02.9: Conditioning

Miễn trừ trách nhiệm

Trang này sử dụng bản nháp tạm thời do AI tạo sinh viết và duy trì. Dù đã được xác minh tính chính xác, bạn vui lòng đối chiếu lại với mã nguồn.

The conditioning contributor injects feedback memories detailing behaviors the persona was thanked or corrected for in past interactions.

Flow and ownership

The contributor buildConditioningContextItem in src/utils/text/context/templates.ts formats reinforcement feedback for the persona:

1. Eligibility:
   • In nativeBuilder.ts, the stage is skipped during user impersonation, when serverId is missing, when persona_lineage_id < 0, or when neither reward nor punish conditioning is enabled on the persona.
2. Database retrieval:
   • Queries conditioningMemoryRepository.loadGroupsForPersona(serverId, personaLineageId, type) for "reward" and "punish" feedback.
   • Filters out entries with blank reason text.
   • Caps each category to CONDITIONING_CONTEXT_MAX_GROUPS_PER_TYPE groups.
3. Formatting and emission:
   • Groups format with action past participles (getConditioningContextPastParticiple), actor user mentions or persona names, reasons, and repetition counts:
     • ## Rewarded Behaviors: “Here are past things {botName} did that got rewarded for. Strive to do them again:”
     • ## Punished Behaviors: “Here are past things {botName} did that got punished for. Avoid doing them again:”
   • Passes assembled sections through convertMentions.
   • Emits a single system-role item tagged KNOWLEDGE_SERVER_CONDITIONING. If both sections produce zero groups, returns null.

Constraints and rationale

interact_with_target shares the localized embed renderer with /reward and /punish, including random title variants and italic reasons. Its workspace capability, conditioning_interactions_enabled, defaults to true and is managed under /config > Plugins > Available Tools. This tool remains available in Deliberate Tool Mode when enabled, independently of stickers. Tool embeds use the invoking persona’s shared author resolution, including its committed appearance and sprite for main and alter personas. One persona delivery receipt records the delivered message and every artifact posted with it.

The tool’s reason is optional. When provided for a persona target, it saves conditioning with the active persona as the actor. Without a reason, the gesture is posted and the target can react, but no conditioning is saved. Rows contain exactly one of user_id or actor_persona_id; repeats are counted separately for each actor. Actor names are resolved when the groups are read, and feedback is included without requiring that actor to be present. Deleting an actor’s user or persona row cascades to their feedback. Managers remove mixed-actor groups through the existing conditioning panel.

After a successful post, persona targets enter the normal chat admission and channel queue with the selected target and the initiating member’s access identity. The queued turn receives an injected context item that names the actor, the selected target, the action, and the reason, with the delivered message ID in metadata. /reward and /punish use the same event builder and take the message they posted as the trigger, so a later message or embed cannot redirect the reaction. That message comes from a single-use webhook named after the member and registered to them, so history, reply targeting, and block checks credit it to that member. Where the bot cannot manage webhooks in the channel, or the webhook send fails, the command posts the embed as its own interaction reply instead. The receiving turn uses the delivered embed’s history text when present and replaces that history entry with one recent user item formatted as [System: <gesture> <tool hint>]. The trigger stays separate from adjacent notices during history merging. If the message is outside the fetched history, the injected event supplies the gesture. For hugs and kisses, the receiving turn resolves {conditioning_interaction_tool} inside {{if tool:interact_with_target}} against its own model capabilities and Deliberate Tool Mode allowlist. When unavailable, the event appears without the hint. The hint exists only for this reaction turn; later history contains the posted gesture without it.

NovelAI uses the event’s separate completionText action without a system wrapper or tool hint. The native context retained during preset reassembly receives the same event, including when the delivered message is outside fetched history. Later classified gesture notices retain their localized descriptions as action text. The stored conditioning block remains excluded for NovelAI.

The interaction tool remains available to the receiving persona and allows one call per turn. No runtime limit bounds the number of reciprocal turns. Persona-triggered replies also skip human engagement counters and short-term memory writes, so persona reactions cannot sustain Natural Nudges or establish a relationship with the initiating member. User targets receive only the embed. Cancellation is checked before writing, posting, and triggering a reply; a completed write is retained if delivery fails or is cancelled, no reaction is queued, and an uncertain send is never reposted as the bot.

• Feedback isolation: Conditioning is persona-lineage scoped. It reinforces character consistency across server turns without bleeding feedback across distinct personas.
• Impersonation suppression: User impersonation turns suppress conditioning so that past bot critiques do not bias the imitated user persona.
• Preset reassembly: Tagged as KNOWLEDGE_SERVER_CONDITIONING, which SillyTavern preset reassembly flushes before the first dialogue anchor (dialogueExamples or chatHistory).

Source pointers

• src/utils/text/context/templates.ts: buildConditioningContextItem.
• src/utils/db/repositories/ConditioningMemoryRepository.ts: loadGroupsForPersona.
• src/utils/conditioning/conditioning.ts: CONDITIONING_CONTEXT_MAX_GROUPS_PER_TYPE and past-participle mappings.
• 03: Server Memories: companion persistent memory contributor.