feat(browser): restore full eval recipe

Author: gregpr07

packages/opencode/src/provider/transform.ts

@@ -1248,7 +1248,7 @@ export function options(input: {
       !input.model.api.id.includes("-chat") &&
       input.model.providerID !== "azure"
     ) {
-      result["textVerbosity"] = "low"
+      result["textVerbosity"] = input.model.api.id.includes("gpt-5.5") ? "medium" : "low"
     }
 
     if (input.model.providerID.startsWith("opencode")) {

packages/opencode/src/session/prompt/gpt.txt

@@ -13,5 +13,5 @@ Do not stop after a plan, progress update, partial result, or first blocker. Bef
 the task field by field and resolve omissions or evidence/result mismatches when possible. Save requested deliverables
 in the user's required location.
 
-Your final response must contain every requested value or an explicit, evidence-backed limitation. Do not ask the user
-questions when a deterministic attempt is possible.
+Your final response must start with `FINAL ANSWER:` and contain every requested value or an explicit, evidence-backed
+limitation. Do not ask the user questions when a deterministic attempt is possible.

packages/opencode/src/tool/browser-execute.ts

@@ -7,10 +7,13 @@ import { Effect, Schema } from "effect"
 import { BrowserExecute } from "@browser-use/bcode-browser/browser-execute"
 import { Global } from "@opencode-ai/core/global"
 import { InstanceState } from "@/effect/instance-state"
+import { Truncate } from "./truncate"
 import * as Tool from "./tool"
 import DESCRIPTION from "./browser-execute.txt"
 
 const MAX_METADATA_LENGTH = 30_000
+const MAX_CONTEXT_BYTES = 8 * 1024
+const MAX_CONTEXT_LINES = 400
 const preview = (text: string) =>
   text.length <= MAX_METADATA_LENGTH ? text : "...\n\n" + text.slice(-MAX_METADATA_LENGTH)
 
@@ -24,6 +27,7 @@ export const BrowserExecuteTool = Tool.define(
   "browser_execute",
   Effect.gen(function* () {
     const impl = yield* BrowserExecute.make(Global.Path.data)
+    const truncate = yield* Truncate.Service
     return {
       // Substitute the resolved skills path so `{{SKILLS_DIR}}` references in
       // the description point at a concrete location. Workspace is
@@ -66,18 +70,28 @@ export const BrowserExecuteTool = Tool.define(
             mime: s.mime,
             url: `data:${s.mime};base64,${s.base64}`,
           }))
+          const fullOutput = [
+            result.output.trimEnd(),
+            result.result === "null" ? "" : `=> ${result.result}`,
+            attachments.length > 0
+              ? `(${attachments.length} screenshot${attachments.length === 1 ? "" : "s"} attached)`
+              : "",
+          ]
+            .filter(Boolean)
+            .join("\n\n")
+          const contextualOutput = yield* truncate.output(fullOutput, {
+            maxBytes: MAX_CONTEXT_BYTES,
+            maxLines: MAX_CONTEXT_LINES,
+          })
           return {
             title: "browser_execute",
-            output: [
-              result.output.trimEnd(),
-              result.result === "null" ? "" : `=> ${result.result}`,
-              attachments.length > 0
-                ? `(${attachments.length} screenshot${attachments.length === 1 ? "" : "s"} attached)`
-                : "",
-            ]
-              .filter(Boolean)
-              .join("\n\n"),
-            metadata: { result: result.result, output: preview(result.output) },
+            output: contextualOutput.content,
+            metadata: {
+              result: result.result,
+              output: preview(result.output),
+              truncated: contextualOutput.truncated,
+              ...(contextualOutput.truncated && { outputPath: contextualOutput.outputPath }),
+            },
             attachments,
           }
         }).pipe(Effect.orDie),

packages/opencode/test/provider/transform.test.ts

@@ -399,6 +399,12 @@ describe("ProviderTransform.options - gpt-5 textVerbosity", () => {
     expect(result.include).toEqual(["reasoning.encrypted_content"])
   })
 
+  test("gpt-5.5 should have textVerbosity set to medium", () => {
+    const model = createGpt5Model("gpt-5.5")
+    const result = ProviderTransform.options({ model, sessionID, providerOptions: {} })
+    expect(result.textVerbosity).toBe("medium")
+  })
+
   test("Bedrock Mantle gpt-5.5 uses OpenAI Responses defaults", () => {
     const model = {
       ...createGpt5Model("openai.gpt-5.5"),
@@ -415,7 +421,7 @@ describe("ProviderTransform.options - gpt-5 textVerbosity", () => {
     expect(result.reasoningEffort).toBe("medium")
     expect(result.reasoningSummary).toBe("auto")
     expect(result.include).toEqual(["reasoning.encrypted_content"])
-    expect(result.textVerbosity).toBe("low")
+    expect(result.textVerbosity).toBe("medium")
   })
 
   test("openai-compatible gpt-5 models omit Responses-only reasoningSummary", () => {