ghostonbutterbread/retard-collaboration

Retard Collaboration — Creative Multi-Agent Brainstorming

What It Does

A 3-agent creative brainstorming harness that builds on zero_day_team findings to generate novel attack chains using deliberate creative chaos.

Agents:

1. Creative Chaos Agent — model=gpt-4.1 (weaker, weirder, more creative). Reads zero_day findings, generates 10+ wild lateral ideas
2. Analyst Agent — model=gpt-5.4. Filters creative output to 2-3 genuine diamonds
3. Synthesizer Agent — model=gpt-5.4. Combines filtered ideas with zero_day findings into novel attack chains

Invocation

/retard_collaboration <program> [--source <path>]

Examples:

/retard_collaboration evernote
/retard_collaboration evernote --source ~/Shared/bounty_recon/evernote/0day_team/

Harness Location

~/projects/bug_bounty_harness/agents/retard_collaboration.py

Workflow Position

zero_day_team → retard_collaboration → chainer

Run zero_day_team first to get findings, then feed the findings to retard_collaboration to generate creative chains.

How It Works

File-based 3-stage pipeline:

/tmp/collab_{program}_{date}/
  creative/findings.txt       ← wild ideas (gpt-4.1)
  analyst/filtered.txt        ← diamonds only (gpt-5.4)
  synthesizer/chains.md       ← novel chains (gpt-5.4)
  shared_context/
    zero_day_findings.txt     ← from zero_day_team
    app_map.txt               ← optional app map
    task.txt

Stage 1 — Creative Chaos (gpt-4.1):

• System prompt: "You are the Creative Chaos Agent. Your sole job is to be CREATIVE. There are NO bad ideas."
• Reads zero_day findings + app map
• Outputs 10+ wild, lateral, unconventional attack ideas

Stage 2 — Analyst (gpt-5.4):

• System prompt: "You are the Senior Security Researcher — the gatekeeper. Be brutal in filtering."
• Reads creative output
• Filters to 2-3 ideas with actual security potential

Stage 3 — Synthesizer (gpt-5.4):

• System prompt: "You are the Synthesis Expert — the builder. Think in chains."
• Reads filtered ideas + zero_day findings
• Outputs 3-5 novel attack chains that neither team produced alone

Output Location

~/Shared/bounty_recon/{program}/ghost/collaboration/{date}/
├── collaborative_chains.md    ← main output (chains)
├── creative/findings.txt      ← raw creative output
├── analyst/filtered.txt       ← filtered ideas
└── synthesizer/chains.md     ← synthesizer output

Usage Pattern

# 1. Run zero_day_team first
python3 agents/zero_day_team.py evernote ~/source/

# 2. Run retard_collaboration on the findings
python3 agents/retard_collaboration.py evernote --source ~/Shared/bounty_recon/evernote/0day_team/

# 3. Run chainer on the collaborative output (optional)
python3 agents/chainer.py evernote --source ~/Shared/bounty_recon/evernote/ghost/collaboration/

Why "Retard"?

The name reflects a real technique: deliberately impaired or unconventional reasoning generates more diverse ideas than pure optimization. A slightly "dumb" agent breaks out of local minima that a fully capable agent never leaves.

The creative chaos agent uses:

• A weaker model (gpt-4.1) that defaults to less obvious paths
• A prompt that explicitly tells it to break conventions
• No penalty for wild ideas — only reward for quantity

Options

• --skip-creative — Skip creative stage (if output already exists)
• --skip-analyst — Skip analyst stage
• --verbose — Print raw agent output

Credits

Concept inspired by research on impaired agents and creative diversity in LLM ensembles.