Repository navigation
fix(dashboard): keep feature-preview switches on the gateway - #21624
mbriones98 wants to merge 2 commits into
Conversation
The Feature previews switches lived only in browser localStorage, so a new origin, the desktop app beside a browser tab, or cleared site data turned them back off. Each switch is now stored in the gateway config under dashboard.feature_previews.<name>, written through the existing owner-only config PATCH, with localStorage kept as a cache. A switch set only in this browser is carried up once with if_unset; a refused toggle names the switch that went back and says why. Fixes #21396
Design Review (Opus 5.5) — ✅ PASSDesign-level review of Design-Verdict: PASS Check that a non-owner dashboard session (token or QR) can still use previews now that every toggle needs the owner-only config PATCH. [DESIGN-REVIEWED] ff3fdb0 Description read[DESCRIPTION-READ] b320da9371ee5dbaf7da9ff2ae55297f2c0470cbd9069c5a4a2b2d1a1cdd0440 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 (Opus 5.5) — 🟡 CONCERNSPremise-level review of First-Principles-Verdict: CONCERNS The cross-surface fix is sound. The per-flag write queue and the one-time carry-up add module-global state that the settings toggles elsewhere don't need. Not justified as shipped
What this change shipsInventory (9 items) — 6 justifiedIntent: keep a Feature previews switch on across every browser and the desktop app that use one gateway. FIX (#21396).
[FIRST-PRINCIPLES-REVIEWED] ff3fdb0 Description read[DESCRIPTION-READ] 3fb5d1801f97857352fb30ec363f96f1201a162402e1fa930649601d90991e70 That sha256 names the PR description this verdict read. Recompute it with |
UX Review (Opus 5.5) — 🟡 CONCERNSUX-level review of UX-Verdict: CONCERNS The new "saved for everyone using this Kiro Crew" wording makes a single-user setting read as a shared change, and it made the cold reader hesitant to flip switches. Watch
Suggestions
[UX-REVIEWED] ff3fdb0 Description read[DESCRIPTION-READ] d4f9ad7a732351b328595336b5322a6f49116db97821c70917b92c37c1170dbc That sha256 names everything this verdict read: the PR description, plus the 7 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 5.5 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 — 🔴 changes requested (blocking)GPT 6.1 found at least one blocking issue that must be resolved before merging This comment is updated in place on each push. BLOCKING -- website/src/hooks/usePreviewFlagServerSync.ts:251 -- A failed carry lets another flag resurrect a deleted config value
Adjudication (Opus 5.5) — is blocking on each finding proportionate?F1 (fenced): I'm flagging this, not upholding it, but it keeps blocking either way. The defect is real. It can only happen when four rare things line up, and the damage is a visible preview switch that one click fixes.
🏷️ Fenced finding(s) machine-flagged as likely edge caseThe security fence keeps these findings blocking regardless of adjudication; the only clearance path is a human override. The annotate-only pass judged the condition combination each one requires extreme, and pre-drafted the override rationale below. A repository writer must independently verify a rationale before posting it — it is machine-authored, and a wrong override on a security-class finding ships exactly the class the fence exists to stop.
False positive or not applicable? A repository writer can comment: |
bolichen97
left a comment
There was a problem hiding this comment.
Not ready yet. The design is sound: gateway config as the source of truth with localStorage as a cache fixes the root cause in #21396, and the PATCH allowlist and loader only accept the six known names as bools. Two things remain.
Head reviewed: ff3fdb0. Bots read: Design PASS, First Principles CONCERNS, UX CONCERNS, Opus no blocking, GPT BLOCK (fenced), readiness action required.
🔴 0 must-fix · 🟡 2 should-fix (inline).
- The GPT fenced finding at
usePreviewFlagServerSync.ts:251keeps readiness blocked. I traced it on this head and it is real: in a browser that has not synced, a failed carry of one switch leaves the next reconcile free to carry up a different switch whose cached ON came from the server. The fix is small (see inline 2), so fixing it is cheaper than an override. - A non-owner session (for example a QR session) reads config, so its cached ON is carried up on every load and gets a 403. The notice then says "Reload the page to try again", which never helps (inline 1).
A question, not a blocker: Design Review asked whether a non-owner dashboard session can still use previews. After this change such a session cannot turn any preview on; the switch bounces back with the owner-only notice. The Backwards compatibility section names the take-away, but no Reader line covers that session. If this is the intended behaviour for the issue owner, please say so in the description and I will withdraw this point.
The First Principles and UX CONCERNS on this head have no reply yet. Please answer or rebut each one.
Verified by reading: the backend allowlist, loader filter and if_unset path on main; that the config GET route is not owner-gated while PATCH is; that a config refresh frame invalidates ['kirocrewConfig']. Not verified: the cause of the Backend Tests (Windows) (1) failure on this head (log not available yet).
| {writeError && ( | ||
| <ErrorNotice | ||
| testId="feature-previews-save-error" | ||
| message={writeError.carry |
There was a problem hiding this comment.
🟡 A carry refused with 403 tells the user to reload, but a reload repeats the same 403.
persist() records { flag, denied, carry: true }, and this branch checks carry before denied. A non-owner session can read config, so each page load finds the cached ON, sends the if_unset carry, and gets a 403. The notice then says the switch "couldn't be saved for everyone" and asks for a reload. The reload sends the same request and fails the same way, so the user is told to repeat a step that cannot work.
message={writeError.carry
? i18nT('pages.settings.developerPanel.feature_previews_carry_failed', { name: previewFlagLabel(writeError.flag) })Why it matters: the only next step offered is a dead end, and each load also logs one more denied config write for that session.
Suggestion: when writeError.carry && writeError.denied, show a message that says only the owner can save it for every browser and that it stays on here, without the reload step. Optionally, do not retry a carry that was refused with 403 on later loads. A test for a 403 carry notice would pin it (the current carry tests use 500 or assert only the error object).
| continue | ||
| } | ||
| const server = fp[previewFlagName(flag)] | ||
| if (typeof server === 'boolean') { |
There was a problem hiding this comment.
🟡 A switch whose cached value came from the server can still be carried up as if it were a pre-upgrade value.
This confirms the GPT fenced finding at line 251 and suggests a fix on these lines. In a browser that has not synced, a failed carry of one switch sets carryFailed, so the marker is never written. If another switch was applied from the server here (line 238) and that key is later removed from config, the next reconcile sees a cache-only ON, finds the switch missing from carried, and sends it back with if_unset. The removed key comes back on.
if (typeof server === 'boolean') {
if (readPreviewFlag(flag) !== server) cachePreviewFlag(flag, server)
continue
}Why it matters: the impact is small (a visible preview switch turns back on), but this fenced finding is the only open BLOCK on readiness. A fix in a few lines clears it without an override.
Suggestion: record that the value came from the server when you apply it, for example carried.add(flag) in this branch. That stops the carry in the same document. To also cover later loads of a browser that never synced, keep a per-flag marker in localStorage when the server value is applied, and treat a cache-only ON as off for a flag with that marker. A test: unsynced browser, carry of switch A fails with 500, the server returns crew: true, then a body without crew; assert no if_unset PATCH for crew.
Rule: GPT 6.1 Review fenced finding on this head.
|
Superseded by #21722. |
Pull request was closed
|
@bolichen97 thanks for the review. This PR is superseded by #21722, which addresses all three points:
|
Problem / Motivation
Goal: A Feature previews switch (such as Crewmates) stays on once you turn it on, in every browser and in the desktop app that use the same Kiro Crew.
The switches under Settings > Developer > Feature previews were kept only in browser localStorage (
mc-preview-*). A new origin or port, the desktop app next to a browser tab, or cleared site data each read the switch as off, so Crewmates and the other previews kept turning themselves off.Why it matters
Anyone trying a preview has to turn it on again on every surface and after every storage reset. A preview that seems to switch itself off reads as broken.
Not a goal
DURABLE_PREF_KEYS//api/ui-prefsbackup, so there is one source of truth.configKey/ settings-registry entries for the new paths (<SettingRef>deep links). That is a separate navigation feature.What changed (motivation → approach → change)
Symptom: a switch turned on in one surface is off in another. Root cause: localStorage was the only copy.
#21396 was filed by its owner asking for the switches to sync across surfaces, so moving them from per-device to gateway-wide is the requested behaviour, not a side effect.
Approach: the gateway config keeps each switch, and localStorage stays as a fast cache. Config rather than adding the keys to
DURABLE_PREF_KEYS: the uiPrefs backup restores only on a cold profile, so a warm desktop app and a warm browser tab would never pick up each other's change. The config route reaches every open dashboard through the configrefreshframe it already receives.sections_dashboard.py,loader_builders.py,loader.py,sections.py,handlers/core.py,change_card_catalog.py):dashboard.feature_previews.<name>, for the 6 live flags:webhooks,artifact_deploy,crew,remote_crew_chat,layout_harness,folder_conductor.configuration.mdandconfig-baseline.jsonare updated.mc-preview-instance-sessionskey is not kept on the gateway. Its ON is carried up asremote_crew_chat.usePreviewFlagServerSync.ts, mounted inApp.tsxbeside the shell's existing reader of the same config query, so boot order is unchanged):if_unset, so it never replaces a stored value.kc-feature-previews-syncedmarker makes a cache-only ON read as off, so a key the operator removed stays off. The marker is not set while a toggle is in flight, or after a refused toggle or a refused carry, so a browser-only value is never wiped.{{productName}}), without the word "gateway".en-XAis regenerated.Overlap: open PR #20420 also edits
DURABLE_PREF_KEYS. This PR does not touch that list, so the two do not conflict.Backwards compatibility
Compatible: the new config fields are optional and default to off, the PATCH allowlist only gains names, and an existing browser-only ON is carried up once instead of being dropped.
The diff takes one thing away: localStorage is no longer the source of truth for a switch. Once synced, a cache-only ON that the gateway does not hold reads as off, and a non-owner's toggle no longer overrides the gateway's stored value.
Reader: website/src/utils/previewFlags.ts:readPreviewFlag -- chat|crew page -- still reads the cache, which reconcile keeps equal to the gateway copy | previewFlagServerSync.test.tsx "lets the server copy win over a stale browser copy"
Reader: website/src/hooks/usePreviewFlag.ts:usePreviewFlag / useRemoteCrewChat -- chat|crew page -- re-renders on the same cache write event | previewFlagServerSync.test.tsx "turns Crewmates on from the server when this browser has no copy"
Reader: website/src/surfaces/registry.ts:surfacePreviewEnabled (nav rail, pagesProvider, appNavHidden, CommandBarOverlay) -- chat|crew page -- reads the cache via readPreviewFlag | previewSurfaces.test.tsx
Reader: website/src/pages/settings/FeaturePreviewsSection.tsx -- crew page -- toggles still write the cache first, then the gateway | previewSurfaces.test.tsx "says why a switch went back when the gateway refuses the write"
Reader: retired mc-preview-instance-sessions -- chat -- carried up as remote_crew_chat, then cleared | previewFlagServerSync.test.tsx "carries the retired Remote crew sessions switch up as Chat on a crew", "never lets the retired key turn Chat on a crew back on after it was turned off"
Tests
website/src/test/previewFlagServerSync.test.tsx(30 tests). The repro: Crewmates turns on from the server when localStorage is empty, and it fails with the sync disabled. The other tests cover:if_unsetcarry-up, not replacing a stored value, a 409 counted as synced, and a refused carry reported and kept on;PREVIEW_FLAGSand every declared flag constant.I removed each of these lines on a scratch copy, and each removal makes a test fail:
pending.sizeguard;website/src/test/previewSurfaces.test.tsx: the section shows the owner-only notice naming the switch when the PATCH returns 403.website/src/test/featurePreviewsSectionNotice.test.tsx(10 tests), with the sync mounted as the shell mounts it:It lifts
FeaturePreviewsSection.tsxabove the 80% per-file coverage floor (93.9% of lines locally, across the tests that render it).test/test_feature_previews_config.py(25 tests):if_unsetwrites an absent switch and never replaces a stored one;Local runs on this commit:
tsc(app),eslint src/ --max-warnings 0andjscpdare clean.i18n-check.mjsand thesrc/i18ntests, including the[changed-values]French and Korean style checks, pass withI18N_BASE_REFset to the base commit.test_feature_previews_config.py,test_facade_mirror_guards.py,test_config_baseline.pyandtest_change_card_setting_classes.pypass (116). The full backend run on the previous cut matchedorigin/main's failure set on this host, which come from the host itself (enterprise profile, liveKIROCREW_HOME, no userns). black, isort, flake8 and mypy are clean on the changed Python.Local AI review: GPT (
gpt-6.1-sol) and Opus read-only reviews ran over several rounds. Each quoted the commit SHA it read, and I checked those against the worktree. Every Critical/High/Medium finding is fixed in this commit, and the final round on the last code change found none open.Known CI timeout, unrelated to this diff:
Backend Tests (3.12, 7)fails ontest/test_session_allocation_bindings.py::test_every_name_patched_through_the_facade_reaches_every_owner_reading_it, which hits pytest-timeout's 120 s limit on all four attempts (latest run). The same test takes 107.65 s onmainin CI (main run) and about 81 s locally on bothmainand this branch. It is the same timeout as #21441.Coverage Gatefails closed only because that backend job failed.Manual verification
I took screenshots against the built SPA, with
/api/config/kirocrewreturningcrew: true:Screenshots / video
Synced from the gateway (Crewmates on, empty browser storage):
A toggle refused with 403, naming the switch:
A toggle refused with 500 (the same notice shows for an expired session's 403), naming the switch:
A browser-only switch whose carry-up failed:
Related Issues
Fixes #21396
Related: #16237, #17424, #14232, #9876
Supersedes #21602 and #21569.
Pattern harvest
Rule candidate: review-prompt
Pattern: a user-visible setting persisted only in browser localStorage, when it is meant to hold across surfaces of the same gateway; check whether it should live in gateway config with localStorage as a cache.
Checklist
feat|fix|docs|style|refactor|perf|test|chore|ci|build|revert: ...)Contribution License Agreement
Contributions are accepted under the project's Apache-2.0 license; see the Licensing section of CONTRIBUTING.md.