05: Buffer Management
此内容尚不支持你的语言。
Buffer management accumulates streamed text into semantic buffers and flushes discrete segments to stage 06 when delivery boundaries are detected. It prevents broken markdown formatting across message splits and isolates private thinking blocks.
Flow and ownership
Section titled “Flow and ownership”The stage 04 orchestrator calls StreamBufferFlusher.processTextChunk() for every normalized chunk of
type "text". Buffering and boundary detection proceed in four steps:
- Chunk deduplication:
deduplicateIncomingTextChunk()inspects incoming text against recent accumulated output. Overlapping characters re-emitted across provider chunk boundaries are trimmed. - Model replay history: Non-empty text deltas append to
context.currentTurnModelPartsso the adapter can replay partial assistant output when reconstructing function interaction history. - Semantic block routing: Incoming text routes to
state.thinkBlockBufferwhile inside<think>tags, tostate.detailsBlockBufferwhile inside<details>tags, or to the mainstate.buffer. Drained thinking content collects instate.thoughtRawSegmentsfor thought logs, while details blocks collect instate.detailsSegmentsfor short-term memory. - Boundary flushing:
processStreamBufferContent()evaluatesstate.bufferin a loop, detecting flush boundaries (code block transitions, newlines, and sentence endings) and forwarding segments toStreamSegmentProcessor.sendBufferSegment().
Flush types
Section titled “Flush types”Beyond the per-chunk evaluation loop, three flush paths handle specific lifecycle events:
- Final flush (
flushFinalBuffer): Invoked when the generator finishes. Sends remaining buffer text, auto-closes unclosed inline markers viaautoCloseStreamBufferMarkers(), drains unterminated think or details blocks, flushes held orphan punctuation, and executes aggregated-mode batch sends. - Pending flush (
flushPendingBuffer): Invoked on tool calls or user stops. Flushes the buffer up to a clean boundary. When called before tool calls, trailing incomplete clauses are trimmed to prevent partial narration from reaching Discord. - Overflow flush: When
state.bufferexceedsFLUSH_BUFFER_SIZE_REGULAR(1,000 characters) orFLUSH_BUFFER_SIZE_CODE_BLOCK(15,000 characters) without hitting a normal boundary,findRegularOverflowFlushIndex()locates a safe sentence or word break to flush segments down to size.
Marker balance and holds
Section titled “Marker balance and holds”To prevent messages from splitting inside styled text spans, hasIncompleteSemanticMarkers() defers
flushes when unbalanced markers are detected:
- Parentheses: Counts unclosed opening parentheses. The count clamps at zero so orphan closers
(such as emoticons
:)orB)) cannot stall stream splitting. - Quotes and brackets: Balances straight quotes and paired brackets from
PAIRED_QUOTE_MARKS. - Spoilers: Holds on odd counts of
||markers in prose to keep multi-line spoilers together. - Emphasis: Tracks open
*and~~runs using renderer flanking rules. Underscores are excluded from stream holds because they appear frequently in identifiers and kaomoji. - Partial tags: Holds the buffer when it ends with a prefix of
<think>or<details>.
Markdown table atomicity
Section titled “Markdown table atomicity”Markdown tables must arrive at stage 06 as intact blocks; partial tables fail parsing and deliver as raw pipe characters rather than rendered PNG graphics:
- Overflow snapping:
findRegularOverflowFlushIndex()passes candidate indices tosnapFlushIndexOutOfMarkdownTable(). If an index lands inside a table, it moves before the table starts or after it ends. If no safe cut exists, the function returns 0 and holds the buffer for the final flush. - Scoped marker auto-close:
autoCloseStreamBufferMarkers()counts unbalanced markers over prose only. Closers land at the end of the last prose segment rather than on table rows, preventing cell count alterations that break table rendering.
Source pointers
Section titled “Source pointers”src/utils/discord/stream/bufferFlusher.ts:StreamBufferFlusherimplementation.src/utils/discord/stream/bufferManager.ts: boundary detection, marker balance checks, and overflow indexing.src/utils/text/markdownTable.ts: table boundary detection and block parsing.src/types/stream/types.ts:StreamStateand buffer threshold constants.