fix(cli): send max-step instruction as user message

Co-authored-by: kiloconnect[bot] <240665456+kiloconnect[bot]@users.noreply.github.com>

.changeset/fix-max-steps-prefill.md

@@ -0,0 +1,5 @@
+---
+"@kilocode/cli": patch
+---
+
+Prevent provider errors when agents reach their step limit by sending the final summary instruction as user input instead of an assistant prefill.

packages/core/src/session/runner/llm.ts

@@ -209,7 +209,7 @@ const layer = Layer.effect(
         system: [agent.info?.system, system.baseline]
           .filter((part): part is string => part !== undefined && part.length > 0)
           .map(SystemPart.make),
-        messages: [...toLLMMessages(context, model), ...(isLastStep ? [Message.assistant(MAX_STEPS_PROMPT)] : [])],
+        messages: [...toLLMMessages(context, model), ...(isLastStep ? [Message.user(MAX_STEPS_PROMPT)] : [])], // kilocode_change - avoid provider-incompatible assistant prefill
         tools: toolMaterialization?.definitions ?? [],
         toolChoice: isLastStep ? "none" : undefined,
       })

packages/core/test/session-runner.test.ts

@@ -3022,7 +3022,7 @@ describe("SessionRunnerLLM", () => {
       expect(requests[1]?.toolChoice).toMatchObject({ type: "none" })
       expect(requests[1]?.tools).toEqual([])
       expect(requests[1]?.messages.at(-1)).toMatchObject({
-        role: "assistant",
+        role: "user", // kilocode_change - max-step instructions must not become assistant prefill
         content: [{ type: "text", text: expect.stringContaining("MAXIMUM STEPS REACHED") }],
       })
       expect(executions).toEqual(["done"])

packages/opencode/src/session/prompt.ts

@@ -1772,7 +1772,7 @@ export const layer = Layer.effect(
             system,
             messages: [
               ...modelMsgs,
-              ...(isLastStep ? [{ role: "assistant" as const, content: MAX_STEPS_PROMPT }] : []),
+              ...(isLastStep ? [{ role: "user" as const, content: MAX_STEPS_PROMPT }] : []), // kilocode_change - avoid provider-incompatible assistant prefill
             ],
             tools,
             model,

packages/opencode/test/session/prompt.test.ts

@@ -550,6 +550,42 @@ it.instance("loop calls LLM and returns assistant message", () =>
   }),
 )
 
+// kilocode_change start - guard provider-compatible max-step request shape
+it.instance(
+  "loop sends max steps instruction as a user message",
+  () =>
+    Effect.gen(function* () {
+      const { llm } = yield* useServerConfig((url) => ({
+        ...providerCfg(url),
+        agent: { build: { steps: 1 } },
+      }))
+      const prompt = yield* SessionPrompt.Service
+      const sessions = yield* Session.Service
+      const chat = yield* sessions.create({ title: "Pinned" })
+      yield* prompt.prompt({
+        sessionID: chat.id,
+        agent: "build",
+        noReply: true,
+        parts: [{ type: "text", text: "finish at the limit" }],
+      })
+      yield* llm.text("summary")
+
+      yield* prompt.loop({ sessionID: chat.id })
+
+      const inputs = yield* llm.inputs
+      const messages = inputs.at(-1)?.messages
+      if (!Array.isArray(messages)) throw new Error("expected LLM messages")
+      expect(messages.at(-1)).toMatchObject({
+        role: "user",
+        content: expect.arrayContaining([
+          { type: "text", text: expect.stringContaining("MAXIMUM STEPS REACHED") },
+        ]),
+      })
+    }),
+  30_000,
+)
+// kilocode_change end
+
 noLLMServer.instance(
   "new prompt dismisses a pending question",
   () =>