Skip to content

Keep narrow Sidekick feedback rows in place when focus settles - #1412

Merged
taras merged 1 commit into
mainfrom
codex/retain-inspection-focus-window
Nov 18, 2026
Merged

taras merged 1 commit into
mainfrom
codex/retain-inspection-focus-window

Conversation

@taras

@taras taras commented Nov 18, 2026 •

Copy link
Copy Markdown
Owner

Why

The narrow Sidekick feedback drawer could move its entry link after it opened, between two frames with no input. A click aimed at the first frame then hit an inert row. This produced the conversation journey's intermittent wait for Entries. Refs #1410.

What changes

The drawer now retains the admitted inspection offset when focus settles. Its focused feedback toggle stays visible, and the entry link keeps its position.

How it works

focusSettled already retains reveal offsets for other drawer owners. Six lines extend that existing behavior to a presented Sidekick inspection while preserving route drawers. No API or shared shape changes.

How to verify it

  • A deterministic narrow-layout regression commits two frames, checks retained offset/focus and compares toggle/link bounds.
  • Removing only the production fix makes that regression fail at the settled-offset assertion. Restoring the exact candidate passes; passing evidence separately covers second-frame geometry.
  • Existing focus and wheel tests passed under Deno, Node and Bun.
  • Preflight and the complete conversation journey passed locally, including wide/medium/narrow layouts, entry-link pointer navigation and reopen.
  • The journey additionally observes the focused feedback toggle. Existing click/destination checks and 75s/20s timeouts are preserved.
  • Original Planner: PASS on signed 581251e8821ba4b0cc88641e38ce256e60684321. A fresh independent Fable review: PASS on the same exact head, with committed-source regression/journey and fix-removal controls.

Scope

Three existing files: application state retention, its regression and the conversation journey. No added test files, runner/exclusion/dependency/weights changes, sleeps, warm-ups or timeout increases.

Risks and limitations

The original CI click coordinate was reconstructed from the retained terminal stream and screenshots; it was not directly recorded. The no-input window shift and stale/current click outcomes were reproduced locally. Local Deno was 2.9.7 versus CI's 2.9.5.

UI captures were collected before commit; the candidate manifest and exact blob/diff comparison bind their tested bytes to the signed head. Scenario capture metadata itself has no Git revision. No approved visual design asset exists for this behavioral repair; the existing window/focus contract and retained failure evidence are the comparison sources.

An existing deliberate expand-and-collapse transition can still hide the narrow feedback toggle. That older reducer is unchanged and is outside this opening-focus correction; the review records it separately.

Scope confirmation

  • Every changed file supports inspection-window retention.
  • Existing pointer, keyboard, scrolling and cleanup coverage is retained.
  • Unrelated cleanup and generated files are excluded.
  • Completed and pending verification are stated separately.

Visual evidence

Both captures are 72×20. Before is the unchanged-source controlled probe after the spontaneous window shift: the feedback toggle is out of view and the link is on row 15. After is the full candidate journey: the focused feedback toggle is visible on row 17 and the link remains on row 14.

The baseline probe skips the earlier wide/medium journey sections. The candidate screenshot precedes its explicit focus wait but already shows the focused toggle; the passing wait and two-frame regression provide the separate settled-state checks. No approved design asset exists for this behavioral correction.

Before: narrow drawer loses its revealed feedback toggle

After: narrow drawer keeps the focused toggle and link position

At 72x20 the narrow feedback drawer opened at offset 1 to reveal its
toggle, then focus settlement spent the reveal without keeping that
offset, so the next frame redrew at offset 0, unmounted the focused
toggle and moved the linked entry down one row. focusSettled now keeps
the admitted start in the presented inspection's own offset.

Refs #1410

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

taras commented Nov 18, 2026

Copy link
Copy Markdown
Owner Author

Independent Fable review: PASS on signed 581251e8821ba4b0cc88641e38ce256e60684321, no blocking findings. Reviewer conversation 2c654599-f8df-4437-ae33-ff859aa07848 is separate from the diagnosis, Planner and Implementor.

The reviewer traced both real focusSettled callers and the inspection offset consumer. In a scratch clone of the committed source it independently ran the Deno regression and real conversation journey. With only the fix removed, it also removed the early state assertion to reach the geometry check: the window changed from 1 to 0, the toggle unmounted and the link moved from row 14 to 15. The behavior assertion failed. The uncorrected journey also failed its focused-toggle wait. Committed source passed both. Wide and medium geometry remained stable; existing Node/Bun and preflight evidence was explicitly reused.

Non-blocking existing behavior, outside this opening-focus fix: at 72×20, expanding and then collapsing recorded feedback still resets the older reducer's offset to 0, hides the toggle and moves focus to the link. A second Enter can therefore activate that link. The reducer is unchanged by this PR and the path existed before it; the reviewer left its product disposition to the coordinator. This PR does not claim to repair that separate collapse transition.

Original Planner also PASSed the exact head. Required CI and maintainer merge remain delivery gates; issue #1410 stays open for the correction's delivery evidence.

@taras
taras marked this pull request as ready for review November 18, 2026 09:00
@taras
taras enabled auto-merge (rebase) November 18, 2026 09:00
@taras
taras merged commit 76535a3 into main Nov 18, 2026
64 checks passed
@taras
taras deleted the codex/retain-inspection-focus-window branch November 18, 2026 09:18

taras commented Nov 18, 2026

Copy link
Copy Markdown
Owner Author

Delivery checks are green on exact reviewed 581251e8821ba4b0cc88641e38ce256e60684321: CI 44270525676, 62 successful jobs and one expected skip. Current base remains green main 6207d248; the PR is already non-draft and mergeable.

Planner and independent Fable are PASS on this exact head. The committed-source regression/journey, fix-removal controls and qualified before/after captures are recorded above. Ready for maintainer merge. Issue #1410 stays open until the merged repair's hosted delivery evidence is checked.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant