diff --git a/apps/cli/src/ui/view/transcript/assistant-message.ts b/apps/cli/src/ui/view/transcript/assistant-message.ts
index 0cf39aaf..4fec0874 100644
--- a/apps/cli/src/ui/view/transcript/assistant-message.ts
+++ b/apps/cli/src/ui/view/transcript/assistant-message.ts
@@ -18,6 +18,37 @@ const OSC133_ZONE_FINAL = "\x1b]133;C\x07";
 /** Cached zone-marked output, keyed by the child lines it was derived from. */
 type ZoneCache = { width: number; native: string[]; lines: string[] };
 
+/**
+ * Normalize a provider-assembled content array so that thinking blocks always
+ * render before the answer. Most models emit the reasoning stream first, but
+ * some turns flush it only after the answer text, which would otherwise paint
+ * the thinking block below the answer in those turns and above it in every
+ * other turn of the same conversation.
+ *
+ * Only the degenerate shape is rewritten: when no thinking block precedes the
+ * first answer block, every thinking block is hoisted ahead of the answer
+ * (stable within each group). Alternating runs (thinking, answer, thinking,
+ * answer) already lead with thinking and are returned untouched, preserving
+ * segmented reasoning where each answer follows its own thinking.
+ */
+export function withThinkingBeforeAnswer(content: AssistantMessage["content"]): AssistantMessage["content"] {
+	let firstText = -1;
+	let firstThinking = -1;
+	for (let i = 0; i < content.length; i++) {
+		const block = content[i];
+		if (block === undefined) continue;
+		if (firstText === -1 && block.type === "text" && block.text.trim().length > 0) firstText = i;
+		if (firstThinking === -1 && block.type === "thinking" && block.thinking.trim().length > 0) firstThinking = i;
+	}
+	if (firstText === -1 || firstThinking < firstText) {
+		return content;
+	}
+	return [
+		...content.filter((block) => block.type === "thinking"),
+		...content.filter((block) => block.type !== "thinking"),
+	];
+}
+
 /**
  * Component that renders a complete assistant message
  */
@@ -130,9 +161,11 @@ export class AssistantMessageComponent extends Container {
 			this.contentContainer.addChild(new Spacer(1));
 		}
 
+		const orderedContent = withThinkingBeforeAnswer(message.content);
+
 		// Render content in order
-		for (let i = 0; i < message.content.length; i++) {
-			const content = message.content[i];
+		for (let i = 0; i < orderedContent.length; i++) {
+			const content = orderedContent[i];
 			if (content.type === "text" && content.text.trim()) {
 				// Assistant text messages with no background - trim the text
 				// Set paddingY=0 to avoid extra spacing before tool executions
@@ -143,8 +176,8 @@ export class AssistantMessageComponent extends Container {
 				);
 			} else if (content.type === "thinking") {
 				const thinkingBlocks: string[] = [];
-				for (; i < message.content.length; i++) {
-					const thinkingContent = message.content[i];
+				for (; i < orderedContent.length; i++) {
+					const thinkingContent = orderedContent[i];
 					if (thinkingContent.type !== "thinking") {
 						break;
 					}
@@ -161,7 +194,7 @@ export class AssistantMessageComponent extends Container {
 
 				// Add spacing only when another visible assistant content block follows.
 				// This avoids a superfluous blank line before separately-rendered tool execution blocks.
-				const hasVisibleContentAfter = message.content
+				const hasVisibleContentAfter = orderedContent
 					.slice(i + 1)
 					.some((c) => (c.type === "text" && c.text.trim()) || (c.type === "thinking" && c.thinking.trim()));
 
diff --git a/apps/cli/src/ui/view/transcript/step-message.ts b/apps/cli/src/ui/view/transcript/step-message.ts
index dbba2ff9..1176b993 100644
--- a/apps/cli/src/ui/view/transcript/step-message.ts
+++ b/apps/cli/src/ui/view/transcript/step-message.ts
@@ -3,7 +3,7 @@ import { getMarkdownTheme, theme } from "@step-harness/coding-agent";
 import { type MarkdownTheme, stripTerminalSequences, truncateToWidth, visibleWidth } from "@step-harness/pi-tui";
 import type { AssistantMessage } from "@step-harness/providers";
 import { estimateTokens, formatElapsedTime, formatTokens } from "../chrome/footer.ts";
-import { AssistantMessageComponent } from "./assistant-message.ts";
+import { AssistantMessageComponent, withThinkingBeforeAnswer } from "./assistant-message.ts";
 import { UserMessageComponent } from "./user-message.ts";
 
 /** OSC 133 prompt-zone markers emitted by pi's message components. */
@@ -107,7 +107,7 @@ function partitionAssistantContent(message: AssistantMessage): AssistantContentR
 	const runs: AssistantContentRun[] = [];
 	let previousKind: AssistantContentRun["kind"] | undefined;
 
-	for (const content of message.content) {
+	for (const content of withThinkingBeforeAnswer(message.content)) {
 		if (content.type !== "thinking" && content.type !== "text") {
 			previousKind = undefined;
 			continue;
diff --git a/apps/cli/test/assistant-message.test.ts b/apps/cli/test/assistant-message.test.ts
index 161fe360..3e19a666 100644
--- a/apps/cli/test/assistant-message.test.ts
+++ b/apps/cli/test/assistant-message.test.ts
@@ -91,6 +91,48 @@ describe("AssistantMessageComponent", () => {
 		expect(rendered).toContain("answer");
 	});
 
+	test("hoists thinking ahead of the answer when the model streams it last", () => {
+		initTheme("dark");
+
+		const component = new AssistantMessageComponent(
+			createAssistantMessage([
+				{ type: "text", text: "answer text" },
+				{ type: "thinking", thinking: "late reasoning" },
+			]),
+		);
+		const lines = stripAnsi(component.render(80).join("\n")).split("\n");
+		const thinkingLine = lines.findIndex((line) => line.includes("late reasoning"));
+		const answerLine = lines.findIndex((line) => line.includes("answer text"));
+
+		expect(thinkingLine).toBeGreaterThanOrEqual(0);
+		expect(answerLine).toBeGreaterThanOrEqual(0);
+		expect(thinkingLine).toBeLessThan(answerLine);
+	});
+
+	test("keeps segmented reasoning in place when thinking already leads", () => {
+		initTheme("dark");
+
+		const component = new AssistantMessageComponent(
+			createAssistantMessage([
+				{ type: "thinking", thinking: "first reasoning" },
+				{ type: "text", text: "first answer" },
+				{ type: "thinking", thinking: "follow-up reasoning" },
+				{ type: "text", text: "final answer" },
+			]),
+		);
+		const lines = stripAnsi(component.render(80).join("\n")).split("\n");
+
+		expect(lines.findIndex((line) => line.includes("first reasoning"))).toBeLessThan(
+			lines.findIndex((line) => line.includes("first answer")),
+		);
+		expect(lines.findIndex((line) => line.includes("follow-up reasoning"))).toBeGreaterThan(
+			lines.findIndex((line) => line.includes("first answer")),
+		);
+		expect(lines.findIndex((line) => line.includes("final answer"))).toBeGreaterThan(
+			lines.findIndex((line) => line.includes("follow-up reasoning")),
+		);
+	});
+
 	test("uses configured output padding for text and thinking", () => {
 		initTheme("dark");
 
diff --git a/apps/cli/test/step-message.test.ts b/apps/cli/test/step-message.test.ts
index 683ebb25..cf3ddec7 100644
--- a/apps/cli/test/step-message.test.ts
+++ b/apps/cli/test/step-message.test.ts
@@ -176,6 +176,24 @@ describe("Step message presentation", () => {
 		expect(finalAnswer).toBeGreaterThan(followUpThinking);
 	});
 
+	test("hoists thinking ahead of the answer when the model streams it last", () => {
+		initTheme("step-blue");
+		const component = new StepAssistantMessageComponent(
+			assistantMessage([
+				{ type: "text", text: "the answer" },
+				{ type: "thinking", thinking: "late reasoning" },
+			]),
+			false,
+			getMarkdownTheme(),
+		);
+		const lines = plain(component.render(80));
+		const thinking = lines.findIndex((line) => line.includes("late reasoning"));
+		const answer = lines.findIndex((line) => line.includes("the answer"));
+		expect(thinking).toBeGreaterThanOrEqual(0);
+		expect(answer).toBeGreaterThanOrEqual(0);
+		expect(thinking).toBeLessThan(answer);
+	});
+
 	test("updates output padding without dropping the first character", () => {
 		initTheme("step-blue");
 		const assistant = new StepAssistantMessageComponent(