diff --git a/changelog.d/workflow-prompt-codepoint-cut.fixed.md b/changelog.d/workflow-prompt-codepoint-cut.fixed.md
new file mode 100644
index 000000000..93e4ab085
--- /dev/null
+++ b/changelog.d/workflow-prompt-codepoint-cut.fixed.md
@@ -0,0 +1 @@
+A workflow node prompt, the task brief handed to a run, and the reply a step hands on now stop on their character bound without splitting the character there: the cut counts code points like a task title already did, so a prompt or reply that ends in an emoji or another character outside the BMP no longer carries half of it as a replacement character.
diff --git a/plugins/workflow/src/graph.js b/plugins/workflow/src/graph.js
index dafb801ac..582ef67d4 100644
--- a/plugins/workflow/src/graph.js
+++ b/plugins/workflow/src/graph.js
@@ -34,7 +34,9 @@ export function cleanPrompt(value, max = 4000) {
     if (code === 0x0a || code === 0x09) { out += ch; continue }
     out += code <= 0x1f || code === 0x7f ? ' ' : ch
   }
-  return out.replace(/[ \t]+\n/g, '\n').trim().slice(0, max)
+  const text = out.replace(/[ \t]+\n/g, '\n').trim()
+  // Same rule as cleanLine above: a UTF-16 cut can end inside one character.
+  return text.length <= max ? text : Array.from(text).slice(0, max).join('')
 }
 
 export function cleanTools(value) {
@@ -78,7 +80,8 @@ export function readStepResult(text, node) {
     if (!match) { kept.push(line); continue }
     outputs[match[1].toLowerCase()] = cleanLine(match[2], 2000)
   }
-  const body = kept.join('\n').trim().slice(0, 2000)
+  const joined = kept.join('\n').trim()
+  const body = joined.length <= 2000 ? joined : Array.from(joined).slice(0, 2000).join('')
   const declared = node?.outputs ?? []
   if (declared.length === 1 && !outputs[declared[0].name] && body) outputs[declared[0].name] = body
   if (declared.length > 0) {
diff --git a/plugins/workflow/test/graph.test.js b/plugins/workflow/test/graph.test.js
index f7484b33a..69cb7ff6e 100644
--- a/plugins/workflow/test/graph.test.js
+++ b/plugins/workflow/test/graph.test.js
@@ -1,8 +1,8 @@
 import assert from 'node:assert/strict'
 import { test } from 'node:test'
 import {
-  addNode, chainLabel, cleanLine, completeNode, connectNodes, doneCount, failNode, freshStatus, layerMarks,
-  moveNode, nextDemoOp, normalizeGraph, reconnectEdge, removeEdge, removeNode, retryNode,
+  addNode, chainLabel, cleanLine, cleanPrompt, completeNode, connectNodes, doneCount, failNode, freshStatus, layerMarks,
+  moveNode, nextDemoOp, normalizeGraph, readStepResult, reconnectEdge, removeEdge, removeNode, retryNode,
   rowPreview, startFrontier, summarizeTask, validateGraph,
 } from '../src/graph.js'
 import { deliveryWorkflow, weeklyWorkflow } from '../src/template.js'
@@ -188,3 +188,24 @@ test('a long task label keeps its whole title and never ends in a lone surrogate
   assert.equal(cleanLine('🚀'.repeat(10), 7), '🚀'.repeat(7))
   assert.equal(hasLoneSurrogate(cleanLine('🚀'.repeat(10), 7)), false)
 })
+
+test('a node prompt and a step reply stop on the bound without splitting the character there', () => {
+  // #1588 moved cleanLine's cut onto code points; the prompt and the step body
+  // still counted UTF-16 units, so a cut inside one astral character left a lone
+  // surrogate — the prompt a Bot was told to follow, and its own reply, carried
+  // half a character. The bound counts code points here too, so a character that
+  // fits stays whole instead of being half-kept or dropped.
+  const fitting = `${'a'.repeat(3999)}🚀`
+  const prompt = cleanPrompt(fitting, 4000)
+  assert.equal(prompt, fitting, 'a character that fits the bound is kept whole')
+  assert.equal(hasLoneSurrogate(prompt), false, 'cleanPrompt left a lone surrogate at the bound')
+
+  const over = cleanPrompt(`${'a'.repeat(4000)}🚀`, 4000)
+  assert.equal(over, 'a'.repeat(4000), 'the character past the bound goes whole, never half')
+  assert.equal(hasLoneSurrogate(over), false, 'cleanPrompt left a lone surrogate when it must drop one')
+
+  const reply = `${'a'.repeat(1999)}🚀`
+  const step = readStepResult(reply, { outputs: [] })
+  assert.equal(step.text, reply, 'the step body keeps the whole last character')
+  assert.equal(hasLoneSurrogate(step.text), false, 'readStepResult left a lone surrogate in the body')
+})