Skip to content

LLM Wiki is a cross-platform desktop application that turns your documents into an organized, interlinked knowledge base — automatically. Instead of traditional RAG (retrieve-and-answer from scratch every time), the LLM incrementally builds and maintains a persistent wiki from your sources。

TypeScript

fix: harden chat image provider support

nash_su08b9289b767d
diff --git a/src-tauri/tauri.conf.json b/src-tauri/tauri.conf.json
index dd08e040..97b94520 100644
--- a/src-tauri/tauri.conf.json
+++ b/src-tauri/tauri.conf.json
@@ -22,7 +22,7 @@
}
],
"security": {
- "csp": "default-src 'self'; connect-src 'self' https: http:; img-src 'self' asset: http://asset.localhost https://asset.localhost blob: data:; media-src 'self' asset: http://asset.localhost https://asset.localhost blob: data:; style-src 'self' 'unsafe-inline'",
+ "csp": "default-src 'self'; connect-src 'self' https: http:; img-src 'self' asset: http://asset.localhost https://asset.localhost blob: data:; media-src 'self' asset: http://asset.localhost https://asset.localhost; style-src 'self' 'unsafe-inline'",
"assetProtocol": {
"enable": true,
"scope": ["**"]
diff --git a/src/components/chat/chat-input.tsx b/src/components/chat/chat-input.tsx
index 7c1288e8..40b89fbf 100644
--- a/src/components/chat/chat-input.tsx
+++ b/src/components/chat/chat-input.tsx
@@ -24,10 +24,18 @@ interface ChatInputProps {
onStop: () => void
isStreaming: boolean
anyTxtAvailable?: boolean
+ imageInputAvailable?: boolean
placeholder?: string
}
-export function ChatInput({ onSend, onStop, isStreaming, anyTxtAvailable = true, placeholder }: ChatInputProps) {
+export function ChatInput({
+ onSend,
+ onStop,
+ isStreaming,
+ anyTxtAvailable = true,
+ imageInputAvailable = true,
+ placeholder,
+}: ChatInputProps) {
const { t } = useTranslation()
const [value, setValue] = useState("")
const [useWebSearch, setUseWebSearch] = useState(false)
@@ -49,6 +57,10 @@ export function ChatInput({ onSend, onStop, isStreaming, anyTxtAvailable = true,
async (files: File[]) => {
const imageFiles = files.filter((f) => f.type.startsWith("image/"))
if (imageFiles.length === 0) return
+ if (!imageInputAvailable) {
+ setImageError(t("chat.imageInputUnavailable"))
+ return
+ }
let error: string | null = null
const accepted: MessageImage[] = []
// Read current count via the functional updater below; here we
@@ -79,7 +91,7 @@ export function ChatInput({ onSend, onStop, isStreaming, anyTxtAvailable = true,
}
setImageError(error)
},
- [images.length, t],
+ [imageInputAvailable, images.length, t],
)
const handlePaste = useCallback(
@@ -129,6 +141,10 @@ export function ChatInput({ onSend, onStop, isStreaming, anyTxtAvailable = true,
const trimmed = value.trim()
// Allow image-only messages: send if there's text OR at least one image.
if ((!trimmed && images.length === 0) || isStreaming) return
+ if (images.length > 0 && !imageInputAvailable) {
+ setImageError(t("chat.imageInputUnavailable"))
+ return
+ }
onSend(trimmed, images, { useWebSearch, useAnyTxtSearch })
setValue("")
setImages([])
@@ -136,7 +152,7 @@ export function ChatInput({ onSend, onStop, isStreaming, anyTxtAvailable = true,
if (textareaRef.current) {
textareaRef.current.style.height = "auto"
}
- }, [value, images, isStreaming, onSend, useWebSearch, useAnyTxtSearch])
+ }, [imageInputAvailable, images, isStreaming, onSend, t, useAnyTxtSearch, useWebSearch, value])
const handleKeyDown = useCallback(
(e: React.KeyboardEvent<HTMLTextAreaElement>) => {
@@ -187,6 +203,11 @@ export function ChatInput({ onSend, onStop, isStreaming, anyTxtAvailable = true,
{imageError && (
<p className="mb-1 px-1 text-xs text-destructive">{imageError}</p>
)}
+ {images.length > 0 && !imageError && (
+ <p className="mb-1 px-1 text-xs text-muted-foreground">
+ {t("chat.imageVisionHint")}
+ </p>
+ )}
<textarea
ref={textareaRef}
value={value}
@@ -210,16 +231,20 @@ export function ChatInput({ onSend, onStop, isStreaming, anyTxtAvailable = true,
/>
<div className="mt-1 flex items-center justify-between gap-3 border-t border-border/50 pt-2">
<div className="flex min-w-0 flex-wrap items-center gap-1.5">
- <button
- type="button"
- onClick={() => fileInputRef.current?.click()}
- disabled={isStreaming || images.length >= MAX_IMAGES_PER_MESSAGE}
- className={searchToggleClass(false)}
- title={t("chat.attachImage")}
+ <span
+ className="inline-flex"
+ title={!imageInputAvailable ? t("chat.imageInputUnavailable") : t("chat.attachImage")}
>
- <ImagePlus className="h-3.5 w-3.5" />
- <span className="hidden sm:inline">{t("chat.attachImage")}</span>
- </button>
+ <button
+ type="button"
+ onClick={() => fileInputRef.current?.click()}
+ disabled={isStreaming || !imageInputAvailable || images.length >= MAX_IMAGES_PER_MESSAGE}
+ className={searchToggleClass(false)}
+ >
+ <ImagePlus className="h-3.5 w-3.5" />
+ <span className="hidden sm:inline">{t("chat.attachImage")}</span>
+ </button>
+ </span>
<button
type="button"
aria-pressed={useWebSearch}
diff --git a/src/components/chat/chat-panel.tsx b/src/components/chat/chat-panel.tsx
index 59e7389b..e181ea75 100644
--- a/src/components/chat/chat-panel.tsx
+++ b/src/components/chat/chat-panel.tsx
@@ -7,6 +7,7 @@ import { ChatInput, type ChatSendOptions } from "./chat-input"
import { useChatStore, chatMessagesToLLM, type MessageReference, type MessageImage } from "@/stores/chat-store"
import { useWikiStore } from "@/stores/wiki-store"
import { streamChat, type ChatMessage as LLMMessage } from "@/lib/llm-client"
+import { supportsImageInput } from "@/lib/llm-providers"
import { executeIngestWrites } from "@/lib/ingest"
import { listDirectory, readFile, deleteFile } from "@/commands/fs"
import { searchWiki } from "@/lib/search"
@@ -160,6 +161,7 @@ export function ChatPanel() {
const llmConfig = useWikiStore((s) => s.llmConfig)
const searchApiConfig = useWikiStore((s) => s.searchApiConfig)
const anyTxtAvailable = hasConfiguredAnyTxt(searchApiConfig.anyTxt)
+ const imageInputAvailable = supportsImageInput(llmConfig)
const setFileTree = useWikiStore((s) => s.setFileTree)
const abortRef = useRef<AbortController | null>(null)
@@ -638,6 +640,7 @@ export function ChatPanel() {
onStop={handleStop}
isStreaming={isStreaming}
anyTxtAvailable={anyTxtAvailable}
+ imageInputAvailable={imageInputAvailable}
placeholder={
mode === "ingest"
? t("chat.ingestPlaceholder")
diff --git a/src/components/settings/llm-presets.ts b/src/components/settings/llm-presets.ts
index bc77c9af..9e88d687 100644
--- a/src/components/settings/llm-presets.ts
+++ b/src/components/settings/llm-presets.ts
@@ -323,21 +323,31 @@ export const LLM_PRESETS: LlmPreset[] = [
baseUrl: "https://open.bigmodel.cn/api/paas/v4",
defaultModel: "glm-4.6",
apiMode: "chat_completions",
- // Current-gen Zhipu BigModel lineup. glm-5 / glm-5.1 are routed
- // through the Z.AI international endpoint (api.z.ai) so they're
- // not on this preset's suggestion list — users targeting those
- // should type the id or point the base URL at api.z.ai.
+ // Current Zhipu BigModel lineup on /api/paas/v4/chat/completions.
+ // Vision-capable models use the same OpenAI-compatible image_url
+ // content blocks as our generic chat-completions wire.
suggestedModels: [
+ "glm-5.1",
+ "glm-5-turbo",
+ "glm-5",
+ "glm-5v-turbo",
+ "glm-4.7",
+ "glm-4.7-flash",
+ "glm-4.7-flashx",
"glm-4.6",
+ "glm-4.6v",
"glm-4.5",
+ "glm-4.5v",
"glm-4.5-air",
"glm-4.5-airx",
"glm-4.5-flash",
+ "glm-4-flash-250414",
+ "glm-4-flashx-250414",
"glm-4-plus",
"glm-4-air",
"glm-4-flash",
- "glm-zero-preview",
"glm-4v-plus",
+ "glm-zero-preview",
],
suggestedContextSize: 128000,
},
diff --git a/src/i18n/en.json b/src/i18n/en.json
index 7d0fab75..c8ee9e48 100644
--- a/src/i18n/en.json
+++ b/src/i18n/en.json
@@ -143,6 +143,8 @@
"attachImage": "Attach image",
"removeImage": "Remove image",
"imageMessage": "Image",
+ "imageVisionHint": "Images are sent to the current LLM. Use a vision-capable model for image questions.",
+ "imageInputUnavailable": "The current provider does not support image input. Switch to a vision-capable provider in Settings.",
"imageTooLarge": "Image too large (max {{max}}MB): {{name}}",
"tooManyImages": "Too many images (max {{max}})",
"unsupportedImageType": "Unsupported image type: {{type}}"
diff --git a/src/i18n/zh.json b/src/i18n/zh.json
index 3316c0df..a595e95a 100644
--- a/src/i18n/zh.json
+++ b/src/i18n/zh.json
@@ -143,6 +143,8 @@
"attachImage": "添加图片",
"removeImage": "移除图片",
"imageMessage": "图片",
+ "imageVisionHint": "图片会发送给当前 LLM。图片提问需要使用支持视觉的模型。",
+ "imageInputUnavailable": "当前 Provider 不支持图片输入。请在设置中切换到支持视觉的 Provider。",
"imageTooLarge": "图片过大(最大 {{max}}MB):{{name}}",
"tooManyImages": "图片数量过多(最多 {{max}} 张)",
"unsupportedImageType": "不支持的图片类型:{{type}}"
diff --git a/src/lib/llm-providers.test.ts b/src/lib/llm-providers.test.ts
index b4e00da3..885fd971 100644
--- a/src/lib/llm-providers.test.ts
+++ b/src/lib/llm-providers.test.ts
@@ -16,7 +16,13 @@
* text wire would break every existing call site.
*/
import { describe, it, expect } from "vitest"
-import { getProviderConfig, type ChatMessage, type ContentBlock } from "./llm-providers"
+import {
+ buildAnthropicUrl,
+ getProviderConfig,
+ supportsImageInput,
+ type ChatMessage,
+ type ContentBlock,
+} from "./llm-providers"
import type { LlmConfig } from "@/stores/wiki-store"
const TINY_PNG_B64 =
@@ -135,6 +141,165 @@ describe("Anthropic buildBody — vision content", () => {
})
})
+describe("MiniMax Anthropic-compatible endpoint", () => {
+ it("normalizes bare MiniMax global host to the Anthropic messages path", () => {
+ expect(buildAnthropicUrl("https://api.minimax.io")).toBe(
+ "https://api.minimax.io/anthropic/v1/messages",
+ )
+ })
+
+ it("normalizes bare MiniMax China host to the Anthropic messages path", () => {
+ expect(buildAnthropicUrl("https://api.minimaxi.com")).toBe(
+ "https://api.minimaxi.com/anthropic/v1/messages",
+ )
+ })
+
+ it("recovers from MiniMax URLs pasted as /v1/messages", () => {
+ expect(buildAnthropicUrl("https://api.minimaxi.com/v1/messages")).toBe(
+ "https://api.minimaxi.com/anthropic/v1/messages",
+ )
+ })
+
+ it("keeps already-correct MiniMax Anthropic URLs intact", () => {
+ expect(buildAnthropicUrl("https://api.minimaxi.com/anthropic/v1/messages")).toBe(
+ "https://api.minimaxi.com/anthropic/v1/messages",
+ )
+ })
+
+ it("uses Bearer auth for bare MiniMax endpoints after normalization", () => {
+ const provider = getProviderConfig(mkConfig({
+ provider: "custom",
+ apiKey: "sk-minimax",
+ model: "MiniMax-M2",
+ customEndpoint: "https://api.minimaxi.com",
+ apiMode: "anthropic_messages",
+ }))
+
+ expect(provider.url).toBe("https://api.minimaxi.com/anthropic/v1/messages")
+ expect(provider.headers.Authorization).toBe("Bearer sk-minimax")
+ expect(provider.headers["x-api-key"]).toBeUndefined()
+ })
+
+ it("allows official MiniMax-M3 Anthropic image input", () => {
+ const body = getProviderConfig(mkConfig({
+ provider: "custom",
+ apiKey: "sk-minimax",
+ model: "MiniMax-M3",
+ customEndpoint: "https://api.minimaxi.com",
+ apiMode: "anthropic_messages",
+ })).buildBody([visionMessage()]) as { messages: Array<{ content: unknown }> }
+
+ const content = body.messages[0].content as Array<{ type: string }>
+ expect(content[1]).toMatchObject({
+ type: "image",
+ source: { type: "base64", media_type: "image/png", data: TINY_PNG_B64 },
+ })
+ })
+
+ it("fails early for official MiniMax M2.x Anthropic image input", () => {
+ const provider = getProviderConfig(mkConfig({
+ provider: "custom",
+ apiKey: "sk-minimax",
+ model: "MiniMax-M2",
+ customEndpoint: "https://api.minimaxi.com",
+ apiMode: "anthropic_messages",
+ }))
+
+ expect(() => provider.buildBody([visionMessage()])).toThrow(
+ /MiniMax image input is supported only by MiniMax-M3/,
+ )
+ })
+
+ it("fails early for native MiniMax M2.x image input", () => {
+ const provider = getProviderConfig(mkConfig({
+ provider: "minimax",
+ apiKey: "sk-minimax",
+ model: "MiniMax-M2",
+ customEndpoint: "https://api.minimaxi.com",
+ }))
+
+ expect(() => provider.buildBody([visionMessage()])).toThrow(
+ /MiniMax image input is supported only by MiniMax-M3/,
+ )
+ })
+
+ it("allows native MiniMax-M3 image input", () => {
+ const body = getProviderConfig(mkConfig({
+ provider: "minimax",
+ apiKey: "sk-minimax",
+ model: "MiniMax-M3",
+ customEndpoint: "https://api.minimaxi.com",
+ })).buildBody([visionMessage()]) as { messages: Array<{ content: unknown }> }
+
+ const content = body.messages[0].content as Array<{ type: string }>
+ expect(content[1]).toMatchObject({ type: "image" })
+ })
+
+ it("still allows text-only MiniMax Anthropic requests", () => {
+ const body = getProviderConfig(mkConfig({
+ provider: "minimax",
+ apiKey: "sk-minimax",
+ model: "MiniMax-M2",
+ customEndpoint: "https://api.minimaxi.com",
+ })).buildBody([{ role: "user", content: "hi" }]) as Record<string, unknown>
+
+ expect(body.model).toBe("MiniMax-M2")
+ expect(body.messages).toEqual([{ role: "user", content: "hi" }])
+ })
+
+ it("reports official MiniMax-M3 Anthropic endpoints as image-capable", () => {
+ expect(supportsImageInput(mkConfig({
+ provider: "custom",
+ model: "MiniMax-M3",
+ customEndpoint: "https://api.minimaxi.com",
+ apiMode: "anthropic_messages",
+ }))).toBe(true)
+ })
+
+ it("reports native MiniMax-M3 as image-capable and native MiniMax M2.x as not image-capable", () => {
+ expect(supportsImageInput(mkConfig({
+ provider: "minimax",
+ model: "MiniMax-M3",
+ customEndpoint: "https://api.minimaxi.com",
+ }))).toBe(true)
+ expect(supportsImageInput(mkConfig({
+ provider: "minimax",
+ model: "MiniMax-M3-Preview",
+ customEndpoint: "https://api.minimaxi.com",
+ }))).toBe(true)
+ expect(supportsImageInput(mkConfig({
+ provider: "minimax",
+ model: "MiniMax-M2.7",
+ customEndpoint: "https://api.minimaxi.com",
+ }))).toBe(false)
+ })
+
+ it("reports Codex CLI as not image-capable because images are text-omitted", () => {
+ expect(supportsImageInput(mkConfig({
+ provider: "codex-cli",
+ model: "gpt-5-codex",
+ }))).toBe(false)
+ })
+
+ it("reports official MiniMax M2.x Anthropic endpoints as not image-capable", () => {
+ expect(supportsImageInput(mkConfig({
+ provider: "custom",
+ model: "MiniMax-M2",
+ customEndpoint: "https://api.minimaxi.com",
+ apiMode: "anthropic_messages",
+ }))).toBe(false)
+ })
+
+ it("does not block non-MiniMax custom Anthropic-compatible endpoints from image input", () => {
+ expect(supportsImageInput(mkConfig({
+ provider: "custom",
+ model: "claude-compatible-vision",
+ customEndpoint: "https://example.com/anthropic",
+ apiMode: "anthropic_messages",
+ }))).toBe(true)
+ })
+})
+
describe("Google buildBody — vision content", () => {
it("emits parts with inline_data for images", () => {
const cfg = mkConfig({ provider: "google", model: "gemini-2.5-pro" })
@@ -150,6 +315,75 @@ describe("Google buildBody — vision content", () => {
})
})
+describe("Zhipu BigModel / GLM vision content", () => {
+ it("uses OpenAI-compatible image_url blocks for GLM vision models", () => {
+ const provider = getProviderConfig(mkConfig({
+ provider: "custom",
+ apiKey: "zhipu-key",
+ model: "glm-4.5v",
+ customEndpoint: "https://open.bigmodel.cn/api/paas/v4",
+ apiMode: "chat_completions",
+ }))
+ const body = provider.buildBody([visionMessage()]) as {
+ messages: Array<{ content: unknown }>
+ }
+ const content = body.messages[0].content as Array<{ type: string }>
+
+ expect(provider.url).toBe("https://open.bigmodel.cn/api/paas/v4/chat/completions")
+ expect(content[1]).toEqual({
+ type: "image_url",
+ image_url: { url: `data:image/png;base64,${TINY_PNG_B64}` },
+ })
+ })
+
+ it("reports BigModel GLM vision models as image-capable", () => {
+ for (const model of ["glm-5v-turbo", "glm-4.6v", "glm-4.5v", "glm-4v-plus"]) {
+ expect(supportsImageInput(mkConfig({
+ provider: "custom",
+ model,
+ customEndpoint: "https://open.bigmodel.cn/api/paas/v4",
+ apiMode: "chat_completions",
+ }))).toBe(true)
+ }
+ })
+
+ it("reports BigModel text-only GLM models as not image-capable", () => {
+ for (const model of ["glm-5.1", "glm-5", "glm-4.7", "glm-4.6", "glm-4.5-flash"]) {
+ expect(supportsImageInput(mkConfig({
+ provider: "custom",
+ model,
+ customEndpoint: "https://open.bigmodel.cn/api/paas/v4",
+ apiMode: "chat_completions",
+ }))).toBe(false)
+ }
+ })
+
+ it("fails early when BigModel text-only GLM models receive image input", () => {
+ const provider = getProviderConfig(mkConfig({
+ provider: "custom",
+ apiKey: "zhipu-key",
+ model: "glm-4.6",
+ customEndpoint: "https://open.bigmodel.cn/api/paas/v4",
+ apiMode: "chat_completions",
+ }))
+
+ expect(() => provider.buildBody([visionMessage()])).toThrow(
+ /Zhipu BigModel image input is supported only by GLM vision models/,
+ )
+ })
+
+ it("keeps the default permissive image capability for OpenAI and Claude Code providers", () => {
+ expect(supportsImageInput(mkConfig({
+ provider: "openai",
+ model: "gpt-4o-mini",
+ }))).toBe(true)
+ expect(supportsImageInput(mkConfig({
+ provider: "claude-code",
+ model: "claude-sonnet-4-5-20250929",
+ }))).toBe(true)
+ })
+})
+
describe("Azure OpenAI provider", () => {
it("uses Azure deployment URL and api-key auth", () => {
const cfg = getProviderConfig(mkConfig({
diff --git a/src/lib/llm-providers.ts b/src/lib/llm-providers.ts
index 9db2fa2b..d96d907e 100644
--- a/src/lib/llm-providers.ts
+++ b/src/lib/llm-providers.ts
@@ -301,6 +301,18 @@ function isXiaomiMimoEndpoint(config: LlmConfig): boolean {
|| /\.?xiaomimimo\.com(?::|\/|$)/i.test(config.customEndpoint)
}
+function isBigModelEndpoint(config: LlmConfig): boolean {
+ return /(?:^|\/\/)open\.bigmodel\.cn(?:[:/]|$)/i.test(config.customEndpoint)
+}
+
+function isGlmVisionModel(model: string): boolean {
+ const normalized = model.trim().toLowerCase()
+ return /(?:^|[-_.])glm[-_.]5v[-_.]turbo(?:[-_.]|$)/i.test(normalized)
+ || /(?:^|[-_.])glm[-_.]4\.?6v(?:[-_.]|$)/i.test(normalized)
+ || /(?:^|[-_.])glm[-_.]4\.?5v(?:[-_.]|$)/i.test(normalized)
+ || /(?:^|[-_.])glm[-_.]4v(?:[-_.]|$)/i.test(normalized)
+}
+
function isOpenAiStrictCompletionModel(config: LlmConfig): boolean {
if ((config.provider === "azure" || (config.provider === "custom" && isAzureOpenAiEndpoint(config.customEndpoint)))
&& config.azureModelFamily === "gpt5") {
@@ -376,6 +388,7 @@ function buildOpenAiCompatibleBody(
messages: ChatMessage[],
overrides?: RequestOverrides,
): Record<string, unknown> {
+ assertBigModelImageSupport(config, messages)
const reasoning = effectiveReasoning(config, overrides)
const body: Record<string, unknown> = buildOpenAiBody(messages, stripWireAgnosticOverrides(overrides))
adaptOpenAiStrictCompletionBody(config, body)
@@ -556,6 +569,12 @@ function buildAnthropicBodyWithReasoning(
return body
}
+function hasImageContent(messages: ChatMessage[]): boolean {
+ return messages.some((m) =>
+ Array.isArray(m.content) && m.content.some((block) => block.type === "image"),
+ )
+}
+
/**
* Some Anthropic-compatible third-party endpoints (MiniMax global + CN)
* serve the Messages API but authenticate with `Authorization: Bearer`
@@ -592,6 +611,73 @@ function requiresBearerAuth(url: string): boolean {
)
}
+function isMiniMaxAnthropicHost(url: string): boolean {
+ try {
+ const parsed = new URL(url)
+ return parsed.hostname === "api.minimax.io" || parsed.hostname === "api.minimaxi.com"
+ } catch {
+ return false
+ }
+}
+
+function normalizeMiniMaxAnthropicBase(base: string): string {
+ if (!isMiniMaxAnthropicHost(base)) return base
+
+ try {
+ const parsed = new URL(base)
+ const path = parsed.pathname.replace(/\/+$/, "")
+ if (path === "" || path === "/" || /^\/v\d+(?:\/messages)?$/i.test(path)) {
+ parsed.pathname = "/anthropic"
+ parsed.search = ""
+ parsed.hash = ""
+ return parsed.toString().replace(/\/+$/, "")
+ }
+ } catch {
+ return base
+ }
+
+ return base
+}
+
+function isOfficialMiniMaxAnthropicUrl(url: string): boolean {
+ if (!isMiniMaxAnthropicHost(url)) return false
+ try {
+ const parsed = new URL(url)
+ return /^\/anthropic(?:\/|$)/i.test(parsed.pathname)
+ } catch {
+ return false
+ }
+}
+
+function isMiniMaxM3Model(model: string): boolean {
+ return /^minimax-m3(?:[-_.]|$)/i.test(model.trim())
+}
+
+function assertMiniMaxImageSupport(url: string, model: string, messages: ChatMessage[]): void {
+ if (!isOfficialMiniMaxAnthropicUrl(url) || !hasImageContent(messages) || isMiniMaxM3Model(model)) return
+ throw new Error(
+ "MiniMax image input is supported only by MiniMax-M3 on the official Anthropic-compatible endpoint. Switch the model to MiniMax-M3 or use another vision-capable provider.",
+ )
+}
+
+function assertBigModelImageSupport(config: LlmConfig, messages: ChatMessage[]): void {
+ if (!isBigModelEndpoint(config) || !hasImageContent(messages) || isGlmVisionModel(config.model)) return
+ throw new Error(
+ "Zhipu BigModel image input is supported only by GLM vision models. Switch to glm-5v-turbo, glm-4.6v, glm-4.5v, or glm-4v-plus.",
+ )
+}
+
+export function supportsImageInput(config: LlmConfig): boolean {
+ if (config.provider === "codex-cli") return false
+ if (config.provider === "minimax") return isMiniMaxM3Model(config.model)
+ if (isBigModelEndpoint(config)) return isGlmVisionModel(config.model)
+ if ((config.provider === "custom") && (config.apiMode ?? "chat_completions") === "anthropic_messages") {
+ const url = buildAnthropicUrl(config.customEndpoint)
+ return !isOfficialMiniMaxAnthropicUrl(url) || isMiniMaxM3Model(config.model)
+ }
+ return true
+}
+
/**
* Build the final POST URL for an Anthropic-wire endpoint given whatever
* base the user provided. Handles every shape we've seen in the wild:
@@ -606,7 +692,7 @@ function requiresBearerAuth(url: string): boolean {
* ".../v1/v1/messages" (404) whenever a user typed a URL ending in /v1.
*/
export function buildAnthropicUrl(base: string): string {
- const trimmed = base.replace(/\/+$/, "")
+ const trimmed = normalizeMiniMaxAnthropicBase(base).replace(/\/+$/, "")
if (/\/v\d+\/messages$/i.test(trimmed)) return trimmed
if (/\/v\d+$/i.test(trimmed)) return `${trimmed}/messages`
return `${trimmed}/v1/messages`
@@ -735,10 +821,13 @@ export function getProviderConfig(config: LlmConfig): ProviderConfig {
return {
url,
headers: buildAnthropicHeaders(apiKey, url),
- buildBody: (messages, overrides) => ({
- ...buildAnthropicBodyWithReasoning(config, messages, overrides),
- model,
- }),
+ buildBody: (messages, overrides) => {
+ assertMiniMaxImageSupport(url, model, messages)
+ return {
+ ...buildAnthropicBodyWithReasoning(config, messages, overrides),
+ model,
+ }
+ },
parseStream: parseAnthropicLine,
}
}
@@ -815,10 +904,13 @@ export function getProviderConfig(config: LlmConfig): ProviderConfig {
return {
url,
headers: buildAnthropicHeaders(apiKey, url),
- buildBody: (messages, overrides) => ({
- ...buildAnthropicBodyWithReasoning(config, messages, overrides),
- model,
- }),
+ buildBody: (messages, overrides) => {
+ assertMiniMaxImageSupport(url, model, messages)
+ return {
+ ...buildAnthropicBodyWithReasoning(config, messages, overrides),
+ model,
+ }
+ },
parseStream: parseAnthropicLine,
}
}
@@ -844,10 +936,13 @@ export function getProviderConfig(config: LlmConfig): ProviderConfig {
return {
url,
headers: buildAnthropicHeaders(apiKey, url),
- buildBody: (messages, overrides) => ({
- ...buildAnthropicBodyWithReasoning(config, messages, overrides),
- model,
- }),
+ buildBody: (messages, overrides) => {
+ assertMiniMaxImageSupport(url, model, messages)
+ return {
+ ...buildAnthropicBodyWithReasoning(config, messages, overrides),
+ model,
+ }
+ },
parseStream: parseAnthropicLine,
}
}
diff --git a/src/lib/markdown-image-resolver.test.ts b/src/lib/markdown-image-resolver.test.ts
index 0bd2ba51..7d0bf7ec 100644
--- a/src/lib/markdown-image-resolver.test.ts
+++ b/src/lib/markdown-image-resolver.test.ts
@@ -61,25 +61,65 @@ describe("resolveMarkdownImageSrc", () => {
).toBe("tauri-asset:/Users/me/MyWiki/wiki/entities/transformer/diagram.png")
})
- it("treats an absolute POSIX path as a literal filesystem path", () => {
- // Some user content may legitimately reference an image outside
- // the wiki root (a system-wide asset, a mounted drive). Don't
- // re-anchor — just convert.
+ it("converts an absolute POSIX path only when it stays inside the project", () => {
expect(
- resolveMarkdownImageSrc("/var/data/screenshot.png", PROJECT),
- ).toBe("tauri-asset:/var/data/screenshot.png")
+ resolveMarkdownImageSrc("/Users/me/MyWiki/raw/assets/screenshot.png", PROJECT),
+ ).toBe("tauri-asset:/Users/me/MyWiki/raw/assets/screenshot.png")
})
- it("treats a Windows drive-letter path as absolute", () => {
+ it("decodes percent-encoded CJK in absolute in-project paths", () => {
expect(
- resolveMarkdownImageSrc("C:/Users/me/Pictures/x.png", PROJECT),
- ).toBe("tauri-asset:C:/Users/me/Pictures/x.png")
+ resolveMarkdownImageSrc(
+ "/Users/me/MyWiki/wiki/media/%E4%B8%AD%E6%96%87/x.png",
+ PROJECT,
+ ),
+ ).toBe("tauri-asset:/Users/me/MyWiki/wiki/media/中文/x.png")
+ })
+
+ it("decodes percent-encoded CJK when the project path itself contains CJK", () => {
+ expect(
+ resolveMarkdownImageSrc(
+ "/Users/me/%E6%88%91%E7%9A%84Wiki/wiki/media/x.png",
+ "/Users/me/我的Wiki",
+ ),
+ ).toBe("tauri-asset:/Users/me/我的Wiki/wiki/media/x.png")
+ })
+
+ it("does not convert an absolute POSIX path outside the project", () => {
+ expect(resolveMarkdownImageSrc("/var/data/screenshot.png", PROJECT)).toBe(
+ "/var/data/screenshot.png",
+ )
+ })
+
+ it("does not convert an absolute POSIX path that escapes via .. segments", () => {
+ const src = "/Users/me/MyWiki/../../../etc/passwd.png"
+ expect(resolveMarkdownImageSrc(src, PROJECT)).toBe(src)
+ })
+
+ it("does not treat a sibling path with the same prefix as inside the project", () => {
+ const src = "/Users/me/MyWikiSecret/screenshot.png"
+ expect(resolveMarkdownImageSrc(src, PROJECT)).toBe(src)
+ })
+
+ it("converts a Windows drive-letter path only when it stays inside the project", () => {
+ expect(
+ resolveMarkdownImageSrc(
+ "C:/Users/me/MyWiki/raw/assets/x.png",
+ "C:/Users/me/MyWiki",
+ ),
+ ).toBe("tauri-asset:C:/Users/me/MyWiki/raw/assets/x.png")
+ })
+
+ it("does not convert a Windows drive-letter path outside the project", () => {
+ expect(
+ resolveMarkdownImageSrc("C:/Users/me/Pictures/x.png", "C:/Users/me/MyWiki"),
+ ).toBe("C:/Users/me/Pictures/x.png")
})
- it("treats a UNC path as absolute", () => {
+ it("does not convert a UNC path outside the project", () => {
expect(
resolveMarkdownImageSrc("\\\\share\\folder\\img.png", PROJECT),
- ).toBe("tauri-asset:\\\\share\\folder\\img.png")
+ ).toBe("\\\\share\\folder\\img.png")
})
it("returns the raw src unchanged when no project is loaded", () => {
@@ -166,19 +206,16 @@ describe("resolveMarkdownImageSrc", () => {
).toBe("tauri-asset:C:/Users/me/MyWiki/raw/assets/x.png")
})
- it("collapses `..` segments against the file dir", () => {
- // From /Users/me/MyWiki/raw/sources, four `..` pop
- // sources, raw, MyWiki, me — landing in /Users.
+ it("does not convert a file-relative path that escapes the project", () => {
expect(
resolveMarkdownImageSrc("../../../../etc/x.png", PROJECT, `${PROJECT}/raw/sources`),
- ).toBe("tauri-asset:/Users/etc/x.png")
+ ).toBe("../../../../etc/x.png")
})
- it("clamps `..` at the filesystem root rather than producing ../ garbage", () => {
- // More `..` than there are segments → clamped at root.
+ it("does not convert paths that climb above the project root", () => {
expect(
- resolveMarkdownImageSrc("../../x.png", PROJECT, "/a"),
- ).toBe("tauri-asset:/x.png")
+ resolveMarkdownImageSrc("../../../x.png", PROJECT, `${PROJECT}/wiki/concepts`),
+ ).toBe("../../../x.png")
})
it("still uses wiki-root fallback when no currentFileDir is given", () => {
@@ -249,10 +286,10 @@ describe("resolveMarkdownImageSrc", () => {
).toBe("tauri-asset:/Users/me/MyWiki/wiki/media/slug/img-2.png")
})
- it("passes absolute srcs through convertFileSrc even with a file dir set", () => {
+ it("does not convert project-external absolute srcs even with a file dir set", () => {
expect(
resolveMarkdownImageSrc("/var/data/x.png", PROJECT, `${PROJECT}/raw/sources`),
- ).toBe("tauri-asset:/var/data/x.png")
+ ).toBe("/var/data/x.png")
})
})
})
diff --git a/src/lib/markdown-image-resolver.ts b/src/lib/markdown-image-resolver.ts
index 9bcad7b6..5cf017a9 100644
--- a/src/lib/markdown-image-resolver.ts
+++ b/src/lib/markdown-image-resolver.ts
@@ -39,6 +39,28 @@ import { normalizePath } from "@/lib/path-utils"
const PASSTHROUGH_RE = /^(https?:|data:|blob:|file:|tauri:)/i
+function trimTrailingSlash(path: string): string {
+ return path.replace(/\/+$/, "")
+}
+
+function comparePath(path: string): string {
+ return /^[a-zA-Z]:/.test(path) ? path.toLowerCase() : path
+}
+
+function isInsideProject(path: string, projectPath: string): boolean {
+ const root = comparePath(trimTrailingSlash(normalizePath(projectPath)))
+ const candidate = comparePath(trimTrailingSlash(normalizePath(path)))
+ return candidate === root || candidate.startsWith(`${root}/`)
+}
+
+function decodePathSrc(src: string): string {
+ try {
+ return decodeURIComponent(src)
+ } catch {
+ return src
+ }
+}
+
/**
* Collapse `.` and `..` segments in a forward-slashed path without
* touching the filesystem. A leading `..` that would escape the
@@ -86,9 +108,16 @@ export function resolveMarkdownImageSrc(
const isAbsolute =
rawSrc.startsWith("/") || /^[a-zA-Z]:/.test(rawSrc) || rawSrc.startsWith("\\\\")
- // Absolute paths get fed straight to convertFileSrc — the user (or
- // some plugin) explicitly chose that path; we don't second-guess.
- if (isAbsolute) return convertFileSrc(rawSrc)
+ // Absolute paths are allowed only inside the current project. This resolver
+ // is used for both generated/imported markdown and normal wiki reading, so
+ // this intentionally trades off rendering arbitrary external local images
+ // for a single safe rule: markdown cannot turn into a project-external local
+ // file read through Tauri's asset server. External web URLs still pass
+ // through via PASSTHROUGH_RE above.
+ if (isAbsolute) {
+ const absolute = collapsePath(normalizePath(decodePathSrc(rawSrc)))
+ return isInsideProject(absolute, pp) ? convertFileSrc(absolute) : rawSrc
+ }
// Strip a leading `./` for cleanliness; treat `media/foo.png` and
// `./media/foo.png` identically.
@@ -106,12 +135,7 @@ export function resolveMarkdownImageSrc(
// "%E6", finds nothing, and the image 404s (showing the alt text).
// Decoding is wrapped because a malformed `%` sequence throws; in
// that case we keep the raw value rather than crash the renderer.
- let cleaned: string
- try {
- cleaned = decodeURIComponent(stripped)
- } catch {
- cleaned = stripped
- }
+ const cleaned = decodePathSrc(stripped)
// Generated-wiki convention takes precedence: ingest always emits
// embedded images as `media/<source-slug>/img-N.png` — a path
@@ -138,7 +162,7 @@ export function resolveMarkdownImageSrc(
dir.startsWith("/") || /^[a-zA-Z]:/.test(dir) || dir.startsWith("\\\\")
const baseDir = dirIsAbsolute ? dir : `${pp}/${dir}`
const absolute = collapsePath(`${baseDir.replace(/\/+$/, "")}/${cleaned}`)
- return convertFileSrc(absolute)
+ return isInsideProject(absolute, pp) ? convertFileSrc(absolute) : rawSrc
}
// Fallback: resolve as wiki-root-relative. Image references in
@@ -146,5 +170,5 @@ export function resolveMarkdownImageSrc(
// so the path is stable regardless of page depth, and callers
// without a file context (chat replies) rely on it too.
const absolute = collapsePath(`${pp}/wiki/${cleaned}`)
- return convertFileSrc(absolute)
+ return isInsideProject(absolute, pp) ? convertFileSrc(absolute) : rawSrc
}
diff --git a/src/lib/persist.integration.test.ts b/src/lib/persist.integration.test.ts
index e5906540..2a986d93 100644
--- a/src/lib/persist.integration.test.ts
+++ b/src/lib/persist.integration.test.ts
@@ -219,6 +219,29 @@ describe("chat persistence — round-trip (new format)", () => {
expect(loaded.messages).toHaveLength(2)
})
+ it("does not persist base64 chat images into conversation JSON", async () => {
+ const convs = [makeConv("c1", "Vision")]
+ const msg: DisplayMessage = {
+ ...makeMsg("m1", "c1", "what is this?"),
+ images: [{ mediaType: "image/png", dataBase64: "A".repeat(1024) }],
+ }
+
+ await saveChatHistory(tmp.path, convs, [msg])
+ expect(msg.images).toHaveLength(1)
+ expect(msg.images?.[0].dataBase64).toBe("A".repeat(1024))
+
+ const raw = await readFileRaw(`${tmp.path}/.llm-wiki/chats/c1.json`)
+ expect(raw).not.toContain("dataBase64")
+ expect(raw).not.toContain("image/png")
+
+ const loaded = await loadChatHistory(tmp.path)
+ expect(loaded.messages[0]).toMatchObject({
+ id: "m1",
+ content: "what is this?",
+ })
+ expect(loaded.messages[0].images).toBeUndefined()
+ })
+
it("caps each conversation's persisted messages at 100 (oldest dropped)", async () => {
const convs = [makeConv("c1")]
const msgs = Array.from({ length: 150 }, (_, i) =>
diff --git a/src/lib/persist.ts b/src/lib/persist.ts
index 1acbb371..cf8ca398 100644
--- a/src/lib/persist.ts
+++ b/src/lib/persist.ts
@@ -46,6 +46,12 @@ interface PersistedChatData {
messages: DisplayMessage[]
}
+function stripPersistedMessageImages(msg: DisplayMessage): DisplayMessage {
+ if (!msg.images || msg.images.length === 0) return msg
+ const { images: _images, ...rest } = msg
+ return rest
+}
+
export async function saveChatHistory(
projectPath: string,
conversations: Conversation[],
@@ -64,7 +70,10 @@ export async function saveChatHistory(
const byConversation = new Map<string, DisplayMessage[]>()
for (const msg of messages) {
const list = byConversation.get(msg.conversationId) ?? []
- list.push(msg)
+ // Images can be multi-megabyte base64 payloads. Keep them in memory for the
+ // current chat turn, but don't persist them into chat JSON where they would
+ // quickly bloat auto-save files and project backups.
+ list.push(stripPersistedMessageImages(msg))
byConversation.set(msg.conversationId, list)
}