02.1: Prompt Items

Aviso

Esta página usa borradores temporales escritos y mantenidos por IA generativa. Consulta el código fuente si necesitas confirmar algún detalle.

Prompt items form the top of the context list, establishing the LLM’s identity framing and behavioral rules.

Flow and ownership

The contributor buildPromptContextItems in src/utils/text/context/templates.ts emits up to four system-role items depending on channel configuration and impersonation status:

1. System prompt slot (SYSTEM_HUMANIZER_RULES):
   • In standard turns, this slot contains the server’s configured system_prompt or the built-in DEFAULT_SYSTEM_PROMPT.
   • When the active channel has a prompt override configured in replace mode, the channel prompt replaces this slot’s content.
   • When a SillyTavern preset is active, presetOwnsSystemPrompt skips both the configured system_prompt and the default, so the preset’s system blocks control the prompt. A replace-mode channel prompt still fills the slot.
   • During user impersonation, this slot instead holds impersonatedUserPrompt.
2. Channel prompt block (SYSTEM_CHANNEL_PROMPT):
   • When the active channel has a prompt override configured in append mode, the channel prompt is emitted as a distinct block immediately after the system prompt. Under SillyTavern preset reassembly, this block accompanies the main marker so it remains adjacent to the system prompt.
3. Persona prompt block (SYSTEM_PERSONA_PROMPT):
   • Emits the persona’s distinctive prompt instructions (personaPrompt). Omitted during user impersonation.
4. Personality attributes (SYSTEM_PERSONALITY):
   • Emits the persona’s personality traits (tomoriAttributes, joined by newlines). Omitted during user impersonation.

Transforms and expansion

Before emission, every prompt text undergoes macro and mention expansion:

• Capability and tool macros: toolPromptMacroResolver.expand prunes inactive {{if capability:...}} and {{if tool:...}} conditionals, and substitutes active tool names such as {short_term_memory_tool} or {memory_tool}. A macro whose tool the turn does not expose renders as “an unavailable tool” rather than the tool’s name, because a model shown a tool it cannot call tends to write the call out as chat text. Conditionals and macros both resolve against provider capabilities and the turn’s Deliberate Tool Mode allowlist, so seeded prompts wrap every macro in a matching {{if tool:...}} block.
• Mention normalization: convertMentions resolves <@id>, <#id>, and {bot} / {user} tokens. Because prompt items instruct the model rather than addressing a specific user, the user name argument is fixed to "User".

Constraints and rationale

• Impersonation isolation: Impersonation turns emit only the impersonated user prompt. Persona prompts, attributes, and default humanizer rules are suppressed so the model adheres strictly to the target identity.
• Channel override scope: Channel overrides alter only the system prompt slot. They never alter or overwrite persona prompts or personality attributes.
• Prompt tag stability: Distinct metadata tags (SYSTEM_HUMANIZER_RULES, SYSTEM_CHANNEL_PROMPT, SYSTEM_PERSONA_PROMPT, SYSTEM_PERSONALITY) allow SillyTavern preset reassembly to map each block to its corresponding preset marker (main, charDescription, charPersonality).

Source pointers

• src/utils/text/context/templates.ts: buildPromptContextItems and DEFAULT_SYSTEM_PROMPT.
• src/utils/cache/channelPromptCache.ts: getCachedChannelPrompt channel override lookup.
• src/utils/tools/toolPromptMacros.ts: toolPromptMacroResolver conditional and macro evaluation.
• src/utils/text/context/mentionNormalizer.ts: convertMentions mention conversion.