Δrhll and Δlasl restore swarming, and at which ra" is cut mid-word at 1440, 1024 and 768 px. No box reports it. Only the screenshot shows it.

The 404

The agent plugin is not the cause. Its two functions behave as their contracts say, checked directly on the built plugin:

• The observer turns an absolute path, the form a live run stores, into analysis/summary_figure.R with state observed. The prefixed form from the example comes out as PA14 quorum sensing/analysis/summary_figure.R with state missing.
• Capture refuses the prefixed path with not_found, "This file is not in the project any more." It refuses the existing .R file with unavailable, because only CSV, PNG, JPEG and SVG can be captured.

The prefix is in the example's recorded data. All 12 distinct observed location paths in example/pa14-qs/.rho/rho/conversations carry it, while the turn's evidence records hold analysis/summary_figure.R. The manifest says its curation "removed … the folder and host the build ran in", which fits. A search for the manifest's modelCatalogues field finds only example.json, so the curation script is not in the repository. The folder name is not the cause: the run under the install name gave the same 404.

The request is the preview's. It has no expectedDigest, and an R script is not capturable, so it can only come from web/previews.ts:32. That line asks the server to capture a node's file whenever capture is available and the node has no resource, with no extension check; web/canvas.ts:54 checks capturable first. On a live run the R script would still be requested and refused as unavailable. That follows from the code and the direct check; I have not seen it in a live browser.

Both were fixed in Rho, not in the agent plugin. The preview asks for a capture only when capturable names the file, the same test canvas.ts already applies. The server accepts exactly those kinds, so no request that could succeed is removed. The example's 256 bare project-file values lose the projects-folder prefix (PA14 quorum sensing/analysis/x.R becomes analysis/x.R). The 38 occurrences inside command text stay as the agent ran them. A leaf-by-leaf comparison with the previous commit shows that only those 256 values changed.

Changed

• pageFindings counts elements past the edge that no box keeps in (elementsPastEdgeUnclipped) and lists them first.
• The report lists each failing response with its status and address (httpErrors).
• The Skill names both fields and the mid-word blind spot.
• test:screen-check builds first, since its Skill test imports dist/.
• playwright-core is a runtime dependency and is listed in bundleDependencies, so an offline install still works. The person accepted its size, about 13 MB unpacked.
• web/previews.ts: a file that no capture can open gets the note "Preview could not be read" without a request, and it no longer uses one of the preview slots.
• example/pa14-qs: the three recorded conversations name their project files relatively, in the form the turns' evidence records already use.

Checks

• npm run test:screen-check in the verification copy: 20 of 20 pass, none skipped, exit 0. Five are Skill tests: the reference matches the served text and digest, every option the Skill names appears in --help, every failure kind it names is emitted, SETUP and SKILL route to the reference, and the service lists the Skill with the same text. Fifteen cover the findings and the command; the ones that open pages use Chrome.
• node --test test/skills.test.mjs in the copy: 3 of 3, with Rho-plugins linked beside it. Without that link, the verify-script paths do not resolve and one check fails. That comes from the copy's layout, not from the change.
• npm pack --ignore-scripts in the copy: 1,439 files, 12.3 MB packed and 36.9 MB unpacked. playwright-core adds 115 files, about 13 MB unpacked.
• The tarball installed offline (npm install -g --offline) into a scratch prefix. Its bins are rho, rho-screen-check and rho-ui. The installed rho-screen-check found the overflow on a fixture at 390 px and exited 0, and reported nothing on the live home page at four widths.
• After the fixes, with a fresh service on the fixed example: the workspace load made 56 API responses, all 200, and no capture call. The four-width check found no HTTP error, page error, console error or outside request, and nothing unkept. Past-edge counts moved to 84, 280, 315 and 307, because the observed paths now resolve; all sit inside scrolling boxes. A direct capture of results/summary.png returns a media resource, and results/summary.md is still refused as unavailable.
• The observer over the fixed records: 98 references, 12 distinct files, all observed. test/example-project.test.mjs passes 4 of 4. The copy's build, including the web type check, passes.

Limits

• Only macOS Google Chrome was used. Edge, Playwright's Chromium and a path given to --browser are covered only by the missing-browser test.
• Screenshots are at scale 1. The tool has no 2× option.
• An Agent trial that runs the tool through Bash has not been done. Auto mode blocks Bash without authorization.
• The release route has not run: pack-local bundles every dependency, so playwright-core also enters the release candidate, and the candidate's size and the Windows install time are unmeasured.
• Installations that copied the example before this keep their copy, which is never overwritten. The preview stops asking for those files, but their observer still reports the prefixed paths missing.
• The mid-word sidebar cut is not detected by the tool.
• Nothing is committed or pushed. The eight probe scripts are unchanged and still untracked.