research-agents

OpenCode plugin that registers three research-oriented subagents: Ideator, Critic, and Sweeper. Together they form an iterative research loop for generating, evaluating, and grounding research ideas.

Install

Add to opencode.json:

{
  "plugin": [
    "research-agents"
  ]
}

Or install from npm:

npm install research-agents

Agents

All agents are registered as OpenCode subagents (mode: subagent) using claude-sonnet-4-20250514 by default.

research-ideator

Research ideation specialist that generates 2-3 novel, grounded research ideas per iteration.

• Temperature: 0.7 (creative)
• Input: Experiment results, related work, previous critiques
• Output: Structured idea proposals with problem statement, approach, expected outcome, novelty justification, required experiments, and effort estimate
• Behavior: Classifies research state (early exploration / promising lead / plateau / near completion) and adjusts idea diversity accordingly. In revision mode, builds on surviving ideas rather than starting from scratch.

research-critic

Evaluates proposed ideas on four dimensions with structured scoring (1-10):

• Temperature: 0.3 (precise)
• Dimensions: Feasibility, Novelty, Impact, Soundness
• Verdicts: ACCEPT (overall >= 7, no dimension below 5), REVISE (4-6 or any dimension 3-4), REJECT (overall < 4 or any dimension below 3)
• Output: Per-idea critique with scores, strengths, weaknesses, revision directions, plus a loop assessment with convergence signal

research-sweeper

Literature search specialist that finds and synthesizes related work from multiple sources.

• Temperature: 0.2 (factual)
• Sources: arXiv (via arxiv-database skill), Semantic Scholar, web search
• Output: Thematic synthesis (not paper-by-paper listing) with research gaps, a structured paper catalog table, and coverage of direct, methodological, foundational, and adjacent angles
• Skills used: arxiv-database, literature-review

Research Loop

The three agents work in a cycle:

Sweeper (find related work)
    → Ideator (generate ideas using results + related work)
        → Critic (evaluate ideas, score, verdict)
            → Ideator (revise based on critique)
                → Critic (re-evaluate)
                    → ... (until ACCEPT or max iterations)

License

MIT