Notes on Claude Code

• not a good idea to use stop hook to orchestrate autonomous agentic work to perform a series of task, unless if there are official support
  • see this discussion on the complexities of using stop hook, decision flag and the stop_hook_active flag
  • there are also complications for making hook logic prompt-based
  • it would be better for autonomous agent logic to have ability to orchestrate not just Claude Code, but also Codex and others
  • although the Ralph for Claude Code looks attractive but it is not sustainable due to reasons above
• an intelligent agent needs to check against limits to maximize the usage, but Claude slash command /usage only works in the interactive mode.
  • Open source solution are also not suitable:
    • claude-monitor only displays in Rich format and only tracks one device
    • ccusage doesn’t show reset time, more useful when paying by API call
    • while forking and enhancing the open source solution is one route, a robust route is to use Claude control over Chrome
• Claude controls Chrome:
  • Claude Code with Chrome
  • the following prompt takes the Claude usage website with existing cookie and session credential and analyzes the screenshot for retrieving key information

      claude --model haiku --chrome --permission-mode acceptEdits -p "open https://claude.ai/settings/usage page, and tell me what percentage is used for the current session, and when does it reset. return the result in yaml with two keys: plan _usage_limit_session_pct and plan_usage_limit_reset_time. plan_usage_limit_session_pct holds a numeric value, plan_usage _limit_reset_time holds a ISO timestamp in my machine's timezone. Save the yaml into a file called ./tmp/claude_usage.yaml"
    

  • other solutions are not suitable:
    • playwright tries to reuse the cookies and session, but runs into cloudflare blocker
    • selenium would not offer reuse of cookies and session information
• claude chat history location: $USERPROFILE\.claude\projects

Back to index