diff --git a/core/dev/perf-baseline.mjs b/core/dev/perf-baseline.mjs
index cb1cc95b..362873cd 100644
--- a/core/dev/perf-baseline.mjs
+++ b/core/dev/perf-baseline.mjs
@@ -23,8 +23,10 @@ const USAGE = 'usage: node dev/perf-baseline.mjs <bench-name> <port> --out <file
   + '[--turns 20] [--warmup 3] [--timeout 120000] [--message <text>]'
 
 // The mock model is scripted; this is a prompt the bench answers with one tool call, which is
-// what makes the tool-call latency sample exist at all.
-const DEFAULT_MESSAGE = '请读一下项目根目录的 README.md 第一段，然后告诉我它讲了什么。'
+// what makes the tool-call latency sample exist at all. It has to be a message `decide()` routes
+// to a tool (`/团队|班底|team/i` answers with `list_bots`), and a test feeds this exact string
+// through `decide()` so a change on either side fails instead of warning once per turn.
+export const DEFAULT_MESSAGE = '看看团队现在在忙什么。'
 
 export function nearestRank(sortedAsc, p) {
   if (!Array.isArray(sortedAsc) || sortedAsc.length === 0) throw new Error('nearestRank: no samples')
diff --git a/core/test/perf-baseline.test.js b/core/test/perf-baseline.test.js
index 6dcb527d..26593e2a 100644
--- a/core/test/perf-baseline.test.js
+++ b/core/test/perf-baseline.test.js
@@ -3,7 +3,10 @@
 import assert from 'node:assert/strict'
 import { test } from 'node:test'
 
+import { decide } from '../dev/mock-llm.mjs'
+
 import {
+  DEFAULT_MESSAGE,
   benchCommit,
   buildReport,
   markdownTable,
@@ -126,6 +129,21 @@ test('arguments: bench name, port and --out are required; flags parse by name',
   assert.equal(flags.timeout, 120_000)
 })
 
+test('the default message makes the scripted model call a tool (#2577)', () => {
+  // The bench answers only a message the scripted model routes to a tool; on any other message
+  // every turn warns "carried no tool call" and the run refuses to record. This is the seam, so
+  // a change to the default message or to the model's branches fails here instead of at run time.
+  const out = decide({
+    tools: [{ name: 'list_bots' }, { name: 'send_message' }],
+    messages: [{ role: 'user', content: DEFAULT_MESSAGE }],
+  })
+  const called = out.filter((block) => block.type === 'tool_use')
+  assert.ok(called.length > 0, `the perf default message carried no tool_use: ${JSON.stringify(out)}`)
+  assert.equal(called[0].name, 'list_bots')
+  // The message is the flag's default, so --message is the only way to override it.
+  assert.equal(parseArgs(['perf', '3097', '--out', '/tmp/perf.json']).message, DEFAULT_MESSAGE)
+})
+
 test('plugin load time is null when the bench never logged a ready line', () => {
   assert.equal(pluginLoadMs('/nonexistent-bench-home'), null)
 })
diff --git a/patches/ds-bot/files/dev/perf-baseline.mjs b/patches/ds-bot/files/dev/perf-baseline.mjs
index cb1cc95b..362873cd 100644
--- a/patches/ds-bot/files/dev/perf-baseline.mjs
+++ b/patches/ds-bot/files/dev/perf-baseline.mjs
@@ -23,8 +23,10 @@ const USAGE = 'usage: node dev/perf-baseline.mjs <bench-name> <port> --out <file
   + '[--turns 20] [--warmup 3] [--timeout 120000] [--message <text>]'
 
 // The mock model is scripted; this is a prompt the bench answers with one tool call, which is
-// what makes the tool-call latency sample exist at all.
-const DEFAULT_MESSAGE = '请读一下项目根目录的 README.md 第一段，然后告诉我它讲了什么。'
+// what makes the tool-call latency sample exist at all. It has to be a message `decide()` routes
+// to a tool (`/团队|班底|team/i` answers with `list_bots`), and a test feeds this exact string
+// through `decide()` so a change on either side fails instead of warning once per turn.
+export const DEFAULT_MESSAGE = '看看团队现在在忙什么。'
 
 export function nearestRank(sortedAsc, p) {
   if (!Array.isArray(sortedAsc) || sortedAsc.length === 0) throw new Error('nearestRank: no samples')
diff --git a/patches/ds-bot/files/test/perf-baseline.test.js b/patches/ds-bot/files/test/perf-baseline.test.js
index 6dcb527d..26593e2a 100644
--- a/patches/ds-bot/files/test/perf-baseline.test.js
+++ b/patches/ds-bot/files/test/perf-baseline.test.js
@@ -3,7 +3,10 @@
 import assert from 'node:assert/strict'
 import { test } from 'node:test'
 
+import { decide } from '../dev/mock-llm.mjs'
+
 import {
+  DEFAULT_MESSAGE,
   benchCommit,
   buildReport,
   markdownTable,
@@ -126,6 +129,21 @@ test('arguments: bench name, port and --out are required; flags parse by name',
   assert.equal(flags.timeout, 120_000)
 })
 
+test('the default message makes the scripted model call a tool (#2577)', () => {
+  // The bench answers only a message the scripted model routes to a tool; on any other message
+  // every turn warns "carried no tool call" and the run refuses to record. This is the seam, so
+  // a change to the default message or to the model's branches fails here instead of at run time.
+  const out = decide({
+    tools: [{ name: 'list_bots' }, { name: 'send_message' }],
+    messages: [{ role: 'user', content: DEFAULT_MESSAGE }],
+  })
+  const called = out.filter((block) => block.type === 'tool_use')
+  assert.ok(called.length > 0, `the perf default message carried no tool_use: ${JSON.stringify(out)}`)
+  assert.equal(called[0].name, 'list_bots')
+  // The message is the flag's default, so --message is the only way to override it.
+  assert.equal(parseArgs(['perf', '3097', '--out', '/tmp/perf.json']).message, DEFAULT_MESSAGE)
+})
+
 test('plugin load time is null when the bench never logged a ready line', () => {
   assert.equal(pluginLoadMs('/nonexistent-bench-home'), null)
 })