st-audit

name	st-audit
model	sonnet
description	Audit current SillyTavern config — explain settings, surface non-defaults, recommend changes for a goal. Read-only.
argument-hint	[<setting-key> | goal "<text>"]
allowed-tools	Bash, Read, mcp__st__st_get_settings, mcp__st__st_save_settings_path

ST Audit — Discover & Explain Current Config

Read-only audit of SillyTavern configuration. Surfaces what's currently configured, why each knob matters, and what to change for a given goal. Does NOT modify anything — feeds context into conversation so user decides next move.

Why this exists: ST has 100KB+ settings.json + 10+ extensions, each with sub-config. Most users don't know which knobs are load-bearing vs cargo-cult. This skill closes the discovery gap.

Usage:

/st-audit                          # full sweep, grouped by domain
/st-audit <setting-key>            # explain one setting (current value + what it does + safe range)
/st-audit goal "<natural language>" # goal-driven: list relevant settings + recommendations

Constants

ST_DATA = /home/haint/Projects/home-server/sillytavern/data/default-user
SETTINGS = $ST_DATA/settings.json   # fallback only; prefer mcp__st__st_get_settings
PLAYBOOK = /home/haint/Projects/home-server/sillytavern/PROMPT-PLAYBOOK.md

Knowledge Source Order

1. PROMPT-PLAYBOOK.md — 33 verified gotchas, the canonical local knowledge base. Always cite gotcha number when relevant.
2. Live settings via mcp__st__st_get_settings() — current state (ground truth). Works while ST runs (no file-lock risk). Falls back to Read SETTINGS if MCP unavailable (ST container down).
3. ST source defaults (read from /home/node/app/public/scripts/extensions/<name>/index.js via podman exec if needed).
4. Built-in skill knowledge below (curated facts about load-bearing knobs).

If knowledge sources conflict → live settings wins for "current value", playbook wins for "why it matters".

---

Mode 1: Full Sweep (/st-audit no args)

Read each domain via path-based MCP calls (full-tree read returns 70KB+, exceeds Claude's MCP token cap). Each call returns a JSON string of the subtree at that path.

import json

# Per-domain reads (compose a full audit without ever loading the full tree)
sd          = json.loads(mcp__st__st_get_settings(path="extension_settings.sd"))
memory      = json.loads(mcp__st__st_get_settings(path="extension_settings.memory"))
conn_mgr    = json.loads(mcp__st__st_get_settings(path="extension_settings.connectionManager"))
power_user  = json.loads(mcp__st__st_get_settings(path="power_user"))
ext_all     = json.loads(mcp__st__st_get_settings(path="extension_settings"))

# Disabled extensions are stored per-extension as <extension>.disabled (true/false),
# not aggregated at a single key. Walk extension_settings for {disabled: true} entries.
disabled_exts = [k for k, v in ext_all.items() if isinstance(v, dict) and v.get("disabled") is True]

Produce a grouped report. Skip categories where everything is at default — focus attention on non-defaults.

Categories to audit

[1] Image Generation (extension_settings.sd)

• source — Forge URL backend
• sampler (expect Euler for NoobAI, NOT Euler a)
• scheduler (expect karras)
• steps, scale (CFG)
• prompt_prefix (must start with quality tags)
• prompts['4'] — Mode 4 template length + first line
• character_prompts — count entries + list keys
• character_negative_prompts — count entries

[2] Memory / Summary (extension_settings.memory)

• source (main = uses primary LLM, extras = separate)
• prompt_builder (0=DEFAULT/generateQuietPrompt, 1=RAW_BLOCKING, 2=RAW_NON_BLOCKING)
• SkipWIAN (true = exclude WIAN from summary prompt)
• promptInterval (0 = manual-only, N = auto every N messages)
• position, depth, role (where summary injects)
• promptWords (max summary length)

[3] Connection Profiles (extension_settings.connectionManager.profiles)

• List each profile: name, preset, model, api
• Active profile (extension_settings.connectionManager.selectedProfile)

[4] RP Behavior (power_user.*)

• instruct.preset_name + instruct.enabled
• context.preset (context template)
• max_context, response_length
• prefer_character_prompt / prefer_character_jailbreak
• user_avatar (active persona)

[5] Extensions State

• Walk extension_settings.<name>.disabled — list every extension with disabled: true
• Highlight notable ones: LALib (slash command lib), GuidedGenerations-Extension, memory

[6] Quick Replies (quickReplyApi.config.setList)

• Per set: name, button count, auto-execute hooks

[7] Persona (power_user.personas, power_user.persona_descriptions)

• Active persona + linked lorebook
• Total persona count

Output format

## ST Config Audit — [date]

### 🟢 Image Gen
- sampler: Euler ✓ (correct for NoobAI epsilon-pred)
- scheduler: karras ✓
- steps: 30 ✓ (>=28 required, gotcha 5.X)
- scale: 5 ✓ (CFG 4-5 for NoobAI)
- prompt_prefix: "masterpiece, best quality..." ✓
- Mode 4 template: 2144 chars (v8.2 — 5 hard rules only)
- char_prompts: 4 entries [Naoko, Parasite, Helena Lin, Klee]

### 🟡 Memory/Summary
- source: main ✓ (uses primary LLM)
- prompt_builder: 1 (RAW_BLOCKING) ✓ (gotcha 5.33)
- SkipWIAN: true ✓
- promptInterval: 0 ✓ (manual-only, gotcha 5.33)
- position: ?, depth: ?, role: ?

### 🔴 Extensions
- DISABLED: GuidedGenerations-Extension (diagnostic state, can re-enable)
- ENABLED: LALib ✓ (required for /dom workaround)

### Connection Profiles
| Name | Preset | Model | API |
|------|--------|-------|-----|
| ...  | ...    | ...   | ... |

Active: DeepSeek daily

### Findings
- Non-default values: N
- Settings flagged: M
- Suggested next checks: [...]

Use 🟢 (looks correct), 🟡 (notable but intentional), 🔴 (worth attention).

---

Mode 2: Single Setting (/st-audit <key>)

Parse <key> — accept dot-path (sd.sampler) or last segment (prompt_builder matches extension_settings.memory.prompt_builder).

If ambiguous (multiple matches) → list candidates, ask user to disambiguate.

Output

## Setting: extension_settings.memory.prompt_builder

**Current value**: 1 (RAW_BLOCKING)

**What it does**: Controls which generation path the Summarize extension uses.
- 0 (DEFAULT) → generateQuietPrompt → routes through prompt manager → injects WIAN, persona, char defs into summary prompt
- 1 (RAW_BLOCKING) → generateRaw → bypasses prompt manager → clean summary prompt only ✓ recommended
- 2 (RAW_NON_BLOCKING) → generateRaw async → faster but less reliable for long chats

**Why current value matters**: Setting 0 caused WIAN contamination in Magnum's summary output (gotcha 5.33). Switched to 1 to route through generateRaw.

**Depends on / affects**:
- Pairs with `SkipWIAN: true` (redundant safety — RAW_BLOCKING already bypasses WIAN)
- If you switch back to 0, restore SkipWIAN check

**Source**: PROMPT-PLAYBOOK.md gotcha 5.33; ST source `/home/node/app/public/scripts/extensions/memory/index.js:507`

**Safe to change**: Only if changing summary architecture. Current value is load-bearing.

---

Mode 3: Goal-Driven (/st-audit goal "<text>")

Parse natural language goal. Match against known goal categories:

Goal pattern	Relevant settings
"image gen quality" / "ảnh đẹp hơn"	sd.sampler, scheduler, steps, scale, prompt_prefix, char_prompts, Mode 4 template
"summary clean" / "summary không bị bẩn"	memory.prompt_builder, SkipWIAN, promptInterval, source
"RP voice" / "ít interrupt" / "tone"	instruct preset, context preset, char card PHI, AN, model temperature
"model switching" / "profile"	connectionManager.profiles, selectedProfile, /profile slash command
"function calling" / "tool calling"	enableFunctionCalling, model api support, prompts
"context window" / "token budget"	max_context, response_length, summary depth, lorebook entry budgets
"persona setup" / "user persona"	power_user.personas, persona_descriptions, user_avatar, linked lorebook
"lorebook" / "world info"	worlds/*.json, character lorebook linkage, depth, position, scanDepth

For each match, output:

1. Current state of relevant settings (live read)
2. Common recommendations (with rationale + gotcha refs)
3. Risks of changing each
4. Suggested order of changes (start with lowest-risk)

Output

## Goal: "summary không bị bẩn"

Matched category: Memory/Summary

### Current state
- prompt_builder: 1 (RAW_BLOCKING) ✓
- SkipWIAN: true ✓
- promptInterval: 0 (manual-only) ✓
- source: main ✓

### Status
**Already optimized.** All 4 load-bearing settings match recommended values from gotcha 5.33.

### If still seeing problems
- Check active model — DeepSeek tends to write prose-style continuations even with RAW_BLOCKING (RP-tuned). Switch to Magnum profile via QR `[📝 Summary]` button.
- Check chat metadata — old `extra.memory` entries from prior bad summaries can contaminate next regen. Inspect `chats/<char>/<chat>.jsonl`.

### Reference
PROMPT-PLAYBOOK.md gotcha 5.33; section 8.1 (Summary Workflow)

---

Built-in Knowledge (curated facts)

Embedded so skill works without re-reading PROMPT-PLAYBOOK every invocation. Update this list when new gotchas added.

Image gen knobs

• NoobAI XL = epsilon-prediction → Euler/Karras/CFG5/steps≥28. NOT Euler a.
• vpred models broken on ai-dock Forge image (gotcha 5.X) — stick with epsilon checkpoints.
• prompt_prefix MUST start with masterpiece, best quality, newest, absurdres, highres,
• Mode 4 (Last Message) template = prompts['4']. Current: v8.2, 2144 chars, 5 hard rules. Magnum picks tags from booru training.
• char_prompts key = char filename WITHOUT .png (gotcha: getCharaFilename() strips ext)

Memory/Summary

• prompt_builder=1 (RAW_BLOCKING) bypasses prompt manager → clean summary prompt
• promptInterval=0 → manual only (recommended). Auto-trigger contaminates context unpredictably.
• DeepSeek RP-tuned → ignores "STOP. END OF ROLEPLAY" directive. Switch to Magnum for clean summary.
• Bad prior summary in extra.memory chat metadata → contaminates next regen. Clear it before retry.

STscript / Slash commands

• is_send_press lock → /summarize via QR pipe silent fails (gotcha 5.33)
• LALib /dom action=click "#memory_force_summarize" bypasses lock via native DOM event
• Profile switch via /profile timeout=5000 <name> — needs delay before next command

Extensions

• Third-party extensions live in data/default-user/extensions/, NOT public/extensions/third-party/ (legacy path)
• LALib provides /dom, /regex, /runc, /db, /fetch, etc.
• GuidedGenerations adds Quick Reply hooks on GENERATION_AFTER_COMMANDS

Persona vs Character

• char_prompts has NO persona equivalent → visual tags must embed in persona_descriptions[avatar].description text
• Active persona = power_user.user_avatar (filename of avatar PNG)
• Persona-bound lorebook = power_user.persona_descriptions[avatar].lorebook

---

Implementation Steps

For any mode:

1. Read live state:

   import json
   with open("/home/haint/Projects/home-server/sillytavern/data/default-user/settings.json") as f:
       s = json.load(f)
   

2. Read PROMPT-PLAYBOOK.md (skim relevant sections only — file is ~700 lines):

   • Mode 1 full sweep → read sections 1-3 + 8 + 8.1 + 10 (gotcha index)
   • Mode 2 single setting → grep playbook for the setting key
   • Mode 3 goal → match goal to playbook section, read that section

3. Cross-reference current value vs recommended:

   • Match → 🟢
   • Different but documented choice → 🟡
   • Different and undocumented → 🔴 (flag for user attention)

4. Output report (markdown table or structured sections, scannable)

5. NEVER modify settings.json — this skill is read-only by design.

---

Edge Cases

Case	Handling
settings.json not found	ST data not at expected path. Check container running, paths correct.
ST container running while reading	Safe — read-only. But warn that values may change if user edits via UI mid-audit.
Setting key ambiguous in mode 2	List candidates, ask user to pick.
Goal text doesn't match any category	Show available categories, ask user to rephrase or pick closest.
PROMPT-PLAYBOOK.md missing	Skill still works using built-in knowledge above; note playbook unavailable.

---

Why This Skill, Not MCP Server

• Pain is discovery (cognitive overload), not control (mechanical access)
• Skill = read knowledge + explain. MCP server = expose ST operations to external agent. Different layers.
• Skill 1-2h to build. MCP server 5-10h. Phase 1 capacity wins.
• Skill outputs feed Claude Code conversation → user decides changes → applies via existing skills (/st-setup, /st-persona, /st-arc-save) or direct edits with confirmation.

When to revisit MCP server: only if specific pain emerges that direct-file + skills can't cover.