PR 330 review, round 4

Dragged to 250 to about 265px (the panel minimum is 250), Cancel and Continue read "Canc/el" and "Continu/e". On main they stay whole. word-break: break-word lets the buttons shrink below a word, and this row doesn't wrap. With overflow-wrap: break-word instead (tried in the browser), the words stay whole and the recommendation buttons still wrap like in torchiaf's screenshot. list-style-position: inside in the same block does nothing on a span.

The temporary rule only sets the text colour. The hover background still comes from the shell's link style (--accent-btn, blue), where main had the green tertiary hover. Either the rule also sets the hover background, or it goes into shell issue 19809.

Since d7cc2d3 it is one button that opens and closes the extra info, so the checklist's state-exposure rule applies. The label already switches, so it is small. Good part: focus now stays on the button (main lost it on every click) and long names wrap (main overflowed by 118px).

With link, .cancel-button { background: transparent; } no longer does anything: the background measures the same as the other four Cancels at rest and on hover. The class itself is still used by the unit tests, so only the rule goes.

• Settings > UI tools still says "Show YAML" and "Console Logs", while the chat now says "View YAML" and "View Console Logs".
• The shell pin is 3.0.15-rc.4. Fine for 1.2.0-dev; is there a follow-up to the final 3.0.15 before 1.2.0?

Left out on purpose: the ShowYaml / ShowYamlDiff guards that can never trigger; the temporary rule being copied into four files (one place would need :deep or global CSS, both against the checklist); no unit test for the toggle (Cypress covers it, gap from PR 338). The bump adds no deprecation warnings (small already warns on 3.0.14), and the MCP Cancel moving off the deprecated tertiary prop removes one.

"It should not scroll when interacting with actionable items (expand/collapse of planning messages)" failed 3 of 3 at 5d64721: after collapsing the plan, the last message was visible. It passed this morning at d7cc2d3, and the two commits since are colour and label wrapping only; the plan in that test has no buttons or extra info. An exact Playwright replay at the Cypress viewport (1000x660) passed 10 of 10 on the PR package and 10 of 10 on main. Likely a race in useScrollComposable: scrollToBottom unhooks the scroll listener until the next frame and then forces "at bottom", so a scroll in that frame is lost. Two other tests failed once and passed on retry. Needs a re-run of the failed job; I haven't triggered it.