Repository navigation
fix(dashboard): read each agent's hidden status as its own sentence - #33206
Conversation
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Nothing blocks: a small, contained fix to screen-reader-only text, with accessibility-tree evidence for a change that moves no pixel. Notes
[DESIGN-REVIEWED] ca22217 Description read[DESCRIPTION-READ] bf056b8cb7785765012f3e418dee93a3b8e631f0f5a8256cbf14f7be03539992 That sha256 names everything this verdict read: the PR description, plus the 1 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
UX Review (Opus 5.5) — ✅ PASSUX-level review of UX-Verdict: PASS Nothing blocks: the diff changes only screen-reader text, so no pixel moves and no screenshot could show it. The PR's accessibility-tree transcript covers each new sentence. [UX-REVIEWED] ca22217 Description read[DESCRIPTION-READ] a7a2d732f9154c112029d0088836a9c4463d592cf52032d77dce993bc418f412 That sha256 names everything this verdict read: the PR description, plus the 3 evidence file(s) the workflow collected from it and handed this review. It covers that evidence because the media strip replaces every attachment URL with the same placeholder, so an edit swapping one attachment for another would leave the description's captured text identical while what this review judged changed. Recompute it with |
First Principles Review (Opus 5.5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Nothing blocks. This fixes a reported defect (#32067), the PR adds a test that fails on base, and each new key and field is used by the hidden block. What this change shipsInventory (4 items) — 4 justifiedIntent: let screen-reader users hear which agents are ready, one clear sentence per agent. This is a FIX.
[FIRST-PRINCIPLES-REVIEWED] ca22217 Description read[DESCRIPTION-READ] 1829f88d4b44b9c137882a44046c1487629b82f1f14ea0ae9e6b9d1b0cab7918 That sha256 names the PR description this verdict read. Recompute it with |
Opus 5.5 Review — ✅ no blocking findingsReviewed Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
GPT 6.1 Review — ✅ no blocking findingsGPT 6.1 completed its review of This comment is updated in place on each push. Review detailsNo findings. False positive or not applicable? A repository writer can comment: |
0268383 to
f8e19bb
Compare
|
Rebased onto main to pick up the fix for main's red |
The screen-reader-only status block under Settings > Agent Harness put one inline span per agent side by side, so browse mode and the accessibility tree read every row's status as one run-on line that named no agent. The in-use row also joined two catalog strings with an English space. Each agent now gets its own paragraph, worded by one catalog message that leads with the agent's name, and the in-use row by one message of its own, translated in every catalog. The span each row points at with aria-describedby keeps the status alone, so a focused row does not read its name twice. Closes #32067
f8e19bb to
ca22217
Compare
|
Rebased: main moved in website/src/pages/developer/AgentBackendTab.tsx and the 13 i18n catalogs plus website/src/i18n/en.context.json since this head's CI ran (6d1e817 -> 39aecc7); GitHub reported the PR as conflicting. The only conflict was the icon import line in AgentBackendTab.tsx, where main added [operator: bolichen97] |
buluoray
left a comment
There was a problem hiding this comment.
Approve on ca22217. Screen-reader-only Agent Harness status block now renders one
per agent led by its name (unless the sentence already names it) and words the in-use line with one catalog message; visible UI and row aria-describedby text unchanged apart from the CJK space; tests green, both guards mutation-verified. Decisive lines: website/src/pages/developer/AgentBackendTab.tsx:1134, website/src/pages/developer/AgentBackendTab.tsx:1772, website/src/pages/developer/AgentBackendTab.tsx:1776-1777, website/src/test/AgentBackendTab.test.tsx:419
KiroCrewCaptain
left a comment
There was a problem hiding this comment.
Approving. On head ca22217 the review traced every branch of status() and the hidden per-agent block, which is its only caller. It checked that the aria-describedby span still holds the status alone, and that both new keys exist in all 13 locale catalogs with their placeholders kept. It confirmed that the new test fails on the base, through the extra ASCII space after "使用中。" and the span children. It also checked the test against the determinism and side-effect rules.
Done well: the separator is now a translatable <name/>: <status/> message, and the name stays outside the described span. Browse mode reads one sentence per agent, and focus announcements do not change.
Optional:
- #34147 adds a
sandbox_refusalbranch to the samestatus(), and its sentences already name the agent. Whichever PR lands second should return{ text: sandboxRefusalLine(value), named: true }. Without that flag the code still compiles, but the hidden block reads the name twice. - Only the ready-to-use
named: trueflag has a test. A wrong-version row and a restart-required row in the new test's fixture would make deleting either of the other two flags fail a test.
Problem / Motivation
Goal: A screen reader reads the Agent Harness list's hidden status text as one separate sentence per agent, each naming that agent.
In Settings > Agent Harness, the screen-reader-only status text under the list reads as one run-on line that names no agent, for example "In use. Default.Missing on this machine: kiro-cliMissing on this machine: strands". Browse mode and the accessibility tree both read it that way, in every language. The in-use line also joins two catalog strings with an English space, so Japanese and Chinese get an ASCII space between full-width sentences.
Why it matters
A screen-reader user choosing an agent on first run hears one garbled line and cannot tell which agents are ready. Automated accessibility checks read the same text.
Not a goal
What changed (motivation → approach → change)
A screen reader now reads the Agent Harness list's hidden status as one sentence per agent, each naming that agent: "Kiro CLI: Missing on this machine: kiro-cli", then "Claude Code: In use. Missing on this machine: claude-agent-acp".
The block used to put one inline
<span>per agent side by side with nothing between them, so their text joined, and the in-use row glued "In use." to its status in code. NowAgentBackendTab.tsxgives each agent a<p>worded by one catalog message,row_status_sentence(<name/>: <status/>), and words the in-use row with one message,in_use_status. A status that already names its agent ("Kiro CLI is installed and ready to use.") is read as it is, so the name is never said twice. Both keys are translated in all twelve catalogs, with full-width colons in Japanese and Chinese, anden.context.jsontells translators what they hold.The name sits outside the
<span>each row points at witharia-describedby, so a focused row still reads its name once and then its status.Backwards compatibility
Compatible: only screen-reader-only text changes; element ids, existing catalog keys and row descriptions stay, except that the in-use row's description loses the ASCII space Japanese and Chinese read after "In use.".
Removes nothing: no key, id, element or behaviour is removed; the hidden block gains a paragraph and a name per row.
Tests
website/src/test/AgentBackendTab.test.tsx—reads the hidden status block as one paragraph per agent, led by its name: in Japanese, the hidden block is one<p>per agent; each is led by its name unless its sentence already names the agent (Kiro CLI installed), the in-use line is one message with no ASCII space, and the in-use row's description is still the status alone. It fails on the base (inline spans, no names, "使用中。 " with a space) and passes with the fix.AgentBackendTab*.test.tsxfiles passed in a shuffled order (--sequence.shuffle --sequence.seed=15862, 130 of 130).Proof: not provable -- frontend test (prove.py runs Python only)
Manual verification
Built the SPA from this head, served it with fixture API answers, opened Settings > Agent Harness in headless Chromium, and read the block's ARIA snapshot plus each radio's computed description over the DevTools accessibility tree. Scene A is the reporter's machine (Kiro Agent in use, the other three missing a binary); scene B has Kiro CLI installed. German and Chinese read the same way as English and Japanese.
Screenshots / video
Why no screenshot: the change is to visually hidden screen-reader text, so no pixel on the page moves; the accessibility-tree text is shown under Manual verification.
Related Issues
Closes #32067
Sibling sweep
Swept: git grep -n -A3 'className="sr-only"' -- website/src, then each hit that maps rows into the block
Sibling: website/src/pages/settings/DecisionsCard.tsx:928 -- tracked in #33065
Sibling: website/src/cards/CardChanges.tsx:412 -- not affected -- one hidden word inside a visible line, with its own trailing space
Pattern harvest
Rule candidate: review-prompt
Pattern: a screen-reader-only container that maps several inline spans with no block element or separator between them, so browse mode reads them as one run-on line.
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)