NovelAI Provider Limitations

This document catalogs every feature, tool, and context block that is deliberately disabled or reduced for the NovelAI provider compared to Google Gemini and OpenRouter. All exclusions fall into one of two root causes:

1. Text-only model — GLM 4.6 has no vision or image understanding capability.
2. Token budget — NovelAI’s API has a hard token cap that creates a practical quality threshold around ~2800 tokens of system prompt. Exceeding it degrades output quality noticeably.

See also: tool-calling.md for how prompt-based tool calling itself works.

---

Built-in Tools Disabled

Some tools are disabled by provider-specific isAvailableFor() checks, while vision-dependent tools are removed by the registry’s first-pass capability filter because NovelAI models have sees_images = false.

Tool	File	Reason
select_sticker_for_response	src/tools/functionCalls/stickerTool.ts:144	GLM 4.6 cannot reliably generate CJK/Japanese sticker names as tool arguments — token-level instability causes garbled output.
update_short_term_memory	src/tools/functionCalls/updateShortTermMemoryTool.ts:56	Token budget too constrained; the tool definition and invocation overhead is not worth the benefit at GLM’s prompt size.
fetch_url	src/tools/fetchUrl/fetchUrlTool.ts	Token budget too constrained for fetched-page payloads; disabled until realistic URL-fetch prompt/tool history behavior is validated.
peek_profile_picture	src/tools/functionCalls/peekProfilePictureTool.ts	Requires requiredModelCapabilities = { sees_images: true }; NovelAI models are text-only.
process_gif	src/tools/functionCalls/processGifTool.ts	Requires requiredModelCapabilities = { sees_images: true }; NovelAI models are text-only.

---

MCP Functions Blocked

Even when MCP servers are running, the following function names are stripped from the tool list before it is passed to GLM. Controlled in src/providers/novelai/novelaiToolAdapter.ts:301.

felo-search
iask-search
monica-search
fetch-url
url-metadata
fetch

Reason: These are either redundant with the unified web_search tool (which routes search through the Brave → SearXNG → DuckDuckGo → Felo engine chain), internal to the unified fetch_url path, or too token-expensive in their argument schemas and response payloads for GLM’s strict prompt budget. The brave_* MCP function names are no longer LLM-visible at all post-unification, so they don’t need to be in this disable list anymore.

---

Context Blocks Excluded from System Prompt

Each provider’s stream adapter defines which ContextItemTag blocks are included in the system instruction. NovelAI’s list omits two tags that all other providers (Google, OpenRouter, Custom) include:

Tag	Other providers	NovelAI
KNOWLEDGE_SERVER_EMOJIS	✅ Included	❌ Excluded
KNOWLEDGE_SERVER_STICKERS	✅ Included	❌ Excluded

Source: src/providers/novelai/novelaiStreamAdapter.ts:107-109

private static readonly SYSTEM_INSTRUCTION_TAGS_TOOLING: ContextItemTag[] = [
    ContextItemTag.SYSTEM_FUNCTION_GUIDE,
    // REMOVED: KNOWLEDGE_SERVER_EMOJIS, KNOWLEDGE_SERVER_STICKERS — GLM 4.6 is text-only
    // and cannot use emojis/stickers. Omitting saves significant tokens toward the ~2800
    // token quality threshold.
];

Practical effect: The model is never told which custom server emojis or stickers exist, so it will not reference or attempt to use them. The emojiUsageEnabled config flag still flows through to the stream adapter (it controls output formatting), but without the knowledge block, the model has no emoji list to draw from.

---

Short-Term Memory Tool Instructions Suppressed

Even though the update_short_term_memory tool is excluded from the tool list for NovelAI (see above), there is a second independent suppression point in context building.

Source: src/utils/text/contextBuilder.ts:519

const isStmToolAvailable = tomoriState.llm.llm_provider !== "novelai";

When isStmToolAvailable is false:

• The hint message "[System: HINT: Use the update_short_term_memory tool...]" is never injected after short-term memory summaries.
• The nudge prompt that encourages the model to call the tool when a conversation goes stale is also suppressed.

The short-term memory data itself (summaries and recent messages) is still included in context when available — only the tool-use instructions around it are removed.

---

Provider Capability Flags

Declared in src/providers/novelai/novelaiProvider.ts:129-141:

supportsImages: false,   // Text-only, no vision
supportsVideos: false,   // Text-only, no vision
supportsFunctionCalling: true, // Prompt-based only, GLM 4.6 models only

These flags gate image/video attachment processing earlier in the pipeline, so media is never forwarded to NovelAI even if a user sends it.

---

Reminder Tool: Auto-Fill Quirk

Source: src/tools/functionCalls/reminderTool.ts:255-266

When the reminder tool is called via NovelAI and repetition_interval_hours is missing from the model’s response (a common GLM omission for simple “remind me in X” requests), the tool automatically defaults the value to 0 (one-time reminder) instead of rejecting the call.

// NovelAI only — other providers require the model to set this explicitly
if (context.provider === "novelai" && typeof repetitionIntervalHoursArg !== "number") {
    repetitionIntervalHoursArg = 0;
}

Other providers do not get this fallback — they must explicitly set the field so the model is “conscious” of whether the reminder is one-time or recurring.

---

Summary Table

Feature	Disabled for NovelAI?	Root Cause
select_sticker_for_response tool	✅ Yes	CJK name generation instability
update_short_term_memory tool	✅ Yes	Token budget
peek_profile_picture tool	✅ Yes	Text-only model
process_gif tool	✅ Yes	Text-only model
Unified fetch_url tool	✅ Yes	Token budget
MCP fetch functions (fetch, fetch-url, url-metadata)	✅ Yes	Token budget / redundant
MCP alternative search engines	✅ Yes	Token budget / redundant
Unified web_search tool	❌ Enabled	Replaces the previous 4-tool Brave surface with one categorized tool
KNOWLEDGE_SERVER_EMOJIS context block	✅ Yes	Text-only + token budget
KNOWLEDGE_SERVER_STICKERS context block	✅ Yes	Text-only + token budget
STM tool hint injections	✅ Yes	Token budget
Image attachment processing	✅ Yes	Text-only model
Video attachment processing	✅ Yes	Text-only model
repetition_interval_hours required	❌ Relaxed	GLM omits it frequently (auto-fills 0)

---

Adding Future Exclusions

When adding a new tool that depends on model capabilities, declare them directly on the tool:

requiredModelCapabilities = {
    sees_images: true,
};

When adding a new tool that should be disabled for NovelAI for provider-specific reasons (token budget, GLM argument instability, etc.), override isAvailableFor() in the tool class:

isAvailableFor(provider: string, _context: unknown): boolean {
    if (provider === "novelai") return false; // text-only / token budget
    return true;
}

When excluding a context block from NovelAI’s system prompt, remove its ContextItemTag from the relevant static array in src/providers/novelai/novelaiStreamAdapter.ts and leave a comment explaining why.