Prevent offline sessions with empty join links from entering the 5-minute join-link notification flow, omit blank link lines from direct reminders, and add offline persistence/reminder regression coverage.
Add format and location steps to the Telegram /newsession wizard, persist offline addresses in sessions.location_address, and render online links/offline addresses in schedule messages.
Bump version to 3.10.0.
After the shared create handler persists sessions, create a Telegram topic when needed, send the schedule/signup message, and store thread_id/batch_message_id/topic_created_by_bot for the batch. Add a Testcontainers regression test for the wizard SubmitDraftAsync happy path. Bump version to 3.9.9.
Add a PostgreSQL integration regression test for new-platform-group session creation. The production failure was a missing Platform parameter in the group_managers insert, leaving @Platform in SQL and causing PostgreSQL 42883. Bump version to 3.9.8.
The previous comment claimed a 0 MaxPlayers would also be blocked, but
the code never actually enforced that. The wizard's text-input path
already guarantees cap >= MinCapacity, so 0 is unreachable. Drop the
inaccurate half-sentence to keep the comment faithful to the code.
After 3.9.6 fixed long-polling, the bot finally reaches the final
'✅ Создать' step. Users pressing '♾ Без лимита' on the Capacity step
get a valid payload where Single.MaxPlayers = null (the legitimate
no-limit choice from GameCreationWizard.ApplyCapacityChoice 'no_limit'),
but CreateSessionHandler.IsComplete then reports 'лимит мест' as
missing, blocking session creation.
This regression existed since 3.9.3 (when 'no_limit' was added) but
stayed invisible because 3.9.4 and 3.9.5 never reached SubmitDraft
(libgssapi-krb5 missing → long-polling hung). Once 3.9.6 restored
polling, the bug surfaced immediately.
Fix: drop the null-MaxPlayers check from IsComplete for Single type.
Null is a valid 'no limit' state and must pass through to BuildCommands
→ shared handler, which already accepts null MaxPlayers correctly.
Closes#131.
Bump version 3.9.6 -> 3.9.7
Telegram bot's long-polling hangs after the first GetUpdates request
because libgssapi-krb5.so.2 is missing from the runtime-deps:10.0-noble
final image. .NET runtime attempts dlopen() of libgssapi during the
HTTPS handshake; without the library the HttpClient connection pool
enters an unrecoverable state and TelegramBotService never receives
new updates, even though SessionSchedulerService keeps sending
outgoing messages successfully.
Symptom (Loki, container gmrelaybot-bot-1):
Telegram bot polling started
Polling error, retrying in 5s
Telegram.Bot.Exceptions.RequestException: Bot API Service Failure
Cannot load library libgssapi_krb5.so.2
After the single Polling error, no Error handling update, no further
Polling error, and getUpdates from outside returns [] forever.
Fix: install libgssapi-krb5-2 alongside wget in the final stage of
src/GmRelay.Bot/Dockerfile. This also future-proofs Npgsql GSS/SSPI
Kerberos authentication for PostgreSQL.
Closes#129.
Bump version 3.9.5 -> 3.9.6
Fix two wizard FSM bugs reported after v3.9.4:
1. Capacity waitlist buttons could still advance the draft without a
numeric MaxPlayers value. The final submit validation then rejected
the draft with 'Не заполнены поля: лимит мест'. Now waitlist:on/off
stay on Capacity until MaxPlayers is set; users must either enter a
numeric limit or explicitly choose '♾ Без лимита'.
2. PickClub computed NextAfterVisibility before SetClubId, so the first
club click left the wizard on PickClub and the second click advanced.
Now ClubId is saved first and NextAfterVisibility is evaluated after
that mutation, so a valid club click advances on the first try.
TDD:
- WaitlistChoiceWithoutCapacity_StaysOnCapacityStep covers waitlist:on/off.
- PickClub_ValidGuid_AdvancesToPublishOnFirstClick covers the single-click club path.
- Stale Capacity waitlist callback test updated to the safer no-advance contract.
Closes#127
Test-only patch release. Sync the four canonical version sources:
- Directory.Build.props
- compose.yaml (bot, discord, web image tags)
- .gitea/workflows/deploy.yml (VERSION env)
- src/GmRelay.Web/Components/Layout/NavMenu.razor (visible nav-version)
The new NavMenu_ShouldExposeCurrentProjectVersion test reads the
version from Directory.Build.props, so this bump does NOT need a
hand-edited test literal — verified locally that the test passes
on 3.9.4 with no manual changes.
Two non-blocking suggestions from the PR #124 code review:
1. Symmetric coverage for the Discord PoolSlotCapacity no-limit button.
The Capacity step already had a customId-shape assertion for the
'♾ Без лимита' button; PoolSlotCapacity only had a label-presence
assertion. If ChoiceButtonCustomId's wire format ever diverged between
the two steps, only Capacity would catch it. Convert the single Fact
to a Theory with two InlineData rows so a regression in either step
produces a targeted test failure.
2. Brittle hard-coded version literal in NavMenu_ShouldExposeCurrent
ProjectVersion. The test had to be hand-edited on every version bump
(we hit this in PR #124 — bumping 3.9.2 -> 3.9.3 broke CI). Read the
version from Directory.Build.props via XDocument instead so the test
only fails when the rendered NavMenu actually disagrees with the
canonical version, not when someone forgot to update a literal.
Tolerant of whitespace, comments and attribute order; the <Version>
element is a plain string body in the MSBuild schema.
No production code changes; production version is bumped in a
follow-up commit.
Closes#125
The version-bump commit changed the visible version in NavMenu.razor
from 3.9.2 to 3.9.3 but this test hard-coded the old literal. Update
the assertion to match the new release tag.
Sync the four canonical version sources for the patch release:
- Directory.Build.props
- compose.yaml (bot, discord, web image tags)
- .gitea/workflows/deploy.yml (VERSION env)
- src/GmRelay.Web/Components/Layout/NavMenu.razor (visible nav-version)
In the session creation wizard (Telegram + Discord), the Capacity step
only exposed waitlist on/off buttons. The 'no waitlist' button silently
advanced to the next step without setting MaxPlayers, so users who tried
to create a session with no player cap were blocked with
'Не заполнены поля: лимит мест'.
The DB contract and CreateSessionCommand already supported null
MaxPlayers (int?, ck_sessions_max_players check in V006), and the web
form already exposes 'Без лимита' as an empty InputNumber — only the
wizard flow was broken.
Changes:
- Add '♾ Без лимита' choice button to Capacity in shared
WizardStepViewBuilder.BuildCapacity (Telegram) and to
RenderCapacity / RenderPoolSlotCapacity in DiscordWizardStep (Discord).
- Add 'no_limit' branch to GameCreationWizard.ApplyCapacityChoice that
sets MaxPlayers to null and advances to Visibility.
- Change GameCreationWizard.SetMaxPlayers signature from int to int? so
the 'no limit' branch compiles.
- Change CreateSessionCommand builder in both Telegram and Discord
submitters to take int? maxPlayers and drop the '?? 0' that would
have turned null into 0 (violating the DB CHECK and the 'no limit'
contract).
- In Discord BuildConfirmDescription, render '👥 Без лимита, waitlist
вкл/выкл' when MaxPlayers is null (the previous code silently
omitted the line).
- Expose BuildCommand as internal in both submitters and add
InternalsVisibleTo('GmRelay.Bot.Tests') to the DiscordBot assembly
for unit-test access.
Tests (9 new):
- WizardStepRenderTests.CapacityStep_HasWaitlistButtons — asserts the
'Без лимита' button is present.
- GameCreationWizardStepTransitionsTests.NoLimitCapacityButton_… —
asserts the choice advances to Visibility and leaves MaxPlayers null
in the JSON draft.
- GameCreationWizardStepTransitionsTests.ChoiceCallback_AdvancesToExpectedStep —
new Theory row for Capacity/no_limit.
- CreateSessionHandlerBuildCommandTests (new) — null/value propagation
through the Telegram submitter's BuildCommand.
- DiscordWizardStepCapacityRenderTests (new) — 'Без лимита' button is
rendered for both Capacity and PoolSlotCapacity, with the expected
custom-id shape.
- DiscordWizardSubmitterBuildCommandTests (new) — null/value
propagation through the Discord submitter's BuildCommand.
Closes#123
Trivy install keeps failing on the deploy workflow:
- empty TAG: install.sh falls back to 'latest', but the
'latest' GitHub release tag is no longer published.
- pin v0.71.0: pin alone is not durable. The release got
unpublished and install.sh now dies with
'unable to find v0.71.0 - use latest'.
Switch to the official aquasec/trivy Docker image:
- Docker Hub tags are content-addressed and rarely removed,
so the pin is durable.
- The image manifest ships linux/amd64, linux/arm64, linux/ppc64le
and linux/s390x, so the same tag works on the GitHub-hosted
runner and on the ARM64 Pi runner.
- We just need docker pull + docker cp /usr/local/bin/trivy.
- Pinned to 0.70.0 (April 2026) for reasonably current CVE data.
Also normalize .gitignore to plain UTF-8. The working copy had
been re-saved as UTF-16 LE by a Set-Content call without
-Encoding UTF8 (PowerShell 5.1 default on the local Windows box),
so git kept reporting 'Binary files differ' even though the rules
themselves were fine. Re-wrote through the editor to drop the
UTF-16 encoding and added rules for showcase-*.png, *.png.local,
and the test scratch dirs that keep creeping in.
The previous commit accidentally pulled in .opencode/tmp/, .playwright-mcp/,
.superpowers/, and a handful of local screenshots/logs because
'git add -A' was used during the 3.9.2 fix. None of these affect the
build or the deploy; the deploy workflow was triggered by the version
bump and ran cleanly. Add them to .gitignore and untrack so the next
contributor doesn't commit them again.
The 3.9.1 hotfix only repaired WizardDraftRepository, the most common
Dapper call in the wizard. The same AOT-unsafe CommandDefinition pattern
remained in 4 other places that the user hit immediately after the
deploy: the 'Choose visibility' wizard step triggers GetOwnerClubsAsync
when the user picks 'Публичная в витрине клуба' or 'Только для членов
клуба'. The wizard swallowed PlatformNotSupportedException, the
callback ack replied with '⚠️ Ошибка', and the next step never rendered.
Privacy 'didn't stick' from the user's perspective.
Two changes to fix the Discord side as well:
1. Switched GetOwnerClubsAsync / LoadClubsAsync / LoadManagerUserIdsAsync
to the direct (sql, params) overload across TelegramWizardMessenger,
DiscordWizardMessenger, DiscordWizardInteractionModule, and
DiscordPermissionLookup — same pattern as the 3.9.1 fix.
2. Added Dapper.AOT module attribute ([module: Dapper.DapperAot]) and
InterceptorsPreviewNamespaces to the DiscordBot project. The
DiscordBot assembly was previously skipped by the AOT source
generator, so even the direct-overload fix wouldn't have produced
interceptors for the Discord-specific Dapper call sites. With this
addition, the generator emits 3 DiscordBot-specific interceptors
(DiscordWizardMessenger, DiscordWizardInteractionModule,
DiscordPermissionLookup) and the AssemblyLoad ships with the right
GmRelay.DiscordBot.generated.cs.
Also expanded the AOT shape regression tests to cover all 4
CommandDefinition sites + added a 'containingClass' parameter to
ExtractMethodBody to disambiguate the duplicated LoadClubsAsync names
in DiscordWizardInteractionModule.
Bumps: 3.9.1 -> 3.9.2.
The previous 'curl ... | sh -s -- -b /usr/local/bin' call passed no
positional tag, so the install script fell back to the GitHub 'latest'
tag. aquasecurity/trivy no longer publishes a 'latest' release tag, so
the CI failed at 'Install Trivy' with:
aquasecurity/trivy crit unable to find '' - use 'latest' or see ...
This blocked the entire 3.9.1 hotfix deploy: build-and-push succeeded
(3 fresh 3.9.1 images pushed to git.codeanddice.ru), but scan-images
never ran and deploy was skipped. Production still runs 3.9.0 with the
broken wizard.
Pass 'v0.71.0' as the positional tag; v0.71.0 has Linux-ARM64 and
Linux-AMD64 builds so both the deploy runner (RPi 5) and pr-checks
runner pick the right tarball.
Production regression in 3.9.0: Telegram bot silently dropped every update.
WizardDraftRepository.GetActiveAsync was called on every Telegram update
(via UpdateRouter -> TryGetWizardContext) and threw
System.PlatformNotSupportedException in NativeAOT, because Dapper.AOT 1.0.48
only generates interceptors for the (sql, object?) extension overloads and
NOT for the (CommandDefinition) overload. The runtime then fell back to
Dapper.SqlMapper.CreateParamInfoGenerator, which uses Reflection.Emit and
fails on AOT. TelegramBotService swallowed the exception, so /newsession
appeared to start but no button press reached the wizard and no session was
created.
Two related changes in WizardDraft:
1. Switched WizardDraftRepository.* from 'new CommandDefinition(sql, params,
cancellationToken: ct)' to the direct 'connection.Query*(sql, params)'
overload, matching the working pattern in JoinSessionHandler. Dapper.AOT
now generates CommandFactory30<WizardDraft> + RowFactory17<WizardDraft> +
QuerySingleOrDefaultAsync37<WizardDraft> for all four methods.
2. WizardDraft.CreatedAt/UpdatedAt/ExpiresAt are now DateTime (UTC) instead
of DateTimeOffset. AOT RowFactory calls reader.GetDateTime() directly and
does not perform DateTime -> DateTimeOffset conversion; the previous type
raised InvalidCastException on the very first wizard_drafts query.
All 588/590 tests pass (2 pre-existing skipped, +5 new AOT regression tests
in WizardDraftRepositoryAotShapeTests). dotnet format clean.
Bumps: 3.9.0 -> 3.9.1.
Note: GetOwnerClubsAsync (Telegram/Discord), DiscordPermissionLookup, and
DiscordWizardInteractionModule.GetOwnerClubsAsync still use CommandDefinition
and will hit the same Reflection.Emit AOT failure when the user reaches the
PickClub visibility step. Follow-up in 3.9.2.
VERDICT from verifier (D:\Projects\Game\docs\review-report.md):
REQUEST_CHANGES — wizard was functionally broken at runtime.
## Critical
C-1. Choice-button customId was missing the 'choice:' segment.
ButtonCustomId emitted 'wizard:btn:<step>:<value>' but the
dispatcher's switch matches parts[1] == 'choice'. Every choice
button (D&D 5e, Pathfinder, Waitlist, Publish, Confirm) fell
into the default branch and showed 'Unknown button'.
Fix: split into 3 customId helpers:
ChoiceButtonCustomId(step, value) -> 'wizard:btn:choice:<step>:<value>'
ControlButtonCustomId(action) -> 'wizard:btn:<action>:1' (back/cancel/skip/create)
ModalTriggerButtonCustomId(modalStep) -> 'wizard:btn:modal:<modalStep>'
Bulk-rewrote all 66 Btn() call sites in DiscordWizardStep.cs.
C-2. "Другое…" modal-trigger buttons were unrouted in dispatcher.
Added 'parts[1] == "modal"' branch that opens the modal via
InteractionCallback.Modal(BuildModal(parts[2], draft.ChatId)).
C-3. DiscordWizardSubmitter was leaking ex.Message from
CreateSessionHandler to the user-visible draft embed. Postgres
exceptions expose schema/constraint names. Replaced with
generic user-facing error; full exception still logged
server-side on the existing catch block.
## I-3 — parser-roundtrip tests (the gap that let C-1/C-2 through)
Added two real behavioural tests (not string-grep) to
DiscordWizardInteractionModuleSourceTests:
- Renderer_And_Dispatcher_Agree_On_Wire_Format
- ControlButtons_Are_Parsed_As_Control_Not_Choice
These mirror NetCord's [ComponentInteraction("wizard")] prefix
strip, run the parser, and assert the dispatcher would route to
the right branch. Catches the entire class of 'renderer and
dispatcher disagree on the wire format' regressions.
## I-6 — BuildResumeRow (cascading fix from C-1)
After C-1, BuildResumeRow's ButtonCustomId('cancel', '1') would
emit the wrong format. Switched to direct format strings
('wizard:btn:cancel:1', 'wizard:btn:resume:continue', etc.) which
match the dispatcher's 'back'/'cancel'/'create'/'resume' cases
directly, not the 'choice' prefix.
## Version sync (3.8.0 -> 3.9.0)
Directory.Build.props: <Version>3.9.0</Version>
compose.yaml: all 3 image tags -> 3.9.0
Version_ShouldBeSynchronizedForDiscordFeatureRelease test now green.
## Stats
build: 0 warnings, 0 errors
format: 0 of 279 files need changes
tests: 583 passed, 2 skipped (pre-existing), 0 failed
files: 7 changed, 226 +, 79 -
- deploy.yml: VERSION 3.8.0 -> 3.9.0 (docker image tag for next push to main)
- NavMenu.razor: visible version v3.8.0 -> v3.9.0
- CampaignTemplatesNavigationTests.NavMenu_ShouldExposeCurrentProjectVersion: expected v3.7.1 -> v3.9.0 (was broken since 3.8.0 bump in commit 71080ae)
- RELEASE_NOTES.md: prepend Minor 3.9.0 entry (Discord wizard, issue #112) with full file inventory
- docs/review-brief.md: code-review spec for the verifier session
Build green (0 warnings, 0 errors), dotnet format clean.
The dispatcher rejected every valid select menu and modal submit
because of an off-by-one in the customId parts-length check
(NetCord strips the matching 'wizard' prefix and passes the
remainder, so 'wizard:select:Visibility' arrives as
'select:Visibility' = 2 parts, not 3).
Also fixed: WizardClubLookup.LoadClubsAsync returned an empty
list, making the PickClub step always show 'no clubs'. Now
queries the DB via NpgsqlDataSource with the same Owner|CoGm
role filter the messenger uses.
Build green, 190/192 wizard+Discord tests pass (2 pre-existing
skips), format clean.
The previous commit (f095209) shipped a DiscordWizardInteractionModule
whose MaybeOpenModalAsync helper was a documented no-op: the handler
called SendResponseAsync(DeferredMessage) early, then tried to swap
the deferred response for a Modal via ModifyResponseAsync, which
NetCord forbids (the response type is locked after the first call).
As a result, the wizard's button click that advances to a text-input
step (Title, Description, Cover, DateTime, Capacity, PoolSlot*…)
edited the draft embed but never popped the modal, leaving the user
stuck on a step that demanded popup input.
This commit restructures the dispatcher:
- Run the wizard FIRST (a separate REST call to edit the draft embed;
no interaction response is touched yet).
- Then send the interaction response as either
InteractionCallback.Modal(modalProperties) when the new step is in
the OpenModal set (Title, Description, Cover, DateTime, Capacity,
PoolSlotDateTime, PoolSlotCapacity, SystemFreeText,
DurationFreeText, PoolSystemDurationFreeText), or
InteractionCallback.DeferredMessage(MessageFlags.Ephemeral) otherwise.
- The Modal handler's draft.Step = wizardStep / originalStep restore
hack is removed: the wizard's mutation of draft.Step must persist
to the DB (the wizard already called _drafts.UpsertAsync before we
get control back), so restoring locally would only mask the truth
from the next interaction's GetActiveAsync.
- The Resume:continue path re-renders the current step via the
messenger and acks the click with a deferred ephemeral.
- The Create path delegates to DiscordWizardSubmitter.SubmitAsync and
acks the click with a deferred ephemeral.
- The constructor now assigns _messenger (was unassigned, caught by
nullable-flow analysis).
Also adds deliverable.md in the repo root describing the full Discord
adapter for issue #112.
Build green. 190/190 Discord+Wizard tests pass (2 pre-existing skipped).
dotnet format clean. The previous 12 source-level smoke tests still
pass — they assert on file shape, not runtime flow.
Adds the missing inbound handlers for the Discord wizard that the
previous commit (b81d865) left out. Three thin NetCord module classes
share one WizardInteractionDispatcher:
- DiscordWizardButtonModule
- DiscordWizardStringMenuModule
- DiscordWizardModalModule
Each registers a single [ComponentInteraction("wizard")] method that
hands the args string to the dispatcher. The dispatcher parses the
custom-id tail (btn:choice:<step>:<value>, btn:back, btn:cancel,
btn:create, btn:resume:continue, btn:resume:restart, select:<step>,
modal:<step>), looks up the active draft by (platform="Discord",
ownerId=userId), and routes through the shared
GameCreationWizard.HandleInteractionAsync. The "create" callback
delegates to DiscordWizardSubmitter.SubmitAsync (3-retry finalize).
Program.cs gets 4 new singleton registrations (the dispatcher plus
the three module classes) and 2 new AddComponentInteractions calls
(StringMenu + Modal). The existing Button registration is unchanged.
12 new source-level smoke tests in DiscordWizardInteractionModuleSourceTests
cover the file shape: 3 handler classes, 3 base classes, 3
[ComponentInteraction] registrations, all 5 callback kinds parsed,
GameCreationWizard wired in, submitter invoked on create, Program.cs
registers all 3 AddComponentInteractions and all 4 module classes,
draft lookup by GetActiveAsync("Discord", ...), modal walks
Components[0] -> TextInput -> .Value, string menu reads
SelectedValues[0].
Build green. 190/190 Discord+Wizard tests pass (2 pre-existing
skipped). dotnet format clean.
Add Discord adapter for the platform-neutral wizard moved to Shared in
the previous commit. Six new files in src/GmRelay.DiscordBot/Features/
Sessions/Wizard/:
- DiscordWizardContextStore: IWizardContextStore abstraction +
thread-safe in-memory impl keyed by draft id. Holds the (guild,
channel, message) coordinates the messenger needs to re-send the
draft after a 15-minute interaction token expires.
- DiscordWizardStep: renderer for all 15 wizard steps. Returns an
embed plus an IReadOnlyList<IMessageComponentProperties> that mixes
ActionRow buttons with StringMenu select menus. Also exposes
BuildModal for the 8 modal-collecting steps.
- DiscordWizardMessenger: IWizardMessenger impl backed by NetCord's
RestClient + NpgsqlDataSource. Edit falls back to re-send on
401/403/404. Toast replies are stashed in the existing
DiscordInteractionReplyCache.
- DiscordWizardSubmitter: 3-retry finalize loop. Builds the shared
CreateSessionCommand and calls CreateSessionHandler; on success
edits the message to "ok Created: N sessions", on failure shows
retry/cancel buttons.
- DiscordWizardCommand: /newsession-wizard slash command with an
optional mode param (single|pool). Owner/co-GM check via the
shared group_managers table.
- DiscordPermissionLookup: small helper that loads DB manager ids
for a guild.
Program.cs gets 5 new singleton registrations (IWizardDraftRepository,
IWizardContextStore, IWizardMessenger, GameCreationWizard,
DiscordWizardSubmitter). The slash command is auto-discovered by
AddApplicationCommands<SlashCommandInteraction, SlashCommandContext>()
+ AddModules(typeof(Program).Assembly).
Build green. All 85 wizard tests + 95 Discord tests pass.
dotnet format clean.
Open: DiscordWizardInteractionModule (button/modal handlers) is not
yet implemented; the bot starts and /newsession-wizard works to the
point of posting the first embed, but subsequent button clicks won't
be handled. A follow-up commit will add the component-interaction
module.
Moves the game-creation wizard state machine, view builder, and
platform-neutral contracts (callback data, step names, storage
exception, club option, step limits) from GmRelay.Bot to GmRelay.Shared.
Telegram continues to work through a new TelegramWizardMessenger
implementing IWizardMessenger and a WizardInteractionMapper that
converts Update → WizardInteraction. Wires the new platform column on
wizard_drafts (V032 migration) and switches chat/owner/thread/message
ids to TEXT so the same table can hold Discord snowflakes later.
- GameCreationWizard: now in Shared, takes IWizardMessenger +
IWizardDraftRepository, dispatches on WizardInteraction.
- New IWizardMessenger contract with Edit/Send/Answer/GetOwnerClubs
(returns string ids so Telegram longs and Discord snowflakes both
fit).
- New WizardStepViewBuilder in Shared returns
(text, IReadOnlyList<WizardAction>); TelegramWizardMessenger
renders actions into InlineKeyboardMarkup via a new Bot-side
ToInlineKeyboard helper.
- New WizardInteractionMapper in Bot (5-case test) converts Telegram
Update to WizardInteraction.
- WizardDraft gains a Platform column; ChatId/MessageThreadId/OwnerId/
DraftMessageId switched to string. V032 migrates existing rows and
rebuilds the owner lookup index on (platform, owner_id).
- All existing wizard / create-session tests updated to the new
contract (HandleInteractionAsync + WizardInteraction). Wizard
callback-data format preserved.
- dotnet build clean, dotnet format --verify-no-changes clean, all
101 wizard tests pass.
Game-creation wizard: replace /newsession text template with step-by-step
inline-keyboard flow for single games and game pools.
- Directory.Build.props: 3.7.1 -> 3.8.0
- compose.yaml: pin bot, discord-bot and web images to 3.8.0
- deploy.yml VERSION env: 3.7.1 -> 3.8.0
- NavMenu.razor: nav-version 3.7.1 -> 3.8.0
🤖 Generated with Claude Code
@
- Extract IWizardDraftRepository interface for testability (NSubstitute cannot
mock sealed classes; the codebase uses fake-style doubles instead).
- Add step-transition, pool-slot, validation, cancel/back, and render-shape tests
using FakeWizardDraftRepository and FakeWizardMessenger.
- Fix wizard payload persistence bug: HandleCallbackAsync and HandleTextAsync
now call SavePayload after ApplyChoice/ApplyText mutations, so subsequent
LoadPayload calls see the user's progress. Previously, local WizardPayload
mutations were discarded and the wizard reset on every step.
- CommitCurrentPoolSlot now auto-creates a slot via EnsureCurrentPoolSlot when
one is missing, so the PoolSlotCapacity → waitlist click is recoverable
even if the user lands on the step without a slot.
Toutsu
Coder