kakashi-ventures/magellan-cli
15 subagents
Subagent
computational-validatorValidates bridge concepts and hypotheses computationally before generation. Runs KEGG pathway cross-checks, STRING interaction scores, PubMed co-occurrence counts, and back-of-envelope quantitative…convergence-scannerScans for independent convergence signals — clinical trials, funded grants, patents, and partial mechanism confirmations from sources not consulted by the main pipeline.criticAggressively attacks scientific hypotheses. Finds counter-evidence via web search, checks novelty against published literature, identifies logical fallacies and mechanism implausibilities.cross-model-validatorValidates hypotheses by calling OpenAI (GPT-5.5 Pro, web search + code interpreter) and Google Gemini (Deep Research Max, google_search + url_context + code_execution via Interactions API) for…dataset-evidence-minerQueries bioinformatics databases (Human Protein Atlas, GWAS Catalog, ChEMBL, UniProt, PDB) to verify specific molecular claims in passing hypotheses. Complements the pre-generation Computational…discovery-orchestratorOrchestrates a full autonomous scientific discovery cycle. Coordinates all agents through 2 complete cycles. Manages state via state/session.json and results/{session-id}/.evolverRecombines and refines top hypotheses through crossover, mutation, specification, and generalization. Enforces diversity constraint to prevent convergence.generatorGenerates novel scientific hypotheses by finding connections between disconnected fields. Uses parametric knowledge as creative engine AND literature context for grounding. Techniques include facet…holdout-evaluatorEvaluates whether MAGELLAN independently rediscovered a known post-cutoff scientific connection. Performs contamination check (did the pipeline find the answer paper?) and mechanism similarity…literature-scoutRetrieves and analyzes scientific literature from databases (Semantic Scholar, PubMed, arXiv via web search) to provide grounding context for hypothesis generation and validation.quality-gateFinal quality check on surviving hypotheses. 9-point rubric + web novelty verification. Determines PASS/FAIL for each hypothesis.rankerScores surviving hypotheses on 6 dimensions including Groundedness. Applies diversity check to prevent convergence. Selects top candidates for evolution.scoutAutonomously identifies the most promising areas for undiscovered scientific connections. Uses 10 strategies including Swanson ABC bridging, structural isomorphism, and serendipity. Launch when user…session-analystPost-pipeline analyst that extracts quantitative patterns from the current session and all past sessions. Produces meta-insights that improve future Scout and Generator performance. Closes the…target-evaluatorAdversarial target evaluator that challenges Scout targets before pipeline investment. Attacks targets on 4 adversarial axes — popularity bias, vagueness, structural impossibility, local-optima —…