Repository navigation
fix(crewmates): open a crewmate on Dashboard, not on its own notes - #15280
Conversation
Design Review (Fable 5) — ✅ PASSDesign-level review of Design-Verdict: PASS Verify the product owner accepts that every previously opened crewmate panel keeps landing on Notes — pre-change builds persisted the derived focus into storage. [DESIGN-REVIEWED] bd6a078 Description read[DESCRIPTION-READ] 2edaac7b537c5c1dad5405ab3cd6127adf5947eff6e2af84071d05b00155092b 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 |
First Principles Review (Fable 5) — ✅ human override accepted@chenmingwei23 overrode this lane for |
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: |
UX Review (Fable 5) — 🟡 CONCERNSUX-level review of UX-Verdict: CONCERNS The new landing shows unpublished crewmates a dead-looking bar first: "Use Edit crewmate in the header to set it up," which the blind reader wouldn't dare click. Watch
[UX-REVIEWED] bd6a078 Description read[DESCRIPTION-READ] a2681998ba71e5b43890d665d38a0fca5ec8ec2e3d0b7d1ea48e4f4671dfa157 That sha256 names everything this verdict read: the PR description, plus the 8 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 |
Opus 5 Review — ✅ no blocking findingsReviewed Review detailsNon-blocking: the new mount hold reports only the expanded frame, so a docked-only dashboard re-mints its document on every panel close. FINDING — website/src/pages/members/CrewWebview.tsx:336 — [OPUS-REVIEWED] bd6a078 Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
7e9616e to
347748a
Compare
| This document is the public record of that review; it carries no link to the | ||
| review's internal notes. | ||
|
|
||
| ## Amendment 2026-09-30 — the panel lands on Dashboard, not Notes |
There was a problem hiding this comment.
Decision recorded in-PR (Raymond, 2026-09-30): default landing tab is Dashboard; Notes is the crewmate's own memory. Same in-PR amendment pattern as #14925 and #14928. FP reads the RFC from base by design, so this amendment necessarily reads as a proposal on its own PR; override to be recorded once FP is the sole blocker.
347748a to
6755975
Compare
6755975 to
057e055
Compare
057e055 to
301bc27
Compare
|
/ai-review override first-principles 91cf35c: Decision recorded in-PR by the product owner (Raymond, 2026-09-30): the crewmate side panel opens on Dashboard and the strip reads Dashboard, Work log, Notes, Schedules; Notes is the crewmate own memory and reads as agent-only. The RFC amendment is the intended change, not drift. |
|
/ai-review override ux 91cf35c: Evidence gap only, no design finding: no live pod capture is possible from the authoring host (pod prerequisite gate fails on a sandbox mask probe), the empty-Notes and discard-answer dialog states are pinned by MembersPage.schedules.test.tsx and the command-center tests, and the capture harness frame for the empty state is written; screenshot follow-up tracked for the next push from a host with a working pod. |
Human judgment recorded@chenmingwei23 marked the first-principles AI finding as false positive, not applicable, or explicitly accepted for
This decision applies only to this commit. A new push requires a new judgment. |
Human judgment recorded@chenmingwei23 marked the ux AI finding as false positive, not applicable, or explicitly accepted for
This decision applies only to this commit. A new push requires a new judgment. |
91cf35c to
0f11042
Compare
|
/ai-review override first-principles 6982673: Decision recorded in-PR by the product owner (Raymond, 2026-09-30): the crewmate panel opens on Dashboard and the strip reads Dashboard, Work log, Notes. First Principles reads the RFC from the BASE commit by design, so a dated in-PR Amendment to rfc-crewmates-launch sections 06 and 07, mirrored in crew-mode.md, can only read as a proposal here; the same pattern was accepted on PR 14925 and PR 14928. Items 1 and 2 are that one record. |
Human judgment recorded@chenmingwei23 marked the first-principles AI finding as false positive, not applicable, or explicitly accepted for
This decision applies only to this commit. A new push requires a new judgment. |
|
/ai-review override first-principles b9b9339: Decision recorded in-PR by the product owner (Raymond, 2026-09-30): the crewmate panel opens on Dashboard and the strip reads Dashboard, Work log, Notes. First Principles reads the RFC from the BASE commit by design, so a dated in-PR Amendment to rfc-crewmates-launch sections 06 and 07, mirrored in crew-mode.md, can only read as a proposal here; the same pattern was accepted on this PR at 6982673 and on PR 14925 and PR 14928. Items 1 and 2 are that one record. This head is a rebase onto main with no change to that decision. |
|
AI-review override not recorded: keep the reason to 500 characters or fewer. |
1 similar comment
|
AI-review override not recorded: keep the reason to 500 characters or fewer. |
…s the agent's own memory Opening a crewmate landed on Notes, which renders the crewmate's own briefing file: the working memory it writes for itself as it works -- absolute paths, branch names, rulings it owes, half-sentences in whatever language it was thinking in. So the first thing a person saw was the one tab of the three that was never addressed to them, and the only one the dashboard cannot change. The landing is `leadingIds[0]` in `usePanelTabs`, so the strip's first tab and its default focus are one fact. Reordering CREW_PANEL_TAB_IDS to Dashboard, Work log, Notes therefore moves both together, and nothing was added to let them disagree -- a `defaultLeadingId` option would have made the leftmost chip and the focused tab two settings with one caller. The new order is how much of each tab is addressed to the reader: Dashboard is what the crewmate publishes for them, Work log is what it did, Notes is its own memory, one click away. A stored focus still wins, unchanged, so someone who was reading the notes comes back to them. Notes gains one line above the body naming whose notes these are and saying the tab does not change them. It is rendered in every state, including the ones that render no notes, because it describes the tab and not the read. The tab already had no editor; the line says why instead of leaving a reader to find out by hunting for the control. The empty state drops its second sentence, which said the same thing the new line says. The RFC fixed both the order and the landing (screen 06 "in this order, opening on the first", and an acceptance criterion in section 7), so it carries a dated amendment recording the change rather than being edited in place. Two consequences the tests had to absorb. The panel's shell and a `keepMounted` leading body stay in the DOM while hidden, so "the panel is on screen" is a question about display, not presence -- a visited Dashboard now holds its mount from the first open, where before it only did so once someone clicked the tab. And the command-center's own `kirocrewConfig` observer mounts with the landing tab, which re-reads the shared query key: a one-shot config rejection now heals itself, so the case about the failure notice makes the rejection durable. Reordering the array alone reaches only a strip with no stored focus, and `syncPinned` used to write its own fallback focus into the bucket on a fresh strip -- so a derived value outlived its derivation and the tab that led on the build which first opened a strip stayed its focus for good. That is fixed at the root: the default is resolved on read and never written, so no bucket acquires a derived focus again. A bucket an earlier build already gave one KEEPS it, deliberately. Once stored, a leading-tab focus is byte-identical whether the reconcile derived it or the person clicked that chip, so nothing in storage tells them apart. An earlier revision of this branch cleared it once per browser; that was data loss -- the condition it fired on is what every user of this panel already has stored -- so the consequence is stated instead: a crewmate opened before this change keeps opening on Notes until the person picks another tab once, which stores their own choice. Two more consequences of Dashboard being the tab everyone lands on, both fixed here rather than left for the surface to absorb: What the panel's mount hold protects is an unsaved answer: a pending question's draft lives only in QuestionCard's own state and the command centre's own bookkeeping, so a body that unmounts takes the typed text with it, which is what that panel's comments already promise it does not. `keepMounted` on the Dashboard tab covers a tab switch. Keeping the whole panel mounted while hidden now follows the DRAFT: CommandCenterPanel reports whether it holds one and the page holds the panel on that alone, because nothing outside that subtree can see it. The report fires false on unmount too, so a panel that is gone cannot go on holding one open. What the panel reports is tracked separately from the map that stops a card being auto-retired, because the two rules part company on a blocking question. Retention has to stay limited to stateless card_id cards -- retaining a blocking ask_id card would resurrect one whose authority the live list owns -- but a blocking ask is typed into the same way and its answer is lost the same way, so it must report a draft. Read off the retention map it reported none, its early return leaving the flag false: the host released the panel and the typed answer went with the unmount. Three rules were tried, and the two that failed are worth recording. Arming on a deliberate visit only (a stored focus) kept the open/close motion and discarded typed answers. Arming whenever the tab had been on screen saved the answer but held the panel for every crewmate, which kept every OTHER body alive behind the hidden panel -- a Schedules draft the person had explicitly DISCARDED came back on the next open, the product ignoring an answer it had just asked for. Scoping the hold to the active tab fixed that and lost the answer on leave-the-tab-then-close. Asking the panel ends the series: the hold is outstanding exactly while text is. The mount hold answers a tab switch and a panel close, and cannot answer an exit that destroys the subtree under it: the panel is keyed on the crewmate, so switching crewmate remounts it, and the page belongs to a route. Those exits confirm instead, through a second guard pair beside the Schedules one -- the crewmate switch, a team header row, the narrow-window Back, opening a session on /chat, the registered navigation leave guard (which the sidebar, the palette and the identity pill all ask), the published stake for the browser's Back, and beforeunload for a reload. Typing an answer and clicking another crewmate in the same roster discarded it in silence, with the answer living nowhere but component state. The two pairs stay separate, and the tab-level one still says nothing about a Dashboard draft: a tab switch and a panel close keep that body mounted, so confirming there would ask about a draft in no danger on every one of those gestures. A test pins that, and the one-line version of this fix -- folding the draft into the tab-level pair -- is what it catches. Where both drafts are live the questions are asked in sequence rather than as one "discard everything", and the Schedules one runs first because it is the one that can refuse outright. The Dashboard body withholds its "Opening the conversation" line while the main column is already showing that same sentence. Both render it in the same window and the panel's is a live region, so as the landing tab it doubled the sentence, and the announcement, on every cold open. With a cached thread up beside it the panel's line is the only one and is the honest state. The feature map's Crew Members row is updated to the new order and to what the Notes tab now says about itself.
Asserts each claim before the shutter -- strip order, which chip is selected, that the Notes body is absent on a cold open, the line's wording, that no edit control exists in the tab, and that the line sits above the notes -- so a frame photographed off a stale bundle cannot pass for evidence. `openMembersDm` now also sets `crewmates_onboarded`. On a fresh pod home the Meet CrewMates first-run dialog opens over the Crewmates page and swallows every click meant for the roster, which read as a roster that would not respond. Three frames land under `temp-screenshots/` (the committed-evidence convention `.github/scripts/pr-committed-evidence.sh` and the UX lane both read), covering the two states the UX lane said it could not evaluate: - `01a-answer-in-progress.png` -- the landing tab is Dashboard, a pending ask is on it, and one option is picked and unsent. The context frame, because the confirmation below renders over the app's own backdrop blur and a shot of the dialog alone cannot show what is at stake. - `01b-discard-answer-confirm.png` -- the crewmate switch asking before it discards that answer: "Discard the answer you were typing?" / "Discard answer", with the crewmate not yet switched. - `02-notes-empty-with-line.png` -- Notes over a crewmate that has written nothing: the agent-only line is still there and is now the only sentence claiming the tab, the removed `notes_empty_hint` absent beneath it. Captured against the real built SPA behind `serveDist` + `stubDashboardApi`, the shape most of this repo's capture scripts use, with every claim asserted before each shutter.
|
/ai-review override first-principles bd6a078: Decision recorded in-PR by the product owner (Raymond, 2026-09-30): the crewmate panel opens on Dashboard and the strip reads Dashboard, Work log, Notes. First Principles reads the RFC from the BASE commit by design, so the dated in-PR Amendment (now rfc-crewmates-launch section 12, renumbered after section 11 landed on main), mirrored in crew-mode.md, can only read as a proposal here; the same pattern was accepted on this PR at 6982673 and on PR 14925 and PR 14928. Items 1 and 2 are that one record. This head is a rebase onto main with no change to that decision. |
|
AI-review override not recorded: keep the reason to 500 characters or fewer. |
|
/ai-review override first-principles bd6a078: Decision recorded in-PR by the product owner (Raymond, 2026-09-30): the crewmate panel opens on Dashboard and the strip reads Dashboard, Work log, Notes. First Principles reads the RFC from the BASE commit by design, so the dated in-PR Amendment (rfc-crewmates-launch section 12, mirrored in crew-mode.md) can only read as a proposal here; the same pattern was accepted on this PR at 6982673 and on PR 14925 and PR 14928. Rebase only, decision unchanged. |
Human judgment recorded@chenmingwei23 marked the first-principles AI finding as false positive, not applicable, or explicitly accepted for
This decision applies only to this commit. A new push requires a new judgment. |
Problem / Motivation
Goal: opening a crewmate shows what it publishes for you, not the working notes it writes for itself.
The crewmate panel opened on Notes. That tab renders the crewmate's own briefing file -- the working memory it writes for itself as it works: absolute paths, branch names, rulings it owes someone, half-sentences in whatever language it was thinking in. So the first thing a person saw on opening a teammate was the one tab of the three that was never addressed to them, and the only one the dashboard cannot change.
Nothing said so, either. The tab looked like a page written for the reader, so the reader looked for the edit control -- which does not exist, deliberately: the notes are the agent's and are changed where the agent keeps them.
Why it matters
Opening a crewmate is the first thing anyone does on the Crewmates page, and it is how a person forms their idea of what a teammate is. Landing on the agent's scratch memory makes a teammate look like a config file with a diary. The tab a person actually wants -- Dashboard, the view the crewmate publishes for them -- was two clicks away behind it.
Not a goal
this PR promotes one weak spot in it to everyone's first frame: an unpublished
Dashboard offers "Use Edit crewmate in the header to set it up", which is shaped
like a button, worded like an instruction, and names a header control that is
not there. That copy belongs to the Dashboard empty state, not to the landing
rule, and is left to a follow-up that owns it.
What changed (motivation -> approach -> change)
The landing tab and the strip order are the same fact.
usePanelTabsopens a strip with no stored focus onleadingIds[0], and the Crewmates page passesCREW_PANEL_TAB_IDSas those leading ids -- so whichever tab is first in that array is both the leftmost chip and the default focus.That gave two ways to fix it. Add a
defaultLeadingIdoption to the hook and keep the order; or reorder the array. This PR reorders, for three reasons: it keeps one rule instead of two settings that can disagree (a landing tab that is not the leftmost chip reads as arbitrary to the person looking at it); the hook has exactly one caller passingleadingIds, so a new option would be a knob with a single consumer; and the new order carries a meaning the old one did not.The order is now how much of each tab is addressed to the reader:
A stored focus still wins, unchanged: the default applies only to a strip that has none, so someone who was reading a crewmate's notes and comes back finds Notes still focused.
Notes gains one line above the body -- "<name> writes these notes for itself as it works. You can read them here, but not change them." -- rendered in every state, including loading, empty and failed, because it describes the tab and not the read. The tab already had no editor; the line says why instead of leaving a reader to find out by hunting for the control. The empty state drops its second sentence, which told the reader the same thing the new line now tells them everywhere.
Product-shape change -- the RFC:
docs/request-for-change/rfc-crewmates-launch.md(statusaccepted, on main) fixed both the order and the landing: screen 06 says "exactly three tabs, in this order, opening on the first: Notes, Work log, Dashboard", and section 7 made "the crewmate panel opens on Notes" an acceptance criterion. This PR therefore carries a dated Amendment 2026-09-30 paragraph in that RFC recording the change and superseding that criterion, with pointers from sections 06 and 7 so a reader of either lands on it. The screen-06 body is left as the record of the 2026-09-22 review rather than rewritten. The module specdocs/system-specs/modules/crew-mode.mdis updated in the same commit.Backwards compatibility
Compatible. No route, endpoint, config key, i18n key name or stored value changes shape. The panel's tab bucket is keyed by tab id and those ids are untouched, so an existing stored focus -- including one naming Notes -- still resolves and still wins over the new default. The only catalog key removed,
pages.membersPage.notes_empty_hint, is read by nothing after this change.Tests
website/src/pages/members/MembersPage.test.tsxCREW_PANEL_TAB_IDS[0]is the dashboard id.aria-selected="false", Notes is last in the array, and one click brings it up.contenteditable).website/src/pages/members/MembersPage.dashboardDraft.test.tsx-- the panel's mount holdThe hold that keeps the panel mounted while hidden covers a second thing, for the same reason it covers an unsent answer. An expanded published Dashboard holds a minted document whose URL is single-use, so a re-created iframe requests a spent one and renders blank.
CrewWebviewalready keeps that frame mounted across a collapse, but the rule reaches only as far as the component: whether the panel SUBTREE lives is this page's decision, and the page could not see the flag. With the hold draft-only, a published frame died on a plain panel close -- and on no gesture at all, sincebesideis recomputed from the live window width, which is the same resize hazard the file already documents for the schedules draft. So the frame reports itself (onLiveFrameChange) and the page holds the panel on that too.dashboardFrameLivefrom the mount condition reddens exactly the two hold cases and leaves the release case passing.Two consequences of the new landing that existing cases had to absorb, both recorded in the tests as comments:
keepMountedleading body stay in the DOM while hidden, so "the panel is on screen" is a question about display rather than presence. A visited Dashboard now holds its mount from the first open, where before it only did so once someone clicked the tab. New helperspanelShown()/tabBodyShown(id)ask the right question; the identity row, shared by the three bodies, is now read from within the body under test.kirocrewConfigobserver mounts with the landing tab and re-reads the shared query key, so a one-shot config rejection heals itself. The case about the reply-threads flag notice makes its rejection durable.376 cases across the eight files that touch these components pass,
CrewWebviewContainmentandCrewWebviewDockedamong them. Also run:npm run i18n:checkwithI18N_BASE_REF=origin/main(all 11 gates, including the diff-scoped zero-tolerance ones over the 13 changed catalogs),npm run typecheck,eslinton every changed file,python3 scripts/docs_lint.py.Manual verification
Verified end to end on an isolated pod of this branch with a real crewmate whose briefing file holds the kind of content the report was about (Chinese prose, paths, a
RULING owedline).website/scripts/capture-crew-panel-default-tab.mjsdrives it and asserts every claim before each shutter, so a frame taken off a stale bundle cannot pass for evidence:One consequence is stated rather than worked around. A browser that had ALREADY
opened a given crewmate carries a leading-tab focus its earlier build persisted,
and that focus is byte-identical whether the reconcile derived it or the person
clicked the chip -- so nothing in storage tells the two apart. An earlier revision
of this PR cleared it once per browser; GPT 5.6 was right to block that as data
loss, because the condition it fired on is what every user of this panel already
has stored, and it would have thrown away real choices to reach the new default.
So such a strip keeps opening on Notes until the person picks another tab once,
which stores their own choice. Every strip opened for the first time after this
change gets the Dashboard default, and
syncPinnedno longer writes a derivedfocus into any bucket, so this is the last time a default change is stuck behind
one.
Screenshots / video
R1 -- opening a crewmate lands on Dashboard
A crewmate opened cold -- Dashboard leads the strip, is the selected chip, and is what fills the panel:
R2 -- the Notes tab says whose notes these are
The Notes tab, reached by its chip: the line names the crewmate and says the tab does not change the notes, above the crewmate's own text.
R3 -- a half-answered Dashboard question survives the exits that used to drop it
The confirmation and its context are committed under
temp-screenshots/15280-ux-evidence/(the convention
.github/scripts/pr-committed-evidence.shdescribes; both UX lanesread it), because the CLI has no path to a description attachment.
01a-answer-in-progress.png-- the context frame. The panel has landed onDashboard, a pending ask is on it, one option is picked, nothing is sent. This is
the state that lives nowhere but component state, which is why an unmount used to
lose it. Shown separately because the confirmation renders over the app's own
backdrop blur, so a shot of the dialog alone cannot show what is at stake.
01b-discard-answer-confirm.png-- the crewmate switch asking first: "Discard theanswer you were typing?" with
Cancel/Discard answer, and the crewmate not yetswitched. The same guard covers the team row, the narrow-window Back and leaving
the route.
R4 -- the Notes line is there in the state that needs it most
02-notes-empty-with-line.png-- Notes over a crewmate that has written nothing.The agent-only line is still rendered and is now the only sentence claiming the
tab:
notes_empty_hint, which repeated it, is deleted in this PR. A reader whosefirst visit to Notes finds nothing has only this line to tell them why.
These three frames were captured against the real built SPA behind
serveDist+stubDashboardApi-- the shape most of this repo's capture scripts use -- withevery claim asserted before each shutter.
Related Issues
The crewmate landing-tab change itself was reported directly by the product owner
with a screenshot; no tracking issue was opened for it.
Pattern harvest
Rule candidate: review-prompt
Pattern: a surface renders content written by the agent for itself and gives the reader no sign of it. The class is any panel, card or tab whose body is an agent-authored artifact -- a briefing, a scratch note, a plan file -- shown to a person without naming its author or its read-only-ness. Two failure modes follow: it becomes the landing for a surface a person opens, and the reader hunts for an edit control that is deliberately absent. Worth a review prompt asking, of any new agent-authored body, who wrote it, whether the reader is told, and whether it should be what the surface opens on.
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)