diff --git a/CI.md b/CI.md
index e9384acf15..ff278592e7 100644
--- a/CI.md
+++ b/CI.md
@@ -144,6 +144,7 @@ Each run retains `report.json` and `report.md` for 90 days in
 `compatibility-review-run-<run-id>-attempt-<attempt>`. The report records both
 revisions, the train, tool/model metadata, coverage gaps, and actionable
 findings. Raw agent transcripts and PR discussions are not published.
+SDK failures retain the exception message in the report.
 The qualification summary links this separate advisory evidence without waiting
 for it; compatibility
 review remains missing from the blocking RFC profile during this rollout.
diff --git a/tasks/scripts/codex_compatibility.py b/tasks/scripts/codex_compatibility.py
index 99d64698e2..59f893f11c 100644
--- a/tasks/scripts/codex_compatibility.py
+++ b/tasks/scripts/codex_compatibility.py
@@ -129,7 +129,7 @@ async def invoke_codex(context: dict, directory: Path) -> dict:
             result = await thread.run(prompt, output_schema=schema)
     except (openai_codex.CodexError, RuntimeError, ValueError) as error:
         raise ValueError(
-            f"Codex SDK review failed ({type(error).__name__}); no assessment is available."
+            f"Codex SDK review failed ({type(error).__name__}): {error}; no assessment is available."
         ) from None
     finally:
         await codex.close()