pantheon-opencode 1.1.3-beta.0 → 1.2.0-beta.9.7bef03e

package/src/pantheon/vision.ts

@@ -0,0 +1,1340 @@
+import { createHash, randomUUID } from 'node:crypto'
+import { existsSync } from 'node:fs'
+import { mkdir, readFile, unlink, writeFile } from 'node:fs/promises'
+import { homedir, tmpdir } from 'node:os'
+import { join } from 'node:path'
+import type { Hooks, PluginInput } from '@opencode-ai/plugin'
+import type { FilePart, Part, TextPart, UserMessage } from '@opencode-ai/sdk'
+import type { ResolvedPreset } from './presets.mjs'
+import { hasVision, resolveActivePreset } from './presets.mjs'
+
+// ─── Constants ────────────────────────────────────────────────────────────
+
+const CONFIG_FILENAME = 'opencode-vision.json'
+const TEMP_DIR_NAME = 'pantheon-vision'
+
+/**
+ * Temp images younger than this survive a `session.idle` fired between
+ * auto-continue turns; abandoned files are swept by the next idle.
+ */
+export const TEMP_FILE_GRACE_MS = 120_000
+
+/**
+ * Global cap on temp images in `/tmp/pantheon-vision`. When the LRU registry
+ * (see `enforceTempFileCap`) exceeds this, the least-recently-used files older
+ * than TEMP_FILE_GRACE_MS are unlinked — never a young file that an active
+ * turn may still reference.
+ */
+export const TEMP_MAX_FILES = 200
+
+// Tool the model is instructed to call. Configurable via:
+//   1. env PANTHEON_VISION_TOOL
+//   2. config file imageAnalysisTool (project > user)
+//   3. dynamic detection of an available MCP vision tool
+//   4. the canonical Pantheon Vision MCP tool
+//
+// OpenCode 1.18.11 generates MCP tool IDs as sanitize(clientName) + '_' +
+// sanitize(name): the `pantheon-vision` MCP (FastMCP name) yields
+// `pantheon_vision_vision_describe` (underscores) — there is no `mcp__` prefix
+// in that version. The legacy `mcp__pantheon-vision__vision_describe` form is
+// still accepted when explicitly configured or detected.
+const DEFAULT_IMAGE_ANALYSIS_TOOL = 'pantheon_vision_vision_describe'
+
+/** True for IDs referencing the Pantheon Vision MCP — hyphen OR underscore server form. */
+const isPantheonVisionServerTool = (id: string): boolean => /[pP]antheon[-_]vision/.test(id)
+
+/** True for the describe/ocr/analyze tool suffix (sanitized MCP tool name). */
+const isPantheonVisionActionTool = (id: string): boolean =>
+  /vision_(describe|ocr|analyze)/i.test(id)
+
+// Enabled for ALL models by default ("paste and ask" universal). The config
+// file can restrict this with wildcard patterns.
+const DEFAULT_MODEL_PATTERNS: readonly string[] = ['*']
+
+// Only PNG/JPEG/WebP are intercepted (data: URLs need a known extension).
+const SUPPORTED_MIME_TYPES = new Set([
+  'image/png',
+  'image/jpeg',
+  'image/jpg',
+  'image/webp',
+  'image/gif',
+])
+const MIME_TO_EXTENSION: Record<string, string> = {
+  'image/png': 'png',
+  'image/jpeg': 'jpg',
+  'image/jpg': 'jpg',
+  'image/webp': 'webp',
+  'image/gif': 'gif',
+}
+
+const PROMPT_TEMPLATE_VARIABLES = [
+  '{imageList}',
+  '{imageCount}',
+  '{toolName}',
+  '{userText}',
+] as const
+
+// ─── Native vision (provider-direct, OpenAI-compatible) ─────────────────────
+// Preferred path: when an API key is available the plugin calls the multimodal
+// model DIRECTLY via the opencode Zen OpenAI-compatible endpoint, so any user
+// with a provider key gets vision WITHOUT an MCP tool. The MCP tool pattern
+// (imageAnalysisTool) remains the fallback for missing keys / provider errors.
+//
+// MVP scope: only opencode-go / opencode providers (the user's stack).
+// anthropic/openai native vision is phase 2 — those presets fall back to the
+// MCP tool pattern.
+
+const DEFAULT_VISION_MODEL = 'opencode-go/mimo-v2.5'
+const ZEN_BASE_URL = 'https://opencode.ai/zen/v1'
+const ZEN_GO_BASE_URL = 'https://opencode.ai/zen/go/v1'
+const NATIVE_PROMPT =
+  'Descreva fielmente esta imagem em detalhes (texto visível, cores, layout, objetos).'
+const NATIVE_TIMEOUT_MS = 20_000
+const NATIVE_MAX_TOKENS = 500
+const MAX_NATIVE_IMAGE_BYTES = 25 * 1024 * 1024
+
+/**
+ * Descriptions produced by the native path are cached per session, keyed by
+ * the user prompt + the content hashes of the attached images (see
+ * `buildDescriptionCacheKey`). TTL bounds staleness: an identical re-ask
+ * within the window reuses the stored text without another gateway call;
+ * after expiry the multimodal model is called again.
+ */
+export const DESCRIPTION_CACHE_TTL_MS = 30 * 60 * 1000
+
+// ─── Types ─────────────────────────────────────────────────────────────────
+
+type VisionConfig = {
+  models?: string[]
+  imageAnalysisTool?: string
+  promptTemplate?: string
+  mode?: VisionMode
+  visionModel?: string
+}
+
+export type ModelInfo = { providerID: string; modelID: string }
+
+export type NativeVisionTarget = { modelID: string; baseURL: string; apiKey: string }
+
+type NativeVisionEndpoint = { modelID: string; baseURL: string }
+
+export type VisionMode = 'native' | 'tool' | 'auto'
+
+/** Return whether OpenCode's current model is known to accept image input. */
+export function modelAcceptsImages(model: ModelInfo | undefined): boolean {
+  if (!model) return true
+  try {
+    return hasVision(`${model.providerID}/${model.modelID}`)
+  } catch {
+    // Unknown providers/models should retain OpenCode's native behavior.
+    return true
+  }
+}
+
+type SavedImage = {
+  path: string
+  partId: string
+  temporary: boolean
+}
+
+type NativeImagePayload = { url: string }
+
+// ─── Config loading (project > user) ──────────────────────────────────────
+
+// ─── OpenCode auth store (fallback credential source) ────────────────────
+// `opencode auth login` stores provider tokens in <data dir>/auth.json.
+// The vision plugin reads that store so connected users don't need to export
+// any env var. Mirrors the usage-bar pattern (src/plugins/tui) but stays
+// local to this module — no TUI imports.
+
+function defaultAuthStorePath(): string {
+  return join(homedir(), '.local', 'share', 'opencode', 'auth.json')
+}
+
+/**
+ * Read a provider token from opencode's auth store (`opencode auth login`).
+ * `statePath` is the opencode data dir (default ~/.local/share/opencode); the
+ * auth.json inside it maps provider IDs to `{ access, expires, ... }`.
+ * Returns the first non-expired `access` token across `providerIDs`, or null
+ * when the file is missing/corrupted or no entry has a usable token — callers
+ * fall back to the MCP tool pattern. The token is never logged.
+ */
+export async function readOpencodeAuthToken(
+  providerIDs: readonly string[],
+  statePath?: string,
+): Promise<string | null> {
+  const authPath = statePath ? join(statePath, 'auth.json') : defaultAuthStorePath()
+  let raw: string
+  try {
+    raw = await readFile(authPath, 'utf8')
+  } catch {
+    return null // missing/unreadable → no crash
+  }
+  let auth: Record<string, { access?: unknown; expires?: unknown } | undefined>
+  try {
+    auth = JSON.parse(raw) as typeof auth
+  } catch {
+    return null // corrupted → no crash
+  }
+  for (const id of providerIDs) {
+    const entry = auth[id]
+    const access = typeof entry?.access === 'string' ? entry.access.trim() : ''
+    if (access === '') continue
+    const expires = entry?.expires
+    if (typeof expires === 'number' && expires > 0 && Date.now() >= expires) continue
+    return access
+  }
+  return null
+}
+
+function userConfigPath(): string {
+  const xdg = process.env.XDG_CONFIG_HOME ?? join(homedir(), '.config')
+  return join(xdg, 'opencode', CONFIG_FILENAME)
+}
+
+async function readConfigFile(configPath: string): Promise<VisionConfig | null> {
+  if (!existsSync(configPath)) return null
+  try {
+    const parsed = JSON.parse(await readFile(configPath, 'utf8')) as Record<string, unknown>
+    const config: VisionConfig = {}
+    if (Array.isArray(parsed.models)) {
+      const models = parsed.models.filter((m): m is string => typeof m === 'string')
+      if (models.length > 0) config.models = models
+    }
+    if (typeof parsed.imageAnalysisTool === 'string' && parsed.imageAnalysisTool.trim() !== '') {
+      config.imageAnalysisTool = parsed.imageAnalysisTool
+    }
+    const template = parsed.promptTemplate
+    if (
+      typeof template === 'string' &&
+      PROMPT_TEMPLATE_VARIABLES.some((variable) => template.includes(variable))
+    ) {
+      config.promptTemplate = template
+    }
+    if (parsed.mode === 'native' || parsed.mode === 'tool' || parsed.mode === 'auto') {
+      config.mode = parsed.mode
+    }
+    if (typeof parsed.visionModel === 'string' && parsed.visionModel.trim() !== '') {
+      config.visionModel = parsed.visionModel.trim()
+    }
+    return Object.keys(config).length > 0 ? config : null
+  } catch {
+    return null
+  }
+}
+
+export async function loadVisionConfig(directory: string): Promise<VisionConfig> {
+  const project = await readConfigFile(join(directory, '.opencode', CONFIG_FILENAME))
+  const user = await readConfigFile(userConfigPath())
+  const merged: VisionConfig = {}
+  if (project?.models) merged.models = project.models
+  else if (user?.models) merged.models = user.models
+  if (project?.imageAnalysisTool) merged.imageAnalysisTool = project.imageAnalysisTool
+  else if (user?.imageAnalysisTool) merged.imageAnalysisTool = user.imageAnalysisTool
+  if (project?.promptTemplate) merged.promptTemplate = project.promptTemplate
+  else if (user?.promptTemplate) merged.promptTemplate = user.promptTemplate
+  if (project?.mode) merged.mode = project.mode
+  else if (user?.mode) merged.mode = user.mode
+  if (project?.visionModel) merged.visionModel = project.visionModel
+  else if (user?.visionModel) merged.visionModel = user.visionModel
+  return merged
+}
+
+// ─── Active-preset resolution (native vision model) ────────────────────────
+// Same candidate order as the plugin config hook, so the native vision model
+// matches the active preset (resolved.vision?.model) exactly.
+
+export function activePresetCandidates(): string[] {
+  const home = process.env.HOME ?? ''
+  const xdg = process.env.XDG_CONFIG_HOME ?? `${home}/.config`
+  return [
+    `${process.cwd()}/.pantheon/active-preset.json`,
+    `${xdg}/opencode/.pantheon/active-preset.json`,
+    `${home}/.opencode/.pantheon/active-preset.json`,
+  ]
+}
+
+/**
+ * Resolve the vision mode: env PANTHEON_VISION_MODE, then the config file's
+ * `mode` field, then default `auto`. `auto` = try native vision first (when a
+ * key exists), fall back to the MCP tool pattern. `tool` forces the legacy
+ * tool pattern. `native` tries native first and still falls back on failure
+ * (never break a user turn).
+ */
+export function getVisionMode(
+  env: Record<string, string | undefined> = process.env,
+  configMode?: string,
+): VisionMode {
+  const envMode = env.PANTHEON_VISION_MODE?.trim().toLowerCase()
+  if (envMode === 'native' || envMode === 'tool') return envMode
+  const cfgMode = configMode?.trim().toLowerCase()
+  if (cfgMode === 'native' || cfgMode === 'tool') return cfgMode
+  return 'auto'
+}
+
+// ─── Wildcard model matching ───────────────────────────────────────────────
+// Patterns: `*`, `prefix*`, `*suffix`, `*contains*`, `provider/*`, `*/model`.
+
+export function matchesWildcardPattern(pattern: string, value: string): boolean {
+  const p = pattern.toLowerCase()
+  const v = value.toLowerCase()
+  if (p === '*') return true
+  if (p.startsWith('*') && p.endsWith('*') && p.length > 2) return v.includes(p.slice(1, -1))
+  if (p.endsWith('*')) return v.startsWith(p.slice(0, -1))
+  if (p.startsWith('*')) return v.endsWith(p.slice(1))
+  return v === p
+}
+
+export function matchesModelPattern(pattern: string, model: ModelInfo): boolean {
+  if (pattern === '*') return true
+  const slash = pattern.indexOf('/')
+  if (slash === -1) {
+    return (
+      matchesWildcardPattern(pattern, model.providerID) ||
+      matchesWildcardPattern(pattern, model.modelID)
+    )
+  }
+  const provider = pattern.slice(0, slash)
+  const modelPattern = pattern.slice(slash + 1)
+  return (
+    matchesWildcardPattern(provider, model.providerID) &&
+    matchesWildcardPattern(modelPattern, model.modelID)
+  )
+}
+
+export function modelMatchesAnyPattern(
+  model: ModelInfo | undefined,
+  patterns: readonly string[],
+): boolean {
+  if (!model) return false
+  return patterns.some((pattern) => matchesModelPattern(pattern, model))
+}
+
+// ─── Type guards ───────────────────────────────────────────────────────────
+
+export function isImageFilePart(part: unknown): part is FilePart {
+  if (!part || typeof part !== 'object') return false
+  const value = part as { type?: unknown; mime?: unknown }
+  return (
+    value.type === 'file' &&
+    typeof value.mime === 'string' &&
+    value.mime.toLowerCase().startsWith('image/')
+  )
+}
+
+function isSupportedImageFilePart(part: unknown): part is FilePart {
+  if (!isImageFilePart(part)) return false
+  return SUPPORTED_MIME_TYPES.has(part.mime.toLowerCase())
+}
+
+function isTextPart(part: Part): part is TextPart {
+  return part.type === 'text'
+}
+
+function syntheticPartId(): string {
+  return `prt${randomUUID()}`
+}
+
+// ─── Image target resolution ───────────────────────────────────────────────
+// file:// → local path used directly; data: → decoded + saved to temp;
+// http(s):// → URL passed through unchanged.
+
+function getExtensionForMime(mime: string): string {
+  return MIME_TO_EXTENSION[mime.toLowerCase()] ?? 'png'
+}
+
+// Same set as the paste-and-ask interception and the Pantheon Vision MCP.
+const EXTENSION_TO_MIME: Record<string, string> = {
+  png: 'image/png',
+  jpg: 'image/jpeg',
+  jpeg: 'image/jpeg',
+  webp: 'image/webp',
+  gif: 'image/gif',
+}
+
+/**
+ * Guess a MIME type from a file extension (.png/.jpg/.jpeg/.webp/.gif).
+ * Returns null for unknown extensions — callers surface a friendly error.
+ */
+export function getMimeForPath(path: string): string | null {
+  const match = /\.([a-z0-9]+)$/i.exec(path)
+  if (!match) return null
+  const extension = (match[1] ?? '').toLowerCase()
+  return EXTENSION_TO_MIME[extension] ?? null
+}
+
+// ─── Temp-image LRU (global disk cap) ─────────────────────────────────────
+// Content-hash filenames mean identical re-pasted images share one file; the
+// LRU registry keeps the total on disk bounded. Insertion order = recency:
+// touching a path deletes + re-sets it so the oldest entries sit at the front.
+
+/** LRU registry: temp image path → last touch time. Exposed for tests. */
+export const tempFileLRU = new Map<string, number>()
+
+/** Mark a temp file as recently used (recency = map insertion order). */
+export function touchTempFile(path: string, now: number = Date.now()): void {
+  tempFileLRU.delete(path)
+  tempFileLRU.set(path, now)
+}
+
+/** Drop a temp file from the registry (after it has been unlinked). */
+export function forgetTempFile(path: string): void {
+  tempFileLRU.delete(path)
+}
+
+/**
+ * Enforce the global temp-file cap: when the registry exceeds `maxFiles`,
+ * unlink the least-recently-used files — but NEVER a file younger than
+ * `graceMs`, because an active auto-continue turn may still reference it.
+ * The temp directory itself is never removed (other sessions/processes may
+ * share it). Returns how many files were evicted.
+ */
+export async function enforceTempFileCap(
+  maxFiles: number = TEMP_MAX_FILES,
+  graceMs: number = TEMP_FILE_GRACE_MS,
+  now: number = Date.now(),
+): Promise<number> {
+  let evicted = 0
+  for (const [path, touchedAt] of tempFileLRU) {
+    if (tempFileLRU.size <= maxFiles) break
+    if (now - touchedAt <= graceMs) continue
+    tempFileLRU.delete(path)
+    try {
+      await unlink(path)
+    } catch {
+      // Best effort.
+    }
+    evicted += 1
+  }
+  return evicted
+}
+
+export async function saveDataUrlImage(dataUrl: string, mime: string): Promise<string | null> {
+  const match = /^data:([^;,]+);base64,(.+)$/s.exec(dataUrl)
+  if (!match) return null
+  const data = Buffer.from(match[2] ?? '', 'base64')
+  if (data.length === 0) return null
+  const dir = join(tmpdir(), TEMP_DIR_NAME)
+  await mkdir(dir, { recursive: true })
+  // Content-hash filename: identical re-pasted images reuse the same file
+  // (dedup — less disk, and it stabilizes the per-image description cache).
+  const filePath = join(dir, `${hashContent(data)}.${getExtensionForMime(mime)}`)
+  if (!existsSync(filePath)) await writeFile(filePath, data)
+  touchTempFile(filePath)
+  await enforceTempFileCap()
+  return filePath
+}
+
+async function resolveImageTarget(filePart: FilePart): Promise<SavedImage | null> {
+  const url = filePart.url
+  if (!url) return null
+  if (url.startsWith('file://')) {
+    return { path: url.slice('file://'.length), partId: filePart.id, temporary: false }
+  }
+  if (url.startsWith('data:')) {
+    const path = await saveDataUrlImage(url, filePart.mime)
+    if (!path) return null
+    return { path, partId: filePart.id, temporary: true }
+  }
+  if (url.startsWith('http://') || url.startsWith('https://')) {
+    return { path: url, partId: filePart.id, temporary: false }
+  }
+  return null
+}
+
+// ─── Native vision (provider-direct) ───────────────────────────────────────
+
+/** First non-empty (trimmed) value across the priority list, else `fallback`. */
+function firstNonEmpty(fallback: string, ...values: Array<string | undefined>): string {
+  for (const value of values) {
+    const trimmed = value?.trim()
+    if (trimmed) return trimmed
+  }
+  return fallback
+}
+
+/**
+ * Resolve the native vision endpoint: model ID + Zen base URL (no key I/O).
+ *
+ * Model precedence: env PANTHEON_VISION_MODEL > config file `visionModel` >
+ * active preset `vision.model` > DEFAULT_VISION_MODEL. Returns null when the
+ * model's provider is not opencode-go/opencode (anthropic/openai are phase 2).
+ */
+function resolveNativeVisionEndpoint(
+  preset: Pick<ResolvedPreset, 'vision'> | null | undefined,
+  env: Record<string, string | undefined>,
+  configVisionModel?: string,
+): NativeVisionEndpoint | null {
+  const resolved = preset ?? resolveActivePreset({ env, candidates: activePresetCandidates() })
+  const modelID = firstNonEmpty(
+    DEFAULT_VISION_MODEL,
+    env.PANTHEON_VISION_MODEL,
+    configVisionModel,
+    resolved?.vision?.model,
+  )
+  const providerID = modelID.slice(0, modelID.indexOf('/'))
+  let baseURL: string | null = null
+  if (providerID.startsWith('opencode-go')) baseURL = ZEN_GO_BASE_URL
+  else if (providerID === 'opencode') baseURL = ZEN_BASE_URL
+  if (!baseURL) return null
+  return { modelID, baseURL }
+}
+
+/**
+ * Resolve the native vision target: model + Zen endpoint + API key.
+ *
+ * Key comes from env only (PANTHEON_OPENCODE_API_KEY then OPENCODE_API_KEY).
+ * Returns null when the model's provider is not opencode-go/opencode or when
+ * no env key is set — callers fall back to the MCP tool. For the auth-store
+ * fallback (`opencode auth login`) use `resolveNativeVisionTarget`.
+ */
+export function resolveNativeVisionConfig(
+  preset: Pick<ResolvedPreset, 'vision'> | null | undefined = undefined,
+  env: Record<string, string | undefined> = process.env,
+  configVisionModel?: string,
+): NativeVisionTarget | null {
+  const endpoint = resolveNativeVisionEndpoint(preset, env, configVisionModel)
+  if (!endpoint) return null
+  const apiKey = env.PANTHEON_OPENCODE_API_KEY ?? env.OPENCODE_API_KEY
+  if (!apiKey || apiKey.trim() === '') return null
+  return { ...endpoint, apiKey }
+}
+
+/**
+ * Full native-vision target resolution for the chat hook (async): model +
+ * endpoint via `resolveNativeVisionConfig`, then the API key from env
+ * (PANTHEON_OPENCODE_API_KEY ?? OPENCODE_API_KEY) falling back to the opencode
+ * auth store (`opencode auth login`) — 'opencode-go' entry first, then
+ * 'opencode' — so connected users don't need to export anything. Returns null
+ * when the provider isn't opencode-go/opencode or no usable key exists —
+ * callers fall back to the MCP tool.
+ */
+export async function resolveNativeVisionTarget(
+  input: {
+    preset?: Pick<ResolvedPreset, 'vision'> | null
+    env?: Record<string, string | undefined>
+    config?: VisionConfig | null
+    authStatePath?: string
+  } = {},
+): Promise<NativeVisionTarget | null> {
+  const env = input.env ?? process.env
+  const endpoint = resolveNativeVisionEndpoint(input.preset, env, input.config?.visionModel)
+  if (!endpoint) return null
+  const envKey = env.PANTHEON_OPENCODE_API_KEY ?? env.OPENCODE_API_KEY
+  if (envKey && envKey.trim() !== '') return { ...endpoint, apiKey: envKey }
+  const authToken = await readOpencodeAuthToken(['opencode-go', 'opencode'], input.authStatePath)
+  if (!authToken) return null
+  return { ...endpoint, apiKey: authToken }
+}
+
+/**
+ * Resolve a single image source into an OpenAI-compatible image_url payload:
+ * - http(s) URL → passed through unchanged
+ * - data: URI → validated + size-capped, used as-is
+ * - local path → bytes read, converted to a data URI with the given MIME
+ * Returns null when the source is invalid or exceeds MAX_NATIVE_IMAGE_BYTES —
+ * callers fall back to the MCP tool pattern (or surface a friendly error).
+ */
+export async function resolveImageSourcePayload(
+  source: string,
+  mime: string,
+): Promise<NativeImagePayload | null> {
+  if (source.startsWith('http://') || source.startsWith('https://')) {
+    return { url: source }
+  }
+  if (source.startsWith('data:')) {
+    const match = /^data:([^;,]+);base64,([A-Za-z0-9+/=]+)$/.exec(source)
+    if (!match) return null
+    const approxBytes = Math.ceil(((match[2]?.length ?? 0) * 3) / 4)
+    if (approxBytes > MAX_NATIVE_IMAGE_BYTES) return null
+    return { url: source }
+  }
+  // Local path (from a file:// FilePart or a standalone tool path).
+  try {
+    const data = await readFile(source)
+    if (data.length === 0 || data.length > MAX_NATIVE_IMAGE_BYTES) return null
+    return { url: `data:${mime.toLowerCase()};base64,${data.toString('base64')}` }
+  } catch {
+    return null
+  }
+}
+
+/**
+ * Build OpenAI-compatible image_url payloads for the native vision call.
+ * - data: URI → used as-is (validated + size-capped).
+ * - file:// path → read bytes, converted to a data URI.
+ * - http(s) URL → passed through unchanged.
+ * Returns null when any image is invalid or exceeds MAX_NATIVE_IMAGE_BYTES —
+ * the caller falls back to the MCP tool pattern.
+ */
+async function buildNativeImagePayloads(
+  resolved: ReadonlyArray<{ filePart: FilePart; target: SavedImage }>,
+): Promise<NativeImagePayload[] | null> {
+  const payloads: NativeImagePayload[] = []
+  for (const { filePart, target } of resolved) {
+    const payload = await resolveImageSourcePayload(target.path, filePart.mime)
+    if (!payload) return null
+    payloads.push(payload)
+  }
+  return payloads
+}
+
+/**
+ * Strip the provider prefix from a model ID before sending it to the Zen
+ * gateway. The endpoint URL already encodes the provider (see
+ * `resolveNativeVisionEndpoint`), and a qualified model ID such as
+ * `opencode-go/mimo-v2.5` is rejected by the gateway with a 401
+ * ("Model ... is not supported"). Models without a `/` are returned unchanged.
+ */
+export function stripProviderPrefix(modelID: string): string {
+  const slash = modelID.lastIndexOf('/')
+  return slash === -1 ? modelID : modelID.slice(slash + 1)
+}
+
+/**
+ * Call the multimodal model directly (OpenAI-compatible /chat/completions on
+ * the opencode Zen endpoint). Returns the raw description text, or null on any
+ * failure (HTTP error, network error, timeout, empty content) — the caller
+ * falls back to the MCP tool pattern. The image never reaches the main
+ * provider: it is sent only here and replaced by the returned text.
+ *
+ * `prompt` overrides the default describe prompt (e.g. OCR extraction);
+ * `maxTokens` caps the completion (OCR needs more headroom); `signal` links
+ * an external abort (e.g. the plugin-tool ToolContext.abort) to the request.
+ */
+export async function describeImagesNative(
+  payloads: readonly NativeImagePayload[],
+  target: NativeVisionTarget,
+  prompt: string = NATIVE_PROMPT,
+  maxTokens: number = NATIVE_MAX_TOKENS,
+  signal?: AbortSignal,
+): Promise<string | null> {
+  const controller = new AbortController()
+  const timer = setTimeout(() => controller.abort(), NATIVE_TIMEOUT_MS)
+  const onExternalAbort = () => controller.abort()
+  if (signal) {
+    if (signal.aborted) controller.abort()
+    else signal.addEventListener('abort', onExternalAbort, { once: true })
+  }
+  try {
+    const response = await fetch(`${target.baseURL}/chat/completions`, {
+      method: 'POST',
+      headers: {
+        'Content-Type': 'application/json',
+        Authorization: `Bearer ${target.apiKey}`,
+      },
+      body: JSON.stringify({
+        model: stripProviderPrefix(target.modelID),
+        messages: [
+          {
+            role: 'user',
+            content: [
+              { type: 'text', text: prompt },
+              ...payloads.map((payload) => ({
+                type: 'image_url',
+                image_url: { url: payload.url },
+              })),
+            ],
+          },
+        ],
+        max_tokens: maxTokens,
+      }),
+      signal: controller.signal,
+    })
+    if (!response.ok) return null
+    const data = (await response.json()) as {
+      choices?: Array<{ message?: { content?: string } }>
+    }
+    const content = data.choices?.[0]?.message?.content
+    if (typeof content !== 'string' || content.trim() === '') return null
+    return content.trim()
+  } catch {
+    return null
+  } finally {
+    clearTimeout(timer)
+    if (signal) signal.removeEventListener('abort', onExternalAbort)
+  }
+}
+
+// ─── Session description cache ─────────────────────────────────────────────
+// The native path stores the gateway's text answer per session, keyed by the
+// user prompt + the sorted content hashes of every image payload. An identical
+// re-ask within the TTL reuses the stored text and never contacts the gateway.
+// Only textual descriptions are cached — never image bytes — and a cached
+// description is injected as text (it is never re-sent as an image payload).
+
+export type DescriptionCacheEntry = { text: string; storedAt: number }
+
+/** sha256 hex of arbitrary content (prompt, data-URI, mime+bytes). */
+export function hashContent(content: string | Buffer): string {
+  return createHash('sha256').update(content).digest('hex')
+}
+
+/**
+ * Composite cache key: hash of the user prompt + the sorted content hashes of
+ * every image payload. The prompt is part of the key so a NEW question about
+ * the same image never reuses a stale description; sorting makes the key
+ * order-insensitive (the same image set in any order maps to one key). For
+ * data-URI/local payloads the URL encodes both mime and bytes, so hashing it
+ * covers "mime+bytes"; remote URLs hash the URL itself.
+ */
+export function buildDescriptionCacheKey(
+  prompt: string,
+  payloads: readonly { url: string }[],
+): string {
+  const promptHash = hashContent(prompt)
+  const imageHashes = payloads.map((payload) => hashContent(payload.url)).sort()
+  return [promptHash, ...imageHashes].join(':')
+}
+
+/**
+ * Look up a cached description. Returns null on miss or when the entry is
+ * older than `ttlMs` — callers then hit the gateway and store the result.
+ */
+export function getCachedDescription(
+  cache: ReadonlyMap<string, DescriptionCacheEntry>,
+  key: string,
+  now: number = Date.now(),
+  ttlMs: number = DESCRIPTION_CACHE_TTL_MS,
+): string | null {
+  const entry = cache.get(key)
+  if (!entry) return null
+  if (now - entry.storedAt >= ttlMs) return null
+  return entry.text
+}
+
+/** Store a description with the current timestamp (overwrites on re-ask). */
+export function setCachedDescription(
+  cache: Map<string, DescriptionCacheEntry>,
+  key: string,
+  text: string,
+  now: number = Date.now(),
+): void {
+  cache.set(key, { text, storedAt: now })
+}
+
+// ─── Prompt generation ─────────────────────────────────────────────────────
+
+export function generateInjectionPrompt(
+  images: Array<{ path: string }>,
+  userText: string,
+  toolName: string,
+  template?: string,
+): string {
+  if (!Array.isArray(images)) {
+    throw new TypeError('Vision image resolution must return an array')
+  }
+  const imageList = images.map((image, index) => `- Image ${index + 1}: ${image.path}`).join('\n')
+  if (template && PROMPT_TEMPLATE_VARIABLES.some((variable) => template.includes(variable))) {
+    return template
+      .replaceAll('{imageList}', imageList)
+      .replaceAll('{imageCount}', String(images.length))
+      .replaceAll('{toolName}', toolName)
+      .replaceAll('{userText}', userText)
+  }
+  const plural = images.length > 1
+  const heading = plural
+    ? `The user has shared ${images.length} images. The images are available at:`
+    : 'The user has shared an image. The image is available at:'
+  const analyze = plural ? 'each image' : 'this image'
+  return [
+    heading,
+    imageList,
+    '',
+    `Use the \`${toolName}\` tool to analyze ${analyze}.`,
+    '',
+    `User's request: ${userText || '(analyze the image)'}`,
+  ].join('\n')
+}
+
+/**
+ * Injection text for the native-vision path: the description returned by the
+ * multimodal model, without any tool instruction — the main model already has
+ * the visual content in text form.
+ */
+export function generateNativeInjection(
+  description: string,
+  userText: string,
+  modelID: string,
+): string {
+  return [
+    `The user shared an image. Vision description (native, ${modelID}):`,
+    description,
+    '',
+    `User's request: ${userText || '(analyze the image)'}`,
+  ].join('\n')
+}
+
+// ─── Intent-calibrated native prompts + structured responses ───────────────
+// The native vision path calibrates its prompt to the user's actual request
+// (compare / ocr / reconstruct / bugs / describe) and asks the multimodal
+// model for a structured, parseable answer: one `<item id="N"><description>…`
+// block per image plus an optional `<context>` block for cross-image analysis.
+// The parser maps each item back to its image in the text injected into the
+// main model; any parse failure falls back to treating the whole gateway
+// answer as a single description — the turn never breaks.
+
+export type VisionIntent = 'compare' | 'ocr' | 'reconstruct' | 'bugs' | 'describe'
+
+const INTENT_PATTERNS: Record<Exclude<VisionIntent, 'describe'>, RegExp> = {
+  compare:
+    /\b(compare|comparar|comparação|diferença|diferenças|difference|differences|qual é melhor|qual e melhor|which is better|versus|vs)\b/i,
+  ocr: /\b(texto|ler|read|ocr|extrair|extraia|extract|transcrever|transcreva|transcribe)\b/i,
+  reconstruct:
+    /\b(código|codigo|code|implementar|implemente|reconstruir|reconstrua|html|css|component|ui|tela|screen)\b/i,
+  bugs: /\b(bug|erro|problema|issue|debug|não funciona|nao funciona|not working)\b/i,
+}
+
+/**
+ * Classify the user's request into a vision intent, in priority order:
+ * `compare` (requires 2+ images) → `ocr` → `reconstruct` → `bugs` → `describe`.
+ * Term matching is case-insensitive, word-boundary based, PT + EN.
+ */
+export function detectVisionIntent(userText: string, imageCount: number): VisionIntent {
+  if (imageCount >= 2 && INTENT_PATTERNS.compare.test(userText)) return 'compare'
+  if (INTENT_PATTERNS.ocr.test(userText)) return 'ocr'
+  if (INTENT_PATTERNS.reconstruct.test(userText)) return 'reconstruct'
+  if (INTENT_PATTERNS.bugs.test(userText)) return 'bugs'
+  return 'describe'
+}
+
+/** Escape `&`, `<`, `>` as XML entities (& first) — for template interpolation. */
+export function escapeXml(text: string): string {
+  return text.replaceAll('&', '&amp;').replaceAll('<', '&lt;').replaceAll('>', '&gt;')
+}
+
+/** Reverse of `escapeXml` — applied to parsed `<description>`/`<context>` bodies. */
+function unescapeXml(text: string): string {
+  return text.replaceAll('&lt;', '<').replaceAll('&gt;', '>').replaceAll('&amp;', '&')
+}
+
+const INTENT_TASK_INSTRUCTIONS: Record<VisionIntent, string> = {
+  describe:
+    'Descreva cada imagem fielmente e em detalhes: texto visível verbatim, cores, layout, objetos.',
+  compare:
+    'Descreva cada imagem no seu próprio bloco <item>; depois coloque uma comparação direta e exaustiva no <context> (diferenças por região/elemento), terminando com um veredito claro.',
+  ocr: 'Transcreva literalmente todo o texto visível em cada imagem, preservando ordem e estrutura. A transcrição vai dentro do <description> da imagem correspondente.',
+  reconstruct:
+    'Descreva cada imagem com detalhe de reconstrução para reprodução em código: layout completo (regiões, alinhamento, espaçamento), todo texto visível verbatim, paleta de cores exata em hex, fontes/tamanhos/pesos por elemento, estrutura de componentes e estados.',
+  bugs: 'Descreva o sintoma visual observado em cada imagem: o que está errado, onde, esperado vs real, com referências de região precisas e contexto suficiente para localizar e reproduzir.',
+}
+
+const INTENT_CONTEXT_INSTRUCTIONS: Record<VisionIntent, string> = {
+  compare:
+    'Depois dos itens, adicione UM bloco <context> com a análise cruzada das imagens (comparação, veredito).',
+  describe: '',
+  ocr: '',
+  reconstruct: '',
+  bugs: '',
+}
+
+/**
+ * Build the calibrated prompt for the native vision gateway call. The model is
+ * instructed to reply with N `<item id="N"><description>…</description></item>`
+ * blocks (N = imageCount) and, for `compare`, one extra `<context>` block.
+ * The user's text is XML-escaped when interpolated so it cannot break the
+ * format instructions. `intent` is auto-detected from `userText` +
+ * `imageCount` when omitted.
+ */
+export function buildNativeVisionPrompt(
+  userText: string,
+  imageCount: number,
+  intent: VisionIntent = detectVisionIntent(userText, imageCount),
+): string {
+  const count = Math.max(1, imageCount)
+  const request =
+    escapeXml(userText.trim()) || '(sem pedido explícito — descreva as imagens com precisão)'
+  const contextInstruction = INTENT_CONTEXT_INSTRUCTIONS[intent]
+  const contextLine = contextInstruction ? ` ${contextInstruction}` : ''
+  return [
+    'You are a multimodal analysis specialist embedded inside a coding assistant. The coding',
+    'assistant CANNOT perceive the images in this turn — your output is its ONLY view of them.',
+    'It will act on exactly what you write.',
+    '',
+    "READ THE USER REQUEST FIRST: infer the user's goal and calibrate the depth of each",
+    'description to it.',
+    '',
+    INTENT_TASK_INSTRUCTIONS[intent],
+    '',
+    'OUTPUT FORMAT — binding, no exceptions:',
+    `Responda com UM bloco <item id="N"> para cada uma das ${count} imagens anexadas`,
+    `(N = 1..${count}), em ordem. Cada bloco contém um <description>.${contextLine}`,
+    '',
+    '<item id="1">',
+    '<description>...apenas fatos sobre a imagem 1...</description>',
+    '</item>',
+    '<item id="2">',
+    '<description>...apenas fatos sobre a imagem 2...</description>',
+    '</item>',
+    '',
+    'RULES',
+    '- O primeiro caractere da sua resposta deve ser "<". Sem preâmbulo, sem "Aqui está",',
+    '  sem observações finais, sem wrap em blocos de código.',
+    '- ESCAPING — obrigatório: dentro dos corpos de <description> e <context>, escape TODO',
+    '  "&", "<" e ">" literais como "&amp;", "&lt;", "&gt;". Isso vale para qualquer texto',
+    '  tag-like reproduzido verbatim (HTML, XML, erros com colchetes angulares). NÃO escape',
+    '  as tags estruturais <item>, <description>, <context>.',
+    '- Nunca invente detalhes que não consegue perceber. Se um texto estiver ilegível,',
+    '  escreva "ilegível". Coloque análises cruzadas (comparação, veredito) no <context>,',
+    '  nunca dentro de um <description>.',
+    '',
+    `User's request: ${request}`,
+  ].join('\n')
+}
+
+export type StructuredVisionItem = { id: number; description: string }
+export type StructuredVisionResponse = { items: StructuredVisionItem[]; context: string | null }
+
+const ITEM_BLOCK_RE = /<item\b[^>]*>([\s\S]*?)<\/item>/gi
+const OPEN_TAG_RE = /<item\b[^>]*>/i
+const DESCRIPTION_BODY_RE = /<description>([\s\S]*?)<\/description>/i
+const CONTEXT_BODY_RE = /<context>([\s\S]*?)<\/context>/i
+const ITEM_ID_ATTR_RE = /id\s*=\s*["']?(\d+)["']?/i
+const BARE_DESCRIPTION_RE = /<description>([\s\S]*?)<\/description>/gi
+
+/**
+ * Parse the gateway's structured answer into per-image `<item>` descriptions
+ * plus an optional `<context>` block. Resistant to formatting variations
+ * (whitespace, quotes, tag case, missing id → sequential ids, markdown fences
+ * around the XML). XML entities in the bodies are unescaped for the main
+ * model. Returns null when no `<item>`/`<description>` block is found —
+ * callers then treat the whole answer as a single description (the
+ * pre-structured behavior, no extra gateway call).
+ */
+export function parseStructuredVisionResponse(text: string): StructuredVisionResponse | null {
+  const items: StructuredVisionItem[] = []
+  let match = ITEM_BLOCK_RE.exec(text)
+  while (match !== null) {
+    const desc = DESCRIPTION_BODY_RE.exec(match[1] ?? '')
+    if (desc) {
+      const openTag = OPEN_TAG_RE.exec(match[0] ?? '')?.[0] ?? ''
+      const rawId = ITEM_ID_ATTR_RE.exec(openTag)?.[1]
+      const description = unescapeXml((desc[1] ?? '').trim())
+      items.push({
+        id: rawId !== undefined ? Number(rawId) : items.length + 1,
+        description,
+      })
+    }
+    match = ITEM_BLOCK_RE.exec(text)
+  }
+  if (items.length === 0) {
+    let bareMatch = BARE_DESCRIPTION_RE.exec(text)
+    while (bareMatch !== null) {
+      items.push({
+        id: items.length + 1,
+        description: unescapeXml((bareMatch[1] ?? '').trim()),
+      })
+      bareMatch = BARE_DESCRIPTION_RE.exec(text)
+    }
+  }
+  const contextMatch = CONTEXT_BODY_RE.exec(text)
+  const context = contextMatch ? unescapeXml((contextMatch[1] ?? '').trim()) : null
+  return items.length > 0 ? { items, context } : null
+}
+
+/**
+ * Build the text injected into the main model from a parsed structured
+ * response: each `<item id>` is mapped to its image (by id, falling back to
+ * position when an id is missing), followed by the optional cross-image
+ * `<context>` block. The fallback single-description path stays in
+ * `generateNativeInjection`.
+ */
+export function buildStructuredInjection(
+  items: readonly StructuredVisionItem[],
+  context: string | null,
+  userText: string,
+  modelID: string,
+  imageCount: number,
+): string {
+  const count = Math.max(1, imageCount)
+  const byId = new Map(items.map((item) => [item.id, item.description]))
+  const header =
+    count > 1
+      ? `The user shared ${count} images. Vision descriptions (native, ${modelID}):`
+      : `The user shared an image. Vision description (native, ${modelID}):`
+  const lines = [header]
+  for (let i = 1; i <= count; i += 1) {
+    const description = byId.get(i) ?? items[i - 1]?.description
+    lines.push(`- Image ${i}: ${description ?? '(no description provided)'}`)
+  }
+  if (context) {
+    lines.push('', 'Cross-image context:', context)
+  }
+  lines.push('', `User's request: ${userText || '(analyze the image)'}`)
+  return lines.join('\n')
+}
+
+// ─── Tool name resolution ──────────────────────────────────────────────────
+// env PANTHEON_VISION_TOOL > config imageAnalysisTool > dynamic detection
+// (prefer the canonical Pantheon MCP) > canonical Pantheon MCP default.
+
+async function resolveImageAnalysisTool(
+  client: PluginInput['client'],
+  directory: string,
+  configPromise: Promise<VisionConfig>,
+): Promise<string | null> {
+  const envTool = process.env.PANTHEON_VISION_TOOL?.trim()
+  if (envTool) return envTool
+  const config = await configPromise
+  if (config.imageAnalysisTool) return config.imageAnalysisTool
+  try {
+    const response = await client.tool?.ids?.({ query: { directory } })
+    const ids = response?.data
+    if (Array.isArray(ids)) {
+      // Exact canonical ID first (fast path), then the underscore form and the
+      // legacy mcp__ hyphen form (both match the pantheon[-_]vision + action
+      // regexes), then the loose describe fallback for older ID shapes.
+      const match =
+        ids.find((id) => id === DEFAULT_IMAGE_ANALYSIS_TOOL) ??
+        ids.find((id) => isPantheonVisionServerTool(id) && isPantheonVisionActionTool(id)) ??
+        ids.find((id) => isPantheonVisionServerTool(id) && /describe/i.test(id))
+      // OpenCode 1.18.11 keeps MCP tools in the MCP service, NOT in
+      // client.tool.ids() — an empty/absent list is the NORMAL runtime state,
+      // so it must resolve to the canonical default tool instead of the
+      // 'Vision fallback unavailable' message. Only a thrown exception (the
+      // detection genuinely failed) returns null below.
+      return match ?? DEFAULT_IMAGE_ANALYSIS_TOOL
+    }
+  } catch {
+    // Detection failed (network/MCP error). Returning null keeps the explicit
+    // 'Vision fallback unavailable' safe-failure message in the turn instead
+    // of guessing at a tool we could not verify — the turn never breaks
+    // (interception is best-effort). Returning DEFAULT here too would hide
+    // real detection failures behind a tool name the model might not have.
+    return null
+  }
+  return DEFAULT_IMAGE_ANALYSIS_TOOL
+}
+
+// ─── Temp image lifecycle ──────────────────────────────────────────────────
+
+/**
+ * Age-guarded cleanup of a session's temp images.
+ *
+ * A `session.idle` can fire between auto-continue turns — not just at real
+ * session end — so a temp file referenced by the next turn (seconds old) must
+ * survive. Only paths older than `graceMs` are unlinked; younger paths stay
+ * registered and are swept by the next idle. `server.instance.disposed` and
+ * `session.deleted` pass `graceMs = 0` to unlink everything while the session
+ * is gone / process is dying. The temp directory itself is never removed: it
+ * may hold files from other sessions or processes. The session's description
+ * cache is always dropped (cache is per-session state).
+ */
+export async function cleanupSessionTempImages(
+  sessionTempFiles: Map<string, Map<string, number>>,
+  sessionVisionText: Map<string, Map<string, string>>,
+  nativeVisionSessions: Set<string>,
+  sessionID: string,
+  graceMs: number = TEMP_FILE_GRACE_MS,
+  sessionDescriptionCache?: Map<string, Map<string, DescriptionCacheEntry>>,
+): Promise<void> {
+  const files = sessionTempFiles.get(sessionID)
+  if (files) {
+    const now = Date.now()
+    for (const [path, createdAt] of files) {
+      // graceMs === 0 means "unlink everything now" (session gone / process
+      // dying); a positive grace skips files younger than the threshold so the
+      // next auto-continue turn can still read them.
+      if (graceMs > 0 && now - createdAt <= graceMs) continue
+      try {
+        await unlink(path)
+      } catch {
+        // Best effort.
+      }
+      forgetTempFile(path)
+      files.delete(path)
+    }
+    if (files.size === 0) sessionTempFiles.delete(sessionID)
+  }
+  sessionVisionText.delete(sessionID)
+  nativeVisionSessions.delete(sessionID)
+  sessionDescriptionCache?.delete(sessionID)
+}
+
+type VisionHistoryMessage = {
+  info?: { sessionID?: string; id?: string }
+  sessionID?: string
+  id?: string
+  parts?: unknown
+  content?: unknown
+  [key: string]: unknown
+}
+
+function syntheticHistoryPart(message: VisionHistoryMessage, text: string): TextPart {
+  return {
+    id: syntheticPartId(),
+    sessionID: message.info?.sessionID ?? '',
+    messageID: message.info?.id ?? '',
+    type: 'text',
+    text,
+    synthetic: true,
+  }
+}
+
+/**
+ * Remove image-bearing parts before a provider serializer can turn them into
+ * image_url content. This hook is deliberately synchronous in spirit: it only
+ * reuses text cached by chat.message and never calls a vision tool/model.
+ */
+export function sanitizeVisionHistory(
+  output: { messages?: unknown },
+  sessionVisionText: ReadonlyMap<string, ReadonlyMap<string, string>>,
+  nativeVisionSessions: ReadonlySet<string> = new Set(),
+): void {
+  if (!Array.isArray(output.messages)) return
+
+  for (const message of output.messages) {
+    if (!message || typeof message !== 'object') continue
+    const historyMessage = message as VisionHistoryMessage
+    const sessionID = historyMessage.info?.sessionID ?? historyMessage.sessionID ?? ''
+    if (nativeVisionSessions.has(sessionID)) continue
+    const descriptions = sessionVisionText.get(sessionID)
+    const sanitize = (value: unknown, inParts: boolean): unknown => {
+      if (Array.isArray(value)) return value.map((item) => sanitize(item, inParts))
+      if (!value || typeof value !== 'object') return value
+
+      const raw = value as { type?: unknown; id?: unknown; mime?: unknown }
+      const isImageUrl = raw.type === 'image_url'
+      const isImageFile =
+        raw.type === 'file' &&
+        typeof raw.mime === 'string' &&
+        raw.mime.toLowerCase().startsWith('image/')
+      if (isImageUrl || isImageFile) {
+        const partID = typeof raw.id === 'string' ? raw.id : ''
+        const text =
+          descriptions?.get(partID) ??
+          'Imagem removida do histórico antes do envio ao provedor textual.'
+        return inParts ? syntheticHistoryPart(historyMessage, text) : { type: 'text', text }
+      }
+
+      const copy: Record<string, unknown> = {}
+      for (const [key, child] of Object.entries(value)) {
+        copy[key] = sanitize(child, key === 'parts')
+      }
+      return copy
+    }
+
+    const sanitized = sanitize(historyMessage, false) as VisionHistoryMessage
+    for (const [key, value] of Object.entries(sanitized)) {
+      historyMessage[key] = value
+    }
+  }
+}
+
+// ─── Handler factory ───────────────────────────────────────────────────────
+
+export function createVisionHandler(input: PluginInput) {
+  const { client, directory } = input
+  const sessionTempFiles = new Map<string, Map<string, number>>()
+  const sessionVisionText = new Map<string, Map<string, string>>()
+  const nativeVisionSessions = new Set<string>()
+  const sessionDescriptionCache = new Map<string, Map<string, DescriptionCacheEntry>>()
+  let configPromise: Promise<VisionConfig> | null = null
+  let toolNamePromise: Promise<string | null> | null = null
+
+  const getConfig = (): Promise<VisionConfig> => {
+    configPromise ??= loadVisionConfig(directory)
+    return configPromise
+  }
+
+  const getToolName = (): Promise<string | null> => {
+    toolNamePromise ??= resolveImageAnalysisTool(client, directory, getConfig())
+    return toolNamePromise
+  }
+
+  const chatMessage: NonNullable<Hooks['chat.message']> = async (hookInput, output) => {
+    try {
+      const message: UserMessage | undefined = output?.message
+      if (!message || message.role !== 'user') return
+      const model = hookInput?.model ?? message.model
+      if (modelAcceptsImages(model)) {
+        nativeVisionSessions.add(hookInput.sessionID)
+        return
+      }
+      nativeVisionSessions.delete(hookInput.sessionID)
+      const config = await getConfig()
+      const patterns =
+        config.models && config.models.length > 0 ? config.models : DEFAULT_MODEL_PATTERNS
+      if (!modelMatchesAnyPattern(model, patterns)) {
+        // Pattern configuration may disable enrichment, but it must never
+        // disable the text-only provider safety boundary.
+        const currentTurn = { messages: [{ ...message, parts: output.parts }] }
+        sanitizeVisionHistory(currentTurn, sessionVisionText)
+        const sanitized = currentTurn.messages as VisionHistoryMessage[]
+        output.parts.splice(0, output.parts.length, ...((sanitized[0]?.parts as Part[]) ?? []))
+        return
+      }
+
+      const parts: Part[] = Array.isArray(output.parts) ? output.parts : []
+      const images = parts.filter(isImageFilePart)
+      if (images.length === 0) {
+        // Some runtimes hand the current turn to plugins already serialized as
+        // provider content. Do not let that shape bypass the final sanitizer.
+        const currentTurn = { messages: [{ ...message, parts: output.parts }] }
+        sanitizeVisionHistory(currentTurn, sessionVisionText)
+        const sanitized = currentTurn.messages as VisionHistoryMessage[]
+        output.parts.splice(0, output.parts.length, ...((sanitized[0]?.parts as Part[]) ?? []))
+        return
+      }
+
+      const userText = parts
+        .filter(isTextPart)
+        .map((part) => part.text)
+        .join('\n')
+        .trim()
+
+      const resolved: Array<{ filePart: FilePart; target: SavedImage }> = []
+      for (const image of images.filter(isSupportedImageFilePart)) {
+        const target = await resolveImageTarget(image)
+        if (target) resolved.push({ filePart: image, target })
+      }
+      const saved = resolved.map((item) => item.target)
+
+      const tempPaths = saved.filter((item) => item.temporary).map((item) => item.path)
+      if (tempPaths.length > 0) {
+        const createdAt = Date.now()
+        const existing = sessionTempFiles.get(hookInput.sessionID) ?? new Map<string, number>()
+        for (const path of tempPaths) existing.set(path, createdAt)
+        sessionTempFiles.set(hookInput.sessionID, existing)
+      }
+
+      // Native-first: call the multimodal model directly when a key exists
+      // (mode auto/native). Any failure — no key, unsupported provider,
+      // oversized/invalid image, HTTP/network error — falls back to the tool.
+      // The prompt is calibrated to the user's intent (compare/ocr/
+      // reconstruct/bugs/describe) and asks for a structured answer
+      // (<item>/<context>), which is parsed and mapped back to each image.
+      // Descriptions are cached per session: the key is the calibrated prompt
+      // + the content hashes of the images, so an identical re-ask within the
+      // TTL reuses the stored (possibly structured) text without another
+      // gateway call, while a new question about the same image misses and is
+      // answered freshly.
+      let injection: string | null = null
+      if (getVisionMode(process.env, config.mode) !== 'tool') {
+        try {
+          const target = await resolveNativeVisionTarget({ config })
+          if (target) {
+            const payloads = await buildNativeImagePayloads(resolved)
+            if (payloads) {
+              let sessionCache = sessionDescriptionCache.get(hookInput.sessionID)
+              if (!sessionCache) {
+                sessionCache = new Map()
+                sessionDescriptionCache.set(hookInput.sessionID, sessionCache)
+              }
+              const intent = detectVisionIntent(userText, payloads.length)
+              const prompt = buildNativeVisionPrompt(userText, payloads.length, intent)
+              const cacheKey = buildDescriptionCacheKey(prompt, payloads)
+              let rawText = getCachedDescription(sessionCache, cacheKey)
+              if (rawText === null) {
+                rawText = await describeImagesNative(payloads, target, prompt)
+                if (rawText !== null) setCachedDescription(sessionCache, cacheKey, rawText)
+              }
+              if (rawText !== null) {
+                const structured = parseStructuredVisionResponse(rawText)
+                injection = structured
+                  ? buildStructuredInjection(
+                      structured.items,
+                      structured.context,
+                      userText,
+                      target.modelID,
+                      payloads.length,
+                    )
+                  : generateNativeInjection(rawText, userText, target.modelID)
+              }
+            }
+          }
+        } catch {
+          // Native vision is best-effort; fall back to the tool pattern.
+        }
+      }
+
+      if (injection === null && saved.length > 0) {
+        const toolName = await getToolName()
+        injection = toolName
+          ? generateInjectionPrompt(saved, userText, toolName, config.promptTemplate)
+          : 'Vision fallback unavailable: pantheon-vision MCP is unavailable. The image was removed safely; please configure the MCP server and retry.'
+      }
+      if (injection === null) {
+        injection = 'A imagem compartilhada foi removida antes do envio ao provedor textual.'
+      }
+
+      const descriptions = sessionVisionText.get(hookInput.sessionID) ?? new Map<string, string>()
+      for (const image of images) descriptions.set(image.id, injection)
+      sessionVisionText.set(hookInput.sessionID, descriptions)
+
+      const injectionPart: TextPart = {
+        id: syntheticPartId(),
+        sessionID: message.sessionID,
+        messageID: message.id,
+        type: 'text',
+        text: injection,
+        synthetic: true,
+      }
+      const kept = parts.filter((part) => !isImageFilePart(part))
+      // Mutate in place: drop image parts, lead with the vision text
+      // (native description or tool instruction).
+      output.parts.splice(0, output.parts.length, injectionPart, ...kept)
+      sanitizeVisionHistory({ messages: [{ ...message, parts: output.parts }] }, sessionVisionText)
+    } catch {
+      // Vision interception is best-effort; never fail a user turn.
+    }
+  }
+
+  const event: NonNullable<Hooks['event']> = async ({ event: ev }) => {
+    if (ev.type === 'session.idle') {
+      // An idle can fire between auto-continue turns: temp files younger than
+      // the grace period survive, but the per-session description cache is
+      // dropped (it is pure optimization, safe to reset on every idle).
+      const sessionID = ev.properties?.sessionID
+      if (sessionID)
+        await cleanupSessionTempImages(
+          sessionTempFiles,
+          sessionVisionText,
+          nativeVisionSessions,
+          sessionID,
+          TEMP_FILE_GRACE_MS,
+          sessionDescriptionCache,
+        )
+    } else if (ev.type === 'session.deleted') {
+      // The session is gone for good: no active turn can reference its temp
+      // images, so unlink them all immediately and drop its cached state.
+      const sessionID = ev.properties.info.id
+      await cleanupSessionTempImages(
+        sessionTempFiles,
+        sessionVisionText,
+        nativeVisionSessions,
+        sessionID,
+        0,
+        sessionDescriptionCache,
+      )
+    } else if (ev.type === 'server.instance.disposed') {
+      const sessions = [...sessionTempFiles.keys()]
+      for (const sessionID of sessions) {
+        await cleanupSessionTempImages(
+          sessionTempFiles,
+          sessionVisionText,
+          nativeVisionSessions,
+          sessionID,
+          0, // The process is dying: unlink everything immediately.
+          sessionDescriptionCache,
+        )
+      }
+      sessionVisionText.clear()
+      sessionDescriptionCache.clear()
+    }
+  }
+
+  const messagesTransform: NonNullable<Hooks['experimental.chat.messages.transform']> = async (
+    _hookInput,
+    output,
+  ) => {
+    try {
+      sanitizeVisionHistory(output, sessionVisionText, nativeVisionSessions)
+    } catch {
+      // History sanitization is best effort; never break runtimes that expose
+      // the experimental hook with a partial output shape.
+    }
+  }
+
+  return { chatMessage, messagesTransform, event }
+}