Desktop agent misread single Windows folder name and repeatedly verified the wrong path

Correction to the original report

The original issue text incorrectly claimed that the user literally entered a nested Windows path ending in 문항별\_DOCX. That was the agent's interpretation, not a verified record of the user's typed or displayed input. The user explicitly states that the visible folder name was 문항별_DOCX (one folder name, with no backslash before the underscore). Please disregard the original reproduction wording.

Observed sequence

• In a Korean-language Codex desktop conversation on Windows, the user wanted six final DOCX files directly inside a single folder named 문항별_DOCX under the working directory.
• The text representation available to the agent contained 문항별\_DOCX at one point. We have not determined where that backslash came from or whether it was a Markdown escape, client serialization, or another transformation. It must not be attributed to the user's typing without evidence.
• The agent treated the backslash in that representation as a Windows path separator and constructed a command for two nested directories, 문항별 then _DOCX.
• After the user objected that the folder should not be under 문항별, the agent checked only whether files existed in its own chosen path and repeatedly asserted that the result was correct. It moved/renamed the folders multiple times before finally placing the files directly in 문항별_DOCX.
• In its explanations, the agent also asserted an unverified cause (Markdown escaping) and described the agent-visible representation as though it were the user's original input, compounding the error.

Expected behavior

Distinguish user-visible/typed text from any agent-facing escaped representation; do not infer unverified provenance for extra characters. Before filesystem mutation, validate the intended destination as a single folder name against the user's explicit corrections. After a correction, compare the actual absolute path with the user's intended structure, not merely with the agent's prior interpretation.

Investigation request

Please investigate both (1) whether the desktop message pipeline can present an escaped underscore differently to the model than to the user and (2) why the agent failed to reconcile repeated explicit corrections with its filesystem commands and validation. The available evidence does not establish whether Korean, model choice, or reasoning-effort setting caused this. The user described the setting as "GPT-5.6 Sol light", but that was not independently verified.

No real username, document content, or full session transcript is included.

Start by tracing the desktop message pipeline and the filesystem command and validation flow; the issue names no specific files or tests. Reproduce the single-folder-name case and compare the user-visible text with the agent-facing representation, then verify correction handling against the actual absolute path. Done means the transformation source and repeated-correction failure are identified without assuming an unverified cause.