SillyTavern Preset System

Import and use SillyTavern (ST) presets to control how TomoriBot assembles LLM prompts.

Overview

SillyTavern presets are JSON files that define:

1. Prompt node ordering — where character info, instructions, chat history, etc. appear in the final prompt
2. Custom prompt injection — additional instructions, output format rules, task descriptions inserted at specific positions
3. Depth-based insertion — placing prompts relative to the end of chat history
4. Template macros — variables like {{setvar::X::Y}} / {{getvar::X}} for dynamic content
5. Per-node enable/disable — users toggle individual prompt nodes on or off

This is distinct from SillyTavern Card Import, which imports character data (description, personality, sample dialogues). Presets control how the prompt is structured, not what character data exists.

How It Works (User Perspective)

1. User imports an ST preset JSON via /st-preset import
2. The preset becomes active for that server
3. On every LLM call, the context builder detects the active preset and rearranges blocks accordingly
4. The /sysprompt and personality settings still apply — the preset controls where they appear, not whether they exist
5. If no preset is active, the system uses the native fixed context assembly (see Context Assembly)
6. Removing or deactivating the preset reverts to native assembly instantly

Current Status

• Phase 1: Import & Visualization — implemented
• Phase 2: Template Engine — implemented
• Phase 3: Context Assembly Override — implemented
• Phase 4: Management Commands — not yet implemented (/st-preset activate, /st-preset deactivate, /st-preset delete)

Commands

For a user-facing explanation of behavior, surprises, and limitations in SillyTavern terms, use /help st-preset.

/st-preset import

Imports a SillyTavern preset JSON file and stores it for the current server.

Flow:

1. User attaches a .json file to the slash command
2. Bot validates the file (format, size <= 2 MB, and a supported preset shape)
3. If the preset already has a Prompt Manager prompts array, parses prompt_order (prefers character_id 100001, falls back to 100000) to determine node sequence and default enabled states
4. If the preset is an older text-completions export with context.story_string + sysprompt.content, converts that legacy layout into synthetic Prompt Manager-style nodes and markers
5. Normalizes legacy post-history fields carried by modern or converted presets into synthetic depth-injection nodes
6. Filters out comment-only nodes (content resolves to empty after macro stripping)
7. Stores preset metadata + raw JSON in st_presets, individual nodes in st_preset_nodes
8. Activates the preset for the server
9. Replies with an import summary (total nodes, markers, toggleable count, and warnings for enabled unsupported macros)

Preset name: Derived from the uploaded filename (minus .json extension), truncated to 100 chars. Must be unique per server.

Legacy compatibility: TomoriBot accepts modern Prompt Manager presets directly. It also accepts older text-completions presets when they provide context.story_string + sysprompt.content; those are converted best-effort into synthetic Prompt Manager-style nodes at import time. In both shapes, extra legacy post_history fields such as root post_history, sysprompt.post_history, or context.post_history are converted into synthetic depth-injection nodes instead of being ignored.

/st-preset node toggle

Shows a modal with checkbox groups representing the preset’s toggleable prompt nodes.

Flow:

1. Loads the active preset for the server (or falls back to the first available)
2. Queries st_preset_nodes for non-marker nodes ordered by node_order
3. Chunks nodes into up to 5 checkbox groups (10 options each, 50 max per modal)
4. If more than 50 nodes, shows page-selection buttons first (up to 9 pages)
5. Modal title = preset name (dynamic, truncated at 45 chars by Discord)
6. On submit, persists changed enabled states and invalidates the preset cache

/st-preset remove

Deletes the currently active SillyTavern preset for this server, reverting context assembly to native fixed-block order.

Flow:

1. Loads the active preset for the server
2. If no active preset, replies with “nothing to remove”
3. Deletes the preset (cascade deletes all nodes) and invalidates the preset cache
4. Replies with confirmation

Missing Commands (Phase 4)

Command	DB Function	Purpose
/st-preset activate	setActivePreset()	Switch between uploaded presets
/st-preset list	loadPresetsForServer()	Show all presets for the server

Template Engine (Phase 2)

The template engine resolves ST-specific macros in preset node content at context build time. Located in src/utils/text/stPresetEngine.ts.

Two-Pass Variable Resolution

Pass 1 — Collect vars: Walk all enabled non-marker nodes in node_order, applying variable declarations into a shared Map<string, string>.

• {{setvar::key::value}} replaces the current value for the key
• {{addvar::key::value}} appends to the current value for the key

Pass 2 — Resolve everything: For each enabled non-marker node:

1. Strip {{// comment }} blocks
2. Remove {{setvar::...}} / {{addvar::...}} declarations (already collected)
3. Replace {{getvar::key}} from the variable map
4. Expand content macros ({{personality}}, {{description}}, {{scenario}}, {{mesExamples}}, {{lastChatMessage}})
5. Evaluate {{random: A, B, C}} / {{random::A::B::C}} — pick a random item
6. Evaluate {{roll: XdY}} — sum X random [1..Y]
7. Process {{trim}} — trim whitespace; if empty, mark node as disabled
8. Detect HTML content — set hasHtmlWarning flag

After preset rearrangement, buildContext() also performs a final random-choice pass over assembled text parts and tail directives. This catches ST preset variants that use single braces, such as {random::apple::banana}, and ensures /tool prompt snapshot shows the same rolled text the provider receives.

Identity Macros

{{user}}, {{char}}, and {{bot}} are intentionally not resolved by the template engine. They are left intact for downstream resolution by convertMentions() in the context builder, which applies the stable “User” placeholder optimization for system-role content.

Content Macro Deduplication

The engine tracks which content macros were expanded with real (non-empty) data in a Set<string> called expandedContentMacros. This is used by the preset context builder to avoid duplication:

• If a custom node expanded {{description}}, the charDescription marker skips pulling the native persona prompt block
• If a custom node expanded {{personality}}, the charPersonality marker skips pulling the native personality block

Macro Reference

Macro	Replacement	Source
{{user}}	Triggerer’s display name	Deferred to convertMentions()
{{char}} / {{bot}}	Bot’s display name	Deferred to convertMentions()
{{personality}}	tomoriAttributes.join("\n")	Server personality settings
{{description}}	personaPrompt	Active persona’s prompt
{{scenario}}	"" (empty)	No TomoriBot equivalent
{{mesExamples}}	Formatted sample dialogues	sample_dialogues_in/out
{{lastChatMessage}}	Most recent user message	Conversation history
{{setvar::key::value}}	(removed from output)	Sets a variable
{{addvar::key::value}}	(removed from output)	Appends to an existing variable
{{getvar::key}}	Variable value or ""	Reads a variable
{{random: A, B, C}}	Random pick from list	Runtime
{{random::A::B::C}}	Random pick from list	Runtime
{random: A, B, C}	Random pick from list	Runtime
{random::A::B::C}	Random pick from list	Runtime
{{roll: XdY}}	Dice roll sum	Runtime
{{trim}}	Trim whitespace	If empty after trim, node is disabled
{{// comment }}	(removed)	Stripped entirely

Compatibility Patches

Some presets use additional placeholder conventions that fall outside the official ST macro spec — often because they rely on ST’s regex post-processing pipeline (which TomoriBot does not implement) to substitute these tokens. We resolve them directly instead.

All compatibility patches are in one location in stPresetEngine.ts for easy auditing.

Placeholder	Replacement	Observed in
<USER>	Triggerer’s display name	Marinara’s Spaghetti Recipe
<BOT>	Bot/persona display name	Marinara’s Spaghetti Recipe

These are case-sensitive (uppercase only) to avoid false positives with lowercase HTML tags.

Context Assembly Override (Phase 3)

When an active preset exists, the context builder uses a Build-Then-Rearrange strategy instead of the fixed native order. Located in src/utils/text/presetContextBuilder.ts.

Transformation Example

To understand what the preset system does, here’s a concrete before/after comparison.

Native assembly (no preset):

 1. System prompt (/sysprompt)                     [SYSTEM_HUMANIZER_RULES]
 2. Persona prompt (/persona)                      [SYSTEM_HUMANIZER_RULES]
 3. Personality attributes                          [SYSTEM_PERSONALITY]
 4. Server info                                     [KNOWLEDGE_SERVER_INFO]
 5. Server memories                                 [KNOWLEDGE_SERVER_MEMORIES]
 6. Emojis                                          [KNOWLEDGE_SERVER_EMOJIS]
 7. Stickers                                        [KNOWLEDGE_SERVER_STICKERS]
 8. Users in conversation                           [KNOWLEDGE_USERS_IN_CONVERSATION]
 9. STM                                             [KNOWLEDGE_SHORT_TERM_MEMORY]
10. RAG documents                                   [KNOWLEDGE_SERVER_DOCUMENTS]
11. Conditioning guidance                           [KNOWLEDGE_SERVER_CONDITIONING]
12. Sample dialogues                                [DIALOGUE_SAMPLE]
13. Conversation history                            [DIALOGUE_HISTORY]

Same blocks after a preset rearranges them (example preset node order):

 1. [main marker]           → System prompt          ← pulled from SYSTEM_HUMANIZER_RULES
 2. ★ Custom node: "You are a creative writing assistant. Always use vivid language."
 3. [charDescription marker] → Persona prompt         ← pulled from SYSTEM_HUMANIZER_RULES
 4. [charPersonality marker] → Personality attributes  ← pulled from SYSTEM_PERSONALITY
    ↳ FLUSH: Server info, memories, emojis, stickers  ← TomoriBot-only blocks injected here
 5. ★ Custom node: "{{setvar::style::narrative}}"     ← (removed from output, variable stored)
 6. [worldInfoBefore marker] → RAG documents           ← pulled from KNOWLEDGE_SERVER_DOCUMENTS
 7. ★ Custom node: "Write in {{getvar::style}} style." ← resolved to "Write in narrative style."
 8. [dialogueExamples marker]                          ← pulled from DIALOGUE_SAMPLE
    ↳ PRE-FLUSH: Users in conversation, STM, conditioning
                                                       ← TomoriBot-only blocks injected here
 9. [chatHistory marker]    → Conversation history     ← pulled from DIALOGUE_HISTORY
    ↳ Depth injection at depth 0: "Remember to stay in character."  ← merged into last history item

Key observations:

• The /sysprompt content still appears — it’s just at the main marker position instead of always being first
• Custom nodes (marked with ★) are new content from the preset, inserted between native blocks
• TomoriBot-only blocks (server info, emojis, etc.) have no ST marker, so they’re auto-flushed at anchor points
• {{setvar}}/{{getvar}} are resolved at build time, not stored in the prompt
• Depth injections merge into existing history items rather than creating new messages
• Marker order is literal: if a preset places chatHistory before dialogueExamples, the sample dialogues will appear after live history
• If dialogueExamples becomes the terminal block, TomoriBot appends a short user-side separator note so strict providers do not treat the final sample assistant turn as the active prompt

Why Build-Then-Rearrange?

The native buildContextNative() remains the fixed-order orchestrator for TomoriBot context blocks, with responsibility-specific helpers extracted under src/utils/text/context/ for memories, RAG, template/conditioning blocks, and history/media formatting. The preset system still treats native output as tagged buckets and does not duplicate those block builders.

Instead, the preset builder:

1. Calls native buildContextNative() to produce all blocks (tagged with metadataTag)
2. Groups items by tag into consumable “buckets”
3. Walks the preset’s node order, pulling from the right bucket at each marker
4. Inserts custom preset nodes at their declared positions

The trade-off: blocks that a preset might not use are still built (minor wasted work). The safety and simplicity gains are massive.

Routing

At the top of buildContext() (the exported entry point):

1. Is user impersonation? → Skip preset, use native (presets are character-centric)
2. Does this server have an active preset? → Check in-memory cache
3. If preset found → build native → rearrange via preset
4. If no preset → use native directly

The check uses the preset cache (getCachedActivePreset()), which avoids a DB query on every call.

Marker-to-Tag Mapping

When the preset walker encounters a marker node, it pulls items from the corresponding native bucket:

ST Marker	ContextItemTag	Native Block	Typical TomoriBot Source
main	SYSTEM_HUMANIZER_RULES (first item only), then SYSTEM_CHANNEL_PROMPT	System prompt + per-channel append prompt	/config system-prompt set (or fallback), plus /server channel-prompt in append mode
charDescription	SYSTEM_PERSONA_PROMPT	Persona prompt	/persona prompt set
charPersonality	SYSTEM_PERSONALITY	Personality attributes	/persona attribute add
dialogueExamples	DIALOGUE_SAMPLE	Sample dialogues	/persona sample-dialogue add
chatHistory	DIALOGUE_HISTORY	Conversation history	Live channel message history
worldInfoBefore	KNOWLEDGE_SERVER_DOCUMENTS	RAG documents	Retrieved document context / uploaded docs
worldInfoAfter	KNOWLEDGE_SERVER_DOCUMENTS	RAG documents	Retrieved document context / uploaded docs

Special case: main — The main marker pulls the first SYSTEM_HUMANIZER_RULES item (the system prompt) and then the SYSTEM_CHANNEL_PROMPT item if present, keeping a per-channel append prompt directly after the system prompt. In replace mode there is no separate channel block — the channel prompt has already taken over the SYSTEM_HUMANIZER_RULES content upstream. The persona prompt is carried by SYSTEM_PERSONA_PROMPT and pulled by charDescription.

These marker-controlled blocks are usually moved, not removed. The real suppressions are narrow:

• The built-in fallback system prompt is removed only when a preset is active and the user has not set /config system-prompt set
• The native charDescription block is skipped only if a custom preset node already expands {{description}}
• The native charPersonality block is skipped only if a custom preset node already expands {{personality}}

TomoriBot-Only Block Flushing

These blocks have no ST marker equivalent. They are flushed at anchor points during the node walk:

Blocks	Flushed at	Timing	Notes
Server info, memories, emojis, stickers	charPersonality, charDescription, or main marker	After the marker’s items	TomoriBot-only automatic context; no ST marker equivalent
Users in conversation, STM, conditioning, remaining RAG	dialogueExamples or chatHistory marker	Before the marker’s items	TomoriBot-only automatic context; still included even if the preset omits explicit ST markers

If the preset doesn’t include these anchor markers, remaining blocks are appended at the end before dialogue history.

Depth Injection (Critical Design)

Nodes with injection_position: 1 are depth-injected — they target a specific position counting from the end of the conversation history.

Key constraint: Depth-injected content is merged into existing dialogue history items, not inserted as new standalone messages. This prevents role-alternation violations that would break providers with strict role ordering (Gemini, Anthropic).

depth 0 = append to last history item (closest to model's response)
depth 1 = append to second-to-last item
depth N = append to Nth-from-last item (clamped to first if exceeds length)

Multiple injections at the same depth are ordered by injection_order (ascending).

Batched Injection

All injections targeting the same depth are batched into a single [System: ...] text part rather than creating one [System: ...] per node. This reduces token waste and closely matches SillyTavern’s contiguous injection behavior.

For example, a preset with 5 depth-0 nodes (XML wrappers + instructions) produces:

# Before (unbatched — each node gets its own wrapper):
\n[System: </chat_history>]
\n[System: <task>]
\n[System: Write the next response.]
\n[System: </task>]
\n[System: <output_format>]

# After (batched — one wrapper per depth target):
\n[System: </chat_history>
<task>
Write the next response.
</task>
<output_format>]

This batching is transparent — the LLM sees the same instructions, just without repeated [System: prefixes.

Role Mapping

ST Role	TomoriBot Role
system	system
user	user
assistant	model

Preset Caching

Active presets are cached in-memory to avoid a DB query on every buildContext() call. Located in src/utils/cache/stPresetCache.ts.

Feature	Detail
Cache key	server_id (numeric)
Cached data	{ preset: StPresetRow, nodes: StPresetNodeRow[] } or null (no active preset)
TTL	Configurable via ST_PRESET_CACHE_TTL_MINUTES env var (default: 10 minutes)
Invalidation	On preset activate, deactivate, node toggle, or preset delete
Graceful fallback	Returns stale cache on DB error
Negative caching	null result is cached to avoid repeated “no preset” queries

Cache invalidation is called from stPresetDb.ts after every successful write operation. The serverId parameter is required (not optional) on all write functions to ensure invalidation cannot be accidentally skipped.

Database Schema

st_presets table

Column	Type	Description
preset_id	SERIAL PK	Auto-incrementing primary key
server_id	INT FK	References servers(server_id), CASCADE on delete
preset_name	TEXT	Display name (unique per server)
raw_json	JSONB	Complete original ST preset JSON for re-parsing
is_active	BOOLEAN	Whether this is the active preset (one per server)
created_at	TIMESTAMP	Import timestamp
updated_at	TIMESTAMP	Last modification timestamp

Unique constraint: (server_id, preset_name)

st_preset_nodes table

Column	Type	Description
node_id	SERIAL PK	Auto-incrementing primary key
preset_id	INT FK	References st_presets(preset_id), CASCADE on delete
identifier	TEXT	ST node identifier (UUID or well-known name)
name	TEXT	Display name from the preset
role	TEXT	Message role: system, user, or assistant
content	TEXT	Raw prompt text (with unresolved ST macros)
is_marker	BOOLEAN	Structural anchor (charDescription, chatHistory, etc.)
is_enabled	BOOLEAN	User-togglable enabled state
node_order	INT	Position in the preset’s prompt_order sequence
injection_position	INT	0 = relative to system prompt, 1 = relative to chat end
injection_depth	INT	Messages from end for depth-based insertion
injection_order	INT	Priority for tie-breaking at same position+depth

Unique constraint: (preset_id, identifier)

ST Preset Anatomy

Node Types

When parsing a preset’s prompts array, nodes fall into three categories:

Category	Detection	Stored?	Toggleable?
Comment-only	Content is purely {{// ... }}{{trim}}	No	No
Marker	marker: true (structural anchor)	Yes	No
Content node	Has real content after macro stripping	Yes	Yes

Comment-Only Detection

A node is comment-only if its content matches: ^(\s*\{\{\/\/[^}]*\}\}\s*|\s*\{\{trim\}\}\s*)+$

Examples of comment-only content:

• {{// Empty for card override. }}{{trim}}
• {{// Choose the narration style. }}{{trim}}
• {{// Enable only one out of the list below. }}{{trim}}

Well-Known Markers

Identifier	ST Purpose	TomoriBot Equivalent
main	Main system prompt	System humanizer rules
charDescription	Character description	Persona description
charPersonality	Character personality	Personality attributes
scenario	Scenario text	(no direct equivalent)
personaDescription	User persona	(no direct equivalent)
dialogueExamples	Example dialogues	Sample dialogues
chatHistory	Conversation log	Conversation history
worldInfoBefore	World info (before char data)	RAG documents
worldInfoAfter	World info (after char data)	RAG documents

Unrecognized markers are logged as warnings and skipped.

Prompt Order

ST presets have a prompt_order array with entries for two scopes:

• character_id: 100000 — System prompt order (well-known markers only)
• character_id: 100001 — User prompt order (custom nodes + markers, preferred when present)

Each entry has { identifier, enabled }. The array order defines the rendering sequence. TomoriBot prefers the 100001 entry and falls back to 100000 only if 100001 is missing.

Parity with SillyTavern

This section documents what our implementation supports versus what native SillyTavern does, organized by category.

Supported Macros

Macro	Status	Notes
{{user}}	Supported	Deferred to convertMentions()
{{char}} / {{bot}}	Supported	Deferred to convertMentions()
{{personality}}	Supported	Maps to /persona attribute add values
{{description}}	Supported	Maps to persona prompt
{{mesExamples}}	Supported	Maps to sample dialogues
{{lastChatMessage}}	Supported	Most recent user message
{{scenario}}	Supported (empty)	Always resolves to "" — no TomoriBot equivalent
{{setvar::key::value}}	Supported	Replaces the variable value
{{addvar::key::value}}	Supported	Appends to the variable value in node order
{{getvar::key}}	Supported	Unknown keys resolve to ""
{{random: A, B, C}}	Supported	Random pick from comma-separated list
{{random::A::B::C}}	Supported	Legacy random choice syntax
{random: A, B, C}	Supported	Single-brace random choice syntax resolved by buildContext()
{random::A::B::C}	Supported	Single-brace legacy random choice syntax resolved by buildContext()
{{roll: XdY}}	Supported	Capped at 100 dice, 1000 sides
{{trim}}	Supported	Node disabled if result is empty
{{// comment}}	Supported	Stripped from output

Unsupported Macros

Macro	ST Purpose	Why Not Supported
{{summary}}	Short-term memory summary text	TomoriBot has STM but doesn’t expose it as a macro. STM is injected as its own context block instead.
{{group}}	Multi-character group RP names	TomoriBot is single-character-per-context. Fundamental design difference.
{{persona}}	User persona description	No user persona system in TomoriBot.
{{input}}	User-provided text for ST frontend helper prompts	Not available in TomoriBot’s preset runtime.
{{mesExamplesRaw}}	Raw example dialogue block	TomoriBot only exposes formatted {{mesExamples}}.
{{#if ...}} / {{/if}}	Legacy conditional template blocks	Not implemented inside modern prompt nodes. Older text-completions presets are converted at import time instead of running these blocks directly.
{{time}}, {{date}}, {{weekday}}, {{isotime}}, {{isodate}}	Date/time formatting	Time/channel info is embedded in the Users in Conversation block automatically, not exposed as macros.
{{idle_duration}}	Time since last message	Not tracked.
{{maxPrompt}}	Max token budget	TomoriBot doesn’t expose token limits to preset macros.
{{exampleSeparator}}, {{chatStart}}	Dialogue formatting tokens	Sample dialogues use their own formatting.
{{banned_tokens}}, {{bias}}	Logit bias control	Logit bias is controlled separately via /config, not from presets.
Nested/recursive macros	e.g., {{getvar::{{getvar::key}}}}	Only single-level resolution.

Unsupported Features

Feature	ST Behavior	TomoriBot Behavior
Regex post-processing	Find/replace rules applied to generated output	Not implemented. Output is sent as-is. Presets that rely on regex formatting (e.g., stripping XML tags, reformatting narration) will look different.
Legacy text-completions presets	Old ST preset shape using context.story_string + sysprompt.content	Best-effort imported by converting the story layout into Prompt Manager-style nodes. Main system prompt, story layout, and post-history map over; ST-only blocks such as persona, scenario, anchors, stop strings, and old backend settings are still ignored.
Preset settings import	Temperature, top_p, frequency_penalty, model overrides embedded in preset JSON	Ignored. These remain server-level settings via /config.
World Info / Lorebook	Static knowledge entries with activation keywords, inserted at worldInfoBefore/worldInfoAfter markers	TomoriBot uses dynamic RAG instead. The worldInfo markers pull RAG results, not static lorebook entries. Content may differ significantly.
Token budgeting	Per-node token limits, total prompt budget management	Not implemented. All enabled nodes are included regardless of token count. Context may exceed provider limits if many large nodes are enabled.
Multiple active presets	Some ST setups layer presets	One preset per server. By design.
HTML rendering	ST frontend renders HTML in chat	Discord cannot render HTML. Nodes with HTML are flagged (hasHtmlWarning) but not auto-stripped.
Assistant prefill	model-role nodes at end of context force the AI to start with specific text	Passed through, but provider-dependent. Works on some providers (Anthropic), ignored by others (Gemini).

Architectural Differences

Area	SillyTavern	TomoriBot
Depth injection	Inserts new standalone messages at target depth	Merges into existing messages as [System: ...] text parts. Same-depth injections are batched into a single [System: ...] block for token efficiency. This prevents role-alternation violations on Gemini/Anthropic but may not match exact ST positioning.
TomoriBot-only blocks	N/A	Server info, server memories, emojis, stickers, user list, STM, and conditioning have no ST markers. They are always included and auto-flushed at anchor points. Users cannot reorder or disable them via the preset.
Variable scope	May support scoped/local variables	All {{setvar}} / {{addvar}} variables are global across enabled nodes. No scoping.
prompt_order parsing	Both character_id: 100000 (system) and 100001 (user)	TomoriBot prefers 100001 and falls back to 100000 only when 100001 is missing.
Character-specific ordering	Different prompt orders per character card	TomoriBot uses one preset per server regardless of active persona.

File Map

File	Purpose
src/db/schema_stpreset.sql	Database table definitions
src/types/db/schema.ts	StPresetRow and StPresetNodeRow type definitions
src/utils/db/stPresetDb.ts	CRUD operations + cache invalidation hooks
src/utils/cache/stPresetCache.ts	In-memory preset cache with TTL
src/utils/text/stPresetEngine.ts	Template macro engine (two-pass resolution)
src/utils/text/presetContextBuilder.ts	Preset-driven context rearrangement
src/utils/text/contextBuilder.ts	Routing wrapper + native context assembly
src/commands/st-preset/import.ts	/st-preset import command
src/commands/st-preset/remove.ts	/st-preset remove command
src/commands/st-preset/node/toggle.ts	/st-preset node toggle command