autoconference-skill

Autoconference — Multi-Agent Research Conference

Spawn a conference of autonomous researchers that compete, collaborate, and synthesize breakthroughs.

Mandatory Start Gate

Before dispatching /autoconference, require a pre-flight confirmation gate:

1. If the user has exactly specified researcher count, iterations/budget, Success Metric or Success Criteria, and Critic/Devil's Advocate inclusion, calculate the run conditions and ask for final confirmation.
2. If any of those values are missing or vague, ask the user for exact values first; then calculate conditions and ask for final confirmation.
3. Do not begin the conference loop until the user explicitly confirms the calculated conditions.

The adversarial Reviewer is part of the protocol every round; the gate's Critic/Devil's Advocate choice controls whether an additional researcher is assigned a contrarian role.

Conference Persistence Directive

The Conference Chair is relentless. Once the conference begins:

1. NEVER STOP a round prematurely. Each round runs all 4 phases to completion.
2. NEVER ASK "should I continue to the next round?" Advance automatically.
3. Use the full budget. If max_rounds is 5, run 5 rounds. Stopping at round 2 wastes 60% of the exploration budget.
4. The conference runs until one of these conditions is met:
   • Target metric achieved (convergence)
   • max_rounds or max_total_iterations exhausted (budget spent — this is normal, not failure)
   • All researchers simultaneously stalled at Level 2+ (early synthesis)
   • The user manually interrupts
5. If none of these conditions are true, begin the next round immediately.

Think of max_rounds as a budget to spend, not a limit to fear.

Commands

Command	Description	Skill Path
/autoconference	Core conference loop — N researchers, 4-phase rounds, synthesis	skills/autoconference/SKILL.md
/autoconference:plan	8-step setup wizard — produces conference.md	skills/plan/SKILL.md
/autoconference:resume	Resume interrupted conference from checkpoint	skills/resume/SKILL.md
/autoconference:analyze	Post-conference insight analysis	skills/analyze/SKILL.md
/autoconference:debate	Adversarial 2-researcher debate mode	skills/debate/SKILL.md
/autoconference:survey	Systematic multi-database literature survey	skills/survey/SKILL.md
/autoconference:ship	Convert conference results to paper-ready output	skills/ship/SKILL.md

Dispatch Logic

• No subcommand (/autoconference) → load skills/autoconference/SKILL.md
• With subcommand (/autoconference:<name>) → load skills/<name>/SKILL.md
• If the target skill file does not exist yet, inform the user and fall back to the core conference protocol.

How It Works

┌─────────────────────────────────────────────────────────┐
│                    CONFERENCE ROUND                       │
│                                                           │
│  Phase 1: INDEPENDENT RESEARCH (parallel)                │
│  ┌──────────┐ ┌──────────┐ ┌──────────┐                 │
│  │Researcher│ │Researcher│ │Researcher│  Each runs N     │
│  │    A     │ │    B     │ │    C     │  autoresearch    │
│  │ (iter×N) │ │ (iter×N) │ │ (iter×N) │  iterations     │
│  └────┬─────┘ └────┬─────┘ └────┬─────┘                 │
│       │             │            │                        │
│  Phase 2: POSTER SESSION (Haiku summarizer)              │
│  Phase 3: PEER REVIEW (Opus adversarial critic)          │
│  Phase 4: KNOWLEDGE TRANSFER (validated → shared)        │
└─────────────────────────────────────────────────────────┘
          │
          ▼  Convergence check → next round or final synthesis

Agent Roles

Role	Model	Count	Responsibility
Conference Chair	Sonnet	1	Orchestrator — follows the protocol end-to-end
Researcher	Sonnet	N	Runs autoresearch 5-stage loop within assigned partition
Session Chair	Haiku	1/round	Lightweight summarizer — poster session from logs
Reviewer	Opus	1/round	Adversarial critic — challenges claims, assigns verdicts
Synthesizer	Opus	1	Combines validated findings into unified synthesis

Dependencies

• Required: Any LLM CLI with subagent support (Claude Code, Codex CLI, OpenCode, Gemini CLI)
• Extends: autoresearch-skill — each Researcher runs the autoresearch 5-stage loop
• Optional: /worktree-dashboard, /compare-worktrees
• Python 3.8+ required only for scripts/init_conference.py (scaffolding helper)

Relationship to Other Skills

Skill	Relationship
autoresearch-skill	Parent. Single-agent? Use autoresearch. Multi-agent with review? Use autoconference.
worktree-dashboard	Complementary. Monitor researcher worktrees in real-time.
compare-worktrees	Complementary. Cross-worktree analysis for the Reviewer.

Backward Compatibility

For manual install without the plugin system, this file works standalone — it dispatches to the skills/ directory. If skills/ is missing, the full protocol previously lived in this file. See skills/autoconference/SKILL.md for the complete conference orchestration protocol.