From 078408daf9f776d8b490dbac5cf2af17b3faf91c Mon Sep 17 00:00:00 2001
From: cuteNBD <206630735+bingdongni@users.noreply.github.com>
Date: Thu, 5 Nov 2026 01:29:03 +0800
Subject: [PATCH] fix(experiment): preserve complete agent output

---
 src/experiment/agentRunner.js      | 11 +++++---
 test/experimentAgentRunner.test.js | 41 +++++++++++++++++++++++++++++-
 2 files changed, 47 insertions(+), 5 deletions(-)

diff --git a/src/experiment/agentRunner.js b/src/experiment/agentRunner.js
index b6c3a1b6..ccaeff0a 100644
--- a/src/experiment/agentRunner.js
+++ b/src/experiment/agentRunner.js
@@ -127,7 +127,8 @@ function mapAgentResult(json, ctx) {
 }
 
 /**
- * Run one task prompt through a headless agent CLI.
+ * Run one task prompt through a headless agent CLI, then parse its complete
+ * UTF-8 output after the output streams close.
  *
  * @param {string} prompt
  * @param {object} [opts] { command, extraArgs, timeoutMs, cwd }
@@ -177,8 +178,10 @@ function runAgentTask(prompt, opts) {
       return;
     }
 
+    child.stdout.setEncoding('utf8');
+    child.stderr.setEncoding('utf8');
     child.stdout.on('data', (d) => {
-      stdout += d.toString('utf8');
+      stdout += d;
       if (stdout.length > MAX_STDOUT_BYTES) {
         // A well-behaved `--output-format json` envelope is tiny; this much
         // stdout means a runaway/malformed CLI — kill it rather than OOM.
@@ -187,7 +190,7 @@ function runAgentTask(prompt, opts) {
       }
     });
     child.stderr.on('data', (d) => {
-      stderr += d.toString('utf8');
+      stderr += d;
       if (stderr.length > MAX_STDOUT_BYTES) stderr = stderr.slice(-RESULT_CAP);
     });
 
@@ -198,7 +201,7 @@ function runAgentTask(prompt, opts) {
       done(makeFailure(msg, command, { durationMs: Date.now() - startedAt }));
     });
 
-    child.on('exit', (code, signal) => {
+    child.on('close', (code, signal) => {
       const durationMs = Date.now() - startedAt;
       // A signal-terminated child reports code=null, and Number(null) is 0 —
       // num(null, -1) would read it as a CLEAN exit and let a killed arm
diff --git a/test/experimentAgentRunner.test.js b/test/experimentAgentRunner.test.js
index cb6d38be..d678bc9f 100644
--- a/test/experimentAgentRunner.test.js
+++ b/test/experimentAgentRunner.test.js
@@ -29,6 +29,24 @@ if (mode === 'kill') {
   // Flush the clean envelope FIRST, then die by signal.
   process.stdout.write(envelope + '\\n', () => { process.kill(process.pid, 'SIGKILL'); });
   setInterval(() => {}, 1000); // keep alive until the kill lands
+} else if (mode === 'split-utf8') {
+  const output = Buffer.from(JSON.stringify({ result: '中文𠀀', is_error: false }) + '\\n');
+  const boundary = output.indexOf(Buffer.from('中文𠀀')) + 1;
+  process.stdout.write(output.subarray(0, boundary), () => {
+    setTimeout(() => process.stdout.write(output.subarray(boundary)), 50);
+  });
+} else if (mode === 'split-stderr') {
+  const output = Buffer.from('失败𠀀');
+  process.stderr.write(output.subarray(0, 1), () => {
+    setTimeout(() => process.stderr.write(output.subarray(1)), 50);
+  });
+} else if (mode === 'late-output') {
+  const script = 'setTimeout(() => process.stdout.write(' + JSON.stringify(envelope + '\\n') + '), 100)';
+  const child = require('child_process').spawn(process.execPath, ['-e', script], {
+    stdio: ['ignore', 'inherit', 'inherit'],
+  });
+  child.unref();
+  process.exit(0);
 } else {
   process.stdout.write(envelope + '\\n');
 }
@@ -51,7 +69,7 @@ function fakeCli(mode) {
   return { command: process.execPath, extraArgs: [cliPath, mode], timeoutMs: 15000 };
 }
 
-describe('runAgentTask — exit-code semantics', () => {
+describe('runAgentTask', () => {
   it('a signal-terminated arm is NEVER ok, even when stdout already holds a clean JSON envelope', async () => {
     // Regression: child 'exit' reports code=null for a signal death, and
     // Number(null) === 0 — the old num(code, -1) read that as a clean exit,
@@ -68,4 +86,25 @@ describe('runAgentTask — exit-code semantics', () => {
     assert.equal(res.exitCode, 0);
     assert.equal(res.error, null);
   });
+
+  it('preserves UTF-8 characters split across output chunks', async () => {
+    const res = await runAgentTask('noop', fakeCli('split-utf8'));
+    assert.equal(res.ok, true);
+    assert.equal(res.resultText, '中文𠀀');
+  });
+
+  it('preserves UTF-8 characters split across stderr chunks', async () => {
+    const res = await runAgentTask('noop', fakeCli('split-stderr'));
+    assert.equal(res.ok, false);
+    assert.equal(res.error, 'agent_output_parse_failed');
+    assert.equal(res.resultText, '失败𠀀');
+  });
+
+  it('waits for output streams after the agent process exits', {
+    skip: process.platform === 'win32' && 'inherited pipe behavior differs on Windows',
+  }, async () => {
+    const res = await runAgentTask('noop', fakeCli('late-output'));
+    assert.equal(res.ok, true, JSON.stringify(res));
+    assert.equal(res.resultText, 'done');
+  });
 });