02: Native Assembly

免责声明

这一页目前使用生成式 AI 撰写与维护的临时草稿。内容虽已确认正确，仍请对照源代码再次查核。

Native assembly builds contributions and orders the completed blocks through src/utils/text/context/promptLayout.ts. Contributor execution order allows RAG retrieval to overlap other work. The registry defines emitted prompt order and system routing; the numbered pages follow build stages and retain their URLs when the prompt order changes.

Persistent instructions, server knowledge, Always On lore, and conditioning enter the system zone. Participant profiles and personal memories remain above history. Changing information is positioned at the tail slot, after preset reassembly when applicable.

Contributor index

Contributor	Relationship
Prompt items	Establishes system and persona instructions.
Server info	Supplies the conversation’s server context.
Server memories	Reads persistent server knowledge for the active persona lineage.
Server emojis and stickers	Describe expressions available for output.
Participants	Resolves participant identity and personal context.
Short-term memory	Supplies recent memory and deferred refresh instructions.
Verbatim tool definitions	Adapts tool schemas for text-based calling.
RAG documents	Retrieves relevant uploaded document excerpts.
Lorebooks	Splits Always On system content from activated tail-slot content.
Conditioning	Adds server-configured reinforcement.
Sample dialogues	Shows persona response examples.
Dialogue history	Supplies recent messages and capability-neutral media descriptors.

Output contract

Directives and deferred short-term memory items are kept out of contextItems so that the downstream chat pipeline can position them relative to assembled dialogue turns:

• tailDirectives receive the impersonation command directive (Imitate <User>, start your message with <User>:). <User> uses the history’s resolved speaker label, and the stream cleaner strips that label as well as the webhook name.
• uncensorDirective receives cleaned uncensor text when enabled.
• nudgeItem holds the cadence-gated update_short_term_memory prompt.
• memoryInjectionItems holds content needing positional placement: fresh or cross-channel STM uses the tail slot or a closer configured depth; stale same-channel content uses its saved depth, with -1 staying above history.

Shared helpers used across contributors

Several contributors import shared text and state utilities:

• convertMentions (src/utils/text/context/mentionNormalizer.ts): Normalizes <@id>, <#id>, <@&id>, and macro placeholders ({bot}, {user}) into readable names. It respects user privacy levels and blacklist settings. In dialogue message bodies, it preserves {bot} and {user} literals to avoid clobbering raw prompts.
• toolPromptMacroResolver (src/utils/tools/toolPromptMacros.ts): Evaluates {{if capability:...}} and {{if tool:...}} conditionals, and substitutes provider-specific tool function names (such as {short_term_memory_tool} or {sticker_tool}) based on active model capabilities and deliberate tool allowlists.
• history.ts (src/utils/text/context/history.ts): Provides media window filtering, timestamp formatting, image limit tracking (MEDIA_IMAGE_MESSAGE_LIMIT), and user presence formatting for participant cards.

Source pointers

• src/utils/text/context/nativeBuilder.ts: buildContextNative pipeline orchestrator.
• src/utils/text/context/mentionNormalizer.ts: Discord mention normalization and placeholder substitution.
• src/utils/tools/toolPromptMacros.ts: tool name and capability macro evaluation.
• src/utils/text/context/history.ts: shared history filtering and presence formatting helpers.