diff --git a/packages/opencode/src/cli/cmd/run.ts b/packages/opencode/src/cli/cmd/run.ts
index 9e025ead380..36ea78e8dc5 100644
--- a/packages/opencode/src/cli/cmd/run.ts
+++ b/packages/opencode/src/cli/cmd/run.ts
@@ -779,6 +779,14 @@ export const RunCommand = effectCmd({
         const drain = KiloRunDrain.create(sessionID)
         if (!args.attach && !args.auto && !skipPermissions) KiloHeadless.mark(sessionID) // kilocode_change - --yolo skips too
         // kilocode_change end
+        // kilocode_change start - remember whether the model produced any assistant output,
+        // so a run that ends without one does not exit 0
+        let assistantOutput = false
+        // the raced request (prompt, command, or summarize) itself failed; the
+        // result.error handler below already reported the real cause, so the
+        // empty-output diagnostic must not claim a silent model on top of it
+        let promptFailed = false
+        // kilocode_change end
 
         function emit(type: string, data: Record<string, unknown>) {
           if (args.format === "json") {
@@ -831,6 +839,14 @@ export const RunCommand = effectCmd({
               // kilocode_change end
               if (part.sessionID !== sessionID) continue
 
+              // kilocode_change start - text, reasoning, and tool parts are the
+              // model's response; step markers are not
+              if (part.type === "tool") assistantOutput = true
+              else if ((part.type === "text" || part.type === "reasoning") && part.time?.end && part.text.trim()) {
+                assistantOutput = true
+              }
+              // kilocode_change end
+
               if (part.type === "tool" && (part.state.status === "completed" || part.state.status === "error")) {
                 if (emit("tool_use", { part })) continue
                 if (part.state.status === "completed") {
@@ -1075,11 +1091,21 @@ export const RunCommand = effectCmd({
                     }),
             )
             if (result.error) {
+              promptFailed = true
               if (!emit("error", { error: result.error })) UI.error(formatRunError(result.error))
               process.exitCode = 1
             }
             await drain.wait(client, cwd)
+            // kilocode_change start - an empty model response must not exit 0: a caller
+            // cannot tell an empty run from a successful one otherwise
             if (await completed) process.exitCode = 1
+            else if (!assistantOutput && !promptFailed) {
+              const message = "run ended without an assistant message; the model returned no output"
+              UI.error(message)
+              emit("error", { error: message })
+              process.exitCode = 1
+            }
+            // kilocode_change end
           } catch (error) {
             const text = error instanceof Error ? error.message : String(error)
             if (!emit("error", { error: text })) UI.error(text)
diff --git a/packages/opencode/test/cli/run/run-empty.process.test.ts b/packages/opencode/test/cli/run/run-empty.process.test.ts
new file mode 100644
index 00000000000..0d7f6ac10f0
--- /dev/null
+++ b/packages/opencode/test/cli/run/run-empty.process.test.ts
@@ -0,0 +1,101 @@
+// kilocode_change - new file
+// Subprocess tests for `kilo run` when the model produces no assistant message.
+// Same harness as run-process.test.ts — see `test/lib/cli-process.ts`.
+import { describe, expect } from "bun:test"
+import { Effect } from "effect"
+import { reply } from "../../lib/llm-server"
+import { cliIt } from "../../lib/cli-process"
+
+// kilocode_change start - a run where the model returns no assistant message used to
+// exit 0, so a caller could not tell an empty run from a successful one
+describe("opencode run with an empty model completion (non-interactive subprocess)", () => {
+  cliIt.concurrent(
+    "exits nonzero with a stderr diagnostic",
+    ({ llm, opencode }) =>
+      Effect.gen(function* () {
+        yield* llm.push(reply().stop())
+        const result = yield* opencode.run("say hi", { timeoutMs: 30_000 })
+        opencode.expectExit(result, 1)
+        expect(result.stderr).toContain("run ended without an assistant message")
+        expect(result.stdout).toBe("")
+      }),
+    60_000,
+  )
+
+  cliIt.concurrent(
+    "emits an error record under --format json",
+    ({ llm, opencode }) =>
+      Effect.gen(function* () {
+        yield* llm.push(reply().stop())
+        const result = yield* opencode.run("say hi", { format: "json", timeoutMs: 30_000 })
+        opencode.expectExit(result, 1)
+        expect(result.stderr).toContain("run ended without an assistant message")
+        const events = opencode.parseJsonEvents(result.stdout)
+        expect(events.at(-1)).toEqual({
+          type: "error",
+          timestamp: expect.any(Number),
+          sessionID: expect.any(String),
+          error: "run ended without an assistant message; the model returned no output",
+        })
+      }),
+    60_000,
+  )
+
+  // The prompt request itself can fail before the model ever runs (here: an
+  // attached image whose base64 exceeds the server's limit → 400 BadRequest,
+  // with no session.error event). The real error is reported once; the
+  // empty-output diagnostic must not claim a silent model on top of it.
+  cliIt.live(
+    "request failure reports the real error without the empty-output diagnostic",
+    ({ home, opencode }) =>
+      Effect.gen(function* () {
+        // 4 MiB binary content with a .png extension: under the CLI's 10 MiB
+        // attach cap, but its ~5.33 MiB base64 exceeds the server's 5 MiB
+        // image limit, so the prompt request fails with a 400 BadRequest.
+        const big = Buffer.alloc(4 * 1024 * 1024, 7)
+        Buffer.from([0x89, 0x50, 0x4e, 0x47]).copy(big, 0)
+        yield* Effect.promise(() => Bun.write(`${home}/big.png`, big))
+        const server = yield* opencode.serve()
+
+        const result = yield* opencode.run("read the attachment", {
+          extraArgs: ["--attach", server.url, `--file=${home}/big.png`, "--"],
+          timeoutMs: 30_000,
+        })
+
+        opencode.expectExit(result, 1)
+        expect(result.stderr).toContain("BadRequest")
+        expect(result.stderr).not.toContain("run ended without an assistant message")
+      }),
+    60_000,
+  )
+
+  cliIt.live(
+    "request failure emits only the real error record under --format json",
+    ({ home, opencode }) =>
+      Effect.gen(function* () {
+        const big = Buffer.alloc(4 * 1024 * 1024, 7)
+        Buffer.from([0x89, 0x50, 0x4e, 0x47]).copy(big, 0)
+        yield* Effect.promise(() => Bun.write(`${home}/big.json.png`, big))
+        const server = yield* opencode.serve()
+
+        const result = yield* opencode.run("read the attachment", {
+          format: "json",
+          extraArgs: ["--attach", server.url, `--file=${home}/big.json.png`, "--"],
+          timeoutMs: 30_000,
+        })
+
+        opencode.expectExit(result, 1)
+        expect(result.stderr).not.toContain("run ended without an assistant message")
+        const events = opencode.parseJsonEvents(result.stdout)
+        expect(events).toHaveLength(1)
+        expect(events[0]).toEqual({
+          type: "error",
+          timestamp: expect.any(Number),
+          sessionID: expect.any(String),
+          error: expect.objectContaining({ _tag: "BadRequest" }),
+        })
+      }),
+    60_000,
+  )
+})
+// kilocode_change end