Repository navigation
feat(members): a Deploy my crew panel, read-only and crew-wide - #12400
Conversation
Design Review (Fable 5) — ✅ PASSDesign-level review of Design-Verdict: PASS Verify the deployed / gone / unknown liveness transitions once against a live gateway teardown — every frame here comes from fixtures mirroring [DESIGN-REVIEWED] f2d645c |
UX Review (Fable 5) — 🟡 CONCERNSUX-level review of UX-Verdict: CONCERNS The feature's only entry, the bare "Cloud" chip, was read correctly but only as a guess — the panel's discoverability rests on that one word. Watch
Suggestions
[UX-REVIEWED] f2d645c |
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: |
Opus 4.8 Review — ✅ no blocking findingsReviewed Review detailsBased on my independent review: the discovery pass reported no candidates, so there is nothing to falsify in Step 1. For Step 2, I traced the production logic in No findings. [OPUS-REVIEWED] f2d645c Verdict parsed from the review's SHA-scoped output markers for commit False positive or not applicable? A repository writer can comment: |
First Principles Review (Fable 5) — ✅ PASSPremise-level review of First-Principles-Verdict: PASS Open the capture frames before merge: the panel's entire justification is at-a-glance legibility, which the DOM tests cannot attest — only the screenshots do. The load-bearing premises were verified in-repo rather than taken from the description: registration is a condition of a What this change shipsInventory (10 items) — 10 justifiedIntent: let a crew owner see whether their crew is deployed in the cloud and what to do next, from the Members page — an ADDITION.
[FIRST-PRINCIPLES-REVIEWED] f2d645c |
2764f00 to
9d2d488
Compare
9586b21 to
566418b
Compare
566418b to
e53c2a3
Compare
GPT 5.6 Review —
|
GPT 5.6 Review —
|
fad0c00 to
ec7dcfa
Compare
ec7dcfa to
2fb36ea
Compare
2fb36ea to
e88c23d
Compare
e88c23d to
f567991
Compare
|
Dispositions for the advisory findings on First Principles, verify note -- UX, watch: "instance" in three senses. Of the three phrasings the blind reader quoted, "Create the instance" is the gateway's own step name ( UX, watch: the failed state explains the failure only through the recorded error. No change: the recorded error is the gateway's own message, and the "Ask the agent" hand-off on that notice is the designed remedy, which the reviewer read as such. Left as residual risk to observe with real users. UX, suggestion: reword |
buluoray
left a comment
There was a problem hiding this comment.
Verified at f5679916. Frontend-only — no .py, no route, no schema, no API-client method added.
The panel issues exactly two calls, api.cloudLaunches() and api.listInstances(), both GET. No POST/PATCH/DELETE anywhere in DeployMyCrew.tsx; every action is navigate('/settings/instances') via goToSettings, and copy uses the local clipboard helper rather than a request.
Both routes gate before reading state. /api/cloud/launch returns from _guard(...) ahead of _astore/store.list (handlers_cloud.py:492-495), rejecting a slack: origin with 403, unauthenticated with 401, and a non-owner via is_owner_dashboard_request. /api/instances rejects slack, unauthenticated and feature-disabled, and the panel treats that 403 as "unknown, do not retry" rather than retrying into it.
On "crew-wide": both GETs are gateway-global and owner-only, and what one crew sees of another is only whole-checkout deployment infrastructure (stack tag, provisioner id, raw status, profile, region, instance id). A launch ships the whole local checkout to one machine, so there is no per-crew deployment to leak. No private memory, per-member store, credential, token, or per-member path is read or rendered — from /api/instances the panel consumes only ssm_target/ssh_host for a liveness compare and never renders the instances list.
Frontend is clean: no dangerouslySetInnerHTML, no iframe, no inline style, no fixed palette or hex — only semantic theme tokens, so both schemes resolve. Every record field renders as React text. The single external link's href is built through the URL API and gated by ^[a-z]{2}(-[a-z]+)+-\d$, so a javascript: or arbitrary host cannot reach it, and it carries rel="noreferrer".
The client-side provisioner ids mirror the gateway (platform/defaults.py:674, platform/interfaces.py:1348), which matters because a mismatch would have offered a Fargate ARN as an SSM target. The dedupe into remoteCrew.ts is behaviour-preserving — the old IN_PROGRESS list equals the new set, and awaiting_signin is still matched before launchIsInFlight so it maps to signin rather than deploying.
One thing to know about the "read-only" framing, not a defect this PR adds: GET /api/cloud/launch goes through _astore, which runs a once-per-process store.reap_orphans() that globs the job dir and save()s stale non-terminal records to terminal on disk (handlers_cloud.py:138-151). It is owner-gated, idempotent-once restart-recovery bookkeeping, and the Settings RemoteCrewPanel already triggers it on the same route today. So the panel's own code never writes, but the read it issues is not side-effect-free at the store layer — worth knowing if anyone later reads "read-only" as "safe to poll from anywhere."
A finished Fargate launch's instance_id is a task ARN, which embeds the account id, and it renders in the collapsed Details line. That is the owner reading their own account id on their own local dashboard, so not a leak — noting it only because it is the kind of string that matters elsewhere.
…er state
The Crew Members page gets a labelled Cloud button that opens a panel
answering "is my crew deployed, and what do I do next?" in plain
language. The crew's faces lead; the newest launch decides one of six
states, each with at most one action: not deployed (a button into
Settings), deploying (step N of M with the name of the step under way,
re-read every few seconds, nothing to press), waiting on the user's
sign-in (a button into Settings), deployed (time since the launch was
created, its region, and on the EC2 lane a copyable Session Manager
target with one line saying what to do with it), finished but not
confirmed live (a past event in the past tense, see below), failed (the
recorded error through ErrorNotice with the agent hand-off, and the
Settings button again). Every navigation lands in Settings > Remote
Instances, which owns creating, repairing and tearing down a launch; the
panel never writes.
Only numbers the launch record can attest are shown; a duration estimate
is not one of them, so none is shown. Sessions running on
the deployment and turns it served are deliberately absent: the deployed
side has no counter and no path to report back, and the local activity
log counts the local crew, so its numbers here would state a scope the
data does not have. An unknown value is an en dash, never 0. Month-to-
date spend was measured and dropped: the costs route belongs to the
aws-control app, which is off by default, and a launch record carries no
account id to key it by.
"Deployed" is said only while the Instances registry still holds the
machine. The launch record cannot carry that claim by itself: a Settings
delete waits for AWS to confirm the stack is gone, unregisters the
instance and removes the uploaded source, and nothing ever re-statuses
the done launch record, so on the record alone the panel would read "is
deployed" forever after the one action its own button leads to. The
registry is the field a teardown does update, and the EC2 launcher
registers the instance (its id as the SSM target) as a hard condition of
done. So the panel also reads GET /api/instances, under the query key the
Settings panels already use, only once some launch has finished, and a
403 (feature off) is not retried. deployLiveness answers live, gone or
unknown; only live earns the present tense. Every other done record is
the finished state, a past event in the past tense ("A deploy finished
<age> ago in <region>."): gone adds that the machine is no longer in the
instances list and offers the Settings button; unknown (the Fargate lane
registers nothing by design, an EC2 record that kept no machine, or an
unreadable registry) adds that this page cannot tell and points at the
AWS console in that region, the sentence itself linking to that region's
console home when the record's region parses as a region id; the link
keeps its underline and carries an external-link glyph, so it reads as a
door without hovering, and its URL is built with the URL API from one
whole-URL literal. Unknown is never rounded to either answer. The panel
stays loading until both reads have answered, and a failure of the
registry read other than a 403 is reported as a read error (the same
notice, hand-off and Try again as a failed launch read), never drawn as
a state.
The target row is offered only for the built-in EC2 provisioner. The
Fargate lane records the task's ARN in the same field, and Session
Manager does not accept an ARN, so that launch gets no row and, where
the row would be, one muted line saying the deploy was a container task
with no Session Manager target; its ARN stays in Details.
The buttons name what they do and where they lead, not the outcome a
reader wants. Both only open Settings, and a reader who takes a "Deploy
to the cloud" button for the act itself (something created outside this
computer, possibly costing money, with no visible undo) does not press
it. So the labels name the navigation and the real destination, the
Settings page called Remote Instances, which opens on its Your crews tab:
"Open Remote Instances in Settings" and "Open Remote Instances in
Settings to sign in". Every pointer on the card names that page by that
one name, so three doors do not read as three places. A muted line under
the deploy button says the one thing the label cannot: nothing is created
until the steps there are confirmed. While a launch is moving or waiting
on the sign-in, a line says closing the window does not stop the deploy,
since the work runs in the gateway, not in the dialog; the moving state
also carries one muted link, "Open Remote Instances in Settings", into
the page whose launch row has the Cancel, so the panel still acts on
nothing but names where to. The header trigger carries a hairline border
like the secondary button style, so an icon-plus-word beside the page
title reads as a button, not a status chip, and a plain cloud glyph
rather than an upload arrow, so it does not read as sending something
up.
Copying the target reports its own failure. When both clipboard layers
refuse (the async API rejects, execCommand answers false, the everyday
shape of a plain-HTTP LAN dashboard), the row says the copy failed and
names the remedy instead of painting a tick over an unchanged clipboard,
and the notice carries the agent hand-off: nothing on the row is unsaved,
so the hand-off can destroy nothing, and the agent has a remedy the
message does not (read the id off the launch record and open the session
itself). The dialog closes first, as for the panel's other notices.
Technical identifiers (stack tag, provisioner id, raw status, profile,
region, instance id) survive in a collapsed Details line, one per launch,
so an owner debugging a deploy still has them. The provisioner id is
drawn verbatim there, so no client-side label map or lint exemption is
needed for it. Read failures and a failed launch's error render through
ErrorNotice with askAgent and a hand-off that closes the dialog first.
The in-flight status list lives once, in utils/remoteCrew beside the
provisioner ids (FARGATE_PROVISIONER_ID is added there, mirroring the
gateway's id; the set itself is not exported), and both this panel and
RemoteCrewPanel classify a launch through launchIsInFlight, so a status
the gateway adds is classified in one place; RemoteCrewPanel drops its
private list for it, with no change in classification. It sits there
rather than in the API client because most test files replace that
module with a hand-written mock.
The headline follows the newest launch, and one line names what that
would otherwise bury: another launch whose machine the registry still
holds keeps running (and billing) behind a failed retry, a launch still
moving, an unconfirmed finish or a re-deploy that finished, since
nothing in the launch flow tears an earlier stack down; so the panel
says so under any headline, says it can be deleted under Remote
Instances in Settings, and points at Details. The line requires the
registry's word; a done record it cannot vouch for is not named, nor is
an older record of the newest launch's own machine.
The feature-map row for this surface takes the shape of its neighbours:
one sentence on what it is for a user, the entry point, the files and
the endpoints. It narrates no mechanism, because every sentence about a
branch, a failure path or what a button re-reads is made false by the
next change to the code it describes.
A capture harness (scripts/capture-deploy-my-crew.mjs) renders the
trigger and every panel state, the failed copy included, against the
built SPA with stubbed API answers, asserting each state before it
writes a frame.
Catalogs: 28 strings added to all 13 languages plus the regenerated
pseudo-locale; French carries U+202F before its double punctuation, and
Hindi and Bengali use the informal register. With them the all-catalog
chunk measures 12281.1 KB, inside the 12870 KB ceiling the bundle-size
gate already carries on main (measured 12254.6 KB there at 34fe0de).
Problem / Motivation
A crew's cloud deployment state is reachable only by hunting through the
Settings tab that owns the whole set-up flow, and the first cut of a Members
page panel for it answered the wrong question. It listed launch records with
their stack tag, provisioner chip, raw status enum and SSM target, behind an
icon-only cloud button nobody could name. A reader who opened it could not
tell whether their crew was deployed or what to do next, and its title
promised an action ("Deploy my crew") that the panel did not offer.
The question a crew owner brings to this panel is plain: is my crew deployed,
is it still being set up, and what is the one thing I should do now?
Why it matters
The Members page is where an owner meets their crew as people, not as launch
jobs. If the one surface that says "your crew now runs somewhere other than
this computer" reads like a CloudFormation console, the feature is invisible
to exactly the people it exists for, and a reader who does find it either
waits on a launch that is waiting on them, or copies an identifier they do not
know where to paste. A panel that names the state in one sentence and offers
one action per state turns "deploy my crew" from a Settings expedition into a
button.
What changed (motivation -> approach -> change)
Goal: persona first, plain language, one obvious next step per state, and
technical identifiers demoted rather than deleted.
Approach: the panel leads with the crew's faces (the roster, rendered through
the existing
CrewAvatar, folded into a "+N" tile past six), then onesentence for the state the NEWEST launch is in, then the one thing a reader
can do about it. Every action is a navigation into Settings > Remote Instances,
which owns creating, repairing and tearing down a launch; the panel never
writes. Crew-WIDE, not per-member: a launch ships the whole local checkout to
one machine and names one stack, so the trigger stays in the page header and
the faces are the whole roster.
The five states, from
deployView(jobs), an exported pure function:awaiting_signinis split from the other moving statuses on purpose: waitingachieves nothing there, because the launch is waiting on the user.
The two buttons name what they do, not the outcome a reader wants. Both only
open Settings, and a first-time reader who takes a "Deploy to the cloud"
button for the act itself (something created outside this computer, possibly
costing money, with no visible undo) does not press it, so the panel's whole
point fails on the state every first-time reader lands on. The labels name
the navigation and the real destination, the Settings page called Remote
Instances (which opens on its Your crews tab): "Open Remote Instances in
Settings" and "Open Remote Instances in Settings to sign in". Every pointer
on the card names that page by that one name, so three doors do not read as
three places. A muted line under the deploy button says the one thing the
label cannot:
"Nothing is created until you confirm the steps there." Every other button on
the card was checked against its handler the same way: Copy copies, Try again
re-reads, Close closes, Details toggles. While a launch is moving or waiting
on the sign-in, a line says closing the window does not stop the deploy: the
work runs in the gateway, not in the dialog, and a reader who fears Close
cancels it would otherwise babysit the window. The moving state also
carries one muted link, "Open Remote Instances in Settings", into the page
whose launch row carries the Cancel, so a reader told that nothing here
stops the deploy is not left with no idea where anything would; the panel
itself still acts on nothing. The step counter names the
step under way in the gateway's own words ("Step 2 of 4 - Create the instance
and install Kiro Crew"), so the count is not a bare number.
Honest numbers only. No duration estimate is shown while a launch moves:
nothing in the repo measures a typical launch, and a figure nothing attests
is not shown here; the step counter is the progress. The deployed state shows
two: the time since the launch
was created (labelled "Since deploy", because nothing here can see whether the
machine ran the whole time) and its region. Both come off the launch record.
Sessions running on the deployment and turns it served are deliberately
absent: the deployed side has no counter and no path to report back, and the
local activity log counts the LOCAL crew, so its numbers here would state a
scope the data does not have. Month-to-date spend was measured and dropped:
the costs route belongs to the aws-control app, which is off by default, and a
launch record carries a profile and a region but no account id to key it by.
An unknown value renders as an en dash, never 0, matching the member stat
cards.
Errors go through
ErrorNoticewithaskAgenton, and a hand-off that closesthe dialog first: a failed launch list read, and a failed launch's recorded
error. Both are read surfaces with nothing to lose.
The target row. A Session Manager target is offered only for a finished launch
on the built-in EC2 provisioner that kept a machine, with one line saying what
the value is for ("Paste this ID into AWS Session Manager (in the AWS console)
to open a terminal on the machine"). The Fargate lane records the task's ARN in
the same
instance_idfield, and Session Manager does not accept an ARN, so afinished Fargate launch gets no row and, where the row would be, one muted
line: "That deploy was a container task, not a machine, so it has no Session
Manager target. Its task ID is in Details." That sentence is keyed on the
Fargate provisioner id
(
deployRunsAsTask), not on "anything but EC2": a provisioner this panel doesnot know is drawn verbatim in Details and gets no sentence, and an EC2 launch
that recorded no machine gets no row and no sentence either. A failed or
cancelled launch that still carries an
instance_idis never offered asreachable (
deployIsReachable).Copying the target reports its own failure. When both clipboard layers refuse
(the async API rejects and
execCommandanswers false, the everyday shape ofa plain-HTTP LAN dashboard), the row says "Copy failed. Select the text and
copy it manually." instead of painting a tick over an unchanged clipboard, and
the target stays readable for that manual copy. The notice carries the agent
hand-off: nothing on the row is unsaved, so the hand-off can destroy nothing,
and the agent has a remedy the message does not (read the id off the launch
record and open the session itself). The dialog closes first, as for the
panel's other notices. The failure sentence is the one the Settings panel's
own copy button already uses.
Technical identifiers (stack tag, provisioner id, raw status, profile, region,
instance id) survive in a collapsed Details line, one per launch, newest first.
The provisioner id is drawn verbatim there, which retires the client-side
label map; the lint exemption that once served its literals had already left
main independently, so this diff does not touch the lint config. The raw
status token is deliberate in that line: it is the name the gateway logs use,
and the sentence above is the plain-language version.
The header trigger is a labelled control (a plain cloud glyph + "Cloud")
rather than an icon alone, with a hairline border in the secondary button
style: an icon-plus-word beside the page title with no edge reads as a status
chip, and this is the feature's only entry. The glyph is the plain cloud, not
the upload arrow: an arrow into a cloud reads as "send something up", and a
reader who takes the button for an action never opens the read-only panel. The dialog is titled by state ("Your crew in
the cloud") rather than by an action it delegates.
The in-flight status list lives once:
IN_FLIGHT_LAUNCH_STATUSESandlaunchIsInFlightinutils/remoteCrew.ts, beside the provisioner ids, andboth this panel and the remote-crew Settings panel classify a launch through
it, so a status the gateway adds is classified in one place. It lives there
rather than on the API client because most test files replace that module
with a hand-written mock (an earlier attempt to export it from
api/clientreddened 105 tests for exactly that reason);
utils/remoteCrew.tsis mockedby none of them, and the three RemoteCrewPanel suites pass unchanged.
Declared plainly, since it reaches outside the panel:
utils/remoteCrew.tsgains
FARGATE_PROVISIONER_ID(mirroring the gateway's id beside the EC2 onealready there) and the status set behind
launchIsInFlight(the set itselfis not exported: the predicate is the one spelling);
RemoteCrewPanel.tsxdrops itsprivate
IN_PROGRESSlist and callslaunchIsInFlightinstead (+6/-5). Eachhas two consumers, this panel and the Settings panel; the classification is
unchanged, so this is a harm-free dedupe riding along, not a behaviour change.
"Deployed" is said only while the Instances registry still holds the machine.
The launch record cannot carry that claim by itself: a Settings delete
(
_teardown_after_deleteinhandlers_cloud.py) waits for AWS to confirm thestack is gone, then unregisters the instance and removes the uploaded source,
and nothing in the gateway ever re-statuses the
donelaunch record, so onthe record alone the panel would read "is deployed" forever after the one
action its own button leads to. The registry is the field a teardown does
update, and the EC2 launcher registers the instance (its id as the SSM
target) as a hard condition of
done, so for that lane "registered" and"exists" move together. The panel therefore reads
GET /api/instancestoo(
api.listInstances, under the same['instances']query key the Settingspanels use, so both read one cache entry), only once some launch has
finished, and a 403 (the Instances feature is off) is not retried.
deployLivenessanswers live / gone / unknown, anddeployViewsays"Your crew is deployed in the cloud." only for live. Anything else a
donerecord supports is a past event, and the new
finishedstate says it in thepast tense: "A deploy finished ago in ." followed by either
"Its machine is no longer in your instances list." (gone: the registry was
read and does not hold it; the set-up button and its line follow, as with no
launch at all) or "This page cannot tell whether it is still running."
(unknown: the Fargate lane registers nothing by design, an EC2 record that
kept no machine, or the registry could not be read) with one line pointing
at the AWS console in that region; that line is itself a link to the
region's console home when the record's region parses as a region id
(
awsConsoleUrl; the value comes off a stored record and becomes ahostname, so anything else gets plain words and no link). Neither form ever says the crew is
deployed, and neither says it is not; the two stats and the target row are
shown only in the live state. Unknown is never rounded to either answer,
because both roundings put a false sentence about a paying machine on the
screen. The panel stays in its loading state until both reads have answered,
so a finished launch is not drawn as unconfirmed and then as deployed. A
failure of the registry read other than a 403 is an error, not a fact about
the machine: it renders through the same
ErrorNotice(with the agenthand-off) and Try again as a failed launch read, Try again re-reads the read
that failed, and no state is drawn behind it.
The headline follows the newest launch, and one line names what that would
otherwise bury. Another launch whose machine the registry still holds keeps
running (and billing) behind a failed retry, a launch still moving, a finish
the registry cannot vouch for, or a re-deploy that finished, since nothing in
the launch flow tears an earlier stack down; so a line under the state reads
"An earlier deploy is still running in . You can delete it under
Remote Instances in Settings; its identifiers are in Details."
(
earlierReachableLaunch), under the deployed headline as much as under theothers, since that is the headline a reader has least reason to look past.
The line requires the registry's word: a
donerecord it does not hold orcannot speak to is not named, because the line's whole claim is that a
machine is running now; and an older record of the newest launch's own
machine is not named either, since two records of one machine are one
machine. It ends with where to act, because a warning about a paying machine
with no way to stop it only frightens.
The feature-map row takes the shape of its neighbours: one sentence on what
the surface is for a user, the entry point, the files and the endpoints. It
narrates no mechanism (no state list, no failure path, no "what Try again
re-reads"): every such sentence is made false by the next change to the code
it describes, while a row that states what the feature is stays true across
them. The panel's behaviour is specified by the header comment of
DeployMyCrew.tsxand pinned by its tests.Catalogs: 28 keys added across all 13 languages plus the regenerated
pseudo-locale. Hindi and Bengali use the informal register the style gates
require, and French carries U+202F before its double punctuation. With them the all-catalog chunk measures 12281.1 KB, inside the
12870 KB ceiling the bundle-size gate already carries on main (main measured
12254.6 KB there at 34fe0de), so the gate file is byte-identical to main.
Backwards compatibility
Compatible: one labelled button and one dialog on an existing page; no route,
endpoint, schema or default changes. The panel is read-only, so no write path
exists to be compatible with. A launch record from an older gateway (empty
steps, noprovider_id) renders: the step total falls back to the foursteps every launch has, and a record with no provisioner id is treated as EC2,
the only kind that predates the field. A provisioner id the panel does not
know is drawn verbatim in Details and gets neither the target row nor the
container-task sentence.
Tests
website/src/pages/members/DeployMyCrew.test.tsx(65 cases, 98.9% of thecomponent's lines). Each behaviour below was mutation-checked: the named
change to the source reddens the suite, and the un-mutated suite is green.
deployIsReachable: true only fordonewith a machine; false for everyin-flight status, for
donewith no machine, and forfailed/cancelledeven when the fixture still carries an
instance_id. Mutation: dropping thestatus check reddens 2 tests.
instance_idis a task ARN renders as finished (past tense, no claim about now) with no
target row, no stats, the console pointer, and the ARN in Details.
Mutation: dropping the provisioner gate reddens the suite.
deployLiveness: live only when the registry holds the finished EC2machine (by SSM target, or by host on a legacy row); gone when the registry
was read and does not; unknown for an unreadable registry, for the Fargate
lane, for an EC2 record with no machine, and for a launch that is not done.
Panel: a registry without the machine renders the gone sentence with the
set-up button; a 403 renders the unknown sentence with the console pointer
and is not retried (one call); a 500 renders the read error with the
hand-off and Try again, no state behind it, and Try again re-reads the
registry (the launch list is not re-read); a running launch reads the
registry not at all. Mutations: letting the record alone earn "deployed"
reddens 6 tests; rounding an unreadable registry to gone reddens 3;
rendering a failed registry read as the unknown state reddens 1.
deployRunsAsTask: true for the Fargate id only; false for EC2, for arecord with no provisioner id, and for an unknown provisioner. The
container-task sentence renders for the Fargate launch, and not for an EC2
launch with a row, nor for an EC2 launch that recorded no machine.
Mutations: dropping the gate reddens 2 tests; widening the predicate back to
"anything but EC2" reddens 2.
awsConsoleUrl: the region's console home for a region id (us-east-1,ap-southeast-2, us-gov-west-1); undefined for a missing region, a
half-id, upper case, a hostname or anything carrying a path or query.
Panel: the console sentence is a
_blank/noreferrerlink to that URLon the Fargate launch, and plain words with no link when the record has
no region. The URL is built with the URL API from one whole-URL literal,
so the i18n strings gate (which reads each template fragment alone) sees
a URL rather than a bare host fragment; the link keeps its underline and
carries an external-link glyph, so it reads as a door without hovering.
Settings" link; no deploy button, no sign-in button, no target row.
newestLaunchpicks bycreated_at, not list order. Mutation:jobs[0]reddens 2 tests.
earlierReachableLaunch: the newest other launch whose machine theregistry still holds, under any headline: behind a failed retry, behind a
newer Fargate launch the registry cannot confirm, and behind a live newest
launch on another machine; undefined when no other launch is confirmed
live (torn down, or the registry unreadable) and for an older record of the
newest launch's own machine. The line renders behind a failed retry with
the earlier region and the delete path, under the deployed headline with
the target row still the newest machine's, and never behind a retry whose
earlier machine the registry no longer holds. Mutations: falling back to
the record reddens 2 tests; silencing the line under a live newest reddens
2; dropping the same-machine guard reddens 2; never rendering the line
reddens 1.
deployView: none / deploying / signin / deployed / finished / failed;deployed only with the registry's word, else finished with gone or unknown;
awaiting_signinis signin, not deploying; with an older finished launch and a newer failedone, the panel describes the failed one. Mutations: folding signin into
deploying reddens 2 tests; picking the oldest launch reddens 1.
deployProgress: finished steps plus the one under way, over the recordedtotal; empty steps fall back to four; never past the total.
deployStepLabel: the step marked active, else the first not done, elseempty for a record with no steps; a skipped step before the active one does
not win. Mutations: silencing the label reddens 1 test; dropping the active
preference reddens 1.
deployAge: en dash for a missing or future timestamp; day/hour,hour/minute and minute bands; a zero part is dropped. Mutation: rendering a
future timestamp as an age reddens 1 test.
renders through
ErrorNotice(role="alert") with a Retry, and thenot-deployed sentence is absent.
buttons call
navigate('/settings/instances')and close the dialog first(mutation: leaving it open reddens 3 tests); deploying offers nothing to
press.
accompanies the deploy button in the not-deployed and the failed state.
Mutation: silencing the line reddens 2 tests.
settled (fake timers), the close line and the step's name are shown, and no
duration estimate is; the sign-in state shows the close line too.
Mutations: removing the interval reddens 1 test; silencing the close line
reddens 1; putting a minutes line back reddens 1.
stubbed) and the button flips to Copied with no failure notice, the Details
disclosure is closed and carries the raw provisioner id and tag; a missing
region is the en dash; a finished launch with no machine offers no target.
execCommandansweringfalse, the failure notice renders with the agent hand-off, Copied is never
painted, and the target text is still on the page. Mutations: silencing the
failure branch, painting Copied regardless, and dropping the hand-off each
redden the suite.
ErrorNotice(mutation: ahand-written
text-dangerdiv reddens 2 tests), no target is offered, thedeploy button navigates; a cancelled launch with no error renders no notice.
reddens 1 test); no face row for an empty roster.
Also run locally on the final tree:
tsc -b(0 errors), eslint on the changedsource files, the i18n vitest suites with
I18N_BASE_REF=origin/main(43files, 703 tests, including every language style gate's changed-values arm,
the dead-key ratchet and the context sidecar; without the base ref those
arms see no changed values and pass vacuously),
check-i18n-strings.mjswith the same base ref,
check-i18n-keys.mjs,check-theme-colors.mjs,check-phantom-classes.mjs, the bundle-size gate on an analyze build (697chunks within budget), the feature-map gate enforced against
origin/main,and the RemoteCrewPanel and MembersPage suites.
Manual verification
Performed against the built SPA, not a live gateway. The capture harness
(
website/scripts/capture-deploy-my-crew.mjs) serveswebsite/distthroughthe repo's shared loopback static server with every
/api/**call answeredfrom fixtures (no gateway, no AWS), drives the real Members page, and asserts
each claim before it writes a frame: the trigger's label and visible border,
the one sentence per state, the single action and the absence of any other
(the moving state's only control being the link into Settings),
both Settings buttons' labels, the cost/undo line under the deploy button and
its position below it, the close hint and the absence of any estimate while deploying, the EC2-only target
row with the reworded hint (with the machine registered in the stubbed
registry), the Fargate launch's past-tense finished sentence, its
container-task line and console pointer with its ARN confined to Details, the
failed copy's notice with its hand-off and no Copied, the landed copy's
Copied with no notice, the read error's Retry, the earlier-launch line with
its delete path behind a failed retry (and its absence when the newest launch
is live), the gone sentence with the set-up button when the registry no
longer holds the machine, the unknown sentence with the console pointer and
no deploy button when the registry answers 403 (the console sentence linking
to that region's console on both), the earlier-launch line
under the deployed headline when a second machine is registered (the target
row still the newest machine's), and the read error with the hand-off and
Try again, no state behind it, when the registry answers 500. 20 frames,
every assertion green.
Not exercised here: a launch record produced by a live gateway, a real
teardown (the gone state rests on
_teardown_after_deleteremoving theregistry row, which the fixture mirrors by answering an empty registry), a
real Fargate launch (the container-task sentence rests on the provisioner
writing the task ARN into
instance_id, which the fixture mirrors), and thebrowser's real clipboard permission prompt. The pod route refuses the pod's token mint with
403 on every start from this sandbox, including fresh ones on the rebased
tree, so no frame comes from a running gateway.
Screenshots / video
From the harness above, dark theme unless named. Fixture data throughout
(fake instance and task ids,
defaultprofile).Header trigger, labelled Cloud, with its border:
Not deployed, one action named for what it does:
Deploying, the step under way named, nothing to press, and Close does not cancel:
Waiting on the sign-in, one action:
Deployed on EC2, dark and light:
Details opened, identifiers one line per launch:
Finished on Fargate, said in the past tense, the container-task line and the console link where the row would be:
Failed, the recorded error, the Settings button and its line:
The list is still being read:
The list could not be read:
The copy failed:
A failed retry over an earlier launch that still runs:
The copy landed:
The launch finished, but a Settings delete has since unregistered its machine: a past event in the past tense, and the set-up flow again:
The Instances feature is off, so the registry answers 403: the panel says it cannot tell, never gone, and points at the console:
A re-deploy finished and the older machine is still registered: the headline and the target belong to the newest machine, and the line names the older one:
The registry read fails outright (not a 403): the same read error, hand-off and Try again as a failed launch read, and no state behind it:
Review dispositions
Current-head review findings and what was done with each:
only navigate. Fixed by changing the words to match the action, never the
other way round: both labels name the navigation and the real destination
("Open Remote Instances in Settings", "Open Remote Instances in Settings to
sign in"), and the line under the deploy button carries the cost/undo fact.
No button on this read-only card triggers a deploy. Every other button was
checked against its handler; none mismatches.
the container-task sentence, keyed on the Fargate id.
hairline border. The
+sibling keeps its ghost style; it is icon-only andout of this change's scope.
use the value. Adopted: "Paste this ID into AWS Session Manager (in the AWS
console) to open a terminal on the machine."
different mechanism sentence (the landing tab; "every button opens that
page"; the registry-failure path and what Try again re-reads), each made
false by the round that changed the code it described. Fixed structurally:
the row now has the shape of its neighbours (one sentence on what the
surface is for a user, the entry point, the files, the endpoints) and
narrates no mechanism, so no branch or failure-path sentence is left to
drift. The panel's behaviour is specified by the header comment of
DeployMyCrew.tsxand pinned by its tests. Whether the buttons should openthe Set up a new one tab directly (a
?tab=read in RemoteCrewPanel, whosetab is local state with no URL read today) is a separate product decision
and is not changed here.
and its comment named no protected draft, which is not one of the rule's
sanctioned states. Fixed:
askAgentis on (nothing on the row is unsaved),the dialog closes first, and the test pins the hand-off.
Tried "Crew in the cloud": at the harness's 1920-wide viewport it pushes
the page title to "Crew ..." in the members header, so the label stays
"Cloud" with the tooltip carrying "Your crew in the cloud". The dialog's
title names the subject the moment it opens.
Fixed: plain cloud glyph.
reassurance. Fixed: the same line renders there, reworded so it is true in
both states ("Closing this window does not stop the deploy").
Adopted: the step under way is named beside the counter from
job.steps.Copy. Deferred: Details is the debugging line for every identifier, and a
copy control there is a separate change to that line, not to this panel's
states; the ARN stays selectable text.
deployment behind a failed retry. Fixed with the line the review asked for:
earlierReachableLaunchnames the earlier deploy and its region when thenewest launch is not the deployed one.
exported set in
utils/remoteCrew.ts, consumed by both panels; theRemoteCrewPanel suites pass unchanged.
the sentence "Your crew ..."; left for a human eye, as the review asks.
measurement behind it. Fixed by deletion, as suggested: the step counter
carries the progress, and the panel's own rule (only attested numbers) now
holds for the deploying state too.
gateway record. Not possible from this sandbox (Manual verification above);
the fixture shapes mirror the gateway's launch record fields.
deployed headline read liveness off a launch record that teardown never
mutates, so a destroyed machine read as live forever. Confirmed in the
gateway (
_teardown_after_deleteunregisters and deletes the source only;no site re-statuses a
donelaunch). Fixed by keying liveness on theregistry, the truth source the review named:
deployLivenessoverGET /api/instances; "deployed" only for a registered EC2 machine; everyother
donerecord is a past-tense finished state (gone / unknown); theearlier-live line requires the registry's word. The Fargate lane registers
nothing by design, so its liveness is unknown and it is never called
deployed.
launchIsInFlight/FARGATE_PROVISIONER_IDextraction and theRemoteCrewPanel reclassification were not in the description. Declared
above (What changed): two consumers each, a harm-free dedupe rider.
way to act. Fixed as suggested: the sentence ends with the path to act
("You can delete it under Remote Instances in Settings", the page and the
verb Settings itself uses).
the review leaves it for a human to weigh, as does this PR (the longer
label truncates the page title, see above).
the set-up flow does with the failed launch. Deferred: whether the failed
launch stays listed there is Settings' behaviour, and this panel only
states what it can read; the button's own line already says nothing is
created until the steps there are confirmed.
rendered the unknown state silently, with no
ErrorNoticeand no hand-off.Fixed as the review asked: the failure routes through the same notice and
Try again as a failed launch read, is excluded from
ready, and draws nostate; a 403 stays the capability answer it is (feature off, unknown).
under a live headline, so two registered machines read as one. Fixed:
the line renders under any headline when another launch's machine is still
registered, the target row stays the newest machine's, and an older record
of the same machine is not named.
?tab=deep-link now. Not filed: therepository's owner keeps no tracking issues for features, and whether the
deep-link lands (initializer-only read of
useLocation().searchinRemoteCrewPanel.tsx) is the conductor's call, measured and recorded inthe review thread; the labels name the destination page truthfully today.
exported with no importer. Fixed as suggested: the
exportis dropped;launchIsInFlightis the one spelling.showed the earlier-live sentence without its delete remedy. Fixed: shot 13
re-captured at HEAD with the shipped sentence (and shot 16 shows the same
sentence under the deployed headline).
a possibly-billing machine with no link. Fixed as suggested: the sentence
links to that region's console home, when the region parses as one.
window does not stop the deploy" read as alarming with no way to stop it
anywhere. Fixed as suggested: one muted link, "Open Remote Instances in
Settings", into the page whose launch row carries the Cancel; the label
names the navigation and the destination, and the panel still acts on
nothing.
The Fargate sentence says there is no Session Manager target, which is what
the row is for, and does not claim the task cannot be reached.
as three places ("the cloud set-up", "the sign-in step", "Your
instances"). Fixed as both asked: every pointer names the one page by its
own name, "Open Remote Instances in Settings" (the sign-in variant adds
"to sign in"), and the earlier-live sentence says "under Remote Instances
in Settings".
Fixed: the link keeps its underline and carries an external-link glyph.
control of its own. Deferred: under the none, failed and sign-in states
the card's one button already opens that page, and under the deployed
headline the sentence names the page by its Settings name; adding a second
navigation control to that state is left for a separate decision.
gone sentence. Declined: the registry says only that the machine is not
listed, not why; the sentence states what the read supports.
Related Issues
None. This is a new surface, so the PR carries its full context.
Pattern harvest
Not generalizable: the redesign is a product decision about this one panel.
One reusable check did come out of it: a class name that exists nowhere
renders as nothing, and a usage count that includes the file under review
cannot detect it, so every class in the new component was counted with that
file excluded, and the one novel utility was confirmed in the built CSS.
Checklist
feat|fix|docs|refactor|perf|test|chore|ci|build|revert: ...)