ideate

You can generate structured research ideas by triggering a five-stage pipeline that combines wiki context with external search services like WebSearch, S2, and DeepXiv to produce, screen, and validate high-quality candidates.

To avoid duplicate ideation, the pipeline writes validated ideas and graph edges back to the wiki while recording eliminated candidates as anti-repetition memory, preventing repeated attempts at the same research direction.

Dual-model brainstorming independently generates ideas using Claude and a Review LLM to maximize diversity and coverage, before filtering for feasibility and conducting rapid novelty checks in subsequent stages.

Yes, you need access to the wiki context graph, specifically the context_brief.md and open_questions.md files, along with the /novelty, /review, and exp-eval tools to execute the full validation pipeline.

Yes, the pipeline's landscape scanning phase leverages the wiki alongside WebSearch and S2 to build a panoramic view of the research domain, identifying knowledge gaps and latest progress.

Rapid novelty screening provides a preliminary feasibility check to filter highly irrelevant directions, but deep validation requires invoking the /novelty and /review tools to output truly actionable candidates.