refactor(opencode): fail the execute tool on program failure (#35180)

packages/opencode/src/session/processor.ts

@@ -192,6 +192,8 @@ const layer = Layer.effect(
             status: "error",
             input: match.part.state.input,
             error: errorMessage(error),
+            // Keep metadata streamed while running so failures retain progress detail (e.g. execute's child calls).
+            metadata: match.part.state.metadata,
             time: { start: match.part.state.time.start, end: Date.now() },
           },
         })

packages/opencode/src/tool/code-mode.ts

@@ -287,35 +287,37 @@ export const CodeModeTool = Tool.define(
 
         const result = yield* Effect.raceFirst(runtime.execute(params.code), abort.pipe(Effect.map(cancelled)))
         const logs = result.logs ?? []
-        const attached = attachments.length > 0 ? { attachments } : {}
-        const hints = result.ok
-          ? []
-          : (result.error.suggestions ?? []).filter((hint) => !result.error.message.includes(hint))
-        const metadata: Metadata = result.ok ? { toolCalls: calls } : { toolCalls: calls, error: true }
-        let output: string
-        if (result.ok) {
-          if (typeof result.value === "string") output = result.value
-          else if (result.value === undefined) output = "undefined"
-          else {
-            try {
-              output = JSON.stringify(result.value, null, 2) ?? String(result.value)
-            } catch {
-              output = String(result.value)
-            }
-          }
-        } else {
-          output = [result.error.message, ...hints].join("\n")
+        const withLogs = (text: string) => {
+          if (logs.length === 0) return text
+          return text.length > 0 ? `${text}\n\nLogs:\n${logs.join("\n")}` : `Logs:\n${logs.join("\n")}`
         }
-        if (logs.length > 0)
-          output = output.length > 0 ? `${output}\n\nLogs:\n${logs.join("\n")}` : `Logs:\n${logs.join("\n")}`
+
+        if (!result.ok) {
+          if (ctx.abort.aborted) {
+            return {
+              title: CODE_MODE_TOOL,
+              metadata: { toolCalls: calls, error: true },
+              output: "Execution cancelled.",
+            } satisfies Tool.ExecuteResult<Metadata>
+          }
+          const hints = (result.error.suggestions ?? []).filter((hint) => !result.error.message.includes(hint))
+          return yield* Effect.fail(new Error(withLogs([result.error.message, ...hints].join("\n"))))
+        }
+
+        // The interpreter validates returned values as plain JSON, so stringify cannot throw;
+        // it yields undefined only for a program that returns undefined.
+        const output =
+          typeof result.value === "string"
+            ? result.value
+            : (JSON.stringify(result.value, null, 2) ?? String(result.value))
 
         return {
           title: CODE_MODE_TOOL,
-          metadata,
-          output,
-          ...attached,
+          metadata: { toolCalls: calls },
+          output: withLogs(output),
+          ...(attachments.length > 0 ? { attachments } : {}),
         } satisfies Tool.ExecuteResult<Metadata>
-      }),
+      }, Effect.orDie),
     }
     return init
   }),

packages/opencode/test/tool/code-mode-integration.test.ts

@@ -17,7 +17,7 @@ import {
   ListToolsRequestSchema,
   type Tool as MCPToolDef,
 } from "@modelcontextprotocol/sdk/types.js"
-import { Effect, Layer } from "effect"
+import { Cause, Effect, Exit, Layer } from "effect"
 
 const PNG = "iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNkYPhfDwAChwGA60e6kgAAAABJRU5ErkJggg=="
 
@@ -160,6 +160,12 @@ async function buildTool() {
 }
 
 const run = (code: string) => Effect.runPromise(tool.execute({ code }, ctx))
+// Program failures die at the tool boundary; recover the defect for message assertions.
+const runFailed = async (code: string) => {
+  const exit = await Effect.runPromise(tool.execute({ code }, ctx).pipe(Effect.exit))
+  if (Exit.isSuccess(exit)) throw new Error("expected the tool to fail")
+  return Cause.squash(exit.cause) as Error
+}
 
 beforeAll(async () => {
   const built = await buildTool()
@@ -248,9 +254,8 @@ describe("code mode integration (real MCP server)", () => {
   })
 
   test("an uncaught MCP error surfaces as a failed execution", async () => {
-    const out = await run("await tools.fixtures.boom({}); return 'unreachable'")
-    expect(out.metadata.error).toBe(true)
-    expect(out.output).toContain("kaboom")
+    const error = await runFailed("await tools.fixtures.boom({}); return 'unreachable'")
+    expect(error.message).toContain("kaboom")
   })
 
   test("console output is captured and appended as a Logs section after the result", async () => {
@@ -265,14 +270,13 @@ describe("code mode integration (real MCP server)", () => {
   })
 
   test("console output is preserved on the error path", async () => {
-    const out = await run(`
+    const error = await runFailed(`
       console.log("before the throw")
       await tools.fixtures.boom({})
       return "unreachable"
     `)
-    expect(out.metadata.error).toBe(true)
-    expect(out.output).toContain("kaboom")
-    expect(out.output).toContain("Logs:\nbefore the throw")
+    expect(error.message).toContain("kaboom")
+    expect(error.message).toContain("Logs:\nbefore the throw")
   })
 
   test("a program that logs nothing gets no Logs section", async () => {

packages/opencode/test/tool/code-mode.test.ts

@@ -10,7 +10,7 @@ import { Session } from "@/session/session"
 import { Tool } from "@/tool/tool"
 import * as Truncate from "@/tool/truncate"
 import { MessageID, SessionID } from "@/session/schema"
-import { Effect, Layer, Schema } from "effect"
+import { Cause, Effect, Exit, Layer, Schema } from "effect"
 
 const ctx: Tool.Context = {
   sessionID: SessionID.make("ses_code-mode"),
@@ -86,6 +86,13 @@ function describeFor(mcpTools: Record<string, MCP.McpTool>, servers?: string[],
   return describeCatalog(Permission.visibleTools(mcpTools, permission), serverNames(mcpTools, servers))
 }
 
+// Program failures die at the tool boundary; recover the defect for message assertions.
+async function failure(effect: Effect.Effect<unknown>) {
+  const exit = await Effect.runPromise(effect.pipe(Effect.exit))
+  if (Exit.isSuccess(exit)) throw new Error("expected the tool to fail")
+  return Cause.squash(exit.cause) as Error
+}
+
 describe("code mode execute", () => {
   test("defines execute input with an Effect schema", async () => {
     const decode = Schema.decodeUnknownEffect(Parameters)
@@ -313,18 +320,16 @@ describe("code mode execute", () => {
     expect(output.metadata.toolCalls.every((c) => c.status === "completed")).toBe(true)
   })
 
-  test("returns a readable error when the program throws", async () => {
+  test("a program failure fails the tool with a readable error", async () => {
     const tool = await build({})
-    const output = await Effect.runPromise(tool.execute({ code: "throw new Error('boom')" }, ctx))
-    expect(output.output).toBe("Uncaught: boom")
-    expect(output.metadata.error).toBe(true)
+    const error = await failure(tool.execute({ code: "throw new Error('boom')" }, ctx))
+    expect(error.message).toBe("Uncaught: boom")
   })
 
   test("reports an unknown tool as a failed execution", async () => {
     const tool = await build({ known_tool: mcpTool("tool", () => "ok") })
-    const output = await Effect.runPromise(tool.execute({ code: "return await tools.known.missing({})" }, ctx))
-    expect(output.metadata.error).toBe(true)
-    expect(output.output).toContain("Unknown tool 'known.missing'")
+    const error = await failure(tool.execute({ code: "return await tools.known.missing({})" }, ctx))
+    expect(error.message).toContain("Unknown tool 'known.missing'")
   })
 
   test("propagates an MCP tool error into the program as a catchable failure", async () => {
@@ -576,8 +581,8 @@ describe("code mode execute", () => {
 
   test("isolates the sandbox from host globals", async () => {
     const tool = await build({})
-    const output = await Effect.runPromise(tool.execute({ code: "return process.env" }, ctx))
-    expect(output.metadata.error).toBe(true)
+    const error = await failure(tool.execute({ code: "return process.env" }, ctx))
+    expect(error.message).toContain("process")
   })
 
   test("cancelling via ctx.abort interrupts the running program", async () => {
@@ -627,12 +632,9 @@ describe("code mode execute", () => {
     )
     expect(ok.output).toBe("done\n\nLogs:\nstep one\n[warn] careful")
 
-    const err = await Effect.runPromise(
-      tool.execute({ code: "console.log('before the throw'); throw new Error('boom')" }, ctx),
-    )
-    expect(err.metadata.error).toBe(true)
-    expect(err.output).toContain("Uncaught: boom")
-    expect(err.output).toContain("Logs:\nbefore the throw")
+    const error = await failure(tool.execute({ code: "console.log('before the throw'); throw new Error('boom')" }, ctx))
+    expect(error.message).toContain("Uncaught: boom")
+    expect(error.message).toContain("Logs:\nbefore the throw")
   })
 })
 
@@ -677,12 +679,9 @@ describe("code mode permission visibility", () => {
       [deny("github_create_issue")],
     )
 
-    const denied = await Effect.runPromise(
-      tool.execute({ code: "return await tools.github.create_issue({ title: 'x' })" }, ctx),
-    )
-    expect(denied.metadata.error).toBe(true)
-    expect(denied.output).toContain("Unknown tool 'github.create_issue'")
-    expect(denied.output).not.toContain("permission")
+    const denied = await failure(tool.execute({ code: "return await tools.github.create_issue({ title: 'x' })" }, ctx))
+    expect(denied.message).toContain("Unknown tool 'github.create_issue'")
+    expect(denied.message).not.toContain("permission")
     expect(called).toEqual([])
 
     const allowed = await Effect.runPromise(tool.execute({ code: "return await tools.github.list_issues({})" }, ctx))