feat(evals): add message log deduper utility (#10286)

packages/evals/src/cli/messageLogDeduper.test.ts

@@ -0,0 +1,35 @@
+import { MessageLogDeduper } from "./messageLogDeduper.js"
+
+describe("MessageLogDeduper", () => {
+	it("dedupes identical messages for same action+ts", () => {
+		const d = new MessageLogDeduper()
+		const msg = { ts: 123, type: "say", say: "reasoning", text: "hello", partial: false }
+
+		expect(d.shouldLog("updated", msg)).toBe(true)
+		expect(d.shouldLog("updated", msg)).toBe(false)
+	})
+
+	it("logs again if payload changes for same action+ts", () => {
+		const d = new MessageLogDeduper()
+		expect(d.shouldLog("updated", { ts: 123, text: "a" })).toBe(true)
+		expect(d.shouldLog("updated", { ts: 123, text: "b" })).toBe(true)
+	})
+
+	it("does not dedupe across different actions", () => {
+		const d = new MessageLogDeduper()
+		const msg = { ts: 123, text: "same" }
+		expect(d.shouldLog("created", msg)).toBe(true)
+		expect(d.shouldLog("updated", msg)).toBe(true)
+	})
+
+	it("evicts oldest entries", () => {
+		const d = new MessageLogDeduper(2)
+
+		expect(d.shouldLog("updated", { ts: 1, text: "a" })).toBe(true)
+		expect(d.shouldLog("updated", { ts: 2, text: "b" })).toBe(true)
+		// causes eviction of ts:1
+		expect(d.shouldLog("updated", { ts: 3, text: "c" })).toBe(true)
+		// ts:1 was evicted so it should log again
+		expect(d.shouldLog("updated", { ts: 1, text: "a" })).toBe(true)
+	})
+})

packages/evals/src/cli/messageLogDeduper.ts

@@ -0,0 +1,50 @@
+export class MessageLogDeduper {
+	private readonly lastLoggedByKey = new Map<string, string>()
+
+	constructor(private readonly maxEntries = 10_000) {}
+
+	/**
+	 * Returns true if this message should be logged.
+	 * Dedupe key: `${action}:${message.ts}`.
+	 * Dedupe rule: skip if payload is identical to the last logged payload for that key.
+	 */
+	public shouldLog(action: string | undefined, message: unknown): boolean {
+		if (!action || !message || typeof message !== "object") {
+			return true
+		}
+
+		const ts = (message as { ts?: unknown }).ts
+		if (typeof ts !== "number") {
+			return true
+		}
+
+		let serialized: string
+		try {
+			serialized = JSON.stringify(message)
+		} catch {
+			// If serialization fails, prefer logging.
+			return true
+		}
+
+		const key = `${action}:${ts}`
+		const prev = this.lastLoggedByKey.get(key)
+		if (prev === serialized) {
+			return false
+		}
+
+		// Refresh insertion order so eviction removes true oldest.
+		if (this.lastLoggedByKey.has(key)) {
+			this.lastLoggedByKey.delete(key)
+		}
+		this.lastLoggedByKey.set(key, serialized)
+
+		if (this.lastLoggedByKey.size > this.maxEntries) {
+			const oldestKey = this.lastLoggedByKey.keys().next().value as string | undefined
+			if (oldestKey) {
+				this.lastLoggedByKey.delete(oldestKey)
+			}
+		}
+
+		return true
+	}
+}

packages/evals/src/cli/runTask.ts

@@ -32,6 +32,7 @@ import { EVALS_REPO_PATH } from "../exercises/index.js"
 import { Logger, getTag, isDockerContainer } from "./utils.js"
 import { redisClient, getPubSubKey, registerRunner, deregisterRunner } from "./redis.js"
 import { runUnitTest } from "./runUnitTest.js"
+import { MessageLogDeduper } from "./messageLogDeduper.js"
 
 class SubprocessTimeoutError extends Error {
 	constructor(timeout: number) {
@@ -305,6 +306,7 @@ export const runTask = async ({ run, task, publish, logger, jobToken }: RunTaskO
 	]
 
 	let isApiUnstable = false
+	const messageLogDeduper = new MessageLogDeduper()
 
 	client.on(IpcMessageType.TaskEvent, async (taskEvent) => {
 		const { eventName, payload } = taskEvent
@@ -330,6 +332,15 @@ export const runTask = async ({ run, task, publish, logger, jobToken }: RunTaskO
 				(payload[0].message.say && loggableSays.includes(payload[0].message.say)) ||
 				payload[0].message.partial !== true)
 		) {
+			// Dedupe identical repeated message events (same message.ts + same payload)
+			if (eventName === RooCodeEventName.Message) {
+				const action = payload[0]?.action as string | undefined
+				const message = payload[0]?.message
+				if (!messageLogDeduper.shouldLog(action, message)) {
+					return
+				}
+			}
+
 			// Extract tool name for tool-related messages for clearer logging
 			let logEventName: string = eventName
 			if (eventName === RooCodeEventName.Message && payload[0]?.message?.ask === "tool") {