fix(ai): require OpenAI Responses terminal events

packages/coding-agent/test/compaction.test.ts

@@ -9,6 +9,7 @@ import {
 	calculateContextTokens,
 	compact,
 	DEFAULT_COMPACTION_SETTINGS,
+	estimateContextTokens,
 	findCutPoint,
 	getLastAssistantUsage,
 	prepareCompaction,
@@ -217,12 +218,43 @@ describe("getLastAssistantUsage", () => {
 		expect(usage!.input).toBe(100);
 	});
 
+	it("should skip all-zero assistant usage", () => {
+		const entries: SessionEntry[] = [
+			createMessageEntry(createUserMessage("Hello")),
+			createMessageEntry(createAssistantMessage("Hi", createMockUsage(100, 50))),
+			createMessageEntry(createUserMessage("continue")),
+			createMessageEntry(createAssistantMessage("Partial", createMockUsage(0, 0))),
+		];
+
+		const usage = getLastAssistantUsage(entries);
+		expect(usage).not.toBeNull();
+		expect(usage!.input).toBe(100);
+	});
+
 	it("should return undefined if no assistant messages", () => {
 		const entries: SessionEntry[] = [createMessageEntry(createUserMessage("Hello"))];
 		expect(getLastAssistantUsage(entries)).toBeUndefined();
 	});
 });
 
+describe("estimateContextTokens", () => {
+	it("uses the last non-zero assistant usage as the context anchor", () => {
+		const messages: AgentMessage[] = [
+			createUserMessage("Hello"),
+			createAssistantMessage("Hi", createMockUsage(100, 50)),
+			createUserMessage("continue"),
+			createAssistantMessage("Partial thinking", createMockUsage(0, 0)),
+		];
+
+		const estimate = estimateContextTokens(messages);
+
+		expect(estimate.usageTokens).toBe(150);
+		expect(estimate.lastUsageIndex).toBe(1);
+		expect(estimate.trailingTokens).toBeGreaterThan(0);
+		expect(estimate.tokens).toBe(150 + estimate.trailingTokens);
+	});
+});
+
 describe("shouldCompact", () => {
 	it("should return true when context exceeds threshold", () => {
 		const settings: CompactionSettings = {