Summary:

This PR adds an end-to-end “manual mode” (human-in-the-loop) workflow: OpenEvolve can enqueue LLM requests as tasks, and the visualizer now exposes a dedicated /manual page to review prompts and submit answers back to the runner.

What changed:

Visualizer:

• Added a manual tasks page at GET /manual (monaco-based prompt viewer + answer editor).
• Added manual mode API endpoints:
  • GET /manual/api/tasks (pending tasks only)
  • GET /manual/api/tasks/<id>
  • POST /manual/api/tasks/<id>/answer

Manual-mode LLM integration:

• When llm.manual_mode=true, OpenAILLM writes <task_id>.json into <openevolve_output>/manual_tasks_queue and waits for <task_id>.answer.json.
• Controller creates/clears <openevolve_output>/manual_tasks_queue at run start so stale tasks from previous runs are not present in the UI.
• Queue path is injected via runtime-only private config field _manual_queue_dir and propagated to all model configs (works in process workers too).
• Manual mode only activates when manual_mode is explicitly True to avoid accidental enablement with mock tests.

Why:

• Provides clear and helpful way to study prompts and experiment with responses.

How to use:

1. Run OpenEvolve with llm.manual_mode: true.
2. Start the visualizer pointing at the run output directory (openevolve_output).
3. Open http://<host>:<port>/manual, answer tasks, and evolution continues once answers are submitted.

Pictures: