Cowork: tool-call markup rendered as raw text + executions stall with no progress indicator

[isolhsolfafa]

Here's the English version with a title — ready to submit.

---

Title: Cowork: tool-call markup rendered as raw text + executions stall with no progress indicator

---

Summary
When the assistant runs a tool (code execution, file create/edit/read), two problems occur: (1) the raw tool-call markup is printed as plain text in the chat instead of a clean "working…" indicator, and (2) executions stall or get cancelled, and the status indicator does not reflect the real state — so the user cannot tell whether it's running or stopped.

Symptoms

1. Rendering / display bug

   • Tool calls should appear as a tidy "running…" element. Instead, the raw markup (e.g. <invoke name="...">, <parameter>…</parameter>) is dumped into the chat as visible text.
   • Happens more often and more severely when the command is long, or contains a lot of Korean text / file paths.

2. Execution reliability / status bug

   • Commands don't finish (or get cancelled) partway through.
   • The status shows "stopped" while it is actually running, or no result returns — leaving the user unsure whether it's in progress or dead.

Impact

• File-generation (PPTX, DOCX, etc.) and code-execution tasks do not complete reliably.
• This undermines Cowork's core value (direct file/system work), forcing a text-only workaround.
• Note: sometimes the file is actually created despite the broken display, and sometimes the command genuinely never completes — so it appears to be two overlapping issues (a display-layer bug and an execution-layer bug).

Steps to Reproduce

1. Ask the assistant to generate a PPTX/DOCX file, or read/edit a file.
2. Especially with tasks involving Korean text + a long script, the raw markup appears and/or the run stalls.
3. Repeats across multiple tasks within the same session.

Environment

• Cowork mode (research preview), macOS desktop app
• Date: 2026-06-18, reproduced repeatedly within a single session

---

Short version (for thumbs-down):
"Tool calls render as raw code (<invoke>...) in the chat instead of a progress card, and executions stall midway with no progress/done indicator (status shows 'stopped' while running). Repeats on Korean + long-script file tasks."