currytang/amadeus/todo-to-node

TODO to Tree Node

Convert a single project TODO (title + hypothesis) into a fully-specified ResearchOps tree node that can be inserted into the project plan.

Workflow

1. Input: Receive todo object (id, title, hypothesis) and optional parentNodeId.
2. Clarification Phase (optional, per-run configurable): Call POST /api/researchops/projects/:projectId/tree/nodes/from-todo/clarify with:

• todo: the TODO object
• messages: conversation so far (starts empty)
• The backend returns { done, question, options? }. Ask the question; when done: true, proceed to generate.
• User can Skip Q&A to bypass this phase entirely.
• Prompts are targeted at the TODO kind: papers referenced, KB code files needed, env assumptions, implementation vs experiment differences.

1. Generate: Call POST /api/researchops/projects/:projectId/tree/nodes/from-todo with:

• todo: the TODO object
• parentNodeId: optional parent node to attach under
• messages: clarification exchange collected in phase 2 (empty if skipped)

1. Review: Present the generated node fields to the user:

• id, title, kind
• assumption[]: what the node assumes
• target[]: measurable success criteria
• commands[]: concrete runnable commands
• checks[]: verification steps

1. Refine (optional): If user wants changes, send a follow-up request with:

• Same todo
• messages: conversation history [{role, content}] where assistant messages include the previous node JSON

1. Insert: Call POST /api/researchops/projects/:projectId/tree/plan/patches with:

``json { "patches": [{ "op": "add_node", "node": <generated_node> }] } ``

Node Schema

{
  "id": "snake_case_slug",
  "parent": "optional_parent_node_id",
  "title": "Human-readable title",
  "kind": "experiment | analysis | knowledge | setup | milestone | patch | search",
  "assumption": ["key assumption 1"],
  "target": ["measurable criterion 1"],
  "commands": [{"cmd": "bash command", "label": "label"}],
  "checks": [{"condition": "verification", "label": "label"}],
  "tags": ["tag1"]
}

Refinement Examples

• "Make commands use python3 instead of python"
• "Add a validation step that checks the model accuracy"
• "Change kind to analysis"
• "Add an assumption that the dataset is already downloaded"

Error Handling

• If LLM fails to return valid JSON: show raw text and ask user to retry
• If insert fails: show patch error and keep generated node for retry