Command System

Nội dung này hiện chưa có sẵn bằng ngôn ngữ của bạn.

TomoriBot loads Discord slash commands dynamically from src/commands/ and dispatches interactions through a unified pipeline in src/events/interactionCreate/handleCommands.ts.

Loader and execution pipeline

Command discovery and registration are managed by src/utils/discord/commandLoader.ts.

Directory structure: The loader scans src/commands/ for root command files (subscribe.ts) and nested folders (export/personal/config.ts). Root files represent top-level commands; nested folders map to {category} {group} {subcommand} or {category} {subcommand} hierarchies.
ES module loading: commandLoader.ts uses asynchronous directory traversal and dynamic ES module imports (await import()). Command files may use top-level await.
Registration payload: Command metadata is compiled into discord.js SlashCommandBuilder structures and registered with Discord at startup via src/events/clientReady/01_registercommands.ts.
Runtime execution map: The loader constructs an in-memory CommandExecutionMap mapping each command path to its target execution function.

Single-flight loading

loadCommandData() is called from two code paths: startup command registration and the first incoming interaction in handleCommands.ts. The loader memoizes evaluation behind a shared promise (cachedCommandDataPromise).

This single-flight pattern prevents startup race conditions. If an interaction arrives while startup registration is importing command modules, concurrent evaluation would interleave ES module loading. A second loader could read an export binding while the module remains in its Temporal Dead Zone, throwing initialization errors and silently dropping commands. The promise clears on catastrophic failure or an empty execution map, allowing subsequent interactions to retry.

Import hygiene

Command modules import database repositories, which can lead to large dependency graphs. Data-layer modules (repositories, caches) must not import high-level runtime subsystems (context building, tools, webhooks, providers). Shared leaf constants must be imported directly from their owning leaf files rather than from barrel modules that re-export heavy subsystems.

Command file contract

Subcommand modules export:

configureSubcommand(subcommand)
execute(client, interaction, userData, locale)

Root command modules export:

configureCommand(command)
execute(client, interaction, userData, locale)

Root command files may export optional boundary flags:

guildOnly = true: restricts the command to Discord guild channels.
managerOnly = true: requires the Discord ManageGuild permission.
nsfw = true: marks the command as age-restricted.
isCommandEnabled(context): evaluated before registration; returning false omits the command from the registration payload and execution map.

Command metadata descriptions and choices are auto-localized by commandLoader.ts. Builders configure English text with localizer("en-US", key). The loader matches keys across authored locale trees using {option_name}_description and {choice_value}_option.

Runtime interaction dispatch

All Discord interactions arrive at src/events/interactionCreate/handleCommands.ts:

Autocomplete: If interaction.isAutocomplete() is true, routes to runAutocompleteCommand().
Chat-input slash commands: If interaction.isChatInputCommand() is true, routes to runChatInputCommand():
- Resolves category, group, and subcommand names into an execution key.
- Evaluates server blacklist status; an unreadable status blocks execution with the temporary database error, while a confirmed blacklist entry receives the blacklist warning.
- Checks category cooldown in cooldownRepository, scaling duration with COMMAND_COOLDOWN_SCALE.
- Opens an ambient error context scope with source: "command", enriching it with user and guild database IDs.
- Calls the registered execute(client, interaction, userData, locale) function.
Globally routed components and modals: If isGlobalRoutableInteraction() is true, passes the interaction to dispatchGlobalInteraction().

Globally routed persistent interactions

TomoriBot separates bounded temporary interactions from persistent UI panels:

Collector-owned workflows: Bounded command sessions that rely on in-memory message component collectors. When the collector times out or the process restarts, the controls deactivate.
Globally routed panels: Persistent controls (such as /help, /config, /moderation, /setup, and /expressions manage) where every button click, select choice, or modal submit arrives as an independent interactionCreate event. State is reconstructed from the custom ID and database records rather than in-memory collectors.

Global dispatch mechanics:

src/utils/discord/interactions/routeRegistry.ts parses custom IDs using the format <namespace>:<version>:<action>:<state...>.
src/utils/discord/interactions/router.ts checks the registered route list.
Unmatched IDs return false, allowing invocation-scoped collectors to process the interaction instead.
Outdated route versions receive a localized stale-panel notice with instructions to re-invoke the command.
Deferred panel branches call beginPanelInteraction() in src/utils/discord/interactions/panelController.ts. This helper calls deferUpdate() before authorization or state loading, preventing asynchronous database reads from exceeding Discord’s acknowledgement deadline. Branches that open modals remain outside beginPanelInteraction() because showModal() must serve as the primary acknowledgement.

Personal settings

Personal provider panels load credentials and model registrations by user ID. The invoking workspace’s persona state is optional for personal operations; server provider operations require it. Without a workspace, new personal snapshots keep sampler overrides unset and use the normal runtime defaults. Personal configuration keeps server-specific controls scoped to existing personas.

Panel failure observability

Panels communicate expected refusals by returning a status object ({ status: "write-failed" }) and repainting with an error or warning receipt. Because these paths do not throw unhandled exceptions, the router error handler does not observe them.

Panel failure metrics maintain visibility:

deliverGuardedPanel() in src/utils/discord/ui/interactionCore.ts is the single choke point for guarded panel delivery.
When delivery includes a receipt with error or warning tone, it emits a panel_failure metric containing the namespace, locale, tone, and machine reason key (PanelReceipt.reason).
Failure metrics emit to two destinations: structured host logs via log.metric, and the database via metricSampleRepository.recordSample("panel_failure", fields). The database write is non-blocking.
Metric aggregation groups on the machine reason key rather than localized heading text, so failures across different languages aggregate into uniform defect categories.

Generation model overrides

Each /generate subcommand and /novelai generate image accepts an optional model autocomplete. The choices combine compatible models from saved server providers and the invoking user’s personal providers, including providers that are not the current default. Personal choices show (via Personal Provider). The value identifies the model, provider, and credential source; execution checks scope access again. Server BYOK rules and media quotas still apply to the selected credential source. Omitting the option preserves normal routing.

The override stays in the invocation and never updates saved selections. Auto image generation keeps its text planner and passes the image selection to its tool calls. A scene carries its text selection through every queued speaker turn, without model randomization or model fallback. Voice generation checks Design and Clone compatibility and keeps an explicit model from falling back to server speech. src/utils/provider/modelCatalog.ts supplies the same scoped catalogs to configuration panels and src/utils/provider/generationModels.ts.

The optional nsfw boolean on /generate image, /generate video, and /novelai generate image controls Discord spoilers only. Auto image generation passes it to image tool delivery; it does not change the planner’s instructions, the generation request, or provider content filters.

Reward and punishment targets

/reward and /punish accept an optional user target alongside the existing persona option. Supplying both is rejected. User targets return an embed before persona loading, persona access checks, conditioning writes, or manual chat dispatch. A reason adds a playful footer without saving anything. Reasons appear in italics within the localized action sentence for both user and persona targets. Persona targets retain conditioning and manual chat behavior, including fallback persona selection when neither target option is supplied.

Interaction timing rules

Discord requires interaction acknowledgement within approximately three seconds of generation (interaction.createdTimestamp). An acknowledgement taking longer than 1.5 seconds emits a rate-limited latency warning (INTERACTION_ACK_WARN_MS). After acknowledgement, the bot has up to 15 minutes to complete work.

Reply and deferral ownership

A fast command can reply directly. Commands that need database, network, or other asynchronous work acknowledge first with deferReply(). Synchronous rejection checks can run before deferral.

Opening a modal acknowledges the original interaction with showModal(), so pre-deferral prevents the modal from opening. Pre-modal reads must fit within the acknowledgement deadline. The modal submission is a new interaction: promptWithRawModal() can auto-defer that submission when passed MessageFlags.Ephemeral, or its caller can defer the returned submission before asynchronous work.

Pattern 3A and 3B: Checkbox management modals

Used for managing collections of configuration settings (such as /moderation whitelists and /config channel rules) using Discord modal checkbox groups:

Up to 50 options per modal (chunked into five groups of 10 options).
Pre-check existing active settings; unchecking indicates removal or disabling.
Reopening preloads current database state; submission writes the complete selection or computed diff.
Database writes must invalidate associated caches in the same execution path upon transaction success.

Pattern 4: Pagination helpers (no pre-defer)

Pagination helpers (replyPaginatedChoices() and promptWithPaginatedModal()) send their controls and own the initial acknowledgement. Pre-deferral conflicts with that ownership.

Pattern 4A: Anchor message workflow

The anchor message workflow in src/utils/discord/ui/personaWorkflow.ts powers multi-step persona pickers and complex configuration transactions.

Single message invariant: One command invocation owns exactly one ephemeral Components V2 message (selection.message.anchorMessageId). That message is edited in place through every phase: selection, range pagination, modal launching, progress indicators, and terminal receipts.
Rationale: Discord emits no event when a user dismisses a modal. Leaving a picker message active while opening a modal would leave dead interactive buttons if dismissed. Rendering all phases onto one anchor message and collapsing controls when a modal opens prevents stranded components.
First-acknowledgement contract: Each phase operation owns its first acknowledgement. Calling raw interaction methods before a phase operation, or calling two first-acknowledgement operations on the same interaction, throws PersonaWorkflowUpdateError with code already-acknowledged.
Anchor message controller: The controller exposes typed operations: replace(), edit(), fetchMessage(), disableControls(), and delete(). Payloads require MessageFlags.IsComponentsV2 and components; legacy content and embeds fields are forbidden.

Timing quick reference

Command style	Defer before work?	Primary API
Simple and fast	No	`interaction.reply(...)`
Async DB or API operations	Yes	`interaction.deferReply(...)`, then helper reply
Modal prompt	No (before modal)	`promptWithRawModal(...)`, defer on submit if heavy
Pagination helper	No (before helper)	`replyPaginatedChoices(...)` / `promptWithPaginatedModal(...)`
Persona workflow	No	`runPersonaPickerWorkflow(...)`; phase operation owns acknowledgement

Response Drafting panel

/config > Plugins > Response Drafting stores workspace settings for guild managers and DM owners. Reviewer and checker dropdowns render directly, like /config > Models; picking a provider opens its model modal without changing the primary text model. Clearing the reviewer restores inheritance from the actual response model; clearing the checker stores None. A removed checker binding stays saved and unavailable.

Content to Check shows enabled review items with green circles and disabled items with red circles. Edit Review Items opens a checkbox group prefilled from the saved selection. Submitting replaces the whole selection, including an empty selection, and invalidates persona snapshots after the write.

The page shows availability, added time/cost and a shortened prompt preview. Set Prompt opens the complete instructions in a prefilled 4,000-character modal; Use Default clears the override. Persistence is described in database settings; runtime ownership is in generation review.

Contributor guides

Procedures for adding commands and authoring UI panels belong to dedicated contributor guides:

Authoring slash commands: Adding a Slash Command
Command architecture selection: Command Archetypes
Building Components V2 panels: Adding a Panel

Source pointers

src/utils/discord/commandLoader.ts: dynamic command discovery and metadata localization.
src/events/interactionCreate/handleCommands.ts: interaction routing, cooldowns, and error context attribution.
src/utils/discord/interactions/router.ts: persistent interaction routing and custom ID dispatch.
src/utils/discord/interactions/panelController.ts: panel interaction acknowledgement timing and guarded delivery.
src/utils/discord/ui/personaWorkflow.ts: anchor message lifecycle and phase operations.
src/utils/discord/ui/interactionCore.ts: guarded panel delivery and panel failure metrics.