Repository navigation
fix(git-panel): scope the filter refusal to the half it reports, and split its two causes - #12254
Conversation
UX Review (Fable 5) — 🟡 CONCERNSUX-level review of UX-Verdict: CONCERNS Scope nouns don't match the panel's visible labels: titles say "history" while the surviving list is headed "Commits" — the blind reader couldn't map the halves. Watch
Suggestions
[UX-REVIEWED] cd3b06c |
GPT 5.6 Review — ✅ no blocking findingsGPT 5.6 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: |
Design Review (Fable 5) — ✅ PASSDesign-level review of Design-Verdict: PASS Native-speaker spot-check of the six new titles in the eleven translated catalogs — hand-written values were this family's measured defect source. [DESIGN-REVIEWED] cd3b06c |
First Principles Review (Fable 5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Verify on #11982/#12080 that the "reader called all three frames crossed-out" report exists — the opacity-60 deviation (item 7) is the only item whose harm rests on evidence outside this repo. Every copy item traces to a numbered defect on the linked reviews, the fix models scope at the cause ( What this change shipsInventory (10 items) — 10 justifiedIntent: make the Git panel's filter-refusal notices report only the failure each actually accounts for, and make the two causes and hand-offs distinguishable — a FIX.
[FIRST-PRINCIPLES-REVIEWED] cd3b06c |
Opus 4.8 Review — ✅ no blocking findingsReviewed Review detailsBoth candidates fail falsification. Candidate 1 (shared body loses the consequence clause on the two title-less surfaces): the PR author deliberately restructured this — the body now carries only the cause, and "what can't be shown" moved to the title. Candidate 2 (declared-cause body beside a live refresh in the partial-healthy state): the candidate itself concedes the No new grounded defect emerged from reading the changed logic; the scope resolution and the 13-catalog key additions are covered by the added tests. No findings. [OPUS-REVIEWED] cd3b06c Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
59d1ae1 to
a7ef1e4
Compare
a7ef1e4 to
c6531b8
Compare
|
Round 1 disposition, head UX Review, Design Review and Screenshot Evidence were all the same finding, and it is now closed. Seven frames are attached to the description as permanent assets. Both lanes said staging the refusal through the query layer was admissible, and they were right that my earlier "not captured" reasoning was wrong: the frame verifies rendering, not the query layer. The two frames that answer the specific asks: the divergent two-notice stack, with the scoped title and the two distinct hand-off labels composited; and the two causes' headers side by side, which is what item 5 needed, showing that the slash reads at Frontend Tests (1) and (2) were two real copy defects, now fixed. Worth recording how they got past a local run:
Backend Tests (Windows) (1) is main-owned, not this PR. Four This head is also rebased onto current main, which had moved to Tailwind CSS v4 in #12220. The changed-file set is unchanged across the rebase and no new main commit touches any file in this diff. |
c6531b8 to
c83760a
Compare
|
Round 2 disposition, head Design Review, First Principles, GPT 5.6 and Opus 4.8 all came back clean on the previous head. UX Review was the only blocker and both of its items are now closed. The BLOCK: two of the six title states had no frame. Correct -- the frame set reached The Watch: Both halves of that are pinned, and both pins were mutation-checked. Removing the cue reddens the catalog gate and the divergent-state GitPanel test; adding a cue to One thing worth recording about the frames, because it nearly produced dishonest evidence. The first capture run after the rebase reported every assertion passing while the build had FAILED, so it photographed the previous |
c83760a to
b1c1ad1
Compare
|
Round 3 disposition, head UX flipped BLOCK to CONCERNS and Design, GPT and Opus are clean. First Principles moved PASS to CONCERNS with a finding that was right, and it is the one real defect this round. First Principles: the shared refusal bodies reach three surfaces and I reasoned about one. Verified before acting, and the lane's grep is accurate: The same reading resolves the other half of that finding in the opposite direction from how it reads. On those two title-less surfaces the body no longer states what became unavailable -- but the clause it lost was "Changes and history can't be shown", and neither of them renders history. So the scope clause was over-claiming there before this PR, and removing it removes an over-claim rather than creating a gap. That is now declared in the description rather than left implicit. Two guards were added for the class, because the defect was a shared string quietly acquiring a surface-specific promise:
The capture harness caught the change on its first run after it -- it had been using the cue as its marker for the unreadable cause and failed two assertions, which is what it is for. Its marker is now the cause sentence, and a new assertion pins that the cue is ABSENT from a frame whose only notice is the refusal. UX's light-theme gap: taken. A tenth frame shoots the inert control in light theme. The item 5 argument is about contrast between a muted glyph and what sits behind it, and a theme changes exactly that, so one theme was half an answer. The slash reads on both. UX's noun-binding suggestion: not taken here, and worth saying why rather than going quiet on it. The proposal is to move the six titles from "changes"/"history" to "the changes list"/"commit history" so the nouns name something visible. The panel's own visible section labels are
|
|
Round 4 disposition, head Opposing retry advice on one screen. Real observation, and mine: the cue I added to Titles say "history" while the section is headed Commits. Re-raised, and I under-weighted it last round: I argued "commit history" names nothing the panel labels, and the better argument is the lane's -- the panel already uses that exact term in Hand-off labels differ only in their last word. Agreed as a reading; same disposition. Front-loading them ("Ask about the changes list" / "Ask about the commit history") would also make the two labels carry the same vocabulary as the retitled headings, so it belongs in the same pass as the item above rather than on its own. No screenshot of the rail or the workspace tree. The lane calls this verification rather than a gap, which is right -- the exact strings are evidenced in the panel frames, and the round-3 guards pin what those surfaces may receive: neither shared body may name a control, and the panel-only strings cannot be referenced outside To be explicit about what I am not doing: I am not asking for an override, and nothing here is being dismissed as a false positive. Three of the four are edits I would make and would defend; the reason they are not in this head is the cost of re-rolling verdicts that are already clean, on findings the lane itself rates as friction on the rarest state of the surface. |
…split its two causes The Git panel's filter-driver refusal rendered one title and one sentence for every state it can be in. Three of those states are not about both halves: the status route can fail on its own terms while the log route refuses, the log route can do the same, and a recovered route leaves the other's cached refusal standing while its list renders underneath. In all three the refusal claimed that changes AND history could not be shown, so it either reported a sibling notice's failure as its own or denied a list that was on screen. Scope now comes from which routes are actually refusing, and the title carries it. The title also carries permanence, so a reader who reads only the bold line gets the right one of the two causes: a declared driver stands while the repo config does, an unreadable config can clear by itself and its sentence says to refresh. The declared sentence leads with Git LFS rather than with the Git config that both causes name. Neither sentence says "checks" any more, because the surface in front of the reader is labelled Changes and Commits. In the divergent state the two stacked hand-offs name the half each one carries, and the status notice marks itself as a separate failure. The inert refresh glyph keeps the slash that distinguishes it legible at the size the header renders it. All eleven values plus the pseudolocale are updated for every string in the family, in one pass rather than one per review round. Closes #12080
b1c1ad1 to
cd3b06c
Compare
|
Round 5, head #12247 is CLOSED COMPLETED and its fix is #12269, merged as A The still-open #12271 targets the same defect with the same quoted assertion. #12269 having merged appears to make it redundant, but that is its author's call, not mine, and nothing here depends on it. Round 4's dispositions stand unchanged: every blocking lane was clean at the previous head, and the four advisory UX items are answered in the comment above rather than in this head. This rebase is not a response to any of them -- it is the only action that could clear the last blocking red. |
iamwhatever
left a comment
There was a problem hiding this comment.
Approved on the current head: CI green, readiness passed, no security-sensitive path in the diff.
…split its two causes (kirodotdev#12254) The Git panel's filter-driver refusal rendered one title and one sentence for every state it can be in. Three of those states are not about both halves: the status route can fail on its own terms while the log route refuses, the log route can do the same, and a recovered route leaves the other's cached refusal standing while its list renders underneath. In all three the refusal claimed that changes AND history could not be shown, so it either reported a sibling notice's failure as its own or denied a list that was on screen. Scope now comes from which routes are actually refusing, and the title carries it. The title also carries permanence, so a reader who reads only the bold line gets the right one of the two causes: a declared driver stands while the repo config does, an unreadable config can clear by itself and its sentence says to refresh. The declared sentence leads with Git LFS rather than with the Git config that both causes name. Neither sentence says "checks" any more, because the surface in front of the reader is labelled Changes and Commits. In the divergent state the two stacked hand-offs name the half each one carries, and the status notice marks itself as a separate failure. The inert refresh glyph keeps the slash that distinguishes it legible at the size the header renders it. All eleven values plus the pseudolocale are updated for every string in the family, in one pass rather than one per review round. Closes kirodotdev#12080
Problem / Motivation
The Git panel's filter-driver refusal shipped with one title and one sentence per
cause, and the UX lane on #11982 then named a different defect in that copy on
every pushed head. Eight items accumulated, all in the same string family:
be shown while the status notice directly above it already reports the changes
half. Reaching that state means both routes failed, so the sentence is true --
but it is reporting on a half that is not its own.
Ask the agentlabel, so nothing says whichfailure each hand-off carries. The reports genuinely differ: each has its own
endpoint, status and code.
This repository's Git config)and are told apart only by a verb several words in, although their retry
advice is opposite.
tell the permanent condition from the one that clears itself, and a reader
habituated to the permanent case gives up on the self-healing one.
size={13}: a reader shown three frames called all three crossed-out, whenonly one was.
filter_refused_unreadablesays checks are kept off while its refresh buttonis deliberately left live, so a reader had no idea whether pressing it would
do anything.
reader is looking at is labelled Changes and Commits.
are about reading or checking changes.
A ninth state was reachable and unreported: when one route refuses and the other
is simply healthy, the panel still coalesces and the refusal claims both
halves while the recovered route's list renders underneath it.
Why it matters
Item 1 and the ninth state are attribution failures with a visible consequence:
the panel either reports a sibling notice's failure as its own or denies a list
that is on screen, which is the same class of wrong answer as drawing a clean
badge over a status nobody read. Item 4 is the one that costs a user something
concrete -- the unreadable-config state clears by itself, and a headline shared
with the permanent state teaches people to stop waiting for it.
The rest is comprehension friction on the rarest states of the surface. The
reason they are batched is mechanical rather than editorial: every one of them
rewrites strings that exist in thirteen catalogs, and settling them one per head
re-rolls every review lane for one string. It had not converged after five
rounds. Hand-written catalog values were also #11982's measured defect source (a
Korean non-word, a malformed Bengali conjunction, Hindi spellings missing a
nukta), and writing the family once is fewer chances to produce another.
What changed (motivation → approach → change)
Symptom -- a refusal notice that claims both halves of the panel in states
where it only accounts for one.
Root cause -- scope was not modelled at all. The copy carried a fixed
"Changes and history can't be shown" clause, so every render site said the same
thing whichever routes had actually refused.
Change -- scope is now derived from the two routes' refusal flags
(
gitFilterRefusalScope:bothwhen both refuse, otherwise the one that does)and the title carries it. Putting it on the title rather than in the body is
what keeps the body one sentence per cause: the cause clause does not vary with
scope, and a title is where a reader looks for what is unavailable. That same
move gives item 4 its home -- the unreadable titles say "right now" and the
declared ones do not, so permanence is on the bold line instead of a body-read
away. Six title keys, indexed by an
as constmap beside the existinggitFilterRefusalCopyKeyso every key stays a greppable literal.With scope off the body, the two sentences were rewritten once for the remaining
editorial items: the declared one leads with
Git LFS or another conversion program(item 3), and neither says "checks" any more (item 7).refresh_unavailablefollows the same vocabulary.Those two bodies are shared, and that constrains what they may say.
gitFilterRefusalCopyKeyhas three rendering consumers:GitPanel, the chatFileBrowserRailandPierreWorkspaceTreeImpl. The latter two render NO title bydeliberate design (an
inlinenotice at rail width stacks a title into two-wordfragments), and the workspace tree has no refresh control at all. So the shared
bodies carry neither the scope clause nor any retry cue: scope is the panel's
title's job, and a cue would name a control that exists one panel over. On the
panel the invitation is still there -- the unreadable titles say "right now" and
the control beside them is live. The retry cue lives only on
status_failed_no_history, which has exactly one consumer, and whose surface'srefresh control is live whenever it renders.
That also removes an over-claim from the two title-less surfaces rather than
adding one: they show changed files and no history, so the old shared clause
"Changes and history can't be shown" was naming a half neither of them renders.
For item 2,
ErrorNoticegained an optionalaskAgentLabel, forwarded toAskAgentButtonas an optionallabel; the Git panel's two notices pass labelsnaming the half each carries. Item 8 is
status_failed_no_historymarking itselfas a separate failure, which is the lane's own proposed wording. It also carries
the retry cue, because it stated what had happened and offered no next step -- a
reader called it the one red box they would least know what to do with. The cue is
safe on that string in particular, which is why it is the only string that has it:
one consumer, it renders only in the branch where at least one route is not
refusing, and the refresh control goes inert only when both are, so refresh is
live wherever those words appear. Its sibling
status_faileddeliberately keepsno cue, since there the commit list is still on screen and the sentence is about
staleness.
Item 5 is the only non-copy item. The inert button moves from
opacity-40toopacity-60and its glyph tostrokeWidth={2.5}. That is a deliberatedeviation from the repo's dominant disabled treatment, argued in a comment at
the call site: everywhere else a disabled glyph carries no information, so the
fade IS the message and 40% costs nothing, whereas here the entire difference
between the two causes' controls is one thin diagonal that 40% erases first. It
is still weaker than the live glyph, so the
text-muted hover:text-textresting convention 78 other buttons share is untouched, and no new colour or
shape language is introduced.
All eleven translated catalogs plus the regenerated
en-XAcarry every value.Backwards compatibility
Compatible. No catalog key is removed or renamed --
filter_refused,filter_refused_unreadable,filter_refused_title,refresh_unavailableandstatus_failed_no_historykeep their keys with new values, and the five newtitles plus two hand-off labels are additions.
askAgentLabelonErrorNoticeand
labelonAskAgentButtonare optional and default to the sharedcomponents.askAgent.ask_the_agent, so the roughly eighty existing call sitesrender exactly as before.
Tests
src/i18n/gitPanelRefusalCopy.test.ts(new) pins the family as STRUCTURE acrossevery shipped catalog rather than as literals, because the wording is allowed to
change and the relationships are what broke: the three titles of a cause are
pairwise distinct (scope is real), the declared and unreadable titles differ at
every scope (permanence is on the headline), the declared body starts with the
do-not-translate term
Git LFS(it leads with its cause), and the two hand-offlabels differ from each other and from the shared one. Two more assertions guard
the sharing itself: neither shared body may name the refresh control, because two
of its three surfaces do not have one; and the six titles plus
status_failed_no_historymust be referenced by no source file other thanGitPanel.tsxandgitStatusError.ts, so a future surface cannot quietly adopt apanel-only promise. A first assertion fails if
it resolved fewer than thirteen catalogs, so it cannot pass vacuously. The rules
that cannot be expressed language-independently -- the retry cue, the panel's
vocabulary, the separate-failure marker -- are asserted on English.
src/test/GitPanel.test.tsxgains the render-level pairing: the log-sidedivergence titles the refusal on history and not on both halves, the status-side
divergence titles it on changes, a refusal beside a healthy sibling names only
its own half, the unreadable cause gets its own headline, and the inert glyph
keeps its stroke weight and fade. The two existing divergent-state tests were
updated from the both-halves title to the scoped one, which is the assertion that
had been pinning the defect. The zh-CN hand-off lookup now resolves its label
through the catalog instead of a hardcoded string.
Nine mutations, each killed by the test that names the property, no survivors
(harness asserted every mutation applied and restored from a file copy taken
before the run):
ErrorNoticedrops the per-notice labelGreen locally:
GitPanel.test.tsx+ the new gate (37 tests), the four othersuites that read the changed keys or props (
PierreWorkspaceTreeImpl,AskAgentButton.cov80,errorToAgent,inlineNoticeTruncate-- 132 tests), andthe whole
src/i18n/catalog gate suite (43 files, 702 tests) which owns thethirteen changed catalogs. Also
tsc -b,eslintat--max-warnings 0on thechanged files,
i18n:check,i18n:render,lint:phantom-classes,jscpd,check_brand_name,check_comment_history,check_focus_cue, the analyze build,check-bundle-size(697 chunks within budget) andcheck-chunk-cycles.Manual verification
N/A -- these are 503-driven error notices, and the two divergent ones need a
second independent failure on the sibling route at the same moment. They are
staged through the query layer instead, both in the unit suite and in the
screenshot harness below, which drives the REAL built SPA with
/api/project/git/*answered from fixtures. The copy's cross-language relationships are verified by
the catalog gate rather than by eye.
Screenshots / video
website/scripts/capture-git-panel-refusal-scope.mjs(new, in this diff) shootsthese against the real built SPA. Ten frames cover all six title states on dark
plus the inert control on light. It asserts, per frame, that the strings the frame
is evidence for are present AND that a wider scope's title, or a cue that belongs
to another string, is absent; and for the header frames that the control's
disabled state, stroke weight and opacity class are what the diff claims. Any of
those failing exits non-zero, so a frame that disagrees with the code is never
produced -- it caught the retry cue's removal on the first run after it.
The divergent state the issue was filed for. Two notices: the outage marks itself
a separate problem and hands off "about the changes", the refusal names history
only and hands off "about the history".
The same divergence on the OTHER cause, which is the pair the UX lane found
missing: "right now" instead of the permanent wording, beside a sibling notice
that owns the changes half.
The state no notice covered before: one route refuses while the other is healthy,
so the commit list renders underneath. The refusal names changes only, which is
what stops it denying something on screen. Both causes.
The two causes' headers, which is the pair item 5 asked for. First: a declared
driver, refresh inert,
opacity-60andstrokeWidth={2.5}. Second: anunreadable config, refresh live at the resting convention, and a body that says
to use it. The slash reads at
size={13}and the two controls are not mistakablefor each other.
The same inert control in light theme, because the argument is about contrast
between a muted glyph and what is behind it, and a theme changes exactly that.
The remaining three frames: both-halves scope per cause, and the mirror divergence
Related Issues
Closes #12080
Pattern harvest
Rule candidate: review-prompt
Pattern: user-facing copy that states an EFFECT ("changes and history can't be
shown") while the code path that renders it only establishes a CAUSE. The copy
then over-claims in every state where the cause holds for less than the whole
effect, and it reads as correct in the common state, so review catches it only
when someone imagines the rare one. The generalisation: when a notice names what
is unavailable, that clause belongs to whatever the renderer actually knows is
unavailable, not to the condition's worst case. Sibling instance already in this
file: the panel drops "Commit history may be out of date" exactly when the log
route has its own notice.