fix: strip benign stderr output for clarity

Author: jahilldev

api/src/agent/agent.utility.spec.ts

@@ -9,6 +9,7 @@ import {
   extractJsonBlock,
   generateHermesConfig,
   incompleteOutputReason,
+  stripShutdownNoise,
 } from './agent.utility.js';
 
 describe('buildVerifySpec', () => {
@@ -282,3 +283,41 @@ describe('eventInsertRows', () => {
     expect(rows.map((r) => r.seq)).toEqual([5, 6]);
   });
 });
+
+describe('stripShutdownNoise', () => {
+  const STANZA = [
+    'Exception ignored in: <generator object Langfuse._create_span_with_parent_context at 0x7f04789a74c0>',
+    'Traceback (most recent call last):',
+    '  File "/usr/local/lib/hermes-agent/venv/lib/python3.11/site-packages/langfuse/_client/client.py", line 1196, in _create_span_with_parent_context',
+    '  File "/usr/local/share/uv/python/cpython-3.11.15/lib/python3.11/contextlib.py", line 158, in __exit__',
+    '  File "/usr/local/lib/hermes-agent/venv/lib/python3.11/site-packages/opentelemetry/trace/__init__.py", line 616, in use_span',
+    'TypeError: isinstance() arg 2 must be a type, a tuple of types, or a union',
+  ].join('\n');
+
+  it('removes the langfuse shutdown stanza while keeping surrounding output', () => {
+    const input = `real warning line\n${STANZA}\nanother real line`;
+    const out = stripShutdownNoise(input);
+
+    expect(out).toContain('real warning line');
+    expect(out).toContain('another real line');
+    expect(out).not.toContain('isinstance() arg 2');
+    expect(out).not.toContain('_create_span_with_parent_context');
+    expect(out).not.toContain('Traceback');
+  });
+
+  it('leaves a genuine error traceback untouched', () => {
+    const realError = [
+      'Traceback (most recent call last):',
+      '  File "app.py", line 10, in <module>',
+      'ValueError: something actually broke',
+    ].join('\n');
+
+    expect(stripShutdownNoise(realError)).toBe(realError);
+  });
+
+  it('is a no-op when the stanza is absent', () => {
+    expect(stripShutdownNoise('plain stderr\n[worker_guard] capped read_file')).toBe(
+      'plain stderr\n[worker_guard] capped read_file',
+    );
+  });
+});

api/src/agent/agent.utility.ts

@@ -370,6 +370,49 @@ function cap(buf: string): string {
   return buf.length > STDOUT_CAP ? `${buf.slice(0, STDOUT_CAP)}\n…[truncated]` : buf;
 }
 
+/**
+ * Strip a confirmed-benign shutdown artifact from captured agent stderr. langfuse leaves an
+ * OpenTelemetry span context-manager open at process exit (typically a threaded `delegate_task`
+ * subagent span); during interpreter-shutdown GC, `opentelemetry.trace.use_span` references a
+ * module global already torn down to `None`, raising `isinstance() arg 2 must be a type`. Python
+ * prints it as an "Exception ignored in: <generator object Langfuse._create_span_with_parent_context
+ * …>" stanza and discards it — it never reaches the agent, so the run is unaffected. We drop the
+ * stanza so it neither clutters every run's stderr nor buries the real error when a run genuinely
+ * fails. Narrowly matched to that exact generator, so no other diagnostic is hidden.
+ */
+export function stripShutdownNoise(stderr: string): string {
+  if (!stderr.includes('Langfuse._create_span_with_parent_context')) {
+    return stderr;
+  }
+
+  const lines = stderr.split('\n');
+  const kept: string[] = [];
+
+  for (let i = 0; i < lines.length; i++) {
+    const isStanzaStart =
+      lines[i].startsWith('Exception ignored in:') &&
+      lines[i].includes('Langfuse._create_span_with_parent_context');
+
+    if (!isStanzaStart) {
+      kept.push(lines[i]);
+      continue;
+    }
+
+    // Drop the whole stanza: the "Exception ignored in:" line, the "Traceback" header, every
+    // indented frame line, and the trailing column-0 exception summary (e.g. "TypeError: …").
+    i++;
+    while (
+      i < lines.length &&
+      (lines[i] === 'Traceback (most recent call last):' || /^\s/.test(lines[i]))
+    ) {
+      i++;
+    }
+    // lines[i] is now the exception summary (or out of range); the for-loop's increment skips it.
+  }
+
+  return kept.join('\n');
+}
+
 /**
  * Spawns a process, captures (capped) stdout/stderr, and hard-kills it after
  * `timeoutMs`. Never rejects — failures surface in the result.
@@ -420,7 +463,7 @@ export function spawnProcess(
       resolve({
         exitCode,
         stdout: cap(Buffer.concat(stdoutChunks).toString()),
-        stderr: cap(Buffer.concat(stderrChunks).toString()),
+        stderr: stripShutdownNoise(cap(Buffer.concat(stderrChunks).toString())),
         durationMs: Date.now() - start,
         timedOut,
       });