• on_ticket
  • if sandbox
    • check_sandbox →
  • fetch_relevant_files
  • get_commit_history
  • post_process_snippets
  • ExternalSearcher.extract_summaries of any links mentioned in issue + replies
  • extract_relevant_docs / DocumentationSearcher: get in-repo docs
  • ContextPruning(HumanMessagePrompt()): chat pruning_prompt to get what context to prune
  • sweet_bot = SweepBot.from_system_message_content(HumanMessagePrompt())
  • bot.get_files_to_change()
    • if is python issue: chat extract_files_to_change_prompt
    • chat files_to_change_prompt
    • returns list of checks/creates/changes
  • bot.validate_sandbox(file change requests)
    • check_sandbox for some file →
  • validate_file_change_requests
    • validate_sandbox →
  • bot.generate_pull_request(): chat PR
  • create_pr_changes()
    • bot.change_files_in_github_iterator()
      • switch
        • handle_create_file_main()
          • chat create_file_prompt
        • handle_modify_file_main()
        • refactor
        • test
        • check
        • delete
        • rename
• check_sandbox(file)
  • run_sandbox(file)
    • start container
      • clone if nec
      • update files based on request
      • for each, check it (i.e. lint/fmt)
• HumanMessagePrompt: just dumps a bunch of context about relevant snippets, paths, history, repo tree, repo & issue metadata
• prompts sequence
  • system prompt
  • pruning_prompt - what paths to keep (prune the rest), what dirs to expand
  • extract_files_to_change_prompt
  • files_to_change_prompt - what files to add/update, whether to use tools
  • in parallel: sandbox
  • draft PR
  • create_file_prompt
  • review_prompt
  • final_review_prompt
• observations / ideas
  • doesn’t work in terms of the higher level approach for a lot of cases, not very impressive
  • doesn’t really work mechanically, often
  • no repeated iteration
  • no deep critiquing, but does review
  • no tree of thought
  • shallow
  • find relevant other similar calls to the used functions