feat(opencode): add task model override

Add an optional `model` parameter to the task tool (provider/model format)
that overrides the subagent's model, gated behind a new `model_override`
permission that defaults to deny. Selection precedence is the `model`
parameter, then the subagent's configured model, then the parent assistant
message model.

Ported from anomalyco#29447 (adapted for the Effect layer).

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

Author: randomvariable

packages/core/src/v1/config/permission.ts

@@ -28,6 +28,7 @@ const InputObject = Schema.StructWithRest(
     question: Schema.optional(Action),
     webfetch: Schema.optional(Action),
     websearch: Schema.optional(Action),
+    model_override: Schema.optional(Rule),
     lsp: Schema.optional(Rule),
     doom_loop: Schema.optional(Action),
     skill: Schema.optional(Rule),

packages/opencode/src/agent/agent.ts

@@ -119,6 +119,7 @@ const layer = Layer.effect(
         const defaults = Permission.fromConfig({
           "*": "allow",
           doom_loop: "ask",
+          model_override: "deny",
           external_directory: {
             "*": "ask",
             ...Object.fromEntries(whitelistedDirs.map((dir) => [dir, "allow"])),

packages/opencode/src/tool/task.ts

@@ -10,6 +10,8 @@ import { Agent } from "../agent/agent"
 import { deriveSubagentSessionPermission } from "../agent/subagent-permissions"
 import type { SessionPrompt } from "../session/prompt"
 import { Config } from "@/config/config"
+import { ModelV2 } from "@opencode-ai/core/model"
+import { ProviderV2 } from "@opencode-ai/core/provider"
 import { Effect, Exit, Schema, Scope } from "effect"
 import { EffectBridge } from "@/effect/bridge"
 import { RuntimeFlags } from "@/effect/runtime-flags"
@@ -44,6 +46,10 @@ const BaseParameterFields = {
   description: Schema.String.annotate({ description: "A short (3-5 words) description of the task" }),
   prompt: Schema.String.annotate({ description: "The task for the agent to perform" }),
   subagent_type: Schema.String.annotate({ description: "The type of specialized agent to use for this task" }),
+  model: Schema.optional(Schema.String).annotate({
+    description:
+      "Override the model for this subagent. Format: provider/model (e.g. anthropic/claude-sonnet-4, openai/gpt-4o). Takes precedence over the agent's configured model.",
+  }),
   task_id: Schema.optional(Schema.String).annotate({
     description:
       "This should only be set if you mean to resume a previous task (you can pass a prior task_id and the task will continue the same subagent session as before instead of creating a fresh one)",
@@ -78,6 +84,19 @@ function renderOutput(input: {
   ].join("\n")
 }
 
+function parseModelOverride(model: string): Effect.Effect<{ modelID: ModelV2.ID; providerID: ProviderV2.ID }, Error> {
+  const slash = model.indexOf("/")
+  if (slash <= 0 || slash === model.length - 1) {
+    return Effect.fail(
+      new Error(`Invalid model format: "${model}". Expected provider/model (e.g. anthropic/claude-sonnet-4)`),
+    )
+  }
+  return Effect.succeed({
+    providerID: ProviderV2.ID.make(model.slice(0, slash)),
+    modelID: ModelV2.ID.make(model.slice(slash + 1)),
+  })
+}
+
 export const TaskTool = Tool.define(
   id,
   Effect.gen(function* () {
@@ -116,6 +135,23 @@ export const TaskTool = Tool.define(
         )
       }
 
+      const modelOverride = params.model
+      const overrideModel = modelOverride === undefined ? undefined : yield* parseModelOverride(modelOverride)
+
+      if (overrideModel && modelOverride !== undefined) {
+        yield* ctx.ask({
+          permission: "model_override",
+          patterns: [modelOverride],
+          always: [modelOverride],
+          metadata: {
+            description: params.description,
+            subagent_type: params.subagent_type,
+            model: modelOverride,
+          },
+        })
+      }
+      }
+
       if (!ctx.extra?.bypassAgentCheck) {
         yield* ctx.ask({
           permission: id,
@@ -178,7 +214,7 @@ export const TaskTool = Tool.define(
       if (msg.info.role !== "assistant") return yield* Effect.fail(new Error("Not an assistant message"))
       const variant = msg.info.variant
 
-      const model = next.model ?? {
+      const model = overrideModel ?? next.model ?? {
         modelID: msg.info.modelID,
         providerID: msg.info.providerID,
       }

packages/opencode/src/tool/task.txt

@@ -8,6 +8,12 @@ When NOT to use the Task tool:
 - If you are searching for code within a specific file or set of 2-3 files, use the Read tool instead of the Task tool, to find the match more quickly
 - If no available agent is a good fit for the task, use other tools directly
 
+Model selection:
+- Each agent has a default model (usually inherited from the parent session).
+- You can override the model by passing the `model` parameter in `provider/model` format (e.g. `anthropic/claude-sonnet-4`, `openai/gpt-4o`, `google/gemini-2.5-pro`).
+- Model overrides require the `model_override` permission. By default this permission is denied. The user can allow specific models or providers in their config (e.g. `"model_override": { "anthropic/*": "allow" }`).
+- Model selection precedence is `model` parameter, then the subagent's configured model, then the parent assistant message model.
+
 
 Usage notes:
 1. Launch multiple agents concurrently whenever possible, to maximize performance; to do that, use a single message with multiple tool uses

packages/opencode/test/tool/__snapshots__/parameters.test.ts.snap

@@ -313,6 +313,10 @@ exports[`tool parameters JSON Schema (wire shape) task 1`] = `
       "description": "A short (3-5 words) description of the task",
       "type": "string",
     },
+    "model": {
+      "description": "Override the model for this subagent. Format: provider/model (e.g. anthropic/claude-sonnet-4, openai/gpt-4o). Takes precedence over the agent's configured model.",
+      "type": "string",
+    },
     "prompt": {
       "description": "The task for the agent to perform",
       "type": "string",

packages/opencode/test/tool/parameters.test.ts

@@ -243,6 +243,10 @@ describe("tool parameters", () => {
       const parsed = parse(Task, { description: "d", prompt: "p", subagent_type: "general", background: true })
       expect(parsed.background).toBe(true)
     })
+    test("accepts optional model override", () => {
+      const parsed = parse(Task, { description: "d", prompt: "p", subagent_type: "general", model: "openai/gpt-4o" })
+      expect(parsed.model).toBe("openai/gpt-4o")
+    })
     test("rejects missing prompt", () => {
       expect(accepts(Task, { description: "d", subagent_type: "general" })).toBe(false)
     })

packages/opencode/test/tool/task.test.ts

@@ -3,7 +3,7 @@ import { SessionV1 } from "@opencode-ai/core/v1/session"
 import { Database } from "@opencode-ai/core/database/database"
 import { LayerNode } from "@opencode-ai/core/effect/layer-node"
 import { SessionProjector } from "@opencode-ai/core/session/projector"
-import { Deferred, Effect, Exit, Fiber, Layer } from "effect"
+import { Cause, Deferred, Effect, Exit, Fiber, Layer } from "effect"
 import { Agent } from "../../src/agent/agent"
 import { BackgroundJob } from "@/background/job"
 import { EventV2Bridge } from "@/event-v2-bridge"
@@ -536,6 +536,243 @@ describe("tool.task", () => {
     },
   )
 
+  it.instance(
+    "execute uses explicit model override before subagent and parent models",
+    () =>
+      Effect.gen(function* () {
+        const { chat, assistant } = yield* seed()
+        const tool = yield* TaskTool
+        const def = yield* tool.init()
+        const calls: unknown[] = []
+        let seen: SessionPrompt.PromptInput | undefined
+        const promptOps = stubOps({ onPrompt: (input) => (seen = input) })
+
+        const result = yield* def.execute(
+          {
+            description: "inspect bug",
+            prompt: "look into the cache key path",
+            subagent_type: "general",
+            model: "anthropic/claude-sonnet-4",
+          },
+          {
+            sessionID: chat.id,
+            messageID: assistant.id,
+            agent: "build",
+            abort: new AbortController().signal,
+            extra: { promptOps },
+            messages: [],
+            metadata: () => Effect.void,
+            ask: (input) =>
+              Effect.sync(() => {
+                calls.push(input)
+              }),
+          },
+        )
+
+        expect(result.metadata.model.providerID as string).toBe("anthropic")
+        expect(result.metadata.model.modelID as string).toBe("claude-sonnet-4")
+        expect((seen?.model?.providerID ?? "") as string).toBe("anthropic")
+        expect((seen?.model?.modelID ?? "") as string).toBe("claude-sonnet-4")
+        expect(calls[0]).toEqual({
+          permission: "model_override",
+          patterns: ["anthropic/claude-sonnet-4"],
+          always: ["anthropic/claude-sonnet-4"],
+          metadata: {
+            description: "inspect bug",
+            subagent_type: "general",
+            model: "anthropic/claude-sonnet-4",
+          },
+        })
+        expect(calls[1]).toEqual({
+          permission: "task",
+          patterns: ["general"],
+          always: ["*"],
+          metadata: {
+            description: "inspect bug",
+            subagent_type: "general",
+          },
+        })
+      }),
+    {
+      config: {
+        agent: {
+          general: {
+            model: "openai/gpt-4o-mini",
+          },
+        },
+      },
+    },
+  )
+
+  it.instance("stops before task permission when model override permission fails", () =>
+    Effect.gen(function* () {
+      const { chat, assistant } = yield* seed()
+      const tool = yield* TaskTool
+      const def = yield* tool.init()
+      const calls: unknown[] = []
+
+      const exit = yield* def
+        .execute(
+          {
+            description: "inspect bug",
+            prompt: "look into the cache key path",
+            subagent_type: "general",
+            model: "anthropic/claude-sonnet-4",
+          },
+          {
+            sessionID: chat.id,
+            messageID: assistant.id,
+            agent: "build",
+            abort: new AbortController().signal,
+            extra: { promptOps: stubOps() },
+            messages: [],
+            metadata: () => Effect.void,
+            ask: (input) =>
+              Effect.sync(() => {
+                calls.push(input)
+              }).pipe(
+                Effect.andThen(
+                  input.permission === "model_override" ? Effect.die(new Error("model override denied")) : Effect.void,
+                ),
+              ),
+          },
+        )
+        .pipe(Effect.exit)
+
+      expect(Exit.isFailure(exit)).toBe(true)
+      expect(calls).toEqual([
+        {
+          permission: "model_override",
+          patterns: ["anthropic/claude-sonnet-4"],
+          always: ["anthropic/claude-sonnet-4"],
+          metadata: {
+            description: "inspect bug",
+            subagent_type: "general",
+            model: "anthropic/claude-sonnet-4",
+          },
+        },
+      ])
+    }),
+  )
+
+  it.instance(
+    "execute uses subagent model when no explicit override is provided",
+    () =>
+      Effect.gen(function* () {
+        const { chat, assistant } = yield* seed()
+        const tool = yield* TaskTool
+        const def = yield* tool.init()
+        let seen: SessionPrompt.PromptInput | undefined
+        const promptOps = stubOps({ onPrompt: (input) => (seen = input) })
+
+        const result = yield* def.execute(
+          {
+            description: "inspect bug",
+            prompt: "look into the cache key path",
+            subagent_type: "general",
+          },
+          {
+            sessionID: chat.id,
+            messageID: assistant.id,
+            agent: "build",
+            abort: new AbortController().signal,
+            extra: { promptOps },
+            messages: [],
+            metadata: () => Effect.void,
+            ask: () => Effect.void,
+          },
+        )
+
+        expect(result.metadata.model.providerID as string).toBe("openai")
+        expect(result.metadata.model.modelID as string).toBe("gpt-4o-mini")
+        expect((seen?.model?.providerID ?? "") as string).toBe("openai")
+        expect((seen?.model?.modelID ?? "") as string).toBe("gpt-4o-mini")
+      }),
+    {
+      config: {
+        agent: {
+          general: {
+            model: "openai/gpt-4o-mini",
+          },
+        },
+      },
+    },
+  )
+
+  it.instance("execute uses parent assistant model when no explicit or subagent model is provided", () =>
+    Effect.gen(function* () {
+      const { chat, assistant } = yield* seed()
+      const tool = yield* TaskTool
+      const def = yield* tool.init()
+      let seen: SessionPrompt.PromptInput | undefined
+      const promptOps = stubOps({ onPrompt: (input) => (seen = input) })
+
+      const result = yield* def.execute(
+        {
+          description: "inspect bug",
+          prompt: "look into the cache key path",
+          subagent_type: "general",
+        },
+        {
+          sessionID: chat.id,
+          messageID: assistant.id,
+          agent: "build",
+          abort: new AbortController().signal,
+          extra: { promptOps },
+          messages: [],
+          metadata: () => Effect.void,
+          ask: () => Effect.void,
+        },
+      )
+
+      expect(result.metadata.model.providerID).toBe(ref.providerID)
+      expect(result.metadata.model.modelID).toBe(ref.modelID)
+      expect(seen?.model?.providerID).toBe(ref.providerID)
+      expect(seen?.model?.modelID).toBe(ref.modelID)
+    }),
+  )
+
+  it.instance("rejects invalid model override strings before asking permissions", () =>
+    Effect.gen(function* () {
+      const { chat, assistant } = yield* seed()
+      const tool = yield* TaskTool
+      const def = yield* tool.init()
+
+      yield* Effect.forEach(["gpt-4o", "openai/"], (model) =>
+        Effect.gen(function* () {
+          const calls: unknown[] = []
+          const exit = yield* def
+            .execute(
+              {
+                description: "inspect bug",
+                prompt: "look into the cache key path",
+                subagent_type: "general",
+                model,
+              },
+              {
+                sessionID: chat.id,
+                messageID: assistant.id,
+                agent: "build",
+                abort: new AbortController().signal,
+                extra: { promptOps: stubOps() },
+                messages: [],
+                metadata: () => Effect.void,
+                ask: (input) =>
+                  Effect.sync(() => {
+                    calls.push(input)
+                  }),
+              },
+            )
+            .pipe(Effect.exit)
+
+          expect(Exit.isFailure(exit)).toBe(true)
+          if (Exit.isFailure(exit)) expect(Cause.pretty(exit.cause)).toContain(`Invalid model format: "${model}"`)
+          expect(calls).toHaveLength(0)
+        }),
+      )
+    }),
+  )
+
   it.instance("rejects background execution when the experiment is disabled", () =>
     Effect.gen(function* () {
       const { chat, assistant } = yield* seed()