07: Discord Delivery

本頁內容尚未翻譯。

Discord delivery formats, paces, and transmits normalized segments to Discord. It manages delivery modes, typing simulation, webhook identity routing, and Discord API rate caps.

Flow and ownership

Delivery is shared by StreamMessageDelivery.sendSegment() and StreamUiUpdater.sendSinglePayload(). The flusher passes normalized segments from stage 06 to sendSegment(), which selects the delivery mode and chunks text:

Aggregated mode (Live Streaming off): Queues segments into state.pendingAggregatedText. Output is withheld until a tool call, table attachment, or final flush, delivering content in a single batch. Identity overrides flush queued text immediately to keep webhook identities aligned with their specific lines.
Streaming mode (Live Streaming on): Splits segments into Discord-safe chunks via chunkMessage(). Chunks send immediately or pace through sendChunksWithTyping().

StreamUiUpdater.sendSinglePayload() executes Discord REST calls, routing payloads through webhooks or standard bot messages based on context.

Message chunking and formatting

chunkMessage() parses text into typed blocks to fit within Discord message limits:

Preserved spans: Protects regions that must not split across message boundaries, including fenced code blocks, URLs, markdown formatting (**, *, ~~, `, markdown links), paired quotes, balanced parentheses, and custom emoji tags. An oversized code block repeats its opening fence and syntax language tag on subsequent messages so syntax formatting survives.
Emoji isolation: Custom emojis isolate into individual messages unless positioned mid-sentence or inside numbered or bulleted list items. Consecutive emojis sharing an initial name prefix merge into emoji runs.
Humanizer splitting: In streaming mode, paragraph breaks divide messages. Sentence Splitting also splits at sentence boundaries and drops closing periods. Force Lowercase and Punctuation Splits run humanizeString(); Punctuation Splits rolls punctuation marks between deletion, splitting, and preservation while keeping styled markup spans intact.

Paced messages

With Paced Messages on, sendChunksWithTyping() calculates delays scaled to character count, bounded by minVisibleDurationMs (750 ms) and maxTypingTimeMs (4,000 ms). The same toggle adds random pauses between follow-up messages. Delays poll the stop registry every 250 ms through interruptibleDelay() and exit promptly if a stop or interrupt arrives. The queue sends Discord’s typing indicator regardless of the humanizer toggles. Force Lowercase also applies when Live Streaming is off.

Dispatch and identity routing

StreamUiUpdater.sendSinglePayload() routes messages through one of two channels:

Webhook delivery: Used when alter personas or render-modifier identity overrides are active. The updater routes sends through dispatchWebhookWithIdentity() with persona avatars and names. The initial message answering a source message sends a standalone reply notice built by buildWebhookReplyNoticePayload() under the same selected identity. Invalid webhook errors (Discord codes 10015 and 50027) trigger cache invalidation and retry with a recreated webhook.
Bot messages: Used for the main persona’s base appearance through single-attempt channel delivery, with native reply references when applicable.

Single-attempt delivery lives in src/utils/discord/webhook/singleAttemptMessage.ts and posts at most once. Neither route carries a nonce, so a timeout, reset, or 5xx after the request left ends as unconfirmed instead of being retried. Rate limits are the exception, because Discord accepted nothing: the module waits them out for up to 30 seconds and runs the caller’s stop check throughout. Its REST client never leaves that module, so a new at-most-once send calls one of its functions instead of building a client.

sendWebhookMessageWithIdentity() and its batch variant use the managed pool dispatcher in webhookCore.ts. Local avatars reuse verified image assignments across threads; a deferred avatar PATCH can create an additional slot or deliver with Discord’s generic avatar. Local-avatar selection and sending hold one parent-channel lock, including the reply notice and its following body. HTTPS overrides bypass that lock, and lifecycle lookup uses separate coordination. The dispatch result includes the actual webhook ID and delivered identity; continuity records that identity, including a generic-avatar fallback. Avatar fallback is decided before the send operation; an uncertain POST failure propagates without another avatar-recovery send.

Streaming commits a PersonaDeliveryReceipt through recordPersonaDelivery() in src/utils/discord/webhook/personaPost.ts. The receipt contains the actual message and destination IDs, bot or webhook route, authoring webhook ID, delivered identity, speaker kind, and available sprite metadata. A generic-avatar fallback records the generic avatar. Bot delivery clears webhook adjacency; an alter’s bot fallback has no visible persona attribution. Sprite mappings for history reconstruction are written after successful webhook delivery.

sendPersonaPost() accepts the acting persona separately from a webhook hint. It resolves the current base and inherits that persona’s last committed own appearance in the destination channel. Ordinary and identity sprites survive intervening user, bot, and other-persona messages. Copied identities and explicit user impersonation cannot replace that memory. A delivered base revert replaces the remembered sprite. With no memory, posts use the current base. The bounded appearance store and its invalidators are described in Caching.

Appearance memory is separate from the channel adjacency used to choose group-break usernames. Adjacent posts retain the exact delivered username; later posts resolve the remembered sprite against current sprite data. Webhook selection still uses the managed pool and the thread’s parent, while receipts and appearance state use the actual destination thread. Expressions, fallback notices, media, and voice resolve the active persona through this same author contract, including tool-only turns. The channel’s last speaker never selects a later artifact’s author.

The shared sender distinguishes sent, pre_dispatch, definite_refusal, and unconfirmed. Avatar preparation finishes before dispatch, and cancellation is checked immediately before each send. Held drafts publish no receipt or appearance until approved delivery. The stream may recreate a definitely invalid webhook; an uncertain POST cannot trigger its webhook retry or bot fallback. Raw voice and expression transports use the same author resolution and pool selection, then commit one receipt containing the selected identity and authoring webhook ID. Standard sends use isolated REST clients with transport retries disabled. A proven invalid-webhook refusal permits one recreation; a definite component refusal permits the media helper’s one component-free downgrade. Avatar recovery finishes before POST. Lost responses, AbortError, and uncertain server failures end recovery.

deliveryState is shared by reference across provider contexts, tool delivery, and approved draft presentation. Once a POST is unconfirmed, the turn clears held text and stops before response review, key rotation, model fallback, or another tool call. The orchestrator suppresses error notices before returning that failure. No successful receipt or appearance is inferred from that outcome. A successful earlier message keeps its own receipt.

context.deliveredMessageRefs retains successful stream-body references across attempts. Button expiry and reply-notice cleanup resolve each message’s authoring webhook ID. The shared deliveryReceipts array also receives successful reply notices, so each recorded delivery describes the message that actually landed.

Rendered markdown tables

Tables detected in stage 06 bypass chunkMessage(). They are rasterized to PNG via renderMarkdownTableToPng() and delivered as attachments with an interactive Show Markdown button:

Clicking the button replies ephemerally with the table’s cached source markdown.
Source text is read on demand from markdownTableCache rather than held in event closures.
The button disables after SHOW_MARKDOWN_BUTTON_TIMEOUT_MS (2 hours), matching the cache TTL.
Table atomicity in stage 05 guarantees the block arrives intact; see stage 05 table atomicity.

Constraints and rationale

Deterministic stops over retries: Hitting server send_message_limit or the safety cap STREAMING_LIMITS.MAX_FLUSH_COUNT (40 messages) queues an internal stop. Raising an error would trigger futile fallback retries against an unchangeable limit.
Quiet teardown for inaccessible channels: If a channel is deleted or access is revoked (Discord error 10003 or 50001), stopForUnreachableChannel() terminates the stream without retrying.
Native reply limitations: Webhooks cannot create native Discord reply references. The standalone reply notice provides the only visual link between an alter persona message and its source prompt.

Source pointers

src/utils/discord/stream/messageDelivery.ts: StreamMessageDelivery and message pacing.
src/utils/discord/stream/uiUpdater.ts: StreamUiUpdater.sendSinglePayload and webhook routing.
src/utils/text/processors/chunkProcessor.ts: chunkMessage and span preservation.
src/utils/discord/markdownTableButton.ts: interactive table button and cache interaction.
src/utils/discord/stream/channelDeliveryContinuity.ts: delivered identity recording.
src/utils/security/rateLimiter.ts: STREAMING_LIMITS.MAX_FLUSH_COUNT definition.