02.6: Participants

此内容尚不支持你的语言。

The participant contributor keeps profiles and personal memories above history while emitting presence and the clock separately for the tail slot.

Preparation and hydration flow

Participant processing separates discovery from prompt rendering across two main phases:

prepareParticipantContext(sanitizedMessages, ...)
  │  (extracts seeds from visible authors, mentions, references, bridges)
  ▼
PreparedParticipantContext { discoveryPlan, seeds, matrixUsers, syntheticUsers, ... }
  │
  ▼
buildParticipantContextItems(params)
  │
  ├─ hydrateParticipantProfiles(seeds, activePersonaScope, ...)
  │    ├─ load user rows, privacy levels, blacklist status
  │    ├─ fetch member roles and presence details
  │    ├─ load personal memories (filtered by lineage and corpus tags)
  │    ├─ load user reminders and persona self-tasks
  │    └─ resolve unique mention handles and collision indices
  │
  ├─ renderParticipantPrompt(hydratedProfiles, ...)
  │    └─ separate profile, presence, and current-context text
  │
  └─ convertMentions(renderedText, ...)
       → profile item plus presence and current-context items

Discovery preparation

Callers invoke prepareParticipantContext in src/utils/text/participants/preparation.ts before calling buildContext. Discovery inspects visible message authors, bridge webhooks, and explicit text references across the history window. It produces an ordered ParticipantDiscoveryPlan containing typed ParticipantSeed records.

A ParticipantRequestScope attached to the turn caches discovery results across multi-persona turns within the same request.

Profile hydration

hydrateParticipantProfiles in src/utils/text/participants/hydration.ts receives seeds alongside an activePersonaScope (active persona ID, lineage ID, alter flag, and impersonation state). Hydration enriches each participant:

Discord member details: Resolves display names, server roles, and live presence details (online status and current activity via getUserPresenceDetails).
Privacy and moderation: Checks cached user privacy levels and blacklist status. Users with PrivacyLevel.FULL have custom nicknames and personal memories hidden.
Personal memories: Loads persona-scoped personal memories from personalMemoryRepository.loadForUserLineage. When memory_tagging_enabled is active, memories are tag-filtered against conversationCorpus.
User reminders and persona tasks: Loads pending reminders for human participants matching the active persona. Independently loads pending self-reminders (self_reminder = true) assigned to the active persona_id.
Mention handles: Assembles alias catalogs (server nicknames, usernames, global names). Normalizes handles and detects collisions so that only unambiguous handles are recommended as ping handles.

Prompt rendering

renderParticipantPrompt in src/utils/text/participants/renderer.ts groups names, safe mention handles, personal memories, reminders, and persona tasks into profile text. It returns presence and the channel/time text separately. buildParticipantContextItems emits those as KNOWLEDGE_USER_STATUS and KNOWLEDGE_CURRENT_CONTEXT; the registry places them at the tail slot on native and preset paths. Profiles retain the tool-target metadata. Hydration decides what is safe to include before rendering.

Constraints and rationale

Profile ownership: KNOWLEDGE_USERS_IN_CONVERSATION remains a user-role reference block above history. Presence durations and clock changes cannot alter its text.
Collision-aware mention handles: Aliases claimed by multiple participants are dropped from the recommended mention handles list. The model is instructed to clarify rather than emit an ambiguous ping.
Privacy enforcement: Personal memories, pronouns, and physical appearance tags require at least PrivacyLevel.MINIMAL. Blacklisted users receive no memory recall or custom nickname expansion.
Reminder and task scoping: User reminders require context membership and an active-persona match (main personas also match legacy unassigned reminders). Persona self-tasks require an exact active persona_id match and do not require human participant membership. Persona self-tasks are omitted during user impersonation.
Downstream target index: The contributor attaches participantTargetIndex and a compatibility conversationUsers projection to the context item alongside text output. Downstream tool execution uses this index to resolve user parameters back to Discord snowflakes.
Preset reassembly: Tagged as KNOWLEDGE_USERS_IN_CONVERSATION, which SillyTavern preset reassembly flushes before the first dialogue anchor (dialogueExamples or chatHistory).

The contributor also attaches a memory-only projection for NovelAI. It uses the hydrated visible personal_memories entries, preserving resolved contents, internal memory identities, and recency. NovelAI receives plain memory contents without added owner labels, tool IDs, or tag prefixes. Other providers receive their existing participant text without a second memory block; the nested system-role memory projection is for NovelAI.

Source pointers

src/utils/text/context/participants.ts: buildParticipantContextItems.
src/utils/text/participants/preparation.ts: prepareParticipantContext and request-scope caching.
src/utils/text/participants/hydration.ts: hydrateParticipantProfiles and privacy/memory loading.
src/utils/text/participants/renderer.ts: renderParticipantPrompt.
src/utils/text/participants/targetIndex.ts: ParticipantTargetIndex and downstream tool targeting.