sciomc_skill

Overview

This skill orchestrates parallel scientist agents to run multi-stage research workflows with an optional AUTO mode for fully autonomous execution. It decomposes a research goal into independent stages, runs agents in parallel, verifies consistency, and synthesizes a final report. The tool is teams-first and built to scale controlled concurrency and persistent session state.

How this skill works

Given a goal, the skill decomposes it into 3–7 stages with explicit focus, hypothesis, scope, and tier (LOW/MEDIUM/HIGH). It launches independent scientist agents with explicit model routing for data gathering, analysis, and deep reasoning, collects raw findings, runs a cross-validation verification stage, and generates a synthesized report and figures. AUTO mode iterates autonomously with loop control, promise tags, and a max-iteration cutoff while persisting session state for resume and inspection.

When to use it

• When you need structured, repeatable research across a codebase or data corpus
• For parallel hypothesis testing or independent dataset analysis to save time
• When you want an autonomous runable workflow (AUTO) with progress persistence
• To produce evidence-backed reports with cross-validation and figures
• When coordinating multiple team agents and managing concurrency limits

Best practices

• Decompose goals into 3–7 focused stages with clear scope and hypotheses
• Always set the model parameter explicitly to match task complexity
• Include evidence and confidence tags for each finding to pass quality checks
• Limit concurrent scientists to avoid resource exhaustion; batch large jobs
• Use AUTO mode for end-to-end runs but monitor iterations and state.json for stuck workflows

Example use cases

• Audit authentication flows by enumerating files, analyzing token handling, and identifying vulnerabilities
• Compare implementation patterns (e.g., state management approaches) across a repository
• Generate a reproducible research report with embedded evidence and visualizations
• Run multiple hypothesis tests in parallel to evaluate performance or correctness trade-offs
• Resume or inspect long-running investigations using persisted session directories and state.json