From 95b542ec287bfdf269278f4416facee740c9f463 Mon Sep 17 00:00:00 2001
From: Pravin Kamble <iampbkamble@gmail.com>
Date: Tue, 10 Nov 2026 09:56:46 +0530
Subject: [PATCH] DRU-1092 -- Strip OpenAI citation markers from Codex output

---
 backend/druks/harnesses/codex.py   |  7 +++++++
 backend/tests/test_cost_capture.py | 23 +++++++++++++++++++++++
 2 files changed, 30 insertions(+)

diff --git a/backend/druks/harnesses/codex.py b/backend/druks/harnesses/codex.py
index b50b71ab..ff17a55d 100644
--- a/backend/druks/harnesses/codex.py
+++ b/backend/druks/harnesses/codex.py
@@ -2,6 +2,7 @@
 import json
 import logging
 import os
+import re
 import shlex
 from collections.abc import Collection, Iterator
 from dataclasses import dataclass
@@ -43,6 +44,9 @@
 _TOKEN_COUNT_MARKERS = ('"type":"token_count"', '"type": "token_count"')
 _CHATGPT_HOST = "chatgpt.com"
 _SUBSCRIPTION_TOKEN = "codex_subscription_token"
+# OpenAI web search leaves this citation in model text. The pattern runs on
+# serialized JSON. It stops at a quote, so a match stays in one string.
+_CITATION_MARKER = re.compile(r' *\ue200cite(?:\ue202[^\ue201"]*)?\ue201')
 
 
 @dataclass(frozen=True)
@@ -566,6 +570,9 @@ def parse(self, result: HarnessRunResult, *, artifact_dir: Path, run_id: str) ->
             call_dir / "output.json",
             name=self.name,
         )
+        payload = json.loads(
+            _CITATION_MARKER.sub("", json.dumps(payload, ensure_ascii=False)),
+        )
         (call_dir / "output.json").write_text(
             json.dumps(payload, indent=2, sort_keys=True),
         )
diff --git a/backend/tests/test_cost_capture.py b/backend/tests/test_cost_capture.py
index e2e90012..155f50ee 100644
--- a/backend/tests/test_cost_capture.py
+++ b/backend/tests/test_cost_capture.py
@@ -138,6 +138,29 @@ def test_failed_codex_call_writes_cost_before_raising(tmp_path: Path):
     assert metadata["output_tokens"] == 1_000
 
 
+def test_codex_parse_strips_citation_markers(tmp_path: Path):
+    call_dir = tmp_path / "call"
+    call_dir.mkdir()
+    citation = "\ue200cite\ue202turn905355view5\ue202turn430332view7\ue201"
+    output = {
+        "summary": f"Calls can require fresh execution. {citation}",
+        "sources": [f"Docs {citation} say so.", "\ue200cite\ue201"],
+    }
+    (call_dir / "output.json").write_text(json.dumps(output))
+    harness = CodexHarness(model="openai/gpt-5.5", fast_mode=False, effort=None)
+
+    payload = harness.parse(
+        HarnessRunResult(returncode=0, stdout=b"", stderr=b""),
+        artifact_dir=tmp_path,
+        run_id="call",
+    )
+
+    assert payload == {
+        "summary": "Calls can require fresh execution.",
+        "sources": ["Docs say so.", ""],
+    }
+
+
 def test_failed_pi_call_writes_cost_before_raising(tmp_path: Path):
     message = {
         "type": "message_end",