From ad361faed757429f03ecfe84ae3c66040f78220a Mon Sep 17 00:00:00 2001
From: Roo Code <chris@roocode.com>
Date: Fri, 20 Jun 2025 17:51:20 +0000
Subject: [PATCH] Fixes #4948: Fix context tokens calculation causing
 doubling/halving

- Fixed getApiMetrics to only use tokensIn for context calculation
- Previously was incorrectly adding tokensOut + cacheWrites + cacheReads
- This caused oscillating context display when condensing occurred
- Added comprehensive tests to verify the fix
---
 src/shared/__tests__/getApiMetrics.spec.ts | 156 +++++++++++++++++++++
 src/shared/getApiMetrics.ts                |   6 +-
 2 files changed, 160 insertions(+), 2 deletions(-)
 create mode 100644 src/shared/__tests__/getApiMetrics.spec.ts

diff --git a/src/shared/__tests__/getApiMetrics.spec.ts b/src/shared/__tests__/getApiMetrics.spec.ts
new file mode 100644
index 0000000000..68ee57616a
--- /dev/null
+++ b/src/shared/__tests__/getApiMetrics.spec.ts
@@ -0,0 +1,156 @@
+import { getApiMetrics } from "../getApiMetrics"
+import type { ClineMessage } from "@roo-code/types"
+
+describe("getApiMetrics", () => {
+	it("should calculate context tokens correctly using only tokensIn", () => {
+		const messages: ClineMessage[] = [
+			{
+				type: "say",
+				say: "api_req_started",
+				text: JSON.stringify({
+					request: "test request",
+					tokensIn: 1000,
+					tokensOut: 500,
+					cacheWrites: 200,
+					cacheReads: 100,
+					cost: 0.01,
+				}),
+				ts: Date.now(),
+			},
+		]
+
+		const result = getApiMetrics(messages)
+
+		// Context tokens should only include tokensIn, not the sum of all token types
+		expect(result.contextTokens).toBe(1000)
+		expect(result.totalTokensIn).toBe(1000)
+		expect(result.totalTokensOut).toBe(500)
+		expect(result.totalCacheWrites).toBe(200)
+		expect(result.totalCacheReads).toBe(100)
+		expect(result.totalCost).toBe(0.01)
+	})
+
+	it("should use newContextTokens from condense_context messages", () => {
+		const messages: ClineMessage[] = [
+			{
+				type: "say",
+				say: "api_req_started",
+				text: JSON.stringify({
+					tokensIn: 2000,
+					tokensOut: 800,
+				}),
+				ts: Date.now() - 1000,
+			},
+			{
+				type: "say",
+				say: "condense_context",
+				text: undefined,
+				contextCondense: {
+					summary: "Context was condensed",
+					cost: 0.02,
+					newContextTokens: 800,
+					prevContextTokens: 2000,
+				},
+				ts: Date.now(),
+			},
+		]
+
+		const result = getApiMetrics(messages)
+
+		// Should use newContextTokens from the most recent condense_context message
+		expect(result.contextTokens).toBe(800)
+		expect(result.totalCost).toBe(0.02) // Only condense cost since api_req_started has no cost
+	})
+
+	it("should handle multiple API requests and use the most recent for context", () => {
+		const messages: ClineMessage[] = [
+			{
+				type: "say",
+				say: "api_req_started",
+				text: JSON.stringify({
+					tokensIn: 1000,
+					tokensOut: 400,
+					cost: 0.01,
+				}),
+				ts: Date.now() - 2000,
+			},
+			{
+				type: "say",
+				say: "api_req_started",
+				text: JSON.stringify({
+					tokensIn: 1500,
+					tokensOut: 600,
+					cost: 0.015,
+				}),
+				ts: Date.now(),
+			},
+		]
+
+		const result = getApiMetrics(messages)
+
+		// Should use context tokens from the most recent API request
+		expect(result.contextTokens).toBe(1500)
+		expect(result.totalTokensIn).toBe(2500) // Sum of both requests
+		expect(result.totalTokensOut).toBe(1000) // Sum of both requests
+		expect(result.totalCost).toBe(0.025) // Sum of both costs
+	})
+
+	it("should handle missing or invalid JSON gracefully", () => {
+		const messages: ClineMessage[] = [
+			{
+				type: "say",
+				say: "api_req_started",
+				text: "invalid json",
+				ts: Date.now(),
+			},
+		]
+
+		const result = getApiMetrics(messages)
+
+		expect(result.contextTokens).toBe(0)
+		expect(result.totalTokensIn).toBe(0)
+		expect(result.totalTokensOut).toBe(0)
+		expect(result.totalCost).toBe(0)
+	})
+
+	it("should prioritize condense_context over api_req_started when both exist", () => {
+		const messages: ClineMessage[] = [
+			{
+				type: "say",
+				say: "api_req_started",
+				text: JSON.stringify({
+					tokensIn: 2000,
+					tokensOut: 800,
+				}),
+				ts: Date.now() - 1000,
+			},
+			{
+				type: "say",
+				say: "condense_context",
+				text: undefined,
+				contextCondense: {
+					summary: "Context was condensed",
+					cost: 0.02,
+					newContextTokens: 1200,
+					prevContextTokens: 2000,
+				},
+				ts: Date.now() - 500,
+			},
+			{
+				type: "say",
+				say: "api_req_started",
+				text: JSON.stringify({
+					tokensIn: 1300,
+					tokensOut: 500,
+				}),
+				ts: Date.now(),
+			},
+		]
+
+		const result = getApiMetrics(messages)
+
+		// Should use the most recent message for context calculation
+		// In this case, the most recent api_req_started
+		expect(result.contextTokens).toBe(1300)
+	})
+})
diff --git a/src/shared/getApiMetrics.ts b/src/shared/getApiMetrics.ts
index 49476fdbb6..8997216da5 100644
--- a/src/shared/getApiMetrics.ts
+++ b/src/shared/getApiMetrics.ts
@@ -72,8 +72,10 @@ export function getApiMetrics(messages: ClineMessage[]) {
 		if (message.type === "say" && message.say === "api_req_started" && message.text) {
 			try {
 				const parsedText: ParsedApiReqStartedTextType = JSON.parse(message.text)
-				const { tokensIn, tokensOut, cacheWrites, cacheReads } = parsedText
-				result.contextTokens = (tokensIn || 0) + (tokensOut || 0) + (cacheWrites || 0) + (cacheReads || 0)
+				const { tokensIn } = parsedText
+				// Only use tokensIn for context calculation - tokensOut, cacheWrites, and cacheReads
+				// are not part of the context window usage
+				result.contextTokens = tokensIn || 0
 			} catch (error) {
 				console.error("Error parsing JSON:", error)
 				continue