feat(desktop): check layout and accessibility in done and preview by wudilyy999 · Pull Request #495 · OpenCoworkAI/open-codesign

wudilyy999 · 2026-10-16T04:14:47Z

Summary

Refs #493 (the first part of the plan, PRs 1 to 4, in one PR). done only saw syntax and runtime errors, so a page that loaded cleanly but scrolled sideways, hid part of its text, set grey text on white, or had an icon button with no name got "ok" and an instruction to stop. done and preview now measure the open page and report what they find as advice. Findings never change ok or the done status, so a false positive cannot cost a repair round.

#493 planned this as four PRs. I put them in one so that the contract, the checks and the tests of the whole loop land together, and kept the commits in the planned order. If you would rather review them separately, they split at these commits: c2c8019 (the contract; behavior unchanged), 993fdab (probe, overflow and clipped text), 31bcfd7 to cf9a565 (contrast, tap targets, names), aeffd6f and b098de8 (the scripted-agent tests and the fake browser they use).

What is measured

At the page's own width (1280 px for done, the viewport option for preview) and at 390 px, after which the viewport is put back. No library: plain DOM, getComputedStyle, getBoundingClientRect and a 1 × 1 canvas to read colours in spaces other than sRGB.

Check	Reported when	Not reported
horizontal overflow	the page scrolls sideways by more than 1 px; it names the outermost elements that cross the edge (at most 3, widest first)	the root or body hides horizontal overflow; content inside a container that scrolls on its own; `position: fixed`
clipped text	a clipping element has a text node that sticks out of its box by more than 2 px on any side; the message names the side that sticks out most (at most 5 per width)	`text-overflow: ellipsis`, a line clamp, a box with no height or width, a running CSS animation
contrast (current width)	text colour over the backgrounds behind it is below 4.5:1, or 3:1 for text of 24 px, or 18.66 px and bold (WCAG 2.x AA). Reported once per colour pair with the number of elements, worst first	text on a background image or gradient, or partly transparent: counted and reported, not guessed; hidden text
tap targets (phone width)	a link, button or input is under 24 × 24 px and another target is within the 24 px square around it (WCAG 2.2 SC 2.5.8, with squares instead of circles)	a lone small target; a link inline in a sentence; disabled, hidden, transparent, `aria-hidden` and visually hidden controls
control names (current width)	a control has no text, `aria-label`, `aria-labelledby`, `<label>`, `title`, image `alt` or SVG `title` (input defaults count; a placeholder does not)	`aria-hidden` and hidden controls

Reporting

preview returns the findings as layoutIssues (at most 20, omitted when empty; ok is unchanged) and counts them in its summary line. done lists them under "Non-blocking warnings". When one is present, the stop instruction changes: the agent may fix it and check it with preview, and is told not to call done again, so advice cannot start a loop.
DoneRuntimeVerifier returns { errors, warnings }. done:verify:v1 keeps errors and gains warnings and a schemaVersion.
Budgets are named constants: preview keeps at most 5 warnings; 20 findings in all, taken from each check in turn when there are more; 5,000 elements and 400 controls per measurement; 5 seconds for both measurements, plus a separate 1-second budget to restore the original viewport before returning after success, failure, timeout or cancellation. Past a limit, or on a page with a background image, the result says so (in warnings for preview, as a layout warning for done). Findings past a per-check limit are counted and reported.
A note about the measurement itself (a truncated page, colours that could not be measured, checks that did not run) is a layout warning too, so done uses the advice wording for it. That wording is conditional ("if you agree one is a real defect") and a note names no defect, so the agent writes its summary.
A measurement that fails is reported as "Layout checks did not run: …" and never read as a clean page or turned into a failed run. A cancelled run still stops.

Not in this PR

The DESIGN.md drift check (it needs the colour parser in #490) and the Electron smoke test (the questions in #493 about the driver, the model and CI).

Size

+3,354 / −92 over 30 files: 1,121 lines of code in 12 files and 2,227 lines of tests. The plan in #493 estimated about 1,950 lines across the four PRs. The difference is mostly browser tests: a fixture for each rule and each exemption in system Chrome, and, for each exemption, a page that fails without it.

Type of change

Linked issue

Refs #493

Checklist

I checked the linked issue / relevant context before starting
pnpm lint && pnpm typecheck && pnpm test passes locally (see Checks)
Added/updated tests for the change
Added a changeset (pnpm changeset) if user-visible
Updated docs if behavior changed

No page in website/ describes done or preview checks, so there is nothing to update.

Checks

pnpm lint, pnpm typecheck (10 packages) and pnpm test pass on this branch at 6c799fd: core 608, desktop 2,802 (190 files), shared 350, exporters 198, and the other packages. The first rebase onto main needed one change: the done repair-limit test that feat(core): stop a run that keeps repeating the same tool call #489 added returned the old verifier shape, so it now returns { errors, warnings } (in the first commit, c2c8019). Merging main later brought fix: tolerate compatible DESIGN.md metadata during verification #496, which added a done test beside mine in done.test.ts; I kept both and changed its verifier to return { errors, warnings }. The PPTX tests in packages/exporters unpack under /tmp, a symlink on macOS, where zip-lib refuses to write; turbo's strict env mode drops TMPDIR, so my Mac falls back to /tmp. With TURBO_ENV_MODE=loose TMPDIR=/private/tmp all of them pass. I did not touch that package.
The browser tests run in system Chrome and skip when it is missing, like preview-interactions.browser.test.ts. agent.design-loop.test.ts runs the production agent with only the model stream and the browser executors replaced.
6c799fd is Sun-sunshine06's commit on this branch: each measurement step is bounded by the shared deadline, and the viewport is restored before the inspection ends, with layout-probe.lifecycle.test.ts (four paths). I did not mutation-check that change.
Mutation checks: I changed the code in about 55 places (each exemption and threshold, the viewport restore, the budgets and counts, each caller's failure and cancellation handling, the warning merge, the stop-instruction choice, the preview result and summary, the IPC fields) and each change made a test fail, except one: when the browser shuts down late, the recovered-shutdown warning is now added to the list instead of replacing it, and no test covers that (the path had none before). Several first runs survived because my fixtures never made the page scroll sideways, or put no target next to a small one; they now do. After the first review I changed the clipped-side code and the warning cap in 12 more places (the sides recorded, the choice of the side that sticks out most, both sides of the 2 px tolerance, the edges, each message, the cap); each fails a test, and the one that first survived, the tolerance at exactly 2 px, now has one.

Principles

Compatibility: done still decides ok / has_errors from errors alone, and preview.ok is unchanged. The wire changes are additive (warnings and schemaVersion on done:verify:v1, layoutIssues), and the heading "Non-blocking design metadata warnings" is now "Non-blocking warnings".
Upgradeability: the measurement and the decisions are separate (layout-probe.ts collects, layout-analysis.ts decides on plain data); every threshold and budget is a named constant passed to the page function.
No bloat: no new dependencies, nothing loaded at start (the probe is imported by the two modules that already launch Chrome).
Elegance: one probe serves done and preview, so the two reports cannot disagree.

DoneRuntimeVerifier now returns { errors, warnings } instead of a bare error list. Warnings never change the done status, so a false positive cannot start a repair round; they are listed under "Non-blocking warnings" with the existing metadata warnings. When a layout warning is present the stop instruction changes: the agent may fix one it agrees with and check it with preview, but is told not to call done again. The preview result gains an optional layoutIssues list, capped at 20 and never part of ok. done:verify:v1 keeps its errors field and gains warnings and a schemaVersion. The desktop verifier reports no warnings yet.

After a page loads, preview and done now measure it at its current width and again at a 390 px phone width, then put the viewport back. They look for a page that scrolls sideways, naming the outermost elements that cross the edge, and for text that its box clips. A page that hides horizontal overflow, content inside a scroller, fixed elements, ellipsis and line clamps, collapsed panels and running animations are not reported. The page-side function only collects measurements; the decisions are made on plain data, so the thresholds and the budgets (20 findings, 5,000 elements, 3 elements and 5 clipped texts per width) are tested without a browser. preview returns the findings as layoutIssues and done lists them as warnings. A failed measurement is reported as a warning, a cancelled run still stops.

The layout probe is split into the shared samples and limits, the page function, and the analysis, which works on plain data. Three checks join overflow and clipped text: - contrast: the text colour over the backgrounds behind it, WCAG 2.x AA (4.5:1, and 3:1 for large text), grouped by colour pair and listed worst first. Text on an image or gradient, or partly transparent, is counted and reported, not guessed. Colours are read through a canvas so oklch() and color-mix() come out as sRGB. - tap targets: a link, button or input under 24 by 24 px, at phone width, unless it is inline in text or has no other target within that square (WCAG 2.2 SC 2.5.8). - names: a control with no text, aria-label, aria-labelledby, label, title, image alt or SVG title. Contrast and names run at the current width, tap targets at phone width. When there are more findings than the budget, one is taken from each check in turn; findings past a per-check limit are counted and the result says so. LayoutIssue gains the three checks, and its measured and limit fields become optional, since a missing name has no number.

A fixture for each rule and each exemption, in system Chrome: grey on white just under 4.5:1 and the 3:1 rule for large text, backgrounds from ancestors and translucent layers, oklch() and color-mix(), pairs reported once with a count, and text on a gradient or partly transparent counted instead of guessed. Small targets pass alone and fail beside another target, inline links are exempt, disabled and visually hidden controls are skipped. Names come from text, aria-label, aria-labelledby, label, title, image alt, SVG title and input defaults; a placeholder is not a name. Alpha in rgb() and rgba() is read exactly: a canvas rounds it to 8 bits and moved a ratio by 0.03.

A small target passes alone, so a test for an exemption has to put a target beside it; otherwise the check would pass without the exemption. Disabled, hidden, transparent, aria-hidden and visually hidden controls and a link inside a sentence now each sit next to a neighbour, and removing the exemption fails the test. The colour reader only handles the comma form of rgb() and rgba() that Chrome writes for computed colours; every other space goes through the canvas.

…ted model Eight scenarios on the real pi Agent, with only the model stream and the browser executors replaced: write, preview, hear a runtime error from done, fix it and finish; a layout warning that does not fail done and changes what the agent is told; fixing the warning and checking it with preview, not done; preview listing layout issues while staying ok; warnings beside runtime errors; five done calls with warnings that never spend a repair round; three error rounds that still end the run with the workspace kept; and a layout measurement that did not run.

The harness document test replaces puppeteer and answered the layout function with an early, partial shape. The analysis now reads every field, so it failed and the verifier reported "Layout checks did not run". It answers with the shared fixture instead.

github-actions

Review mode: initial

Findings

[Minor] Text clipped on the left or top is reported as a bottom clip with self-contradictory numbers — collectLayoutSamples flags a clip when text.left < box.left - tolerance or text.top < box.top - tolerance (apps/desktop/src/main/layout-probe.ts, the checks.clipped block), but the ClippedText sample only carries textRight/boxRight/textBottom/boxBottom (apps/desktop/src/main/layout-samples.ts), and clippedIssues picks the vertical branch whenever text.textRight <= text.boxRight (apps/desktop/src/main/layout-analysis.ts, clippedIssues()). A left/top-only overflow (negative text-indent, RTL, a negative inline margin) therefore emits "…the text reaches N from the top but the box ends at M" where N < M, so the warning contradicts itself and points the agent at the wrong side. Impact is limited because findings are advisory and never change ok/done, but the mismatch is real and untested (no left/top case exists in layout-analysis.test.ts or the browser tests).
Suggested fix: record the overflowing side (or textLeft/textTop) in ClippedText and describe that side, or drop the text.left/text.top terms from the probe if left/top clipping is intentionally out of scope.
[Nit] preview can exceed its declared warning budget — PreviewResult.warnings is declared Type.Array(Type.String(), { maxItems: 5 }) (packages/core/src/tools/preview.ts:163) and trimPreviewResult slices consoleErrors, assetErrors and layoutIssues but not warnings. runPreview sets result.warnings from inspection.notes (up to 5: element/control truncation, unmeasured contrast, dropped colour groups, omitted findings, over-budget) and the recovered-shutdown branch now appends one more instead of replacing (apps/desktop/src/main/preview-runtime.ts, report.warnings = [...(report.warnings ?? []), message]). A page that both truncates the probe and fails graceful shutdown therefore yields 6 warnings. Nothing currently runtime-validates maxItems, so this is contract drift rather than a hard failure.
Suggested fix: slice warnings in trimPreviewResult the same way as the other arrays.

Questions

None blocking.

Summary

Review mode: initial
The change is internally consistent with the stated contract: layout findings are advisory, done derives status from errors alone, and preview.ok is untouched (verified in packages/core/src/tools/done.ts and packages/core/src/tools/preview.ts plus the new agent.design-loop.test.ts / layout-wiring.*.test.ts). The done:verify:v1 change is additive (schemaVersion: 1 + warnings) and the preload type is updated in step (apps/desktop/src/main/ipc/generate.ts, apps/desktop/src/preload/index.ts).
Constraints hold: no new dependencies, no direct provider SDK imports, probe is plain DOM/getComputedStyle/canvas, measurement separated from decision (layout-probe.ts vs layout-analysis.ts), and cancellation re-throws rather than being swallowed (signal?.throwIfAborted() in done-verify.ts; the preview path is covered by layout-wiring.failure.browser.test.ts). A changeset is present.
Linked-issue scope: the PR uses Refs #493 and explicitly describes this as part of the plan, so the completion criteria are not claimed and no issue-closure validation is needed.
Residual risk: hasLayoutAdvice also fires for the truncation/partial-measurement notes in done, so a non-defect note (e.g. ">5,000 elements") switches the agent from the STOP instruction to the fix-and-preview wording. That looks deliberate (any layout-sourced warning enters advice mode) but is worth confirming, since the note is not an actionable defect.

Testing

Not run (automation). Coverage is strong: pure colour-maths/selection units (layout-analysis.test.ts, layout-probe.test.ts), system-Chrome fixtures per rule and exemption (layout-probe.browser.test.ts, layout-probe.contrast.browser.test.ts, layout-probe.accessibility.browser.test.ts), end-to-end wiring (layout-wiring.browser.test.ts, layout-wiring.failure.browser.test.ts), and a production-agent design loop (agent.design-loop.test.ts). The one gap is a left/top-only clip case (see the Minor finding).

Open-CoDesign Bot

PreviewResult.warnings declares at most five entries, but trimPreviewResult cut only the console errors, asset errors and layout issues. The layout notes can already fill the list, and a late browser shutdown adds one more, so a result could carry six. It is cut like the other arrays now. Refs OpenCoworkAI#493

The probe flagged text that sticks out of a clipping box on the left or top, but a clipped-text sample held only the right and bottom edges, and the analysis chose the bottom message whenever the text did not reach past the right edge. Text pushed off the left or top by a negative text-indent, a negative margin or a flex alignment was reported as "reaches N from the top but the box ends at M" with N below M, which names the wrong side. A sample now records the side that sticks out the most, with the text's edge and the box's edge on that side, and the message describes that side. The right edge no longer wins by default either: a 1 px overhang on the right used to choose the right-edge message even when the bottom was cut by 30 px. Refs OpenCoworkAI#493

github-actions

Review mode: follow-up after new commits

Findings

No new findings. Both items from the previous review are resolved in the commits since b098de8:
- [Minor → resolved] Left/top clipping is no longer misreported. ClippedText now carries side: ClippedSide plus textEdge/boxEdge (apps/desktop/src/main/layout-samples.ts), the probe picks the worst of the four sides with beyond = text.right - box.right | text.bottom - box.bottom | box.left - text.left | box.top - text.top (apps/desktop/src/main/layout-probe.ts, the checks.clipped block), and clippedEdges() switches on that side so the message and the measured/limit sign agree (apps/desktop/src/main/layout-analysis.ts). Unit coverage now includes left/top (layout-analysis.test.ts) and the browser test drives all four sides plus the 2 px/3 px tolerance boundary (layout-probe.browser.test.ts). The reported numbers are now self-consistent (Math.sign(measured - limit) is asserted per side), so the contradictory "reaches N from the top …" string can no longer be produced.
- [Nit → resolved] preview warning budget is now enforced. MAX_PREVIEW_WARNINGS = 5 is the single source for both the PreviewResult.warnings maxItems and the new slice in trimPreviewResult(), and preview.test.ts asserts the cap.

Questions

None blocking.

Summary

Review mode: follow-up after new commits
No material issues. The change stays advisory-by-construction: done derives status from errors alone (packages/core/src/tools/done.ts) and preview.ok is untouched, so a heuristic finding cannot cost a repair round; the done:verify:v1 change remains additive (schemaVersion: 1 + warnings) with the preload type updated in step (apps/desktop/src/main/ipc/generate.ts, apps/desktop/src/preload/index.ts).
Constraints hold: no new dependencies, no direct provider SDK imports, probe is plain DOM/getComputedStyle/canvas, measurement (layout-probe.ts) is separated from decisions (layout-analysis.ts), and cancellation still re-throws rather than being swallowed. The changeset is present and the changes remain covered by the linked issue as partial work (Refs #493).
Residual observations (not merge blockers):
- The recovered-shutdown warning is now appended rather than replacing the list, but trimPreviewResult() keeps the first 5 warnings, so on a page that already fills the budget with measurement notes the shutdown message is dropped from the tool result. The browser was terminated regardless and the path still emits preview.cleanup.recovered via getLogger('preview'), so this reads as acceptable cleanup degradation, not a silent user-visible failure.
- Carried over from the initial review and unchanged by these commits: hasLayoutAdvice keys off source === LAYOUT_WARNING_SOURCE, so a non-defect note (e.g. the ">5,000 elements" truncation note) also switches done from the STOP wording to the fix-and-preview wording. This looks deliberate (any layout-sourced entry enters advice mode); worth a maintainer confirmation only.

Testing

Not run (automation). Coverage for the touched paths is strong and proportionate: colour maths/selection units (layout-analysis.test.ts, layout-probe.test.ts), system-Chrome fixtures per rule and exemption including the new per-side clip matrix (layout-probe.browser.test.ts, layout-probe.contrast.browser.test.ts, layout-probe.accessibility.browser.test.ts), end-to-end wiring and failure/cancellation (layout-wiring.browser.test.ts, layout-wiring.failure.browser.test.ts), the production-agent design loop (agent.design-loop.test.ts), and the new MAX_PREVIEW_WARNINGS cap assertion in preview.test.ts.

Open-CoDesign Bot

main gained OpenCoworkAI#496, which added a done test next to the runtime warnings tests in done.test.ts. Both blocks are kept. The new test passed a verifier that returned an array, so it returns { errors, warnings } like the others.

github-actions

Review mode: follow-up after new commits

Findings

No new findings. The commits since ea14719 are a merge of main's DESIGN.md compatibility work (.changeset/design-md-generated-output-compatibility.md, packages/shared/src/design-md.ts, apps/desktop/resources/templates/skills/design-system-baton.md) plus the matching test-shape updates; the probe/analysis code (layout-probe.ts, layout-analysis.ts, layout-samples.ts, done-verify.ts, preview-runtime.ts) is unchanged from the last reviewed state, and both previously reported items stay resolved.

Summary

Review mode: follow-up after new commits
The merge was checked for interaction with this PR's contract and is consistent: DESIGN.md extension findings keep source: 'DESIGN.md', layout findings use LAYOUT_WARNING_SOURCE (packages/core/src/tools/done.ts), so hasLayoutAdvice still keys only off source === 'layout'. A DESIGN.md warning alone therefore keeps the STOP wording (covered by the PR's keeps the stop instruction when no layout warning is present), and a page with both lists them together under the single "Non-blocking warnings" heading while status is still derived from errors alone.
The done:verify:v1 wire change stays additive (schemaVersion: 1 plus warnings) with the preload type in step (apps/desktop/src/main/ipc/generate.ts, apps/desktop/src/preload/index.ts), and preview.ok remains untouched by layoutIssues.
Project constraints hold: no new dependencies, no direct provider SDK imports, probe is plain DOM/getComputedStyle/canvas, decisions stay separated from measurement, and cancellation re-throws (signal?.throwIfAborted() in done-verify.ts).
Residual observations (not merge blockers, unchanged from the previous review):
- preview keeps only the first 5 warnings after trimPreviewResult(), so on a page already at the note budget the recovered-shutdown message can be dropped from the tool result; the browser is still terminated and preview.cleanup.recovered is still logged.
- A layout note (e.g. the ">5,000 elements" truncation note or "contrast was not checked") also flips done into the fix-and-preview wording even though it names no defect. Looks deliberate; worth a maintainer confirmation only.
- The layout catch in preview-runtime.ts does not call opts.signal?.throwIfAborted() the way done-verify.ts does, so cancellation during measurement relies on the later signal checks in runPreview to keep the run cancelled. layout-wiring.failure.browser.test.ts asserts ok: false / reason: 'Preview cancelled', so this reads as intentional; just worth keeping in mind if the post-measurement flow changes.

Testing

Not run (automation). Coverage remains proportionate: unit tests for the colour maths, selection and warning mapping (layout-analysis.test.ts, layout-probe.test.ts), system-Chrome fixtures per rule and exemption (layout-probe.browser.test.ts, layout-probe.contrast.browser.test.ts, layout-probe.accessibility.browser.test.ts), end-to-end wiring and failure/cancellation (layout-wiring.browser.test.ts, layout-wiring.failure.browser.test.ts), and the production-agent loop (agent.design-loop.test.ts). The merged DESIGN.md tests exercise the warning path alongside it.

Open-CoDesign Bot

Signed-off-by: Sun-sunshine06 <Sun-sunshine06@users.noreply.github.com>

github-actions

Findings

No new findings in the commits since 76118ead. The only functional change since the last review is in apps/desktop/src/main/layout-probe.ts:471-502: runInspection now bounds each step with a shared LAYOUT_BUDGET_MS deadline (withinBudget) and restores the preview viewport under its own LAYOUT_RESTORE_BUDGET_MS (apps/desktop/src/main/layout-samples.ts:14), instead of wrapping the whole inspection in one boundedPreview with an unbounded finally restore. That ordering is correct: the finally await finishes before the caller resumes, and on timeout or cancellation the inspection still rejects, so done-verify.ts / preview-runtime.ts report the failure as a warning and the run stays cancelled. The new apps/desktop/src/main/layout-probe.lifecycle.test.ts exercises the four paths that matter (phone-measurement timeout, already-cancelled call, desktop overrun, cancellation mid-phone-measurement).
The rest of the compare diff is a merge of main: @types/node 22.20.5, postcss 8.5.30 / nanoid lockfile churn, and the SHA-pinned anchore/sbom-action and ossf/scorecard-action bumps in .github/workflows/. Those are not author changes and are already on the base branch, so they do not need review here.

Summary

Review mode: follow-up after new commits
The bounded-restore change is consistent with the PR's stated contract: findings remain advice (ok and the done status still derive from errors only), a failed measurement still surfaces as Layout checks did not run: … rather than a clean page or a failed run, and no new wire, budget, or dependency changes were introduced.
Previously reported behaviour stays intact: done:verify:v1 is still additive (schemaVersion: 1 + warnings), preview.layoutIssues does not change preview.ok, and LAYOUT_WARNING_SOURCE still drives hasLayoutAdvice in packages/core/src/tools/done.ts independently of DESIGN.md warnings.
Residual observations (non-blocking, unchanged unless noted):
- trimPreviewResult keeps only the first MAX_PREVIEW_WARNINGS (5) warnings (packages/core/src/tools/preview.ts), so the recovered-shutdown message appended in apps/desktop/src/main/preview-runtime.ts can still be dropped when the note budget is already full; the browser is still terminated and preview.cleanup.recovered is still logged.
- A layout note (the >5,000 elements truncation note, “contrast was not checked”, or the per-check-limit note) also flips done into the fix-and-preview wording even though it names no defect. Reads as deliberate.
- The catch around inspectLayout in apps/desktop/src/main/preview-runtime.ts still does not call opts.signal?.throwIfAborted(); cancellation relies on the later signal checks in runPreview, which the failure browser test asserts.
- New, low risk: if page.setViewport(original) itself exceeds LAYOUT_RESTORE_BUDGET_MS, that rejection propagates as the inspection failure and a vision preview still proceeds to the screenshot after the catch. Only reachable with a wedged renderer, and no test covers a restore timeout (the lifecycle tests cover measurement timeouts and cancellation only).
- The PR body still describes the budget as “5 seconds for both measurements”; the effective worst case is now 5 s of measurement plus up to 1 s of restore. Worth a one-line update in the description, not a code change.
Apart from those notes, the PR looks ready to merge.

Testing

Not run (automation). Coverage is proportionate: unit tests for the colour maths, issue selection, warning mapping and lifecycle (layout-analysis.test.ts, layout-probe.test.ts, layout-probe.lifecycle.test.ts), system-Chrome fixtures per rule and exemption (layout-probe.browser.test.ts, layout-probe.contrast.browser.test.ts, layout-probe.accessibility.browser.test.ts), end-to-end wiring and failure/cancellation (layout-wiring.browser.test.ts, layout-wiring.failure.browser.test.ts), and the production-agent loop (agent.design-loop.test.ts).

Open-CoDesign Bot

wudilyy999 added 7 commits October 16, 2026 01:56

github-actions Bot added docs Documentation area:desktop apps/desktop (Electron shell, renderer) area:core packages/core (generation orchestration) labels Oct 16, 2026

github-actions Bot reviewed Oct 16, 2026

View reviewed changes

wudilyy999 added 2 commits October 16, 2026 13:50

github-actions Bot reviewed Oct 16, 2026

View reviewed changes

github-actions Bot reviewed Oct 18, 2026

View reviewed changes

fix(desktop): restore viewport before ending layout inspection

6c799fd

Signed-off-by: Sun-sunshine06 <Sun-sunshine06@users.noreply.github.com>

github-actions Bot reviewed Oct 19, 2026

View reviewed changes

Sun-sunshine06 merged commit b0a9ee1 into OpenCoworkAI:main Oct 20, 2026
6 checks passed

Navigation Menu

Repository navigation

feat(desktop): check layout and accessibility in done and preview - #495

feat(desktop): check layout and accessibility in done and preview#495
Sun-sunshine06 merged 11 commits into
OpenCoworkAI:mainfrom
wudilyy999:feat/layout-checks

wudilyy999 commented Oct 16, 2026 •

edited

Loading

Uh oh!

github-actions Bot left a comment

Uh oh!

github-actions Bot left a comment

Uh oh!

github-actions Bot left a comment

Uh oh!

github-actions Bot left a comment

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Conversation

wudilyy999 commented Oct 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

What is measured

Reporting

Not in this PR

Size

Type of change

Linked issue

Checklist

Checks

Principles

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

github-actions Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

wudilyy999 commented Oct 16, 2026 •

edited

Loading