From b80b9d82719455c9a88b32259c701e211b78b871 Mon Sep 17 00:00:00 2001
From: lEWFkRAD <SJWATTS89@OUTLOOK.COM>
Date: Thu, 27 Aug 2026 22:48:59 -0400
Subject: [PATCH] fix(codex): preserve assistant image slots in replay

---
 agent/codex_responses_adapter.py            | 17 +++++++++++++++--
 tests/agent/test_codex_responses_adapter.py | 20 ++++++++++++++++++++
 2 files changed, 35 insertions(+), 2 deletions(-)

diff --git a/agent/codex_responses_adapter.py b/agent/codex_responses_adapter.py
index df90e81388..909437a6ab 100644
--- a/agent/codex_responses_adapter.py
+++ b/agent/codex_responses_adapter.py
@@ -194,8 +194,13 @@ def _chat_content_to_responses_parts(content: Any, *, role: str = "user") -> Lis
             continue
         if ptype in {"image_url", "input_image"}:
             if role == "assistant":
-                # Responses API rejects input_image on assistant messages
-                # (HTTP 400 on every replay) — drop, see docstring (#96816).
+                # Responses output messages cannot carry input_image. Keep a
+                # text marker so image-only assistant turns still survive in
+                # replay and later references retain their conversational slot.
+                converted.append({
+                    "type": "output_text",
+                    "text": "[Assistant image omitted during replay]",
+                })
                 continue
             image_ref = part.get("image_url")
             detail = part.get("detail")
@@ -1167,6 +1172,14 @@ def _preflight_codex_input_items(
                             text = str(text or "")
                         validated.append({"type": text_type, "text": sanitize_text(text)})
                     elif ptype in {"input_image", "image_url"}:
+                        if role == "assistant":
+                            # Enforce the same output-message invariant for
+                            # raw request overrides as for normal history.
+                            validated.append({
+                                "type": "output_text",
+                                "text": "[Assistant image omitted during replay]",
+                            })
+                            continue
                         image_ref = part.get("image_url", "")
                         detail = part.get("detail")
                         if isinstance(image_ref, dict):
diff --git a/tests/agent/test_codex_responses_adapter.py b/tests/agent/test_codex_responses_adapter.py
index c9a6d7696c..d8da67a67b 100644
--- a/tests/agent/test_codex_responses_adapter.py
+++ b/tests/agent/test_codex_responses_adapter.py
@@ -30,6 +30,8 @@ def test_chat_content_drops_images_from_assistant_role():
 
     assert _chat_content_to_responses_parts(content, role="assistant") == [
         {"type": "output_text", "text": "generated image"},
+        {"type": "output_text", "text": "[Assistant image omitted during replay]"},
+        {"type": "output_text", "text": "[Assistant image omitted during replay]"},
     ]
 
 
@@ -46,6 +48,24 @@ def test_chat_content_keeps_images_on_user_role():
     }]
 
 
+def test_preflight_rewrites_raw_assistant_images_to_text_markers():
+    raw = [{
+        "role": "assistant",
+        "content": [{
+            "type": "input_image",
+            "image_url": "https://example.invalid/p.png",
+        }],
+    }]
+
+    assert _preflight_codex_input_items(raw) == [{
+        "role": "assistant",
+        "content": [{
+            "type": "output_text",
+            "text": "[Assistant image omitted during replay]",
+        }],
+    }]
+
+
 def _harmony_token(name: str) -> str:
     """Build a literal Harmony token without spelling it contiguously here."""
     return f"<\x7c{name}\x7c>"