feat: Initial commit of vision-bridge plugin and agent

agents/vision.md

@@ -0,0 +1,18 @@
+---
+description: Анализ изображений и скриншотов (OCR текста или разбор UI)
+mode: subagent
+model: ludmila-ai/gemini/gemini-2.5-flash
+temperature: 0.1
+permission:
+  read: allow
+  glob: allow
+  external_directory: allow
+  edit: deny
+  bash: deny
+---
+Ты анализируешь изображение. Сначала определи тип контента:
+1) Преимущественно ТЕКСТ/документ → извлеки текст ДОСЛОВНО, сохрани структуру
+  (заголовки, списки, колонки, код). Не пересказывай и не исправляй.
+2) ЭЛЕМЕНТЫ ИНТЕРФЕЙСА (веб/моб/десктоп) → опиши: тип экрана, компоненты и
+  расположение, тексты кнопок/полей, состояния, интерактивность, стили.
+3) Смешанное → сделай оба пункта раздельно.

src/vision-bridge.ts

@@ -0,0 +1,280 @@
+import type { Plugin, Hooks } from "@opencode-ai/plugin"
+
+const VISION_PROMPT = `You are a vision assistant. Analyze the image in detail and produce a thorough text representation.
+
+Rules:
+1. If the image is predominantly TEXT (code, error message, document, chat, log, terminal output) — transcribe the text VERBATIM. Preserve structure (headings, lists, columns, indentation). Do not paraphrase or "fix" typos.
+2. If the image shows a UI / INTERFACE (web, mobile, desktop) — describe: type of screen, components and their layout, exact button/label/field texts, states, interactivity, visual style.
+3. If mixed — do BOTH, in separate labeled sections.
+4. If the image is a graph/diagram/chart — describe axes, labels, values, and the overall message.
+
+Be precise and exhaustive. This text will be given to another model that cannot see the image.`
+
+type ImagePart = {
+  id: string
+  sessionID: string
+  messageID: string
+  type: "file"
+  mime: string
+  filename?: string
+  url: string
+}
+
+type TextPart = {
+  id: string
+  sessionID: string
+  messageID: string
+  type: "text"
+  text: string
+  synthetic?: boolean
+}
+
+type AnyPart = { type: string; [key: string]: any }
+
+interface ProviderInfo {
+  baseURL: string
+  apiKey: string
+  visionModel: string
+}
+
+const IMAGE_MIME_RE = /^image\//
+
+function isVisionModel(modelID: string | undefined, visionModels: Set<string>): boolean {
+  if (!modelID) return false
+  // Normalize: strip provider prefix if present, and also handle "provider/model" form
+  for (const m of visionModels) {
+    if (modelID === m) return true
+    if (modelID.endsWith("/" + m)) return true
+    if (m.endsWith("/" + modelID)) return true
+  }
+  return false
+}
+
+function mimeToDataUrl(url: string, mime: string): string {
+  // url is already data:image/png;base64,...
+  if (url.startsWith("data:")) return url
+  return `data:${mime};base64,${url}`
+}
+
+async function recognizeImage(
+  provider: ProviderInfo,
+  modelID: string,
+  dataUrl: string,
+  timeoutMs = 60000,
+): Promise<string> {
+  const controller = new AbortController()
+  const timer = setTimeout(() => controller.abort(), timeoutMs)
+  try {
+    const res = await fetch(`${provider.baseURL}/chat/completions`, {
+      method: "POST",
+      headers: {
+        "Content-Type": "application/json",
+        Authorization: `Bearer ${provider.apiKey}`,
+      },
+      body: JSON.stringify({
+        model: modelID,
+        messages: [
+          {
+            role: "user",
+            content: [
+              { type: "text", text: VISION_PROMPT },
+              { type: "image_url", image_url: { url: dataUrl } },
+            ],
+          },
+        ],
+        max_tokens: 4096,
+        temperature: 0.1,
+      }),
+      signal: controller.signal,
+    })
+    if (!res.ok) {
+      const body = await res.text().catch(() => "")
+      throw new Error(`Vision API ${res.status}: ${body.slice(0, 300)}`)
+    }
+    const data = await res.json()
+    const text = data?.choices?.[0]?.message?.content
+    if (typeof text !== "string" || !text.trim()) {
+      throw new Error("Vision API returned empty content")
+    }
+    return text.trim()
+  } finally {
+    clearTimeout(timer)
+  }
+}
+
+export const VisionBridge: Plugin = async () => {
+  let config: any = null
+  let providerInfo: ProviderInfo | null = null
+  let visionModels = new Set<string>()
+
+  // Image content cache keyed by image data hash to avoid re-analyzing
+  // the same image on every turn.
+  const cache = new Map<string, string>()
+
+  // Track which sessions are running a vision-capable model.
+  const visionSessions = new Map<string, boolean>()
+
+  const providerID = "ludmila-ai"
+
+  function loadProvider(configAny: any) {
+    const p = configAny?.provider?.[providerID]
+    const baseURL = p?.options?.baseURL
+    const apiKey = p?.options?.apiKey
+    if (!baseURL || !apiKey) {
+      providerInfo = null
+      return
+    }
+    providerInfo = { baseURL: baseURL.replace(/\/$/, ""), apiKey, visionModel: "" }
+    const models = p?.models ?? {}
+    visionModels = new Set<string>()
+    let visionApiName = ""
+    for (const [id, m] of Object.entries<any>(models)) {
+      if (m?.attachment === true) {
+        visionModels.add(id)
+        if (!visionApiName) visionApiName = typeof m?.id === "string" && m.id ? m.id : id
+      }
+    }
+    // The vision model used for recognition = the model flagged attachment:true.
+    // Change it in the config by moving the `attachment: true` flag.
+    providerInfo.visionModel = visionApiName
+  }
+
+  function hashImage(dataUrl: string): string {
+    // Use a small fixed part of the base64 as a cheap cache key
+    const i = dataUrl.indexOf(",")
+    const b64 = i >= 0 ? dataUrl.slice(i + 1) : dataUrl
+    let h = 0
+    const step = Math.max(1, Math.floor(b64.length / 512))
+    for (let k = 0; k < b64.length; k += step) {
+      h = (h * 31 + b64.charCodeAt(k)) | 0
+    }
+    return h.toString(16) + ":" + b64.length
+  }
+
+  async function processImagePart(
+    part: ImagePart,
+    modelID: string | undefined,
+    isVision: boolean,
+  ): Promise<AnyPart> {
+    if (isVision || !providerInfo) {
+      // Vision model sees the image natively, or no provider config → keep original part
+      return part
+    }
+    const base: TextPart = {
+      id: part.id,
+      sessionID: part.sessionID,
+      messageID: part.messageID,
+      type: "text",
+      synthetic: true,
+    }
+    const dataUrl = mimeToDataUrl(part.url, part.mime)
+    const key = hashImage(dataUrl)
+    let text = cache.get(key)
+    if (!text) {
+      try {
+        if (!providerInfo.visionModel) {
+          throw new Error('no model with "attachment": true configured for ' + providerID)
+        }
+        text = await recognizeImage(providerInfo, providerInfo.visionModel, dataUrl)
+        cache.set(key, text)
+        if (cache.size > 200) {
+          const first = cache.keys().next().value
+          if (first !== undefined) cache.delete(first)
+        }
+      } catch (err) {
+        const msg = err instanceof Error ? err.message : String(err)
+        text = `[Image analysis failed: ${msg}]`
+      }
+    }
+    return { ...base, text: `[Image: ${text}]` }
+  }
+
+  const hooks: Hooks = {
+    config: async (cfg: any) => {
+      config = cfg
+      loadProvider(cfg)
+    },
+
+    "chat.message": async (input: any, output: any) => {
+      if (!config) return
+      const parts: AnyPart[] = output.parts ?? []
+      const modelID = input.model?.modelID
+      const isVision = isVisionModel(modelID, visionModels)
+      if (input.sessionID) visionSessions.set(input.sessionID, isVision)
+      let changed = false
+      const newParts: AnyPart[] = []
+      for (const part of parts) {
+        if (
+          part.type === "file" &&
+          typeof part.mime === "string" &&
+          IMAGE_MIME_RE.test(part.mime) &&
+          typeof part.url === "string"
+        ) {
+          const result = await processImagePart(
+            {
+              id: part.id,
+              sessionID: part.sessionID,
+              messageID: part.messageID,
+              type: "file",
+              mime: part.mime,
+              filename: part.filename,
+              url: part.url,
+            },
+            modelID,
+            isVision,
+          )
+          newParts.push(result)
+          changed = true
+        } else {
+          newParts.push(part)
+        }
+      }
+      if (changed) output.parts = newParts
+    },
+
+    "experimental.chat.messages.transform": async (_input: any, output: any) => {
+      if (!config || !providerInfo) return
+      const msgs: { info: any; parts: AnyPart[] }[] = output.messages ?? []
+      for (const msg of msgs) {
+        if (!msg?.parts?.length) continue
+        // If this message belongs to a session known to use a vision model,
+        // keep its media intact (native multimodal experience).
+        const sessionID = msg.info?.sessionID ?? msg.parts[0]?.sessionID
+        if (sessionID && visionSessions.get(sessionID)) continue
+        let changed = false
+        const newParts: AnyPart[] = []
+        for (const part of msg.parts) {
+          if (
+            part.type === "file" &&
+            typeof part.mime === "string" &&
+            IMAGE_MIME_RE.test(part.mime) &&
+            typeof part.url === "string"
+          ) {
+            const result = await processImagePart(
+              {
+                id: part.id,
+                sessionID: part.sessionID,
+                messageID: part.messageID,
+                type: "file",
+                mime: part.mime,
+                filename: part.filename,
+                url: part.url,
+              },
+              undefined,
+              false,
+            )
+            newParts.push(result)
+            changed = true
+          } else {
+            newParts.push(part)
+          }
+        }
+        if (changed) msg.parts = newParts
+      }
+    },
+  }
+
+  return hooks
+}
+
+export default VisionBridge