fix(look-at): resolve multimodal models from vision-capable providers

Ultraworked with [Sisyphus](https://github.com/code-yeongyu/oh-my-opencode)

Co-authored-by: Sisyphus <clio-agent@sisyphuslabs.ai>

src/tools/look-at/multimodal-agent-metadata.test.ts

@@ -0,0 +1,115 @@
+/// <reference types="bun-types" />
+
+import { afterEach, beforeEach, describe, expect, mock, spyOn, test } from "bun:test"
+import type { PluginInput } from "@opencode-ai/plugin"
+import { resolveMultimodalLookerAgentMetadata } from "./multimodal-agent-metadata"
+import { setVisionCapableModelsCache, clearVisionCapableModelsCache } from "../../shared/vision-capable-models-cache"
+import * as connectedProvidersCache from "../../shared/connected-providers-cache"
+import * as modelAvailability from "../../shared/model-availability"
+
+function createPluginInput(agentData: Array<Record<string, unknown>>): PluginInput {
+  const client = {} as PluginInput["client"]
+  Object.assign(client, {
+    app: {
+      agents: mock(async () => ({ data: agentData })),
+    },
+  })
+
+  return {
+    client,
+    project: {} as PluginInput["project"],
+    directory: "/project",
+    worktree: "/project",
+    serverUrl: new URL("http://localhost"),
+    $: {} as PluginInput["$"],
+  }
+}
+
+describe("resolveMultimodalLookerAgentMetadata", () => {
+  beforeEach(() => {
+    clearVisionCapableModelsCache()
+  })
+
+  afterEach(() => {
+    clearVisionCapableModelsCache()
+    ;(modelAvailability.fetchAvailableModels as unknown as { mockRestore?: () => void }).mockRestore?.()
+    ;(connectedProvidersCache.readConnectedProvidersCache as unknown as { mockRestore?: () => void }).mockRestore?.()
+  })
+
+  test("returns configured multimodal-looker model when it already matches a vision-capable override", async () => {
+    // given
+    setVisionCapableModelsCache(new Map([
+      [
+        "rundao/public/qwen3.5-397b",
+        { providerID: "rundao", modelID: "public/qwen3.5-397b" },
+      ],
+    ]))
+    spyOn(modelAvailability, "fetchAvailableModels").mockResolvedValue(
+      new Set(["rundao/public/qwen3.5-397b"]),
+    )
+    spyOn(connectedProvidersCache, "readConnectedProvidersCache").mockReturnValue(["rundao"])
+    const ctx = createPluginInput([
+      {
+        name: "multimodal-looker",
+        model: { providerID: "rundao", modelID: "public/qwen3.5-397b" },
+      },
+    ])
+
+    // when
+    const result = await resolveMultimodalLookerAgentMetadata(ctx)
+
+    // then
+    expect(result).toEqual({
+      agentModel: { providerID: "rundao", modelID: "public/qwen3.5-397b" },
+      agentVariant: undefined,
+    })
+  })
+
+  test("prefers connected vision-capable provider models before the hardcoded fallback chain", async () => {
+    // given
+    setVisionCapableModelsCache(new Map([
+      [
+        "rundao/public/qwen3.5-397b",
+        { providerID: "rundao", modelID: "public/qwen3.5-397b" },
+      ],
+    ]))
+    spyOn(modelAvailability, "fetchAvailableModels").mockResolvedValue(
+      new Set(["openai/gpt-5.4", "rundao/public/qwen3.5-397b"]),
+    )
+    spyOn(connectedProvidersCache, "readConnectedProvidersCache").mockReturnValue(["openai", "rundao"])
+    const ctx = createPluginInput([
+      {
+        name: "multimodal-looker",
+        model: { providerID: "openai", modelID: "gpt-5.4" },
+        variant: "medium",
+      },
+    ])
+
+    // when
+    const result = await resolveMultimodalLookerAgentMetadata(ctx)
+
+    // then
+    expect(result).toEqual({
+      agentModel: { providerID: "rundao", modelID: "public/qwen3.5-397b" },
+      agentVariant: undefined,
+    })
+  })
+
+  test("falls back to the hardcoded multimodal chain when no dynamic vision model exists", async () => {
+    // given
+    spyOn(modelAvailability, "fetchAvailableModels").mockResolvedValue(
+      new Set(["google/gemini-3-flash"]),
+    )
+    spyOn(connectedProvidersCache, "readConnectedProvidersCache").mockReturnValue(["google"])
+    const ctx = createPluginInput([])
+
+    // when
+    const result = await resolveMultimodalLookerAgentMetadata(ctx)
+
+    // then
+    expect(result).toEqual({
+      agentModel: { providerID: "google", modelID: "gemini-3-flash" },
+      agentVariant: undefined,
+    })
+  })
+})

src/tools/look-at/multimodal-agent-metadata.ts

@@ -1,6 +1,11 @@
 import type { PluginInput } from "@opencode-ai/plugin"
 import { MULTIMODAL_LOOKER_AGENT } from "./constants"
-import { log } from "../../shared"
+import { fetchAvailableModels } from "../../shared/model-availability"
+import { log } from "../../shared/logger"
+import { readConnectedProvidersCache } from "../../shared/connected-providers-cache"
+import { resolveModelPipeline } from "../../shared/model-resolution-pipeline"
+import { readVisionCapableModelsCache } from "../../shared/vision-capable-models-cache"
+import { buildMultimodalLookerFallbackChain } from "./multimodal-fallback-chain"
 
 type AgentModel = { providerID: string; modelID: string }
 
@@ -19,6 +24,20 @@ function isObject(value: unknown): value is Record<string, unknown> {
   return typeof value === "object" && value !== null
 }
 
+function getFullModelKey(model: AgentModel): string {
+  return `${model.providerID}/${model.modelID}`
+}
+
+function parseAgentModel(model: string): AgentModel | undefined {
+  const [providerID, ...modelIDParts] = model.split("/")
+  const modelID = modelIDParts.join("/")
+  if (!providerID || modelID.length === 0) {
+    return undefined
+  }
+
+  return { providerID, modelID }
+}
+
 function toAgentInfo(value: unknown): AgentInfo | null {
   if (!isObject(value)) return null
   const name = typeof value["name"] === "string" ? value["name"] : undefined
@@ -33,22 +52,83 @@ function toAgentInfo(value: unknown): AgentInfo | null {
   return { name, model, variant }
 }
 
+async function resolveRegisteredAgentMetadata(
+  ctx: PluginInput,
+): Promise<ResolvedAgentMetadata> {
+  const agentsResult = await ctx.client.app?.agents?.()
+  const agentsRaw = isObject(agentsResult) ? agentsResult["data"] : undefined
+  const agents = Array.isArray(agentsRaw) ? agentsRaw.map(toAgentInfo).filter(Boolean) : []
+
+  const matched = agents.find(
+    (agent) => agent?.name?.toLowerCase() === MULTIMODAL_LOOKER_AGENT.toLowerCase()
+  )
+
+  return {
+    agentModel: matched?.model,
+    agentVariant: matched?.variant,
+  }
+}
+
+async function resolveDynamicAgentMetadata(
+  ctx: PluginInput,
+  visionCapableModels = readVisionCapableModelsCache(),
+): Promise<ResolvedAgentMetadata> {
+  const fallbackChain = buildMultimodalLookerFallbackChain(visionCapableModels)
+  const connectedProviders = readConnectedProvidersCache()
+  const availableModels = await fetchAvailableModels(ctx.client, {
+    connectedProviders,
+  })
+
+  const resolution = resolveModelPipeline({
+    constraints: {
+      availableModels,
+      connectedProviders,
+    },
+    policy: {
+      fallbackChain,
+    },
+  })
+
+  const agentModel = resolution ? parseAgentModel(resolution.model) : undefined
+  return {
+    agentModel,
+    agentVariant: resolution?.variant,
+  }
+}
+
+function isConfiguredVisionModel(
+  configuredModel: AgentModel | undefined,
+  dynamicModel: AgentModel | undefined,
+): boolean {
+  if (!configuredModel || !dynamicModel) {
+    return false
+  }
+
+  return getFullModelKey(configuredModel) === getFullModelKey(dynamicModel)
+}
+
 export async function resolveMultimodalLookerAgentMetadata(
   ctx: PluginInput
 ): Promise<ResolvedAgentMetadata> {
   try {
-    const agentsResult = await ctx.client.app?.agents?.()
+    const registeredMetadata = await resolveRegisteredAgentMetadata(ctx)
-    const agentsRaw = isObject(agentsResult) ? agentsResult["data"] : undefined
+    const visionCapableModels = readVisionCapableModelsCache()
-    const agents = Array.isArray(agentsRaw) ? agentsRaw.map(toAgentInfo).filter(Boolean) : []
 
-    const matched = agents.find(
+    if (registeredMetadata.agentModel && visionCapableModels.length === 0) {
-      (agent) => agent?.name?.toLowerCase() === MULTIMODAL_LOOKER_AGENT.toLowerCase()
+      return registeredMetadata
-    )
-
-    return {
-      agentModel: matched?.model,
-      agentVariant: matched?.variant,
     }
+
+    const dynamicMetadata = await resolveDynamicAgentMetadata(ctx, visionCapableModels)
+
+    if (isConfiguredVisionModel(registeredMetadata.agentModel, dynamicMetadata.agentModel)) {
+      return registeredMetadata
+    }
+
+    if (dynamicMetadata.agentModel) {
+      return dynamicMetadata
+    }
+
+    return registeredMetadata
   } catch (error) {
     log("[look_at] Failed to resolve multimodal-looker model info", error)
     return {}