Tool results arrive delayed/invisibly — model appears frozen during long Bash/SSH calls (Opus 4.8, VSCode extension)
Summary
During longer sessions, tool-call results (Bash, Read, Grep, and especially SSH/scp + php artisan commands) arrive in the UI delayed or seemingly not at all, so the assistant appears frozen/stuck even though it is actively working. The user cannot tell whether work has started, is in progress, or finished. On multi-step deploy tasks this badly erodes trust.
Environment
- Model: Claude Opus 4.8 (claude-opus-4-8)
- Surface: Claude Code — VSCode native extension
- OS: macOS (darwin)
- Workload: SSH/scp to a remote Laravel server,
php artisancache/lint commands, multi-file edits
Symptoms
- After a tool call (especially SSH/scp or a batch of reads), the UI shows no visible progress for ~10–30s.
- The user repeatedly perceives the assistant as 'frozen' / 'stuck in a loop' and interrupts it.
- Because results surface late, the assistant sometimes re-sends the same tool call (thinking it failed), producing duplicate/echoed tool calls — amplifying the 'looping' perception.
- Side-effect notices like "Wasted call — file unchanged" and "Cancelled: parallel tool call" appear.
Impact
User quote: "I don't know if you started or not — the work might be half-done and I know nothing. This is a serious problem." Interrupting mid-tool-call leaves the user unsure of state even when nothing was actually mutated.
Expected
- A clear, incremental 'running…' indicator per tool call and an explicit completed state.
- Timely/streaming delivery of tool results so the model's activity is observable in real time.
Notes
Filed at a user's explicit request. The user is a solo developer running live deploy workflows through Claude Code and relies on seeing each step's status.
This issue has 3 comments on GitHub. Read the full discussion on GitHub ↗