Idea Tree autonomous-research engine

Idea Tree runs inside the existing evolve Python service through an ASGI worker. The API handles session authentication, model proxying, and usage recording; Python runs the research loop; and the existing Web tree panel starts, observes, and controls it. The chat Agent does not run the tree workflow, create an execution plan, or dispatch Subagents.

Use

In System settings → Idea Tree, configure the default budget and role prompts. Then enter /idea-tree <research task> or /idea-tree-team <research task> in a session to start it. The frontend has no separate form for an objective, source text, or files. objective and materials remain backend inputs for creating research. Complete retrieval and parsing before starting: the engine has no external retrieval, code-execution, or tool capability. The workspace tree panel shows running state, pause, resume, stop, and node details without opening a separate dialog.

One round consists of a batch of candidates, individual design and assessment, and feedback. By default, a research run has at most three rounds with at most three candidates in each. Candidate count, node count, and depth limits also apply. Maximum depth is a limit, so shallower candidates can still be assessed. Later rounds use existing insights to explore new directions or improve existing candidates. At the depth limit, the engine can create a sibling improvement.

You can replace the design, three assessment, aggregation, and propagation role prompts in System settings. Changes apply to new research. The backend manages ideation prompts and does not expose an additional editor for them. Default scoring retains activity, stability, and sustainability dimensions, and Python calculates the weighted score.

Pausing does not cancel a model request that has already been sent. The panel first shows Pausing, then Paused after the in-flight response finishes. The in-flight call can still incur usage, but it does not start the next stage. Resuming always requires an explicit user action. After a process restart, research is Interrupted and does not resume automatically. Stopping requires confirmation, retains existing results, and cannot be reversed by resuming.

Implementation and state

• vendor/idea_tree/research.py: ideation, design, independent assessment, aggregation, and level-by-level propagation. It saves each completed stage and runs only unfinished stages after recovery.
• vendor/idea_tree/research_tree.py: reuses node and ancestor queries from the root tree.py.
• vendor/idea_tree/prompts.py: role instructions and scoring criteria.
• vendor/idea_tree/research_service.py: create, query, pause, resume, and stop. It ensures that each session runs at most one research job in a process.
• services/api/src/idea-tree/research.ts: research HTTP client and temporary model proxy credentials.
• apps/web/src/IdeaResearchPanel.tsx: research controls in the existing tree panel.

The only persistent state is $SCIENCE_AGENT_DATA_DIR/idea-research/<projectId>/<sessionId>/<researchId>.json in the Python service, saved through atomic replacement. It includes materials, budget, nodes, stage results, and usage, but not model credentials. This path does not use external revisions, leases, childDigest, SHA verification, or result handles. The old tree remains viewable through its existing read API. Its write operations return a read-only error and its prior execution state is not migrated automatically.

The public endpoint is POST /api/sessions/:sessionId/idea-tree/research, where operation accepts create, get, list, pause, continue, end, and defaults. GET on the same path lists research. Sending an Idea Tree command directly to the chat-message API returns a 409 that directs the caller to the research entry point; it does not start Lead silently.

Boundaries

Idea Tree currently uses the session's OpenAI-compatible model through the API's model proxy. When the API or model is unavailable, research is saved as interrupted and can resume only after those services are available again. Model output must be JSON, with at most one correction for each formatting error. Network timeouts are not retried automatically. Model requests use the proxy's roughly 20-minute limit.

An optional total-token budget relies on usage reported by the provider. Before a request, the service conservatively reserves input and output tokens by UTF-8 byte count, so it can pause early. This is not an exact billing limit. When a provider does not report usage, the display is unknown and research with a total budget is interrupted. Materials and each stage's output have length limits. Ideation carries only direction summaries plus recent and stronger candidates, not the chat history.

Research scores and material recommendations are model assessments. Do not treat them as experimental conclusions without experiments or supplied evidence.