Merge pull request #2371 from code-yeongyu/fix/issue-2323

fix: respect multimodal provider vision capabilities
2026-03-11 20:22:35 +09:00
parent 29ff2e1ff1 bb899dfa97
commit 2f4687d9b6
8 changed files with 430 additions and 13 deletions
@@ -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,
+    })
+  })
+})
@@ -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 agentsRaw = isObject(agentsResult) ? agentsResult["data"] : undefined
-    const agents = Array.isArray(agentsRaw) ? agentsRaw.map(toAgentInfo).filter(Boolean) : []
+    const registeredMetadata = await resolveRegisteredAgentMetadata(ctx)
+    const visionCapableModels = readVisionCapableModelsCache()

-    const matched = agents.find(
-      (agent) => agent?.name?.toLowerCase() === MULTIMODAL_LOOKER_AGENT.toLowerCase()
-    )
-
-    return {
-      agentModel: matched?.model,
-      agentVariant: matched?.variant,
+    if (registeredMetadata.agentModel && visionCapableModels.length === 0) {
+      return registeredMetadata
    }
+
+    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 {}
@@ -0,0 +1,31 @@
+describe("buildMultimodalLookerFallbackChain", () => {
+  it("builds fallback chain from vision-capable models", async () => {
+    // given
+    const { buildMultimodalLookerFallbackChain } = await import("./multimodal-fallback-chain")
+    const visionCapableModels = [
+      { providerID: "openai", modelID: "gpt-5.4" },
+      { providerID: "opencode", modelID: "gpt-5.4" },
+    ]
+
+    // when
+    const result = buildMultimodalLookerFallbackChain(visionCapableModels)
+
+    // then
+    const gpt54Entries = result.filter((entry) => entry.model === "gpt-5.4")
+    expect(gpt54Entries.length).toBeGreaterThan(0)
+  })
+
+  it("avoids duplicates when adding hardcoded entries", async () => {
+    // given
+    const { buildMultimodalLookerFallbackChain } = await import("./multimodal-fallback-chain")
+    const visionCapableModels = [{ providerID: "openai", modelID: "gpt-5.4" }]
+
+    // when
+    const result = buildMultimodalLookerFallbackChain(visionCapableModels)
+
+    // then
+    expect(result.length).toBeGreaterThan(0)
+    expect(result[0].model).toBe("gpt-5.4")
+    expect(result[0].providers).toContain("openai")
+  })
+})
@@ -0,0 +1,49 @@
+import type { FallbackEntry } from "../../shared/model-requirements"
+import { AGENT_MODEL_REQUIREMENTS } from "../../shared/model-requirements"
+import type { VisionCapableModel } from "../../plugin-state"
+
+const MULTIMODAL_LOOKER_REQUIREMENT = AGENT_MODEL_REQUIREMENTS["multimodal-looker"]
+
+function getFullModelKey(providerID: string, modelID: string): string {
+  return `${providerID}/${modelID}`
+}
+
+export function isHardcodedMultimodalFallbackModel(model: VisionCapableModel): boolean {
+  return MULTIMODAL_LOOKER_REQUIREMENT.fallbackChain.some((entry) =>
+    entry.providers.some((providerID) =>
+      getFullModelKey(providerID, entry.model) === getFullModelKey(model.providerID, model.modelID),
+    ),
+  )
+}
+
+export function buildMultimodalLookerFallbackChain(
+  visionCapableModels: VisionCapableModel[],
+): FallbackEntry[] {
+  const seen = new Set<string>()
+  const fallbackChain: FallbackEntry[] = []
+
+  for (const visionCapableModel of visionCapableModels) {
+    const key = getFullModelKey(visionCapableModel.providerID, visionCapableModel.modelID)
+    if (seen.has(key)) continue
+
+    seen.add(key)
+    fallbackChain.push({
+      providers: [visionCapableModel.providerID],
+      model: visionCapableModel.modelID,
+    })
+  }
+
+  for (const entry of MULTIMODAL_LOOKER_REQUIREMENT.fallbackChain) {
+    const providerModelKeys = entry.providers.map((providerID) =>
+      getFullModelKey(providerID, entry.model),
+    )
+    if (providerModelKeys.every((key) => seen.has(key))) {
+      continue
+    }
+
+    providerModelKeys.forEach((key) => seen.add(key))
+    fallbackChain.push(entry)
+  }
+
+  return fallbackChain
+}