fix(cli): add scrolling, Ctrl+C pause/resume, and enhanced model detection (#1887) #2593
Conversation
… model support
|
| PR | Title | Merged | Author |
|---|---|---|---|
| #2442 | Fix incorrect pip install prompt for CLI addon | Jul 14, 2025 | @fureigh |
| #2358 | hot-fix-Increase-timeouts & fix remove message | Jul 8, 2025 | @MagMueller |
| #2347 | Always use subprocess + CDP for local browser inst... | Jul 8, 2025 | @pirate |
📊 Summary
The command-line interface (CLI) agent control and model detection logic in browser_use/cli.py has undergone several key changes since early 2024, particularly around user interaction (scrolling, pause/resume), API key/model selection, and configuration handling.
MagellaX’s PR #1887 (May 2025) originally introduced a richer TUI with task panel scrolling and attempted to unify model selection, but users reported issues with scrolling conflicts and Ctrl+C causing abrupt exits. This led to feedback and further iteration in unmerged PR #1891 (June 2025), where the team debated how to make agent pausing/resuming and manual scroll coexist smoothly.
In PR #2358 (Jul 2025), MagMueller increased LLM call timeouts and improved message management, which indirectly affected the CLI’s responsiveness and error handling during agent operations. This change made it clearer that the CLI needed more graceful interruption and recovery mechanisms, especially when users pressed Ctrl+C during long-running tasks.
The model and API key detection logic has evolved from a static config-based approach to a more dynamic, environment-variable-prioritized system. Earlier versions only checked config files, but as of late July 2025, the CLI now checks environment variables first, then config files, and finally user config at ~/.config/browseruse/config.json, as seen in the current diff and recent commits.
Manual scrolling in the tasks panel was previously tied to auto-scroll, which frustrated users who wanted to review history while the agent was running; PR #2593 (Aug 2025) decouples this by setting auto_scroll=False and only triggering scroll-to-end when the agent is actively running. This addresses longstanding usability complaints and aligns with feedback from #1887 and #1891.
Ctrl+C handling has shifted from immediate application exit to a two-stage pause/quit flow, with Enter now mapped to resume the agent—this was a direct response to user confusion and accidental session loss, as discussed in #1887 and #1891. The new logic ensures that pressing Ctrl+C once pauses the agent, and a second press quits, while Enter resumes, improving both safety and clarity.
This history is relevant now because PR #2593 consolidates and finalizes these iterative changes, resolving prior usability issues and implementing the pause/resume and scrolling behaviors that were debated and prototyped across multiple earlier PRs.
There was a problem hiding this comment.
cubic analysis
No issues found across 1 file. Review in cubic
PR Description
Addresses CLI issues in #1887 and feedback from unmerged #1891: fixed TUI scrolling conflicts, non-destructive Ctrl+C handling, and model param errors.
Scrolling: Set
auto_scroll=FalseonRichLogin tasks panel; conditionalscroll_end()inupdate_tasks_panel()only whenagent.running == True, enabling manual history navigation during execution.Pause/Resume: Remapped
Ctrl+Ctopause_agent(callsagent.pause()if running, else quits); addedEnterbinding forresume_agent(handles async/syncagent.resume()); secondCtrl+Cduring pause invokesaction_quit(). Updatedon_key()to routeEnterfor resume when input unfocused andagent.state.paused == True.Model Support: Prioritized API key detection (
env> config) for OpenAI/Anthropic/Google/DeepSeek/Groq; load preferences from~/.config/browseruse/config.json; fixedmodelvsmodel_namekwargs inChat*constructors.Verified via
test_cli_improvements.py(bindings, detection, config loading) and manual runs. Passes ruff lint/format.Related: #1887, #1891.
Summary by cubic
Improved the CLI by adding manual scrolling in the tasks panel, non-destructive Ctrl+C pause/resume controls, and better model/API key detection for multiple providers.