Recursive Chain of Task Reasoning

[carabistouflette]

It basically works like that:

• We start as a "task"
• Roocode detects if an "issue" is happening (could be missing informations, a bug or a complex thing to do)
• Do a subtask (or plan a list of subtask) to solve this specific issue and nothing else
  • This subtask can have "issues" itself
  • When subtask is finished, it makes a summary of what she has done to the parent task/subtask
• Continue with the knowledge of the subtask being resolved and what have changed since

To this we could give the user the possibility to add "step" to this routine. Like for example after the subtask is resolved, do a subtask especially for making/updating the documentation.

[carabistouflette]

Also it would make it easier to navigate in conversations as we could just close subtasks we don't want to know about.
Example:

https://excalidraw.com/#json=hUsEhIV2M-ufspSj7-ZxL,HcWoJQtRanDOCVygMQYadA

[carabistouflette]

Also, dividing tasks with self responsible sub tasks could be very good for low context window APIs. Could be good for smaller models too (i'm thinking of Gemini in specific)

[Yikai-Liao]

Great idea @carabistouflette! I’d like to add some thoughts on how this recursive chain of task reasoning could be further enhanced by leveraging LLMs for dynamic subtask scheduling and high-level orchestration.

---

Enhancement: Model-Driven Subtask Scheduling & Robust Agent Execution

• Model-Driven Subtask Scheduling:
  Instead of just statically breaking down tasks, the system could use a LLM to actively analyze the current state, detect when a subtask or plan is needed, and dynamically generate, schedule, and monitor these subtasks. This allows the agent to adapt to new issues or missing information in real time, not just at the start.

• High-Level Task Orchestration:
  By letting the model manage the flow between parent and child tasks, we can ensure that:

  • Each subtask is summarized and its results are passed up the chain.
  • The agent always has an up-to-date, concise context, minimizing context window usage and API costs.
  • The system can automatically insert extra steps (like documentation, testing, or review) as needed.

• Robustness for Long-Running Tasks:
  This approach is especially valuable for models with limited context windows.

  • It prevents the agent from losing track or prematurely ending a long-running task by checkpointing progress and summarizing completed subtasks.
  • If the model session ends unexpectedly, the system can resume from the last completed subtask, avoiding wasted work.

• Agentic Execution & Cost Efficiency:

  • This not only reduces context size, but also improves the agent’s ability to execute complex, multi-step goals.
  • For users with limited budgets, it means you can run sophisticated workflows without incurring huge API costs from long, uncompressed contexts.

---

In summary:
Model-driven subtask scheduling and orchestration can both lower context/token usage and make agentic workflows more robust and reliable—especially for long, complex tasks or when using models with strict context limits.

Would love to see this direction explored further!

[nickchomey]

I wonder if a lot of this is already done by Claude-task-master, which has some active PRs for roo code support.

https://github.com/eyaltoledano/claude-task-master