Thanks for the contribution! The feature direction is great — live Bash output is a real UX improvement. However, a few issues need to be addressed before this can be merged:

1. Tool card expand/collapse behavior (UX)

• On tool_end, please remove the expanded class so the card returns to its default collapsed state, consistent with the current behavior (auto-collapse after execution; user clicks to expand).
• The tool_progress handler calls classList.add('expanded') on every event, so if the user manually collapses the card during execution it keeps popping back open. Please expand only on the first progress event and respect the user's manual toggle afterwards.

2. Process tree is not killed on Windows (functional defect)

In _kill_process, the Windows branch only calls process.kill(), which kills cmd.exe itself but not its descendants (e.g. pip, npm). Orphaned children keep the stdout pipe open, which directly triggers the hang described in issue 3. Please use taskkill /F /T /PID on Windows to kill the whole process tree, matching the os.killpg behavior on POSIX:

if self._IS_WIN:
    subprocess.run(["taskkill", "/F", "/T", "/PID", str(process.pid)], capture_output=True)

3. reader.join() without timeout can hang forever (critical)

If the command spawns a background process (e.g. nohup xxx &, or start /b xxx on Windows) that keeps stdout/stderr open, os.read never returns EOF after the main process exits, so reader.join() in the finally block blocks forever. The entire tool-execution thread hangs and the agent session is stuck permanently. The old subprocess.run path did not hang the calling thread, so this is a regression. Please use reader.join(timeout=5) — the readers are daemon threads, so abandoning them is safe and no already-captured output is lost.

4. Frontend memory / replay buffer growth (performance)

Each progress event can carry up to 32KB, emitted every 0.5s, and every event is appended to the buffer.items replay buffer (capped at 5000 entries, not bytes). A long-running command could grow browser memory to 100MB+. Suggestions: send only the last ~4KB (or last N lines) per snapshot, and keep only the latest progress event per tool_call_id in the replay buffer.

5. i18n inconsistency (minor)

The disabled-state tooltips are hardcoded English strings ("Reply is being generated; editing is temporarily unavailable."). Please route them through the existing t() / I18N system and add Chinese translations.

6. Empty output still renders the "Output" label (minor)

On tool_end, the "Output" label is rendered even when result is empty. The old logic skipped the output section when there was no result — please keep that behavior.

7. Tests (recommend)

This PR rewrites the core execution path of the Bash tool but ships no automated tests. Please add unit tests covering: fast commands, timeout kill, commands spawning background processes, and output without trailing newlines. Also, the test checklist in the PR description doesn't mention Windows — please verify timeout and long-running command scenarios on a real Windows environment.

Thanks for the contribution! The feature direction is great — live Bash output is a real UX improvement. However, a few issues need to be addressed before this can be merged:

1. Tool card expand/collapse behavior (UX)

• On tool_end, please remove the expanded class so the card returns to its default collapsed state, consistent with the current behavior (auto-collapse after execution; user clicks to expand).
• The tool_progress handler calls classList.add('expanded') on every event, so if the user manually collapses the card during execution it keeps popping back open. Please expand only on the first progress event and respect the user's manual toggle afterwards.

2. Process tree is not killed on Windows (functional defect)

In _kill_process, the Windows branch only calls process.kill(), which kills cmd.exe itself but not its descendants (e.g. pip, npm). Orphaned children keep the stdout pipe open, which directly triggers the hang described in issue 3. Please use taskkill /F /T /PID on Windows to kill the whole process tree, matching the os.killpg behavior on POSIX:

if self._IS_WIN:
    subprocess.run(["taskkill", "/F", "/T", "/PID", str(process.pid)], capture_output=True)

3. reader.join() without timeout can hang forever (critical)

If the command spawns a background process (e.g. nohup xxx &, or start /b xxx on Windows) that keeps stdout/stderr open, os.read never returns EOF after the main process exits, so reader.join() in the finally block blocks forever. The entire tool-execution thread hangs and the agent session is stuck permanently. The old subprocess.run path did not hang the calling thread, so this is a regression. Please use reader.join(timeout=5) — the readers are daemon threads, so abandoning them is safe and no already-captured output is lost.

4. Frontend memory / replay buffer growth (performance)

Each progress event can carry up to 32KB, emitted every 0.5s, and every event is appended to the buffer.items replay buffer (capped at 5000 entries, not bytes). A long-running command could grow browser memory to 100MB+. Suggestions: send only the last ~4KB (or last N lines) per snapshot, and keep only the latest progress event per tool_call_id in the replay buffer.

5. i18n inconsistency (minor)

The disabled-state tooltips are hardcoded English strings ("Reply is being generated; editing is temporarily unavailable."). Please route them through the existing t() / I18N system and add Chinese translations.

6. Empty output still renders the "Output" label (minor)

On tool_end, the "Output" label is rendered even when result is empty. The old logic skipped the output section when there was no result — please keep that behavior.

7. Tests (recommend)

This PR rewrites the core execution path of the Bash tool but ships no automated tests. Please add unit tests covering: fast commands, timeout kill, commands spawning background processes, and output without trailing newlines. Also, the test checklist in the PR description doesn't mention Windows — please verify timeout and long-running command scenarios on a real Windows environment.

Thanks for the detailed review. I’ve addressed all requested changes:

• Tool cards now expand only on the first progress event, respect manual collapse, and automatically collapse on tool_end.
• Windows process termination now uses taskkill /F /T /PID to terminate the complete process tree, with fallback handling.
• Reader threads use a bounded shared join timeout to prevent background processes holding pipes open from blocking indefinitely.
• Progress snapshots are limited to 4 KiB, and the replay buffer retains only the latest progress event per tool_call_id.
• Disabled Edit/Delete tooltips now use the existing English and Chinese i18n translations.
• Empty results no longer render an Output section.
• Added automated tests covering fast commands, timeout handling, background processes, output without trailing newlines, and Windows process-tree termination.

I also added a windows-latest GitHub Actions workflow and verified the timeout and long-running command scenarios successfully on a real Windows runner. All Windows Bash streaming tests passed.

thanks for the quick turnaround! Verified all the fixes locally, the bounded reader.join(), taskkill /T process-tree termination, 4KB snapshots, replay buffer dedup, i18n, and the collapse behavior all look good, and the test suite passes on macOS. Nice work on the Windows tests as well

merged. Thanks again for the contribution! 🎉